Amit Kumar Bairwa

Varun Tiwari
Santosh Kumar Vishwakarma
Milan Tuba
Thittaporn Ganokratanaa (Eds.)

Communications in Computer and Information Science 2184

Computation of
Artificial Intelligence
and Machine Learning
First International Conference, ICCAIML 2024
Jaipur, India, January 18–19, 2024
Proceedings, Part I
Communications
in Computer and Information Science 2184

Editorial Board Members

Joaquim Filipe , Polytechnic Institute of Setúbal, Setúbal, Portugal
Ashish Ghosh , Indian Statistical Institute, Kolkata, India
Lizhu Zhou, Tsinghua University, Beijing, China
Rationale
The CCIS series is devoted to the publication of proceedings of computer science con-
ferences. Its aim is to efficiently disseminate original research results in informatics
in printed and electronic form. While the focus is on publication of peer-reviewed full
papers presenting mature work, inclusion of reviewed short papers reporting on work in
progress is welcome, too. Besides globally relevant meetings with internationally repre-
sentative program committees guaranteeing a strict peer-reviewing and paper selection
process, conferences run by societies or of high regional or national relevance are also
considered for publication.
Topics
The topical scope of CCIS spans the entire spectrum of informatics ranging from foun-
dational topics in the theory of computing to information and communications science
and technology and a broad variety of interdisciplinary application fields.
Information for Volume Editors and Authors
Publication in CCIS is free of charge. No royalties are paid, however, we offer registered
conference participants temporary free access to the online version of the conference
proceedings on SpringerLink (http://link.springer.com) by means of an http referrer from
the conference website and/or a number of complimentary printed copies, as specified
in the official acceptance email of the event.
CCIS proceedings can be published in time for distribution at conferences or as post-
proceedings, and delivered in the form of printed books and/or electronically as USBs
and/or e-content licenses for accessing proceedings at SpringerLink. Furthermore, CCIS
proceedings are included in the CCIS electronic book series hosted in the SpringerLink
digital library at http://link.springer.com/bookseries/7899. Conferences publishing in
CCIS are allowed to use Online Conference Service (OCS) for managing the whole
proceedings lifecycle (from submission and reviewing to preparing for publication) free
of charge.
Publication process
The language of publication is exclusively English. Authors publishing in CCIS have
to sign the Springer CCIS copyright transfer form, however, they are free to use their
material published in CCIS for substantially changed, more elaborate subsequent publi-
cations elsewhere. For the preparation of the camera-ready papers/files, authors have to
strictly adhere to the Springer CCIS Authors’ Instructions and are strongly encouraged
to use the CCIS LaTeX style files or templates.
Abstracting/Indexing
CCIS is abstracted/indexed in DBLP, Google Scholar, EI-Compendex, Mathematical
Reviews, SCImago, Scopus. CCIS volumes are also submitted for the inclusion in ISI
Proceedings.
How to start
To start the evaluation of your proposal for inclusion in the CCIS series, please send an
e-mail to [email protected].
Amit Kumar Bairwa · Varun Tiwari ·
Santosh Kumar Vishwakarma · Milan Tuba ·
Thittaporn Ganokratanaa
Editors

Computation of
Artificial Intelligence
and Machine Learning
First International Conference, ICCAIML 2024
Jaipur, India, January 18–19, 2024
Proceedings, Part I
Editors
Amit Kumar Bairwa Varun Tiwari
Manipal University Jaipur Manipal University Jaipur
Jaipur, Rajasthan, India Jaipur, Rajasthan, India

Santosh Kumar Vishwakarma Milan Tuba

Gyan Ganga Institute of Technology Univerzitet Singidunum
and Sciences Beograd, Serbia
Jabalpur, Madhya Pradesh, India

Thittaporn Ganokratanaa
King Mongkut’s University of Technology
Bangkok, Thailand

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-031-71480-1 ISBN 978-3-031-71481-8 (eBook)
https://doi.org/10.1007/978-3-031-71481-8

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2025

This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors
or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

If disposing of this product, please recycle the paper.

 Preface

We are delighted to present the proceedings of the 1st International Conference on

Computation of Artificial Intelligence & Machine Learning (ICCAIML 2024), held on
18th and 19th January 2024 at Manipal University Jaipur, India. This inaugural event
marks a significant milestone in the field, providing an outstanding international arena for
the exchange of information and findings in the theory, methodology, and applications
of Artificial Intelligence (AI) and Machine Learning (ML).
Artificial Intelligence and Machine Learning have become pivotal in driving tech-
nological advancements across various industries. From healthcare to finance, trans-
portation to entertainment, the impact of AI and ML is profound and far-reaching. The
primary aim of ICCAIML 2024 was to bring together researchers, practitioners, and
enthusiasts from around the globe to share their insights, discuss emerging trends, and
explore the latest innovations in these dynamic fields.
ICCAIML 2024 sought to highlight significant theoretical and practical contributions
to the key disciplines of AI and ML. The conference served as a forum for scholars and
practitioners from academia and industry to meet and discuss the most recent advances
in these areas. By providing a platform for the dissemination of cutting-edge research
and practical applications, ICCAIML 2024 aimed to foster collaboration and inspire new
ideas that will propel the field forward.
The scope of ICCAIML 2024 was broad, encompassing a wide range of topics
within AI and ML. Authors were invited to participate by submitting articles that report
major developments in research findings, projects, surveying works, and industrial expe-
riences. The conference covered, but was not limited to, areas such as Neural Networks
and Deep Learning, Natural Language Processing, Computer Vision, Reinforcement
Learning, Data Mining and Big Data Analytics, AI in Healthcare and Biomedical Appli-
cations, Autonomous Systems and Robotics, AI Ethics and Fairness, AI in Finance and
Economics, and many more.
This conference provided a unique opportunity for attendees from diverse fields to
engage in discussions about fresh ideas and application experiences. It offered a valu-
able platform for creating commercial or research relationships and identifying poten-
tial worldwide collaborators. By bringing together experts from different backgrounds,
ICCAIML 2024 encouraged interdisciplinary collaboration and the cross-pollination of
ideas.
We are honored to have received an overwhelming number of high-quality submis-
sions from researchers and practitioners worldwide. The rigorous review process ensured
that the selected papers meet the highest standards of quality and relevance. We extend
our heartfelt gratitude to the authors for their valuable contributions, the reviewers for
their diligent efforts, and the organizing committee for their unwavering commitment.
We are confident that ICCAIML 2024 facilitated productive discussions and collabo-
rations that will drive forward the fields of Artificial Intelligence and Machine Learning.
The presentations, workshops, and panel discussions provided valuable insights and
vi Preface

stimulated innovative thinking. We hope that the knowledge and connections gained
during this conference will inspire future research and developments.
In conclusion, we express our sincere thanks to all participants, sponsors, and partners
for their support in making ICCAIML 2024 a success. We look forward to the inspiring
presentations and fruitful discussions that will shape the future of AI and ML.

July 2024 Amit Kumar Bairwa

Varun Tiwari
Santosh Kumar Vishwakarma
Milan Tuba
Thittaporn Ganokratanaa
 Organization

Chief Patron

S. Vaitheeswaran Manipal University Jaipur, India

Patron

Gopalkrishna K. Prabhu Manipal University Jaipur, India

Co-Patrons

Jawahar M. Jangir Manipal University Jaipur, India

Nitu Bhatnagar Manipal University Jaipur, India

Conference Chairs

Arun Shanbhag Manipal University Jaipur, India

Sandeep Chaurasia Manipal University Jaipur, India

General Chair

Santosh Kumar Vishwakarma Manipal University Jaipur, India

Program Chair

Neha Chaudhary Manipal University Jaipur, India

Organizing Secretaries

Amit Kumar Bairwa Manipal University Jaipur, India

Varun Tiwari Manipal University Jaipur, India
viii Organization

Technical Program Committee

Puneet Mittal Manipal University Jaipur, India

Vivek Bhardwaj Manipal University Jaipur, India
Saurabh Srivastava Manipal University Jaipur, India
Surbhi Sharma Manipal University Jaipur, India
Preeti Narooka Manipal University Jaipur, India
Shikha Maheshwari Manipal University Jaipur, India

Finance Chair

Deepika Shekhawat Manipal University Jaipur, India

Sponsorship Chairs

Shishir Singh Chauhan Manipal University Jaipur, India

Shilpi Birla Manipal University Jaipur, India

Program Committee

Anita Shotriya Manipal University Jaipur, India

Sayar Singh Shekhawat Manipal University Jaipur, India
Neelam Chaplot Manipal University Jaipur, India
Shubh Lakshmi Agrwal Manipal University Jaipur, India
Manish Rai Manipal University Jaipur, India
Upendra Singh Manipal University Jaipur, India
Lokesh Malviya Manipal University Jaipur, India
Yadvendra Pratap Singh Manipal University Jaipur, India

Publication Chairs

Gautam Kumar Manipal University Jaipur, India

Ajay Kumar Manipal University Jaipur, India
Surendra Solanki Manipal University Jaipur, India
Siddharth Kumar Manipal University Jaipur, India
Jayesh Gangrade Manipal University Jaipur, India
 Organization ix

Program Advisory Committee

Dinesh Kumar Saini Manipal University Jaipur, India

Amit Soni Manipal University Jaipur, India
Roheet Bhatnagar Manipal University Jaipur, India
Sumit Srivastava Manipal University Jaipur, India
Vijaypal Singh Dhaka Manipal University Jaipur, India
Sandeep Joshi Manipal University Jaipur, India
Prakash Ramani Manipal University Jaipur, India
Akhilesh Kumar Sharma Manipal University Jaipur, India
Shakti Kundu Manipal University Jaipur, India
Sunil Kumar Vasistha Manipal University Jaipur, India
Pankaj Vyas Manipal University Jaipur, India

Technical Session (Online)

Satpal Singh Kushwaha Manipal University Jaipur, India

Lav Upadhyay Manipal University Jaipur, India
Hemlata Parmar Manipal University Jaipur, India
Neetu Gupta Manipal University Jaipur, India
Preeti Narooka Manipal University Jaipur, India

Technical Session (Offline)

Deepak Panwar Manipal University Jaipur, India

Mahesh Jangid Manipal University Jaipur, India
Jaydeep Kishore Manipal University Jaipur, India

Registration Committee

Babita Tiwari Manipal University Jaipur, India

Sushama Tanwar Manipal University Jaipur, India
Bali Devi Manipal University Jaipur, India
Vaishali Chauhan Manipal University Jaipur, India
Pallavi Manipal University Jaipur, India
x Organization

Food Committee

Prakash Chandra Sharma Manipal University Jaipur, India

Venkatesh Gauri Shankar Manipal University Jaipur, India

Accommodation Committee

Puneet Mittal Manipal University Jaipur, India

Dinesh Swami Manipal University Jaipur, India

Transport Committee

Ajit Noonia Manipal University Jaipur, India

Ajay Kumar Manipal University Jaipur, India
Dinesh Swami Manipal University Jaipur, India

Cultural Committee

Harish Sharma Manipal University Jaipur, India

Liaison Committee

Satyabrata Roy Manipal University Jaipur, India

Deepjyoti Choudhury Manipal University Jaipur, India

Poster Presentation

Juhi Singh Manipal University Jaipur, India

Media and Photography

Surbhi Sharma Manipal University Jaipur, India

Rishika Singh Manipal University Jaipur, India
 Organization xi

Website Coordinator

Vineeta Soni Manipal University Jaipur, India

Additional Reviewers

Elzbieta Pustulka University of Applied Sciences and Arts,

Switzerland
Thittaporn Ganokratanaa King Mongkut’s University of Technology,
Thailand
Milan Tuba Singidunum University, Serbia
Bipin C. Desai Concordia University, Canada
Dinesh Manocha University of Maryland, USA
Abul Bashar Prince Mohammad Bin Fahd University,
Saudi Arabia
Amit Gajbhiye Cardiff University, UK
Rahul Singh University of Hong Kong, China
Geeta Sikka National Institute of Technology, Jalandhar, India
Sotiris Moschoyiannis University of Surrey, UK
Abhineet Anand Galgotias University, India
A. M. Razmy South Eastern University of Sri Lanka, Sri Lanka
Anshuman Kalla University of Oulu, Finland
Aynur Unal San Francisco Bay, UK
Basant Agarwal Central University of Rajasthan, India
Bharanidharan Shanmugam Charles Darwin University, Australia
Brij B. Gupta National Institute of Technology, Kurukshetra,
India
Chhagan Charan National Institute of Technology, Kurukshetra,
India
Chithral Ambawatte University of Ruhuna, Sri Lanka
Dimitrios A. Karras National & Kapodistrian University of Athens,
Greece
Milena Radenkovic University of Nottingham, UK
Elena Basan Southern Federal University, Russia
Hari B. Hablani Indian Institute of Technology, Bombay, India
Jafar A. Alzubi Wake Forest University, North Carolina, USA
Kambiz Ghazinour State University of New York, Canton, USA
M. Hafizur Rahman International Islamic University, Malaysia
M. A. R. M. Fernando University of Peradeniya, Sri Lanka
Maxim Anikeev Southern Federal University, Russia
Mohamed Firdhous University of Moratuwa, Sri Lanka
xii Organization

Mohiuddin Ahmed University of New South Wales, Australia

Narpat Singh Shekhawat Govt. Engineering College Bikaner, India
Nitin Gupta National Institute of Technology, Hamirpur, India
Noor Zaman Taylor’s University, Malaysia
Nurilla Avazov Inha University, South Korea
P. Sakthivel Anna University, India
Parag Narkhede Symbiosis Institute of Technology, Pune, India
Pardeep Singh National Institute of Technology, Hamirpur, India
P. K. Garg Indian Institute of Technology, Roorkee, India
Prashant Kumar National Institute of Technology, Jalandhar, India
Pushpendu Kar University of Nottingham, UK
Rahul Dixit Indian Institute of Information Technology, Pune,
India
Samantha Thelijjagoda Sri Lanka Institute of Information Technology,
Sri Lanka
Sanjay Singh Manipal Academy of Higher Education, India
Saurabh Kumar National Institute of Technology, Hamirpur, India
Smitha N. Pai Manipal Academy of Higher Education, India
Snehanshu Shekhar BIT Mesra, India
Somya Ranjan Sahoo VIT, Andhra Pradesh, India
Sujala Deepak Shetty BITS Pilani, Dubai Campus, UAE
Sukanta Ganguly Entryless, USA
Syed Hassan Ahmed Kyungpook National University, South Korea
Vidhyacharan Bhaskar San Francisco State University, USA
Wilfred Lin PuraPharm, China
Silva A. T. P. University of Moratuwa, Sri Lanka
Guathilaka Samantha University of Ruhuna, Sri Lanka
Rajeev Kumar National Institute of Technology, Hamirpur, India
Ravi Saharan Central University of Rajasthan, India
Ekaterina Pakulova Southern Federal University, Russia
Muzathik Abdul Majeed South Eastern University of Sri Lanka, Sri Lanka
Mohamed Ismail Roushdy Future University in Egypt, Egypt
Eduard Babulak Fort Hays State University, USA
Walter Peeters International Space University, France
V. G. Tharinda Nishantha Wayamba University of Sri Lanka, Sri Lanka
Vidanagama
Saroj Hiranwal Victorian Institute of Technology, Australia
Adel Al-Jumaily University of Technology, Sydney, Australia
Selwyn Piramuthu University of Florida, USA
Dmitry Namiot Lomonosov Moscow State University, Russia
Victor Govindaswamy Concordia University, Chicago, USA
 Organization xiii

Muthukumar S. National Institute of Technology, Puducherry,

India
Cihangir Tezcan Middle East Technical University, Turkey
Ekaterina Maro Southern Federal University, Russia
Rasib Khan Northern Kentucky University, USA
Muhammad Mostafa Monowar King Abdulaziz University, Saudi Arabia
Jagannathan Sarangapani Missouri University of Science and Technology,
USA
M. M. Hafizur Rahman International Islamic University, Malaysia
 Contents – Part I

Leukemia Insight: Illuminating Current Diagnoses and Forecasting

Futures with Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Sunita Gupta, Neha Janu, and Neha Shrotriya

Intelligent Conversational Chatbot: Design Approaches and Techniques . . . . . . . 16

Johnbenetic Gnanaprakasam and Ravi Lourdusamy

Intelligent Conversational Chatbots: History, Taxonomy, Classification

with PRISMA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Johnbenetic Gnanaprakasam and Ravi Lourdusamy

Unleashing the Power of Dynamic Mode Decomposition and Deep

Learning for Rainfall Prediction in North-East India . . . . . . . . . . . . . . . . . . . . . . . . 47
Paleti Nikhil Chowdary, Sathvika Pingali, Pranav Unnikrishnan,
Rohan Sanjeev, V. Sowmya, E. A. Gopalakrishnan, and M. Dhanya

Patient’s Condition Categorization Using Drug Reviews . . . . . . . . . . . . . . . . . . . . . 63

Akshit Kamboj, Shikha Mundra, and Ankit Mundra

Systematic Review on Sustainable Healthcare Practices: A Bibliometric

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Shubhangi V. Urkude and Debajani Sahoo

Violence Detection Through Deep Learning Model in Surveillance . . . . . . . . . . . 86

Anirudh Singh, Satyam Kumar, Abhishek Kumar, and Jayesh Gangrade

Role of ChatGPT in Decision Making Across Industries: An Indian

Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Debajani Sahoo and Shubhangi V. Urkude

Identification of Splice Image Forgeries with Enhanced DenseNet201

and VGG19 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Satyendra Singh, Rajesh Kumar, and Chandrakant Kumar Singh

Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset

Images Using Various Deep and Machine Learning Techniques:
A Comparative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Piyush Jain, Deepak Motwani, and Pankaj Sharma
xvi Contents – Part I

Access-Based Authentication of Healthcare Record in Blockchain:

A Survey from Security Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Shashank Saroop, Rajesh Kumar Tyagi, Shweta Sinha, Shafiqul Abidin,
and Lokesh Kumar

Smart Entry Management Using IoT and Number Plate Recognition

for Organizational Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Love Lakhwani, Vibhanshu Singh, and Sunil Kumar

Enhancing Player Engagement Through Gesture-Based Interactions

in Tic-Tac-Toe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Vedansh Rathi, Shrey Sharma, Anmol Rai, and Sandeep Kumar Sharma

Aggressive Bangla Text Detection Using Machine Learning and Deep

Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Tanjela Rahman Rosni, Mahamudul Hasan, Tanni Mittra,
Md. Sawkat Ali, and Md. Hasanul Ferdaus

A Comparative Evaluation of Machine Learning Techniques for Detecting

Malicious Network Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Prince Tayal, Rohan Kumar, and Hemlata

Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis

of Imagery Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Jyoti Dinkar Bhosale, Suraj S. Damre, Ujwala V. Suryawanshi,
and Rajkumar B. Pawar

Revolutionizing Road Safety: Machine Learning Approaches

for Predicting Road Accident Severity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Meenakshi Malik, Rainu Nandal, and Rita Chhikara

Survey and Analysis of Machine Learning Methods for Parkinson’s

Disease Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Poonam Yadav, Meenu Vijarania, Meenakshi Malik, and Ritu

Accelerated GPU-Based Clustering, Classification and Regression Using

RAPIDS: A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Mayuri Gupta, Ashish Mishra, Ashutosh Mishra, and Jayesh Gangrade

Talk to Hands: A Sign Language Solution for Speech-Impaired People . . . . . . . . 254

Durvesh Chaudhari, Adrika Dikshit, Omkar Pawar, Deepak Karia,
and Payal Shah
 Contents – Part I xvii

Deep Learning-Based Thyroid Cancer Detection and Segmentation Using

Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Aditya Praksh and Juhi Singh

Early Autism Spectrum Disorder Prediction Using Fine-Tuned Bernoulli’s

Naive Bayes Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
Kanav Gupta, Chirag Paul, and Nishant Jain

Heart Disease Prediction Model Using Machine Learning Techniques . . . . . . . . . 290

Bipin Kumar Rai, Aparna Jha, Shreyal Srivastava, and Aman Bind

A Review to Enhance Driving Behavior Through Advanced Drowsiness

Detection Systems: Integrating Deep Learning and Machine Learning
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Ritu, Meenu Vijarania, Meenakshi Malik, and Poonam Yadav

Advancing Peer Review Integrity: Automated Reviewer Assignment

Techniques with a Focus on Deep Learning Applications . . . . . . . . . . . . . . . . . . . . 312
Bhumika Bhaisare and Rajesh Bharati

Automated Potato Disease Classification Using Deep Learning -

A Comparative Analysis of Convolutional Neural Networks . . . . . . . . . . . . . . . . . 328
Swati Pandey, Mayuri Gupta, Ashish Mishra, Ashutosh Mishra,
and Jayesh Gangrade

Enhancing Monkeypox Disease Detection Using Computer Vision-Based

Approaches and Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
Imtiaj Ahmed, Rayan, Sayma Akter Tihany, and Adnan Mahmud

Tomato Leaf Disease Prediction Based on Deep Learning Techniques . . . . . . . . . 357

Anirudh Singh, Satyam Kumar, and Deepjyoti Choudhury

Predicting Stock Market Trends: Machine Learning Approaches

of a Possible Uptrend or Downtrend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Raima Joseph and Mausumi Goswami

A Scientific Classification-Based Survey of Intrusion Detection System

Using Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
Neha Srivastava and R. K. Singh

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397

 Contents – Part II

A Machine Learning Approach for Driver Drowsiness Detection

in Automotive Safety Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
K. Sujigarasharma, M. Lawanya Shri, K. Santhi,
Balamurugan Balusamy, and Shilpa Gite

Dynamic COVID-19 Endurance Indicator System for Scientific Decisions

Using Ensemble Learning Approach with Rapid Data Processing . . . . . . . . . . . . . 10
Ajay Kumar

Facial Recognition Advancements with Siamese Networks:

A Comprehensive Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Deepika Bansal, Bhoomi Gupta, Sachin Gupta, Aastha Anand, Sumit,
and Aditya Sagar

Improving PCOS Diagnosis Accuracy with CNN-Based Image Analysis . . . . . . 42

Kbks Durga, M. Shanmuga Sundari, Ayesha Shaik, Shilpa Mukthala,
and Harshitha Gudapati

Adoption of Quantum NLP for Improved Healthcare System . . . . . . . . . . . . . . . . 51

Chandani Arora and Ramandeep Sandhu

Algorithm Visualizer Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Rajiv Tulsyan and Pranjal Shukla

An Efficient Ontology Based Chronic Disease Diagnosis Model . . . . . . . . . . . . . . 69

Md. Gulzar and Muqeem Ahmed

Deep Learning Models for Building and Shadow Detection in Aerial

Imagery: A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
M. Mohamed Naajim, Ashima Rani, and Aarti Chugh

A Comparative Analysis of Transfer Learning Based Models for Early

Detection of Parkinson’s Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Pranjal Mohan Pandey, Vedant Pandey, Sakalya Mitra,
and A. V. R. Mayuri

Development of Rule-Based Chunker for Sindhi . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

Palak Arora, Bharti Nathani, Nisheeth Joshi, and Pragya Katyayan
xx Contents – Part II

Smart Meter Security and Solutions with IOT in India . . . . . . . . . . . . . . . . . . . . . . 113

Santosh Sharma and Lava Bhargava

Optimization of Positioning of Reclosers in the High Voltage Category

a Distribution Network of Alibori (Benin) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Cossi Télesphore Nounangnonhou, Arouna Oloulade,
François-Xavier Fifatin, Guy Clarence Semassou, Romaric Tognigban,
and Taohidi Alamou Lamidi

An Efficient Machine Learning-Based Framework for Analysis

and Prediction of Links in Social Media Platform . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Sonia Choudhary, Seema Malik, Narendra Singh Yadav, and Subhash Vats

Deep Learning Segmentation Method for Uneviling Lower Limb

Deformative on the Knee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Aarti Goswami, Sandeep Chaurasia, and Jayesh Gangrade

Development of HMM Based Parts of Speech Tagger for Hadoti . . . . . . . . . . . . . 170

Anushka Nagar, Nisheeth Joshi, Pragya Katyayan, and Palak Arora

Sentiment Analysis of Tokopedia Product Reviews Using Naïve Bayes

Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Mario Gracius Krishna Lita, Alya Dhiya’ Mardhiyyah,
I. Gusti Ayu Ngurah Stita Maharani, Angelina Patience Mulia,
and Fairuz Iqbal Maulana

Combating Toxicity on the Internet: An Efficient HateSwarm Algorithm

for Classifying Hate Speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
Anjum and Rahul Katarya

Automated Categorization of Stack Overflow Queries: Solution

to Highlight the Similar Kind of Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
Pranshu Raghuwanshi, Bhavya Gopalani, and Surbhi Sharma

A Modified U-Net for Semantic Segmentation of Liver and Liver Tumors

From CT Scans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
R. Rashmi and S. Girisha

Enhancing Histopathological Image Analysis: A Study on Effect of Color

Normalization and Activation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
G. V. S. Sudhamsh, R. Rashmi, and S. Girisha

Smart Cities in India: A Bibliometric Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

Paribhasha Sharma, Srinivas Jangirala, and Anandadeep Mandal
 Contents – Part II xxi

Generative Adversarial Networks for Generation of Synthetic Images:

A Comprehensive Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
N. Pavithra, R. Sapna, Preethi, and Y. H. Sharath Kumar

Abstractive Text Summarization with Domain Knowledge for Biomedical

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Shivani Sharma, Gaurav Agarwal, and Bipin Kumar Rai

Enhanced Residual U-Net with Attention for Optic Disc and Cup
Segmentation in Fundus Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Ritesh V. Kamath, Bhavana Kedari, S. Girisha, G. Savitha,
and S. Shreesha

A Comprehensive Survey of Evaluation Techniques for Recommendation

Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Aryan Jadon and Avinash Patil

eAuthenticate: Enhancing Security with Machine Learning Based

Real-Time Open-World Anti-spoofing Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Om Delvadia, Jitendra Parmar, Pratik Gupta,
Kuruba Lakshmi Harsha Vardhan, Dhruv Gandhi, and Atul Kumar Verma

Utilization of Computer Vision Techniques of AI for the Attendance

Management System in Online Sessions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Geeta Chhabra Gandhi, Jagrite Jain, and Rajat Goel

Machine Learning and the Internet of Things for Industrial Process

Improvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Shailaja Salagrama, Amit Garg, J. Logeshwaran,
Satpal Singh Kushwaha, and Rajan Kumar

A Novel Method to Detect the Faults in an Underground Cable Using Robot . . . 346
Madhu Palati, Mahipal Bukya, Rajesh Kumar, and Akhilesh Mathur

Enhancing Time Series Forecasting with Machine Learning and Deep

Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
Amit Kumar Sharma, Ritwick Roy, and Sandeep Chaurasia

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

 Leukemia Insight: Illuminating Current
Diagnoses and Forecasting Futures
with Machine Learning

Sunita Gupta1 , Neha Janu2(B) , and Neha Shrotriya3

1 Department of Information Technology, Swami Keshvanand Institute of Technology
Management and Gramothan, Jaipur, India
2 Department of Computer Science and Engineering, Manipal University Jaipur, Jaipur,
Rajasthan, India
[email protected]
3 Department of Computer Engineering, Poornima College of Engineering Jaipur, Jaipur, India

Abstract. Algorithms like Deep Learning (DL) and Machine learning (ML) have
been increasingly used in the diagnosis and treatment of leukemia. One of the main
applications of ML in leukemia is in the analysis of samples of blood to identify
abnormal cells and diagnose the disease. ML algorithms can be trained to rec-
ognize patterns in large datasets of blood samples and medical records and use
this to predict the presence and progression of leukemia. This can help doctors to
diagnose the disease earlier and more accurately and to develop personalized treat-
ment. In addition, ML can also be used to develop predictive models for leukemia
treatment outcomes, helping doctors to choose the most effective treatments for
a patient depending on his unique medical history. This can lead to better out-
comes and fewer side effects from treatment. Overall, ML has revolutionized in
the analysis and cure of leukemia and other types of cancer and is already mak-
ing a significant impact in the field of oncology. ML algorithms can be used for
detection of leukemia and it is easy to detect due to dark-purple color of cells. But
detecting leukemia without ML techniques is time-consuming and challenging.
Utilization of ML techniques for its diagnosis is increasing day by day. In this,
the literature of leukemia, causes, symptoms, different types of leukemia and ML
algorithms used for leukemia diagnosis are given.

Keywords: leukemia · ML · DL white blood cells (WBC) · red blood cells

(RBC) · convolutional neural network (CNN)

1 Introduction

Blood cancer is also called as hematological cancer. It refers to cancer affecting the
bone marrow, blood and the lymphatic system. Leukemia is also a kind of blood cancer
that occurs when there is an unusual production of immature white blood cells (WBC)
or blasts, in the bone marrow. These abnormal cells do not function normally and can
interfere with the construction of normal blood cells, causing to symptoms such as

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 1–15, 2025.
https://doi.org/10.1007/978-3-031-71481-8_1
2 S. Gupta et al.

anemia, bleeding and susceptibility to infections. Blood cancers can also affect other
types of blood cells, containing red blood cells (RBC) and platelets. There are Numerous
types of blood cancer. Each type of blood cancer has different characteristics and may
require different treatment approaches. In recent years, immunotherapy has also emerged
as a promising treatment option for some types of blood cancer. Treatment decisions are
based on several issues like the kind and phase of the cancer, the age of patient’s, overall
health and other individual factors. There are different types and forms of leukemia that
are found and common in kids and adults. Leukemia generally occurs in the WBC which
are strong fighters of infection and usually they develop and split in a systematic way
as our body needs them. But if a person suffers from leukemia, then his bone marrow
yields an extreme quantity of unusual WBC and they generally don’t work well.
The (IoMT) Internet of Medical Things is a web of connected sensors, medical
devices and software applications that can gather and transfer real-time data about a
patient. These devices can be connected to the body of the patient, such as wearables
or implanted sensors or be located in a hospital or clinic [1–3]. IoMT devices captures
patient physiological signs and patient’s activity effectively [4]. An IoMT device is a
small chip implanted in a watch, cloths or any type of same items. These are connected
to a transmission device and collects data from sensor [5]. IoMT devices can also be used
to collect cancer data, containing changes in breast, skin, lung and other abnormalities
[6]. Detection of leukemia and other types of cancer can take time and it is challenging
process. Early detection is critical for effective treatment and improved patient outcomes.
ML and DL techniques have shown promise in helping to advance the precision and
proficiency of cancer diagnosis and classification.
IoMT-enabled DL models can be used to analyze large volumes of medical data,
including blood smear images and to identify patterns and features that are characteristic
of different types of cancer. These models can be trained using large datasets of labeled
images. New images are then classified with a high degree of accuracy. By using IoMT-
enabled DL models, medical professionals can diagnose and classify leukemia and other
types of cancer more quickly and accurately than traditional methods. This can help to
reduce the risk of late-stage diagnosis. These models are like a tool to support their
decision-making process. Any medical diagnosis or treatment plan should always be
made by a qualified healthcare professional.

2 Literature Survey

Leukemia cancer has a very high death rate [7]. It is also called as a malignant hematolog-
ical tumor [8] as it causes cloning of abnormal WBC. Analysis of WBC is an important
part of medical diagnostics as WBCs can provide important information about a patient’s
immune system and can help to identify a range of diseases, including leukemia [9].
Analysis of WBC is an significant part of medical diagnostics and the use of automated
methods and ML/DL techniques can benefit to increase the correctness and proficiency
of this process. Assessment and further processing of WBC can be challenging due to
their natural variations in shape, size and texture as well as their proximity to other
elements of the blood. In the case of leukemia, different kinds of WBC can have charac-
teristic morphologies and features that can be used to identify and classify the disease.
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 3

However, accurately identifying these features can be challenging due to the existence
of other elements in the blood and the natural variations in cell shape and size.
Acute lymphocytic leukemia (ALL) patients often have lymphocytes with a minimal
uniform border and small holes in the cytoplasm called vacuoles, as well as spherical
particles inside the nuclei. These features can become more prominent as the disease
progresses and early detection is critical to improve the prognosis and reduce the risk
of premature death. Age is also a significant risk factor for ALL, with a higher proba-
bility of the disease occurring in children aged 7 to 8 years. The incidence of ALL in
the United States was around 5930 new cases in 2018 and approximately 1500 persons
(containing both children and adults), were expected to expire from the disease. Improv-
ing the accuracy and efficiency of leukemia diagnosis and classification using advanced
technologies such as ML and DL could help to improve patient outcomes and reduce
mortality rates.
Wearable IoMT devices proficiently track variations in the body of a human without
causing allergies [10]. One such device is wearable iTbra, an IoT gadget. iTbra is a bra
that detects breast cancer. That’s why the iTBra comes into play. It’s a wearable and smart
device used to detect changes in the temperature in breast tissues. Seventy biopsies have
been recognized by ML from the data that is being generated from iTbra [11]. Using
this data, healthcare providers measure the risk of breast cancer based on the patient’s
daily practices, habits, cavities, health and other data [12]. One model proposed by the
authors in [13] and it aims to diagnose the stage of breast cancer using data collected
through Internet of Medical Things (IoMT) devices.
To accomplish this, the authors have used a machine learning technique, specifically a
convolutional neural network (CNN), to analyze the medical data. One of the challenges
in analyzing medical data is identifying anomalous patterns and attributes, which can
be difficult to accurately calculate. To address this, the authors have used a CNN with
augmented features, which allows for more accurate and robust analysis of the data [14].
Additionally, the authors have employed hyper-parameter optimization techniques to
fine-tune the model and improve its accuracy. By using this model, the authors hope to
improve the diagnosis and prediction of cancer, as well as provide support for disease-
related psychological choices. Overall, the proposed model represents an important step
forward in ML techniques uses for medical analysis and treatment.
Several strategies been proposed for identification of automated leukemia including
traditional ML classifiers and DL algorithms. In the case of traditional machine learn-
ing approaches, the methods typically involve numerous steps such as preprocessing,
feature extraction and classification. These approaches have been shown to be effective
in identifying leukemia cancer on microscopic images. Deep learning approaches have
added popularity in recent years because of their ability to learn and extract features
automatically from raw data. Many researchers have designed various architectures for
the automated classification of leukemia cancer using deep learning. In addition to these
approaches, some researchers used ensemble ML and hybrid DL methods for detection
of leukemia cancer. These approaches have shown promising results in accurately iden-
tifying leukemia cancer subtypes. Overall, the combination of traditional ML and deep
learning approaches has led to significant progress in the automated identification of
leukemia cancer on microscopic images.
4 S. Gupta et al.

The authors of [15] have proposed a framework based on Internet of Medical Things
(IoMT) to enhance the identification of leukemia, which addresses the limitations of
existing methods. It also includes comparison of the different approaches with respect
to average accuracy, used for automated leukemia detection and its subtypes. In the pro-
posed system, medical devices are connected to network resources via cloud computing.
The proposed framework also offers a solution for patients having critical conditions
in pandemics such as COVID-19. Three methods, namely Dense Convolutional Neural
Network (DenseNet-121), Residual Convolutional Neural Network (ResNet-34) and a
hybrid model combining both DenseNet-121 and ResNet-34, are used for the identifi-
cation of subtypes of leukemia in the proposed framework. Two datasets for leukemia
that are publicly available are ALL-IDB and ASH image bank and are used in this study.
The results of the experiments demonstrate that the suggested models outperform other
famous ML algorithms that can be used for identification of healthy or leukemia infected
person. The accuracy achieved is 99.91, which shows promising results for the leukemia
identification.
A comparative study of ML algorithms like Support Vector Machines, Naïve Bayes,
Neural Networks, k-Nearest Neighbour and Deep Learning algorithms used to classify
leukemia is given in [16]. IoT based Leukemia Detection Techniques is given in [17].
The use of deep learning algorithms for detecting leukemia from microscopic images of
blood samples has gained significant attention in recent years. The proposed model in
[18] utilizes squeeze and excitation learning to enhance the representation ability of the
model at every level of feature representation. This model was tested on two publicly
available datasets, ALL_IDB1 and ALL_IDB2 with promising results when compared
to the traditional deep learning model. Accuracies, precision and other metrics have been
compared in this proposed model on both datasets ALL_IDB1 and ALL_IDB2.
The use of computer-aided diagnosis can help in quick and accurate detection of
leukemia, which can aid in early diagnosis and prompt treatment. A comparative analysis
of image processing algorithms for detecting cancer and analysis of patient’s health
monitored using IoT is given in [19]. This work is very useful in developing the new
image processing applications that are IoT based.

3 Symptoms, Risk Factors and Causes of Leukemia Formation

Leukemia symptoms are different based on the types of leukemia. Some of the main
leukemia signs and indicators are: weakness, feeling tired and persistent fatigue, anemia,
loss in weight and appetite, extreme and mysterious staining, light-headedness, pale
skin color, small red spots because of low platelet levels, testicular expansion, lethargy,
stiffness in the neck, vomiting and headache, recurrent infection and fever, pain in bones
and joint pain, large lymph nodes, recurrent nosebleeds, extreme sweating, especially at
night [20–22]. Blood is important for preserving homeostasis. This means temperature
regulation, hydration and ion concentration. The main tasks of blood are: oxygen supply
from the lungs to body parts, carrying of CO2 to the lungs, distribution of nutrients to all
body parts, transportation of Waste products from cells to the outside through kidneys,
maintain electrolyte balance, defending the body against attack from extraneous entities
via WBC and antibodies, guarding the body against wound or illness, Stopping serious
haemorrhage by the clotting process, maintaining the body’s temperature etc.
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 5

WBC gives significant data and plays a vital role in the analysis of different dis-
eases. WBC are classified into 5 kinds: Basophil, Monocyte, Lymphocyte, Neutrophil,
Eosinophil. These cells gives defense against infections. Generally leukemia occurs
when some of the cells of blood get changes in their genetic material or DNA. DNA has
instructions for a cell. In leukemia, the blood cells to continue grow and divide. Blood
cell production becomes out of control at this stage.
After some time, the number of abnormal cells in the bone marrow become more as
compared to the healthy blood cells. So fewer healthy WBC, RBC and platelets remains
in the body. This all are the causes of leukemia. Some of the main risk factors for
leukemia are given below. Exposure to high radiation, like those experienced in nuclear
accidents, can damage the DNA in blood cells, which can lead to the development of
cancer. Smoking can cause blood cancer, as the chemicals in cigarette smoke, including
benzene, can damage the DNA in blood cells and increase the risk of cancer. Exposure
to some solvents and benzene in industry, can also increase the risk of developing blood
cancer, as these chemicals can damage the DNA in blood cells and lead to cancer. Some
cancer treatments like radiotherapy and chemotherapy can also surge the danger of rising
blood cancer, particularly if certain drugs are used in combination with radiation. This is
known as treatment-related blood cancer. Blood disorder problems like myelodysplasia
or myeloproliferative disorders and genetic disorders, increases the risk of developing
AML.

4 Classification of Leukemia
Leukemia is a cancer of WBC. Hematologists can use blood counts to help in the diagnose
of leukemia. The French-American-British (FAB) classification system is a widely used
method for classifying leukemia [23]. Mainly four kinds of leukemia are there. Acute
Lymphoblastic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lympho-
cytic Leukemia (CLL) and Chronic Myeloid Leukemia (CML). ALL is found more
in children, while AML is mostly found in adults. AML can be classified into eight
sub-types (M0-M7) and ALL can be classified into three sub-types (L1-L3) using this
system. Morphological characteristics, such as size of cell, importance of nuclei, cell
color, and quantity and form of cytoplasm, are used to differentiate between the subtypes
of leukemia [23]. Figure 1 provides a pictorial representation of blood structure and the
different types of leukemia. Table 1 provides an overview of the FAB classification
system.
Leukemia is classified based on its speed of development and the type of cells
involved. According to speed of leukemia progression, it is classified into two types:
Acute Leukemia: The unusual cells in the blood are immature blood cells in acute
leukemia. They cannot perform their usual tasks and they increase speedily. Acute
Leukemia degrades so rapidly that it needs destructive and timely treatment.
Chronic Leukemia: Chronic leukemia are of several kinds. Too many cells or too few
cells to be created in chronic leukemia. Chronic leukemia has more-mature blood cells
which duplicate or add more slowly and work normally for a period of time. No early
symptoms are produced initially in some forms of chronic leukemia. So it can lead to
unobserved or undiagnosed for many years.
6 S. Gupta et al.

Fig. 1. Blood structure and different leukemia types [23]

According to type of WBC affected, it is classified into two types:

Lymphocytic Leukemia: It affects the lymphocytes. Lymphocytes forms lymphatic
tissue which help in making strong immune system.
Myelogenous Leukemia: It affects the myeloid cells which increases the RBC, WBC
and platelet-producing cells.
Based on the speed of leukemia progression and type of WBC affected, there are
four major types of leukemia:
Acute Lymphocytic Leukemia (ALL): ALL is identified by the speedy growth of
immature WBC in the bone marrow and blood. Symptoms of ALL may comprise fever,
fatigue, weakness, and easy bruising or bleeding [8].
Acute Myelogenous Leukemia (AML): It starts in the bone marrow and rapidly pro-
gresses. It is the most common type of acute leukemia in adults and can also occur in
children. Symptoms may include fatigue, fever, and easy bruising or bleeding [24].
Chronic Lymphocytic Leukemia (CLL): This is very slow-growing cancer and it
affects the WBC. CLL is the most common type of leukemia in adults. In the early
stages, a person with CLL may not have any symptoms, but as the disease progresses,
they may experience fatigue, swollen lymph nodes, and recurrent infections.
Chronic Myelogenous Leukemia (CML): It starts in the bone marrow and can
progress slowly or rapidly. It mainly affects adults and is characterized by the pres-
ence of a specific genetic mutation. In the early stages, a person with CML may not have
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 7

any symptoms, but as the disease progresses, they may experience fatigue, fever, and
abdominal discomfort.

Table 1. FAB Classification of leukemia

Category Name General Description

ALL L1 Small monotonous lymphocyte
ALL L2 Small monotonous lymphocytes
ALL L3 Large homogeneous cells
AML M0 Acute myeloblastic leukemia, undifferentiated
AML M1 Acute myeloblastic leukemia, without maturation
AML M2 Acute Myeloblasts with maturation
AML M3 Acute promyelocytic leukemia
AML M4 Acute myelomonocytic leukemia
AML M5 Acute monocytic leukemia
AML M6 Erythroleukemia/DiGuglielm o syndrome
AML M7 Acute megakaryoblastic leukemia

It is important for individuals to discuss treatment options with their healthcare

provider and to have a support system in place throughout their treatment journey.
There may be other types of leukemia including myelodysplastic syndromes,
myeloproliferative disorders, hairy cell leukemia etc.

5 Risk Factors in Leukemia

Some factors are there that can also increases the risk of developing some leukemia types
[25]. These includes:
Previous Cancer Treatment: Use of chemotherapy and radiation therapy in cancer
treatment can increase risk of leukemia. These treatments can damage the DNA in blood-
forming cells in the bone marrow, leading to the development of leukemia. The risk of
developing leukemia after cancer treatment varies depending on the type of treatment
received and other factors, such as age and overall health. However, it’s important to note
that the profits of cancer treatment generally outweigh the risk of developing leukemia.
Patients who have undergone cancer treatment should discuss their risk of developing
leukemia with their healthcare provider and follow up regularly for any signs or symp-
toms of leukemia. Regular follow-up can help detect leukemia early and improve the
chances of successful treatment.
Genetic Disorders: Certain genetic irregularities and disorders can rise a risk of devel-
oping leukemia in a person. Down syndrome is one example of a genetic disorder and
8 S. Gupta et al.

it is related with an improved threat of leukemia. Some genetic syndromes like Bloom
syndrome, Fanconi anemia and ataxia-telangiectasia also increases the risk of leukemia.
These disorders can affect the DNA repair mechanisms in cells and increase the risk
of DNA damage and mutations, which can lead to the development of leukemia. In
some cases, leukemia can also be caused by genetic mutations that occur spontaneously
during the development of blood cells in the bone marrow. These mutations can have
uncontrolled growth and division of irregular white blood cells, which can eventually
develop into leukemia.
Exposure to Chemicals: Exposure to some chemicals can rise the danger of some
kinds of leukemia. For example exposure to benzene (Found in gasoline) by the chemical
industry can increases the risk of particular type of leukemia.
Smoking: Smoking cigarettes increases the risk of developing AML, which is a type of
leukemia that starts in the bone marrow and speedily progresses. Studies have found that
smoking cigarettes can increase the risk of developing AML by up to 30%. The chemicals
in cigarette smoke can damage the DNA in blood cells, leading to genetic mutations that
can contribute to the development of leukemia. In addition to AML, smoking cigarettes
increases the risk of lung cancer, bladder cancer and pancreatic cancer.
Family History of Leukemia: If a member of a family has been detected with leukemia,
the risk of developing the leukemia may be increased in other members. Studies have
shown that having a first-degree relative (parent, sibling, or child) with leukemia can
increase the risk of developing the disease by two to four times.

6 Diagnosis of Leukemia, Observation and Results

A physical test or examination is the first process to see and confirm for leukemia. The
physical test can be attentive on observing or feeling for bump in the lymph nodes,
liver and spleen. Sometimes swelling can also be observed in eyes, mouth and skin.
Doctor might monitor for symptoms of infection bleeding because these are associated
with some kinds of leukemia. A test of nervous system can be done to check balance,
sensation, reflexes etc. in the body [24].
Blood Tests: Blood tests and bone marrow tests, are the most significant indicative
tools for leukemia. A complete blood count (CBC) test is often the first step in detecting
leukemia. This test measures the number and types of cells in the blood, including
RBC, WBC, and platelets. In leukemia, the number and appearance of these cells can
be abnormal, which can provide clues to the presence of the disease. If the CBC test
suggests the presence of leukemia, a bone marrow biopsy and aspiration may be used
to confirm the diagnosis. In this, a small sample of bone marrow and a small amount
of fluid from inside the bone are removed and observed under a microscope to look for
abnormal cells.
Complete Blood Count (CBC): CBC is a common blood test to measures the different
types of cells in the blood, including RBC, WBC, and platelets. RBC carry oxygen
throughout the body, and a low red blood cell count (anemia) can cause fatigue, weakness,
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 9

and other symptoms. WBCs are part of the body’s immune system, and abnormal levels
of these cells can be a sign of infection or other conditions. Platelets help the blood clot
and prevent excessive bleeding, and low platelet counts can increase the risk of bleeding.
In leukemia, the CBC may show abnormal levels of these blood cells, which can provide
important clues to the presence and type of the disease. For example, acute lymphocytic
leukemia may cause low RBC counts and platelet counts, while chronic lymphocytic
leukemia may cause high white blood cell counts. However, CBC alone is not sufficient
to diagnose leukemia, and further testing is usually needed to confirm the diagnosis. If
leukemia is suspected, additional blood tests and a bone marrow biopsy may be done to
look for abnormal cells and other signs of the disease.
Differential: A differential blood count, or simply a “diff,” is blood test to measures
the proportions of different types of WBC in the blood. WBC are significant part of the
body’s immune system, and they help in fighting infections and other foreign invaders.
Different types of WBC are there and they plays a different role in the immune response.
In leukemia, the differential may show abnormal levels of one or more types of white
blood cells, which can provide important information about the type and stage of the
disease. For example, acute myeloid leukemia (AML) may cause high levels of immature
white blood cells called blasts, while chronic lymphocytic leukemia (CLL) may cause
high levels of mature lymphocytes.
Peripheral Blood Smear: A peripheral blood smear is a blood test that involves placing
a drop of blood on a slide, spreading it out in a thin layer, and then examining it under
a microscope. This test allows doctors to look at the size, shape, and structure of the
different types of blood cells, including RBC, WBC and platelets. In leukemia, the
appearance of the blood cells on the peripheral blood smear can provide important clues
about the type and stage of the disease. For example, in acute lymphocytic leukemia
(ALL), the blood cells may appear immature and abnormal, while in chronic lymphocytic
leukemia (CLL), the cells may appear normal but there may be a high number of mature
lymphocytes. The results of the CBC, differential and peripheral blood smear, along
with other diagnostic tests, can help doctors to diagnose leukemia and determine the
best course of treatment. It is important to note that while these tests are important in
the diagnostic process, a definitive diagnosis of leukemia often requires additional tests,
such as bone marrow biopsy and genetic testing.
Flow Cytometry: A flow cytometry test gives an important awareness about normal or
abnormal amount of DNA in the tumor cells. It also shows the speed at which the tumor
is increasing.
Flow cytometry test is a reliable way to determine the exact type of lymphoma or
leukemia a patient has. The test involves using special antibodies to attach to specific
proteins on bone marrow or blood cells collected during a biopsy or blood test. If anti-
bodies attaches to most of the cells in the sample, it is likely that cancer initiated from
one abnormal cell. If different antibodies attach to different cells, it is less likely to be
cancer. The test may also reveal the amount of DNA in cancer cells.
Biopsy: A biopsy is used to decide the category of leukemia that arises, the progress
of the tumor and level of spread. Commonly two types of biopsy procedures are used
for leukemia. A bone marrow biopsy is a process to eliminate a sample of bone marrow,
10 S. Gupta et al.

which is important for diagnosing leukemia as the disease originates in bone marrow.
Two procedures are usually done together to obtain the sample. During bone marrow
aspiration, a patient lies on a table and a doctor injects a numbing medication in the area
that is being tested, typically the hip bone, and then inserts a needle to collect a sample
of liquid bone marrow. During bone marrow biopsy, a small piece of the bone and bone
marrow is collected by inserting a larger needle into the bone and twisting it to collect the
sample. While these procedures may cause some pain, the numbing medication can help
in reducing pain. Most of the leukemia types require bone marrow testing for diagnosis.
However, CLL leukemia can only be diagnosed using blood tests.
Lymph Node Biopsy: It involves the removal of some part or complete lymph node.
This procedure is necessary for diagnosing leukemia, as blood and bone marrow tests are
much more significant. There are two types of lymph node biopsies: excisional, which
involves removing the complete lymph node and incisional: which involves removing a
portion of a lymph node. General anesthesia may be necessary to remove lymph nodes or
part of a lymph node that are deep in to the body, while local anesthetic may be sufficient
for more accessible lymph nodes.
Imaging Tests: Imaging tests are essential tools in diagnosing leukemia, as they can
give information about the level of leukemia, presence of infections or other problems
in the body. Some of the imaging tests used for leukemia diagnosis are given below:
X-rays: Inside images of the body is generated in X-Rays and may be used in some
leukemia patients to check for lung infections.
Computed Tomography (CT) Scans: CT Scans create detailed image of the body and
can help in finding swelling in the lymph nodes and organs. Doctors may use this test
to identify cancer, but it is not typically necessary for diagnosing leukemia.
Positron Emission Tomography (PET) Scan: PET scans are used together with CT
scans to give more precise cancer images. Before starting a PET scan, a material is
injected into the body that in areas of cancer cells. CT scans are more detailed than PET
scans alone.
Magnetic resonance imaging (MRI) scans: MRI scans create detailed pictures of the
body and are needed if a patient has symptoms of cancer that has spread to the brain or
spinal cord.
Ultrasounds: This imaging test produces pictures by sending sound waves into the
body. Ultrasound is helpful to check if a patient’s organs like liver and lymph nodes are
engorged and indicate the existence of cancer.
Pulmonary Function Test: A pulmonary function test is a useful tool to evaluate
lung function before or after treatment for a medical condition. It helps the care team
in determining whether the lungs are strong enough for treatment and how current
treatments may affects the lungs. There are different methods of pulmonary function
testing like lung diffusion testing, spirometry and lung plethysmography.
Lumbar Puncture: A medical procedure commonly referred to as a spinal tap is used
to determine the extent of leukemia. This test involves collecting cerebrospinal fluid
from the lower back using a needle. In addition to diagnosis this test can also be used to
inject medications for leukemia treatment.
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 11

Following blood tests and bone marrow biopsy to diagnose leukemia, patients should
follow up with their doctor. The can doctor describes the results, containing the levels of
different types of blood cells in the samples and whether leukemia cells were detected and
the type of leukemia. This information is important for creating a personalized treatment
plan after diagnosis. If the initial tests do not provide enough detail to determine the type
of leukemia, additional testing may be necessary.

7 ML Algorithms Used in Leukemia Treatment

By microscopic study of human blood and a set of methods (like color imaging, image
segmentation, classification and clustering), identification of leukemia in a patient is
possible. Cancer cells are detected in images using ML. Classification of leukemia of
a particular type is the main concern as different types of leukemia requires different
treatment. All subtypes of Leukemia shares wide similarities. Due to which identification
of subtypes of leukemia under a microscope is challenging and time taken. Comparison
of different ML Algorithms for diagnosis of Leukemia, methodology used for leukemia
diagnosis, limitations of the existing methodology etc. are given in [16].
SVM:- It is a binary classifier algorithm and it classifies sample blood images and mark
them as Leukemia. Authors in [9] Uses SVM and achieved an accuracy of 92%. For every
nucleus its Shape and other features are taken and noted down. The utmost important
features are selected to train the SVM. Some relevant features are selected like the ratio
of nuclei to cell, the number of nuclei lobes, entropy etc.
K-Nearest Neighbour (KNN): It is a lazy classifier learning algorithm. It is a non-
parametric technique. Instead, it finds the k nearest neighbors of a new data point based
on a similarity metric and allots the class label of the majority of these neighbors to
the new data point. The study by Subhan et al. [26] found that k-NN is a scalable and
effective classifier tool for differentiating between leukemic cells and normal blood cells.
This is because the k-NN algorithm is able to identify similarities between cells based
on their features, which can help to accurately classify the cells. Similarly, the study by
Supardi et al. [27] uses k-NN to classify blasts and achieved an accuracy of 80%. This
study takes out 12 main features from leukemic blood images to denote size, color, and
shape, and tested various values of k and distance metrics. The results showed that a
value of k = 4 and cosine distance metric provided good classification results. Overall,
k-NN is a powerful algorithm for classification tasks, including in the field of medical
imaging for identifying and classifying different types of cells.
Neural Networks: Neural networks are a type of ML algorithm and are motivated by
the function of the human brain. They comprise of layers of interconnected nodes, which
can learn to extract features from data and make predictions based on those features. In
the study by I. Vincent et al. (2014) [28], neural networks were used for classification
of blood smear images into normal and leukemic blood cells.
Similarly, the study by M. Adjouadi et al. [29] also used an artificial neural network
(ANN) to categorize blood images as normal, AML or ALL. This study found that accu-
racy in the classification improves as data size increases, suggesting that larger datasets
12 S. Gupta et al.

can lead to more accurate classification results. Overall, the use of neural networks and
artificial intelligence (AI) in the leukemia diagnosis and other medical conditions is
a promising area of research. These techniques have the potential to progress in the
accuracy and speed of diagnosis, and can aid in the development of more personalized
treatment plans for patients.
Naïve Bayes: The Naïve Bayes classifier is a probabilistic classifier that makes use of
Bayes Theorem and Naïve assumptions. In the Naïve assumption value of one features
is independent of other features and each contributes individually to the probability. This
makes it a simple and efficient classifier and it needs very small datasets for training, and
it can guess many factors necessary for classification. In the study by A. Gautam et al.
[30], the Naïve Bayes classifier was used to classify leukocytes. The classifier works by
first calculating the mean and variance for each feature of each class. These values are
assumed to be independent variables and saved. By using the Naïve Bayes classifier, the
study was able to classify leukocytes with very high accuracy. This approach provides a
useful tool for diagnosing various diseases and conditions that can be detected through
analysis of leukocyte morphology.
Deep Learning: Convolutional Neural Networks (CNNs) have become a popular
choice for processing biological image data due to their ability to learn discrimina-
tive features from the image data without requiring manual feature extraction. In the
work done by A. Rahman et al. (2019) [31], CNNs were used to classify blood images
into different subtypes of ALL or normal types. The dataset used in this study included
blood cell images from patients with ALL and normal individuals. The Alexnet model
was used with CNN for classifying normal types and ALL subtypes. The last fully con-
nected layer combines all the learned features and is trailed by the output layer which
uses the softmax function for classification. The CNN model learns the important fea-
tures directly from the input image data and classifies the blood cell images into different
subtypes of ALL or normal types with high accuracy.
Authors in [32], designed a DL model with a modified architecture for sensing
acute leukemia by using lymphocytes and monocytes mages. With 99% accuracy, it
detectes acute leukemia types, including ALL and AML. Authors in [33], recommend
an ensemble automated prediction method that uses four ML algorithms Support Vector
Machine (SVM), K-Nearest Neighbor (KNN), Naive Bayes (NB) and Random Forest
(RF). Dataset used is C-NMC leukemia from the Kaggle repository. Results show that
SVM with 90.0% accuracy outperforms as compared to other algorithms.

8 Conclusion
Leukemia is a deadly form of blood cancer affecting people of all ages and is one of the
leading causes of cancer-related deaths worldwide. It is associated with an increase in
immature lymphocytes, which damages the bone marrow and blood. Therefore, early
and accurate analysis is crucial for effective treatment and improving survival rates. Cur-
rently, manual analysis of microscopic blood samples is used to diagnose leukemia, but it
is slow, time-consuming, and less precise. Additionally, the similarity in appearance and
shape between leukemic and normal cells makes detection challenging. Deep learning
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 13

techniques utilizing (CNN) have shown promise in image classification, but there is still
room for improvement in their effectiveness, learning process, and performance.
The researchers use machine learning, deep learning and image processing to under-
stand the detection of leukemia cancer. Speed of analysis of leukemia samples can be
increased by combining deep learning and image processing. It will also reduce the
chance of causing more harm to cancer patients. By combining deep learning and image
processing, researchers can also categorize ALL cells according to their subtypes like
l1, l2, and l3.
Recent research has utilized deep learning strategies for leukemia detection, but
continuous improvement in deep learning algorithms is necessary. In future, leukemia
dataset can be comprehensive by adding new blood images of samples. For better per-
formances, new augmentation techniques can be utilized. Furthermore, the new func-
tionality of identifying the subclasses of each leukemia type and its treatment can be
added.

Disclosure of Interests. The authors declare that they have no conflicts of interest.

References
1. Juneja, S., Dhiman, G., Kautish, S., Viriyasitavat, W., Yadav, K.: A perspective roadmap for
IoMT-based early detection and care of the neural disorder, dementia. J. Healthc. Eng. 2021,
6712424 (2021)
2. Qureshi, F., Krishnan, S.: Wearable hardware design for the internet of medical things (IoMT).
Sensors 18(11), 3812 (2018)
3. Awotunde, J.B., Jimoh, R.G., Folorunso, S.O., Adeniyi, E.A., Abiodun, K.M., Banjo, O.O.:
Privacy and security concerns in IoT-based healthcare systems. In: Siarry, P., Jabbar, M.A.,
Aluvalu, R., Abraham, A., Madureira, A. (eds.) The Fusion of Internet of Things, Artifi-
cial Intelligence, and Cloud Computing in Health Care, pp. 105–134. Springer International
Publishing, Cham (2021). https://doi.org/10.1007/978-3-030-75220-0_6
4. Awotunde, J.B., Ayoade, O.B., Ajamu, G.J., AbdulRaheem, M., Oladipo, I.D.: Internet of
things and cloud activity monitoring systems for elderly healthcare. In: Scataglini, S., Imbesi,
S., Marques, G. (eds.) Internet of Things for Human-Centered Design: Application to Elderly
Healthcare, pp. 181–207. Springer Nature Singapore, Singapore (2022). https://doi.org/10.
1007/978-981-16-8488-3_9
5. Nayyar, A., Puri, V., Nguyen, N.G.: BioSenHealth 1.0: a novel internet of medical things
(IoMT)-based patient health monitoring system. In: Bhattacharyya, S., Hassanien, A.E.,
Gupta, D., Khanna, A., Pan, I. (eds.) International Conference on Innovative Computing and
Communications: Proceedings of ICICC 2018, Volume 1, pp. 155–164. Springer Singapore,
Singapore (2019). https://doi.org/10.1007/978-981-13-2324-9_16
6. Dwivedi, R., Mehrotra, D., Chandra, S.: Potential of Internet of Medical Things (IoMT) appli-
cations in building a smart healthcare system: a systematic review. J. Oral Biol. Craniofacial
Res. 12(2), 302–318 (2022)
7. Agaian, S., Madhukar, M., Chronopoulo, A.T.: Automated screening system for acute myel-
ogenous leukemia detection in blood microscopic images. IEEE Syst. J. 8(3), 995–1004
(2014)
8. Shafique, S., Tehsin, S.: Acute lymphoblastic leukemia detection and classification of its
subtypes using pretrained deep convolutional neural networks. Technol. Cancer Res. Treat.
17(3), 1–7 (2018)
14 S. Gupta et al.

9. Laosai, J., Chamnongthai, K.: Acute leukemia classification by using SVM and K-Means
clustering. In: Proceedings of the International Electrical Engineering Congress, pp. 1–4
(2014)
10. Legner, C., Kalwa, U., Patel, V., Chesmore, A., Pandey, S.: Sweat sensing in the smart wear-
ables era: towards integrative, multifunctional and body-compliant perspiration analysis. Sens.
Actuators A: Phys. 296, 200–221 (2019)
11. Manogaran, G., Shakeel, P.M., Hassanein, A.S., Kumar, P.M., Babu, G.C.: Machine learning
approach-based gamma distribution for brain tumor detection and data sample imbalance
analysis. IEEE Access 2, 12–19 (2018)
12. Yang, B., et al.: Nuclear magnetic resonance spectroscopy as a new approach for improvement
of early diagnosis and risk stratification of prostate cancer. J. Zhejiang Univ. Sci. B 8(11),
921–933 (2017)
13. Awotunde, J.B., et al.: An enhanced hyper-parameter optimization of a convolutional neural
network model for leukemia cancer diagnosis in a smart healthcare system. Sensors 22(24),
9689 (2022). https://doi.org/10.3390/s22249689
14. Ghaderzadeh, M., Asadi, F., Hosseini, A., Bashash, D., Abolghasemi, H., Roshanpour, A.:
Machine learning in detection and classification of leukemia using smear blood images: a
systematic review. Sci. Programm. 2021, 1–14 (2021). https://doi.org/10.1155/2021/9933481
15. Bibi, N., Sikandar, M., Din, I.U., Almogren, A., Ali, S.: IoMT-based automated detection and
classification of leukemia using deep learning. J. Healthc. Eng. 2020, 1–12 (2020). https://
doi.org/10.1155/2020/6648574
16. Maria, I.J., Devi, T., Ravi, D.: Machine learning algorithms for diagnosis of leukemia. Int. J.
Sci. Technol. Res. 9(1), 267–270 (2020)
17. Ananth Kumar, I., Suresh Kumar, K.: Analysis of IoT based Leukemia Detection Techniques.
Global J. Appl. Data Sci. Internet Things 6(1), 11–19 (2022)
18. El Hussein, S., et al.: Artificial intelligence strategy integrating morphologic and archi-
tectural biomarkers provides robust diagnostic accuracy for disease progression in chronic
lymphocytic leukemia. J. Pathol. 256(1), 4–14 (2022)
19. Bukhari, M., Yasmin, S., Sammad, S., Abd El-Latif, A.A.: A deep learning framework
for leukemia cancer detection in microscopic blood samples using squeeze and excitation
learning. Math. Prob. Eng. 2022, 1–18 (2022). https://doi.org/10.1155/2022/2801227
20. Logan Yagi, T., Sindhu, K., Sripriyadharshini, S.: Comparative analysis of white blood cell
cancer detection using image processing and IoT. Int. J. Adv. Res. Sci. Commun. Technol.
12(2), 198 (2021)
21. http://www.medicinenet.com/leukemia/article.Htm
22. http://en.wikipedia.org/wiki/Leukemia#Research_directions
23. Ugandhar, C., Hapalamadugu, D., Ojochenemi, R.C.: Leukemia – brief review on recent
advancements in therapy and management. Asian J. Res. Pharm. Sci. Biotechnol. 3(1), 12–26
(2015)
24. https://www.cancercenter.com
25. Kuntal, B., Prasun, C.: Detection and classification for blood cancer – a survey. Int. J. Comput.
Trends Technol. 36(2), 65–70 (2016)
26. Subhan, P.K.: Significant analysis of leukemic cells extraction and detection using KNN and
hough transform algorithm. Int. J. Comput. Sci. Trends Technol. 3(1), 27–33 (2015)
27. Supardi, N.Z., Mashor, M.Y., Harun, N.H., Bakri, F.A., Hassan, R.: Classification of blasts
in acute leukemia blood samples using k-nearest neighbor. In: IEEE 8th International
Colloquium on Signal Processing and its Applications, pp. 461–65 (2012)
28. Vincent, I., Kwon, K.-R., Lee, S.-H., Moon, K.-S.: Acute lymphoid leukemia classification
using two-step neural network classifier. In: 21st Korea-Japan Joint Workshop on Frontiers
of Computer Vision (FCV) (2015)
 Leukemia Insight: Illuminating Current Diagnoses and Forecasting Futures 15

29. Adjouadi, M., Ayala, M., Cabrerizo, M., et al.: Classification of Leukemia Blood Samples
Using Neural Networks. Ann. Biomed. Eng. 38, 1473–1482 (2010)
30. Gautam, A., Singh, P., Raman, B., Bhadauria, H.: Automatic classification of leukocytes
using morphological features and Naïve Bayes classifier. In: IEEE Region 10 Conference
(TENCON), pp. 1023–1027 (2016)
31. Rehman, A., Abbas, N., Saba, T., Rahman, S.I., Mehmood, Z., Kolivand, H.: Classification of
acute lymphoblastic leukemia using deep learning. Microsc. Res. Tech. 81(11), 1310–1317
(2018)
32. Sanam, A., Ahmad, H.N., Amin, B.S., Jalil, V.G., Sebelan, D.: A customized efficient deep
learning model for the diagnosis of acute leukemia cells based on lymphocyte and monocyte
images. Electronics 12(2), 322 (2023)
33. Almadhor, A., Sattar, U., Al Hejaili, A., Mohammad, U.G., Tariq, U., Chikha, H.B.: An
efficient computer vision-based approach for acute lymphoblastic leukemia prediction. Front.
Comput. Neurosci. 16, 1083649 (2022)
 Intelligent Conversational Chatbot: Design
Approaches and Techniques

Johnbenetic Gnanaprakasam(B) and Ravi Lourdusamy

Department of Computer Science, Sacred Heart College (Autonomous), Affiliated to

Thiruvalluvar University, Tirupattur District, Tamil Nadu, India
[email protected], [email protected]

Abstract. The simulation of human intelligence in machines is known as artifi-

cial intelligence. A conversational AI-powered program is referred to as a chat-
bot. Rule-based systems, machine learning, natural language processing, and deep
learning are employed in the creation of chatbots. It attempts to investigate the
design strategies and methods employed in the construction of conversational chat-
bots with intelligence. In this paper, we give an outline of the components involved
in building chatbots that can interact meaningfully with people. These elements
include natural language generation, natural language understanding, and natural
language dialogue management. The creation of personas, conversation flow, and
user experience are all thoroughly discussed. The study also examines several
chatbot-building methodologies, such as machine learning algorithms and dia-
logue management tools. The study outlines the benefits and drawbacks of various
design strategies, offering information on their applicability and effects on chat-
bot functionality. Researchers and practitioners can make wise choices to create
intelligent conversational chatbots that provide better user experiences by know-
ing these design methods and strategies. These chatbots will have conversational
skills similar to humans.

Keywords: Intelligent conversational chatbots · Design approaches ·

Development techniques · Natural language understanding · Dialogue
management

1 Introduction
Chatbots are conversational agents driven by AI that mimic human-like interactions
and provide automated support and engagement across a range of domains [1]. Various
techniques and methodologies are used in chatbot development approaches to design,
create, and deploy chatbot systems [2]. The rule-based approach, in which responses are
predefined based on rules and patterns [3], and the machine learning approach, in which
algorithms are used to train chatbots using data [4], are two popular ways. For flexibility
and accuracy, a hybrid approach combines rule-based and machine-learning approaches
[5, 6].
To create interesting and user-friendly conversational experiences, design strategies
like intent recognition, dialogue management, and context maintenance are used [7].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 16–29, 2025.
https://doi.org/10.1007/978-3-031-71481-8_2
 Intelligent Conversational Chatbot 17

While dialogue management makes ensuring that conversations move smoothly, intent
detection aids the chatbot in understanding the user’s request or intention. The chatbot
can remember and refer to earlier encounters thanks to context maintenance, resulting
in a seamless and customized experience. The methods for developing and designing
chatbots change as Artificial Intelligence (AI) and Natural Language Processing (NLP)
technologies mature [8]. Organizations can use chatbots to automate activities, enhance
customer interactions, and provide effective support services by utilizing these strategies
and methods.

2 Chatbot System
An artificial intelligence (AI) program called a chatbot is made to mimic human-like
discussions with users through text or speech interfaces [9]. To comprehend user inputs,
produce responses, and participate in meaningful interactions, it makes use of machine
learning and natural language processing techniques [10]. Virtual assistants, messaging
platforms, and customer service and support systems are just a few of the applications
and platforms where chatbots can be used.
The employment of chatbots has many advantages across a variety of industries.
Chatbots in customer service can offer round-the-clock assistance, respond to frequently
asked queries, and handle straightforward inquiries, lightening the load on human agents
and speeding up response times [9]. They can improve user ease and experience by help-
ing users find information, book reservations, or complete transactions [11]. Chatbots in
the healthcare industry can offer individualized health advice, support for mental health,
or even triage symptoms to allow quick access to information and help [10]. Educa-
tional chatbots can help students access course materials, deliver interactive lectures,
and respond to questions, enabling personalized and independent learning [11].
Intelligent conversational chatbots have drawn a lot of interest recently because of
their potential to completely change how people and computers interact. To improve
the capabilities of chatbots, researchers have concentrated on a variety of design and
development methods.
Natural language understanding (NLU), a vital component of conversational chatbot
creation, comes first. To increase the accuracy of language comprehension, Clark and
Perez-Beltrachini [12] emphasized the need for deep learning techniques in NLU. Effec-
tive NLP in chatbots is described using neural networks, such as sequence-to-sequence
models [13].
Smooth and contextually relevant talks are made possible by effective dialogue man-
agement. The importance of reinforcement learning approaches in enhancing conversa-
tion policies was underlined by Bordes et al. [14]. It was suggested that dialogue man-
agement should use a task-oriented approach to better enable the chatbot to comprehend
user goals and produce pertinent responses.
Numerous studies have looked into various methods for chatbot response develop-
ment. For human-like talks, Lewis and Fan [14] suggested a convolutional sequence-
to-sequence model that successfully collected contextual information. Ritter et al. [15]
examined the use of statistical models to produce replies based on user input, with a
focus on data-driven response generation.
18 J. Gnanaprakasam and R. Lourdusamy

Researchers have looked into several data sources and evaluation criteria to improve
chatbot performance. To train and assess chatbots, Budzianowski et al. [16] estab-
lished the MultiWOZ dataset, a sizable multi-domain task-oriented dataset. It enables
academics to create chatbots that can manage intricate interactions across multiple
domains.
Intelligent conversational chatbots have a crucial component called user modeling.
Different techniques for user modeling in dialogue systems were surveyed by Kim and
Eskenazi [17]. More individualized and customized interactions are now possible thanks
to the exploration of statistical approaches and machine learning techniques to model
user behavior, preferences, and intents.
Conversational chatbots must have a contextual understanding to interpret user inputs
and respond properly. A context-aware chatbot architecture was suggested by Li et al.
[18] that makes use of deep learning methods like recurrent neural networks to record
the situation and produce contextually appropriate responses. To enhance contextual
knowledge, it is crucial to represent long-term dependencies in discussions.
The majority of interactions with traditional chatbots take place via text. Recent stud-
ies, however, have looked at how to improve chatbot functionality by using multimodal
inputs like text, graphics, and audio. Zhang et al.’s [19] multimodal conversational agent
mixes textual and visual data to produce more insightful and interesting responses. To
efficiently handle and integrate multimodal inputs, deep neural networks, and multimodal
fusion approaches were investigated.
Chatbot performance is enhanced through transfer learning, particularly in situations
with little labeled data. The use of transfer learning techniques in chatbot generation was
examined by Chen et al. [20]. To transfer knowledge from extensive datasets to particular
chatbot topics, strategies for fine-tuning and the usage of pre-trained language models,
such as BERT, were investigated. The study showed how transfer learning might enhance
chatbot performance even with sparse data.
As chatbots proliferate across a range of industries, ethical questions about their
creation and application have surfaced. The biases and possibly discriminatory behav-
iors that chatbots may display as a result of biased training data or insufficient ethical
requirements were emphasized by Raji and Buolamwini [21]. To reduce ethical hazards,
it was determined how crucial it is to incorporate fairness, openness, and accountability
principles into the design and development of chatbots.
User experience (UX) is a key component of conversational chatbot success. Usabil-
ity, engagement, satisfaction, and other UX variables are all taken into account in the
complete framework Li et al. [22] established in their discussion of the evaluation of
chatbot UX. To measure the efficiency and user satisfaction of chatbot interactions, a
variety of assessment techniques were used, including user surveys, interviews, and
interaction log analysis.
In conclusion, the literature review emphasizes improvements in design and develop-
ment techniques for conversational intelligent chatbots. Deep learning for comprehend-
ing natural language, reinforcement learning for controlling discourse, and data-driven
response generation are some of the strategies that have been investigated in the area.
 Intelligent Conversational Chatbot 19

The creation and assessment of chatbot systems have been made easier by the avail-
ability of large-scale datasets and evaluation benchmarks. Additionally, user modeling
approaches have made it possible for discussions to be more flexible and personalized.
Future research in this area can concentrate on overcoming the constraints and dif-
ficulties in developing intelligent conversational chatbots, such as handling confusing
user input, strengthening natural and sympathetic responses, and improving long-term
context comprehension. Researchers can improve the creation of very powerful and
interesting conversational chatbot systems by continuously examining and enhancing
these methods and approaches.

3 Chatbots Structure
The fundamental components and interactions within a chatbot system are illustrated in
Fig. 2 [61]. A user interacts with the chatbot, which fundamentally comprises three key
elements: Natural Language Understanding (NLU) [25], Dialogue Management (DM)
[34], and Response Generation (RG) [62]. The user’s input is evaluated by the NLU, the
DM chooses the best response based on the input, and the RG generates the response.
As part of the dialogue, the chatbot may also communicate with external systems like
databases or APIs to retrieve or save information [63]. The important steps of processing
user input and producing responses are highlighted in Fig. 2, which offers a simplified
perspective of a chatbot’s operation [64].

Fig. 2. Chatbot Structure.

A variety of design methods and strategies for intelligent conversational chatbots

are covered in the literature review. The capabilities of natural language understanding
(NLU) have been improved through the use of deep learning techniques [23]. For efficient
dialogue management, reinforcement learning techniques like Q-learning and policy
gradients have been used [6]. Convolutional sequence-to-sequence models are useful for
response creation [25]. A useful tool for developing and testing multi-domain chatbots
is the Multi WOZ dataset [26]. Chatbot user modeling has made use of statistical and
machine-learning approaches [27]. For training dialogue policies, hybrid architectures
that combine supervised learning and reinforcement learning have been developed [13].
20 J. Gnanaprakasam and R. Lourdusamy

Chatbot systems have used information representation and reasoning techniques such as
semantic networks, ontologies, and knowledge graphs [28].
Chatbots have been tested to recognize and respond to human emotions via sentiment
analysis employing sentiment lexicons and machine learning models [29]. Utilizing rule-
based systems and attention mechanisms has improved interpretability [30]. Through
language detection and machine translation approaches, cross-lingual and multilingual
assistance has been made possible [31]. Data anonymization and safe communication
have been taken into account for privacy and security in chatbot systems [32]. To evaluate
chatbot performance, evaluation measures and approaches have been used [28]. Chat-
bot systems have been built on neural network architectures, including recurrent neural
networks and transformers [33]. Multimodal interactions and adaptive dialogue manage-
ment have been the main focuses of conversational UX design [34]. To improve chatbot
adaptability to new domains, domain adaptation approaches like transfer learning and
domain adaptation models have been used [35].

4 Chatbot Design Approaches and Techniques

Figure 1 presents a visual representation of the various design approaches and strategies
used in the context of chatbot development. The tasks that the chatbot must complete are
the focus of the goal-oriented design method. The chatbot’s objectives are first defined
during the design phase, after which conversation flows supporting these objectives are
created [54]. The user’s requirements and expectations are the main emphases of the
user-centered design approach. Understanding the user’s context, needs, and preferred
mode of interaction is necessary. Making personas, user stories, and user journey maps
is a common step in the design process [55]. Making the chatbot interaction feel natural
is the main goal of the conversational design method. It entails creating the personality,
timbre, and manner of talking for the chatbot. Natural language understanding (NLU)
and turn-taking are often employed techniques [56].

Fig. 1. Chatbot Design Approaches and Techniques.

Data and analytics are used in the data-driven design approach to guide the design
process. It entails reviewing user interactions with the chatbot and making use of the
resulting data to enhance the chatbot’s functionality. Machine learning approaches, user
feedback, and A/B testing are frequently employed [57]. The chatbot is regularly tested
and improved using an iterative design methodology. Launching the chatbot with a
minimally viable product (MVP), gathering user input, and then making adjustments in
response to this feedback are the steps involved [58]. The multimodal design method
 Intelligent Conversational Chatbot 21

entails creating chatbots that can interact via several media, including text, speech,
and visual components. Designing the chatbot to move between modes as necessary
requires understanding the advantages and disadvantages of each mode [59]. The goal
of the ethical design approach is to make sure the chatbot respects users’ privacy, offers
transparency, and steers clear of bias. It entails adding protections to protect users and
taking ethical considerations into account during the design process [60].

4.1 Chatbot Development Approaches

The different approaches and methodologies used to plan, create, and implement chatbot
systems are referred to as “chatbot development approaches.” Through the use of artifi-
cial intelligence and natural language processing, chatbots are computer programs that
mimic human communication. They are frequently employed in a variety of businesses
to perform duties like information retrieval, customer service, and personal help [23].
Understanding user requirements, establishing conversational flows, implementing
natural language understanding and generating capabilities, and interfacing with perti-
nent data sources or backend systems are just a few of the crucial factors to take into
account when developing a chatbot [24]. Different strategies can be used, depending on
the complexity of the chatbot and its intended functionality.
The rule-based technique, which predefines chatbot responses based on a set of
rules and patterns, is one of the most conventional methods. This method is simple and
effective for performing jobs with clear guidelines, but it may not be flexible enough to
handle complex or unusual questions [25].
The machine learning approach, which involves teaching chatbots with machine
learning algorithms, is an alternative strategy. These algorithms provide chatbots with
the ability to analyze user inputs, learn from data, and produce pertinent responses.
To improve the chatbot’s comprehension and response creation abilities, techniques
including natural language processing, sentiment analysis, and machine translation might
be used [26, 27].
Combining rule-based and machine-learning strategies, a hybrid approach takes
advantage of each methodology’s advantages. While machine learning methods allow
the chatbot to manage more complicated interactions and learn from user data, rule-based
components give precise and predetermined responses [13].
Additionally, effective chatbot design strategies are essential for developing con-
versational interfaces that are both user-friendly and interesting. The effective flow of
conversations and user pleasure are facilitated by strategies including intent recognition,
dialogue management, and context maintenance [28].
In Table 1, major chatbot design methods and techniques are summarized together
with their benefits, drawbacks, and potential applications. The rule-based method works
effectively for particular domains with clearly defined rules since it is straightforward
to build and maintain but is restricted to specified rules. Machine learning algorithms,
which are frequently utilized in general-purpose chatbots, excel at handling complicated
inputs but necessitate large amounts of training data. Natural Language Processing (NLP)
approaches are appropriate for chatbots that need advanced language processing since
they can comprehend human language nuance but may struggle with uncommon or
domain-specific terminology.
22 J. Gnanaprakasam and R. Lourdusamy

Table 1. Chatbot Development Techniques, Approaches, Advantages, and Limitations.

Technique/Approach Merit Limitations Applicability

Rule-based approach Simple to implement Limited to predefined Well-defined
[65, 66] and maintain rules domains with
specific rules
Machine learning (ML) Ability to handle Requires substantial General-purpose
[67, 68] complex and training data chatbots
ambiguous inputs
Natural Language Understands human May struggle with rare Chatbots with
Processing (NLP) [69, language nuances or domain-specific sophisticated
70] terms language
processing
Deep learning [71, 72] Capable of learning Requires significant Complex and
from vast amounts of computational open-ended
data resources conversational
scenarios
Reinforcement learning Ability to learn and Training can be Chatbots that
[9, 10] adapt through time-consuming improve over time
interactions through feedback
Generative models [73, Can generate coherent Training can be Chatbots focused
74] and contextually computationally on generating
relevant responses intensive responses
Hybrid approach [75, Combines the strengths Increased complexity Versatile chatbots
76] of multiple techniques and development effort that adapt to
various scenarios

Deep learning algorithms, which are good for complicated and open-ended dialogues
but require a lot of processing resources, learn from enormous volumes of data. Although
training can take some time, reinforcement learning enables chatbots to learn and adapt
through interactions. Although they require computationally demanding training, gener-
ative models produce coherent responses, making them appropriate for chatbots that are
response-focused. The versatility of hybrid systems comes at the expense of complexity
and extra development work.
The requirements, resources, and objectives of the project will determine the best
technique to use. The choice of an efficient chatbot system is guided by an evaluation
of benefits, constraints, and applicability in the project context.
In general, when new developments in artificial intelligence and natural language
processing technology appear, chatbot development methods and design strategies are
always changing. These methods seek to develop chatbots that can deliver precise and
contextually appropriate responses, enhance user experiences, and ultimately achieve
the goals of the chatbot system.
The rule-based approach in chatbot development, according to Ismail et al. [36],
entails developing a series of established rules and patterns to choose the suitable
 Intelligent Conversational Chatbot 23

response based on the user’s input. Shishow [37], on the other hand, contrasts rule-
based and self-learning approaches in chatbot development, highlighting the benefits
and drawbacks of each method.
Zhang et al.’s [38] investigation of several speech conversation system techniques It
was addressed how crucial it is to take into account various design strategies, such as rule-
based, machine-learning, and retrieval-based approaches, to create efficient chatbots.
An extensive survey of deep learning-based chatbot models is provided by Luan et al.
[39]. Deep learning algorithms like recurrent neural networks (RNNs) and transformer
models are specifically highlighted in this study’s analysis of the structures and methods
used in the creation of chatbots.
A thorough manual on chatbot development, including design methods, deployment
tactics, and best practices, is provided by the SnatchBot Team [40].

4.2 Chatbot Development Techniques

Cuadros-Munoz [41] claims that a lot of research has been done on chatbot creation meth-
ods for speech communication systems. These methods seek to enhance chatbots’ con-
versational ability and the user experience. Using conversational user interfaces (CUIs)
is a crucial component of chatbot design [2].
Design strategies that are specifically suited for CUI-based chatbots are covered by
Patra and Rath [42]. These methods center on developing engaging and interactive user
interfaces that encourage organic and meaningful user discussions.
Al-Subari and Ramli [43] emphasize the significance of adopting design strategies
that captivate people to create an engaging user experience. To increase user engagement,
the study emphasizes the necessity for chatbots to include qualities like empathy, humor,
and personalization.
Ezenwoke [44] also emphasizes the value of design strategies in enhancing user
engagement with chatbots. To keep people engaged in the dialogue, the study advises
using strategies like gamification, interactive storytelling, and tailored recommendations.
A thorough overview of design strategies for conversational agents, including chat-
bots, is provided by Jain [45]. The study covers a wide range of methods that help create
conversational agents that are useful and easy to use, including dialogue management,
user modeling, and natural language creation.

4.3 Machine Learning Approaches for Chatbot Development

An enormous dataset of conversational data is used by MC Approaches to train the

chatbot using machine learning methods. The data is used for training the chatbot, which
then can produce responses based on patterns it has identified. With ML, there is more
flexibility and a greater range of user inputs that can be processed.
Zhou, Cao, and Xiong [46] assert that machine learning techniques are essential for
chatbot development. Chatbots may understand patterns in data and make wise decisions
thanks to machine learning (ML) techniques.
Natural language processing (NLP) is a common machine-learning technology used
in chatbots [47]. NLP methods like tokenization, part-of-speech tagging, and syntactic
24 J. Gnanaprakasam and R. Lourdusamy

parsing aid in the comprehension and processing of human language by chatbots. These
methods are necessary for deriving meaning from user inputs and producing suitable
responses.
Algorithms for machine learning are frequently used in chatbot development. For
instance, text categorization and sentiment analysis tasks in chatbots have been per-
formed using support vector machines (SVMs) [46]. On the other hand, Naive Bayes
classifiers are frequently used to group user queries into predetermined intents or cate-
gories [46]. RNNs: Recurrent neural networks for sequence modeling, language produc-
tion, and dialogue management in chatbots, RNNs—in particular, variations like Long
Short-Term Memory (LSTM) and Gated Recurrent Units (GRU)—are frequently used.
The creation of chatbots has seen a rise in the popularity of deep learning systems.
For tasks like machine translation and conversation synthesis, Seq2Seq models with an
encoder and a decoder have shown promise [46]. To produce more realistic and coher-
ent chatbot responses, generative adversarial networks (GANs) have been utilized [46].
Deep learning architectures are essential for chatbot development because they allow for
more sophisticated language production and understanding capabilities. Several typical
chatbot architectures are as follows: Machine translation, summarization, and dialogue
synthesis in chatbots are all common uses for Seq2Seq (sequence-to-sequence) models,
which consist of an encoder and a decoder. Using a generator and discriminator network,
generative adversarial networks (GANs) can be used to generate chatbot responses that
are more believable and coherent. Chatbots can be trained through trial-and-error user
interactions using reinforcement learning (RL). By producing the right responses, chat-
bots can learn to maximize a reward signal, such as user satisfaction. Another crucial
method that can improve chatbot effectiveness is transfer learning. Chatbots can take
advantage of existing information and enhance their understanding and generation capa-
bilities by utilizing pre-trained models on large-scale datasets [48]. The use of artificial
neural networks (ANNs) is essential in chatbot development methods. Computer mod-
els known as ANNs were influenced by biological neural networks. The feedforward
neural network (FNN), which enables data to flow from input to output layers, is the
most popular type used in chatbots. Backpropagation can be used to train ANNs, altering
weights to reduce mistakes. The creation of chatbots also makes use of variations like
convolutional neural networks (CNNs) and recurrent neural networks (RNNs). While
CNNs are appropriate for grid-like data, RNNs can handle sequential data [49–52].
It’s vital to remember that the choice and mix of machine learning algorithms depend
on a number of variables, including the particular needs of the chatbot, the data at
hand, and the functionality sought. Together, these strategies help chatbots become more
capable and enable more efficient and organic interactions with users.

5 Discussions
The finest chatbot design strategies and methodologies are determined by many factors,
including the use case in question, the intended audience, the resources at hand, and
the desired results. For clearly defined cases, rule-based techniques provide simplic-
ity and precision, but they lack flexibility. Chatbots can handle a larger range of user
inputs and adapt to different circumstances thanks to machine learning and natural lan-
guage processing algorithms, but they need a lot of data and computational power to
 Intelligent Conversational Chatbot 25

do so. Although they can create responses that are similar to those of a human, gen-
erative models can also be imprecise and require careful adjustment. Although they
can be difficult in complex relationships, contextual awareness approaches can assist in
maintaining coherent talks. Although it takes more work, integrating external knowl-
edge improves precision and domain-specific understanding. Based on actual usage and
ongoing improvement, a strategy or approach’s effectiveness can be measured.

6 Conclusions and Future Works

The study on “Design or Development Approaches and Techniques for Building Intelli-
gent Conversational Chatbots” has drawn attention to the important developments made
in the area. It is clear from a thorough literature analysis that several strategies and
techniques have been investigated to improve the capabilities of chatbot systems. The
ability of chatbots to comprehend and produce more contextually relevant responses
has been greatly improved by the integration of natural language processing, machine
learning, and deep learning. Additionally, efforts have been made to fix issues, improve
conversation management, and incorporate outside information sources. Nevertheless,
there is still room for development in several sectors.
Several directions for further research in the design and development of conversa-
tional chatbots with intelligence may be identified based on the current body of research.
First and foremost, there is a need for more improvements in natural language creation
and processing methods. To help chatbots have more natural-sounding and contextu-
ally appropriate dialogues, this involves investigating sophisticated language models,
semantic understanding techniques, and dialogue coherence models. Second, it’s crucial
to investigate how multi-modal skills might be included in chatbots. Future studies might
concentrate on processing text, audio, pictures, and videos in a fluid manner, enabling
rich and dynamic discussions and letting people communicate using their preferred
modality. The user experience must also be improved via personalization.
Future research can look into ways to personalize chatbot interactions based on user
choices, behavior, and historical data, resulting in a more personalized and engaging
experience. Furthermore, ethical factors like as bias reduction, fairness, and privacy
preservation can be taken into account. Researchers can create frameworks and recom-
mendations to address these ethical issues and ensure responsible chatbot creation and
deployment. Future studies could concentrate on improving natural language compre-
hension and generation to enable more accurate and coherent conversations. To guarantee
responsible and trustworthy chatbot systems, ethical aspects such as bias mitigation and
privacy protection can be addressed.

References
1. Seneviratne, S.: Chatbot systems: a comprehensive review. IEEE Access 9, 8986–9005 (2021)
2. Jain, R., Gupta, R., Singh, P.: A survey on chatbot design techniques in e-learning environment.
Int. J. Adv. Res. Comput. Sci. 11(5), 92–96 (2020)
3. Devendra, B., Wadkar, A.: Comparative study of rule-based and machine learning approaches
for Chatbot development. In: 2018 International Conference on Communication, Informa-
tion & Computing Technology (ICCICT), pp. 1–6 (2018)
26 J. Gnanaprakasam and R. Lourdusamy

4. Dhawan, S., Khosla, N., Kumar, R.: Machine learning approaches for chatbot develop-
ment: a review. In: 2020 International Conference on Smart Electronics and Communication
(ICOSEC), pp. 1–6 (2020)
5. Villena, J., Collada-Pérez, S., Serrano, S., Gonzalez-Cristobal, J.: Hybrid approach combining
machine learning and a rule-based expert system for text categorization. In: Proceedings of
the 24th International Florida Artificial Intelligence Research Society, FLAIRS – 24 (2011)
6. Nigam, A., Khanna, R.: A hybrid approach for rule-based and machine learning-based
Chatbots. Int. J. Comput. Appl. 160(11), 1–5 (2017)
7. Serban, I.V., et al.: A Survey of Available Corpora for Building Data-Driven Dialogue
Systems. arXiv preprint arXiv:1512.05742 (2017)
8. Tsiakas, K., et al.: On the design of context-aware Chatbots. IEEE Internet Comput. 24(3),
24–33 (2020)
9. Camps, R.J.G.B., Esquivel, C.N.Z., Cruz, M.E.C., Herrera, A.T.: Design and implementation
of a chatbot system for customer service support. In: Proceedings of the IEEE International
Conference on Artificial Intelligence and Machine Learning, pp. 123–128 (2021)
10. Kumar, S.S., Swamy, R.M.: Intelligent chatbot for healthcare applications. IEEE Trans. Hum.-
Mach. Syst. 50(2), 112–118 (2020)
11. Carvalho, M.A.L., Prado, R.B.R., Rocha, A.M.: Chatbot-based educational system for person-
alized learning. In: Proceedings of the IEEE International Conference on Advanced Learning
Technologies, pp. 124–129 (2019)
12. Clark, J., Perez-Beltrachini, L.: What deep learning really needs to become useful. Commun.
ACM 62(4), 56–63 (2019)
13. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. In:
Advances in Neural Information Processing Systems, pp. 3104–3112 (2014)
14. Lewis, M., Fan, A.Y.: Convolutional sequence to sequence model for human-like dialogues.
In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics,
pp. 4394–4405 (2019)
15. Ritter, A., et al.: Data-driven response generation in social media. In: Proceedings of the
Conference on Empirical Methods in Natural Language Processing, pp. 583–593 (2011)
16. Budzianowski, P., et al.: MultiWOZ-a large-scale multi-domain wizard-of-Oz dataset for task-
oriented dialogue modelling. In: Proceedings of the 2018 Conference on Empirical Methods
in Natural Language Processing, pp. 5016–5026
17. Kim, Y., Eskenazi, M.: A survey of user modeling in spoken dialogue systems. Speech
Commun. 113, 63–82 (2019)
18. Li, Y., et al.: A context-aware chatbot framework with recurrent neural networks. IEEE Access
7, 148646–148657 (2019)
19. Zhang, Y., et al.: A multimodal conversational agent for chatbot-based customer service. In:
Proceedings of the 2018 IEEE International Conference on Multimedia & Expo Workshops
(ICMEW), pp. 1–6 (2018)
20. Chen, Q., et al.: A survey of transfer learning for chatbots: challenges, methods, and future
directions. IEEE Access 8, 21918–21930 (2020)
21. Raji, J., Buolamwini, T.: Actionable auditing: Investigating the impact of publicly nam-
ing biased performance results of commercial AI products. In: Proceedings of the 2019
Conference on Fairness, Accountability, and Transparency (FAT*), pp. 63–72 (2019)
22. Li, X., et al.: Towards a comprehensive framework for evaluating user experience with chat-
bots. In: Proceedings of the 2018 International Conference on Information Systems (ICIS),
pp. 1–10 (2018)
23. Rajput, N., Shukla, A.: Chatbot Systems: a Survey. In: 2018 2nd International Conference
on Trends in Electronics and Informatics (ICOEI), pp. 1119–1123. Chennai, India (2018).
https://doi.org/10.1109/ICOEI.2018.8553786
 Intelligent Conversational Chatbot 27

24. Chen, X., Liu, A., Yin, D.: A survey on dialogue systems: recent advances and new frontiers.
ACM Trans. Inform. Syst. 36(4), 1–44 (2017). https://doi.org/10.1145/3124895
25. Jurafsky, D., Martin, J.H.: Speech and Language Processing. Pearson (2019)
26. Sordoni, A., et al.: A neural network approach to context-sensitive generation of conversa-
tional responses. In: Proceedings of the 2015 Conference of the North American Chapter of
the Association for Computational Linguistics: Human Language Technologies, pp. 196–205.
Denver, Colorado, USA (2015). https://doi.org/10.3115/v1/N15-1152
27. Mihalcea, R., Strapparava, C.: Computational Models of Emotion and Personality in Natural
Language Processing. Cambridge University Press (2012)
28. Liu, J., et al.: How NOT to evaluate your dialogue system: an empirical study of unsupervised
evaluation metrics for dialogue response generation. In: Proceedings of the 2016 Conference
on Empirical Methods in Natural Language Processing, pp. 2122–2132. Austin, Texas, USA
(2016). https://doi.org/10.18653/v1/D16-1230
29. Liu, B.: Sentiment analysis and opinion mining. Synth. Lect. Hum. Lang. Technol. 5(1),
1–167 (2012)
30. Ribeiro, M.T., Singh, S., Guestrin, C.: Why can I trust you? Explaining the predictions of any
classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 1135–1144. San Francisco, CA (2016)
31. Arivazhagan, N., et al.: Massively multilingual sentence embeddings for zero-shot cross-
lingual transfer and beyond. Trans. Assoc. Comput. Linguist. 7, 597–610 (2019)
32. Cavoukian, A.: Privacy by Design: The 7 Foundational Principles. Information and Privacy
Commissioner of Ontario, Canada (2009)
33. Vaswani, A., et al.: Attention is all you need. In: Guyon, I., et al. (eds). Advances in Neural
Information Processing Systems 30. Curran Associates, Inc., pp. 5998–6008 (2017)
34. McTear, M., Callejas, Z., Griol, D.: The Conversational Interface: Talking to Smart Devices.
Springer (2016)
35. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22(10),
1345–1359 (2010)
36. Ismail, et al., N.M.: Design and evaluation of a rule-based Chatbot architecture. In: Proceed-
ings of the 2nd International Conference on Computer Science and Computational Intelligence
(ICCSCI 2017), pp. 285–290 (2017)
37. Shishow, A.: Chatbot Development: Rule-Based vs. Self-Learning. Chatbots Magazine
(2019). https://chatbotsmagazine.com/chatbot-development-rule-based-vs-self-learning-1f1
6f16c893
38. Zhang, X., et al.: A survey on Chatbot design techniques in speech conversation systems.
Appl. Sci. 10(7), 2357 (2020)
39. Luan, Y., et al.: Deep Learning-Based Chatbot Models: A Survey. arXiv preprint arXiv:1910.
05402 (2019). https://arxiv.org/abs/1910.05402
40. SnatchBot Team: Chatbot Development: A Comprehensive Guide. SnatchBot, 2020. https://
snatchbot.me/blog/chatbot-development-a-comprehensive-guide
41. Cuadros-Munoz, A.: A literature review of chatbot design techniques in speech conversation
systems. Int. J. Adv. Comput. Sci. Appl. 9(1), 83–92 (2018)
42. Patra, G.K., Rath, S.K.: Design techniques for conversational user interface (CUI) based
Chatbots. In: Proceedings of the 2018 International Conference on Advances in Computing,
Communications and Informatics (ICACCI), pp. 1134–1139 (2018)
43. Al-Subari, H.B., Ramli, A.R.: Chatbot design techniques for an engaging user experience. In:
Proceedings of the 2019 International Conference on Electrical Engineering and Informatics
(ICEEI), pp. 521–526 (2019)
44. Ezenwoke, A.: Design techniques for improving user engagement with Chatbots. In: Proceed-
ings of the 2019 12th International Conference on Developments in eSystems Engineering
(DeSE), pp. 231–235 (2019)
28 J. Gnanaprakasam and R. Lourdusamy

45. Jain, S.: Design techniques for conversational agents: a survey. ACM Trans. Interact. Intell.
Syst. 9(4), 1–40 (2020)
46. Zhou, J., Cao, Z., Xiong, C.: A review of machine learning methods for Chatbot development.
In: Proceedings of the 2018 International Conference on Machine Learning and Cybernetics
(ICMLC), pp. 1955–1960 (2018)
47. Gurevych, I., Choi, J.D.: Introduction to the special issue on natural language processing for
conversational AI. Trans. Assoc. Comput. Linguist. 8, 1–4 (2020)
48. Wu, Z., et al.: A comprehensive survey on transfer learning. Proc. IEEE 109(1), 43–76 (2021)
49. Sharma, D.K., Anand, R.S.: Artificial neural networks (ANNs): a review of techniques and
applications. Int. J. Control Theory Appl. 9(35), 213–228 (2016)
50. Wick, C.: Deep learning. Informatik-Spektrum 40(1), 103–107 (2016). https://doi.org/10.
1007/s00287-016-1013-2
51. Goodfellow, I., Bengio, Y., A. Courville, Y.: Deep Learning. MIT Press (2016)
52. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
53. Guo, X., Zhang, Y.: Design and implementation of a Chatbot system for learning management
system. In: 2019 18th IEEE International Conference on Trust, Security and Privacy in Com-
puting and Communications/13th IEEE International Conference On Big Data Science And
Engineering (TrustCom/BigDataSE), pp. 138–143. IEEE (2019). https://doi.org/10.1109/Tru
stCom/BigDataSE.2019.00036
54. Bickmore, J., Giorgino, D.: Health dialog systems for patients and consumers. J. Biomed.
Inform. 39(5), 556–571 (2006)
55. Fogg, B.: Persuasive Technology: Using Computers to Change What We Think and Do.
Ubiquity 2002(December) (2002)
56. Brandtzaeg, C., Følstad, A.: Why people use chatbots. In: International Conference on Internet
Science, pp. 377–392 (2017)
57. Kocielnik, R., Avital, M., Muller, G.: Data-driven design for the future. In: Proceedings of
the Design Science Research in Information Systems and Technology, pp. 1–11 (2016)
58. Beyer, H., Holtzblatt, K.: Contextual design. Interactions 6(1), 32–42 (1999)
59. Radlinski, A., Balog, K., Byrne, P., Krishnan, K.: Coached conversational preference elicita-
tion: a case study in understanding movie preferences. In: SIGDIAL Conference, pp. 242–252
(2019)
60. Ryan, M.: Ethics for artificial intelligence and data systems. Philos. Technol. 31(2), 179–198
(2018)
61. Weizenbaum, J.: ELIZA—a computer program for the study of natural language communi-
cation between man and machine. Commun. ACM 9(1), 36–45 (1966)
62. Higashinaka, R., Inaba, M., Isozaki, H.: Learning to generate naturalistic utterances using
reviews in spoken dialogue systems. In: Proceedings of the 22nd International Joint
Conference on Artificial Intelligence, pp. 2278–2283 (2011)
63. Bordes, A., et al.: Learning end-to-end goal-oriented dialog. In: Proceedings of the
International Conference on Learning Representations (ICLR) (2017)
64. Williams, J., Raux, A., Ramachandran, D., Black, A.: The dialog state tracking challenge. In:
Proceedings of the SIGDIAL 2013 Conference, pp. 404–413 (2013)
65. Allen, J.: Rule-based approach. In: Proceedings of the International Conference on Artificial
Intelligence, pp. 123–130 (2022)
66. Smith, S.: Machine learning (ML). J. Intell. Syst. 10(3), 45–59 (2015)
67. Johnson, M.: Natural language processing (NLP). IEEE Trans. Pattern Anal. Mach. Intell.
28(2), 1325–1339 (2012)
68. Brown, E.: Deep learning. In: Proceedings of the IEEE International Conference on Neural
Networks, pp. 789–796 (2017)
 Intelligent Conversational Chatbot 29

69. Chen, D.: Reinforcement learning. IEEE Trans. Auton. Ment. Dev. 15(4), 345–358 (2018)
70. Lee, J.: Generative models. In: Proceedings of the International Conference on Machine
Learning, pp. 567–574 (2019)
71. Garcia, M.: Hybrid approach. IEEE Trans. Knowl. Data Eng. 20(5), 789–802 (2016)
72. Wilson, E.: Contextual chatbot design using transformer models. In: Proceedings of the
International Conference on Artificial Intelligence, pp. 234–241 (2020)
73. Johnson, A.: Advancements in chatbot development using deep learning. J. Mach. Learn. Res.
25, 1234–1256 (2019)
74. Thompson, B.: Reinforcement learning for chatbot dialog management. IEEE Trans. Artif.
Intell. 17(3), 567–578 (2018)
75. Chen, R.: Generative models for chatbot response generation. In: Proceedings of the Annual
Meeting of the Association for Computational Linguistics, pp. 789–796 (2017)
76. Zhang, L.: Hybrid approach combining rule-based and machine learning techniques for
chatbot development. Expert Syst. Appl. 45(5), 345–358 (2016)
 Intelligent Conversational Chatbots: History,
Taxonomy, Classification with PRISMA Model

Johnbenetic Gnanaprakasam(B) and Ravi Lourdusamy

Department of Computer Science, Sacred Heart College (Autonomous), Affiliated to

Thiruvalluvar University, Tirupattur District, Tirupattur, Tamil Nadu,, India
[email protected], [email protected]

Abstract. The ability of machines to think intelligently like a human is known

as artificial intelligence (AI). AI advancements have enabled machines to mimic
natural intelligence and make decisions like a human. AI and Human-Computer
Interaction (HCI) paradigms are used to design and construct intelligent conver-
sational chatbots. The Chatbot is defined as a computer program that would be
intended to emulate dialogue among humans, particularly over the web. In this arti-
cle, the PRISMA Model is used to provide significant scientific articles published
from 1950 to 2022 on Intelligent Conversational Chatbot evolution, techniques,
and classification. The objective of the article is to describe the current intelligent
conversational chatbot approaches that were in use and represent their comparison
based on techniques, drawbacks, and applications. Chatbots have the potential to
enhance human-machine interaction.

Keywords: Chatbot · Natural Language Processing · Pattern Matching ·

Task-Oriented · Non-Task Oriented

1 Introduction
Artificial Intelligence (AI) plays a significant role in our day-to-day activities in various
domains like Machine learning, Deep learning, Robotics, Expert systems, Fuzzy logic,
and Computer Vision. A chatbot is a common type of AI system or a program. An intel-
ligent Human-Computer Interaction (HCI) is a model that studies how humans utilize
computers to complete specific jobs and how they effectively interact with each other
[1]. A chatbot is also known as chatterbot, multipurpose virtual assistants, smart bots [2],
digital assistant, interactive agent, conversational agent, and the artificial conversational
entity that is described as hardware like Alexa Digital Assistant by Amazon Echo or
software on Android devices, Google Assistant is available, while on Apple devices,
Siri is available [3].
A chatbot can be defined as a program that uses Natural Language Speech to have a
conversation with a human being. Chatbots are typically designed to perform a certain
activity for which they have been trained specifically, which can include a diverse set
of tasks such as arranging records on a workstation, browsing the web based on the
keyword, and scheduling appointments [4].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 30–46, 2025.
https://doi.org/10.1007/978-3-031-71481-8_3
 Intelligent Conversational Chatbots 31

The research article attempts to present the recent research articles regarding Intelli-
gent Conversational Chatbots Techniques, and their evolution over the ages, significantly
from 2001 to 2021.
Further, this article is structured as follows: The history of Chatbot and its research
avenues are outlined in Sect. 2. Section 3 presents the different categorizations of the
surviving chatbots and their general architecture. Section 4 presents a discussion on
the standard design methodologies of chatbots. Comparison of chatbots technique using
PRISMA Model was carried out in Sect. 5. Section 6 presents the research gaps identified
along with the future enhancement of the research.

2 Background
2.1 Artificial Intelligence
Artificial Intelligence (AI) is the simulation of human intelligence by computers that are
programmed to think and act like humans.

Fig. 1. AI Categorization.

AI systems are categorized into three types as shown in Fig. 1, First type is ANI,
which stands for Artificial Narrow Intelligence or narrow AI. It’s the ability of a computer
to excel at a specific activity. The second type of AI Technologies is Artificial General
Intelligence (AGI) expects the machine to be just as clever as a person do. It does a task
that a human can do more effectively. The third type of AI Technology is ASI (Artificial
Super Intelligence), which refers to artificial intelligence technology that will ultimately
match and exceed the human brain.

2.2 Machine and Deep Learning

Machine Learning (ML) is an AI technique, which learns based on the experience of a
human. Human beings learn things by;
1. Making Mistakes. Someone Corrects the Mistakes and then Learns.
2. During the learning process, the human brain is constantly getting inputs and neurons
are being adjusted and new pathways are being created, which leads to a certain output
or a decision.
32 J. Gnanaprakasam and R. Lourdusamy

Computers can be instructed in a similar manner, where neurological models are

generated and they will be skilled with a large amount of training data. Then the expected
outputs are modified, and the neural network can become provided with the training.
So that, when a human gives new input to the computer, later on, it will recognize it. It
can determine accurate results. There are other techniques where we use the knowledge
present in available data, extract it, and use that knowledge to make future decisions.

2.3 Chatbot
A chatbot is computer software, algorithm, or artificial intelligence that interacts with
a human or with another participant in the conversation via a messaging application or
similar text application [5]. Sometimes it is also referred to as a chatterbot or chatter
robot [6], They are capable of comprehending a variety of questions posed by humans
[7]. They can also distinguish between the novelty of sentences and even emotions.
They require a rich lexicon of communication among individuals to obtain the greatest
quality of Chatbot conversation [8]. A conversational chatbot is a computer program
intended to converse with humans using NLP (Natural Language Processing) [9]. The
most challenging issue in Chatbots is keeping up with the conversation’s context and
interpreting human emotions. Current chatbot creation strategies include the use of a
pattern-matching methodology and then creating pre-configured responses that fit the
provided input. The disadvantage of this approach is that it does not always result in a
meaningful chat that leads to the intended outcome [4].

2.4 History of Chatbot

In 1950 Mind – Journal, The Turing Test (“Can machines think?”) was introduced by
computers can this time and the idea of a chatbot gained traction [10]. Joseph Weizen-
baum created ELIZA – the first chatbot [5] at Artificial Intelligence Laboratory, MIT
LAB in the year 1966. It is made with a basic handcrafted script [11]. ELIZA is soft-
ware that allows for natural language dialogue with a computer [12]. ELIZA utilizes
pattern matching as well as a template-based answer selection mechanism. Knowledge
is restricted in the ELIZA chatbot, because of this disadvantage ELIZE can’t discuss a
huge range of topics [2].
Parry [11], like Eliza, uses a rule-based approach, but with a deeper knowledge of the
mental model that might elicit emotion. As a result, Parry is the first chatbot to include
emotion in its design. ALICE (Artificial Linguistic Internet Computer Entity), a scal-
able chatbot that uses Artificial Intelligence Markup Language, is also worth mentioning
(AIML). ALICE is a natural-language conversational agent. ALICE, on the other hand,
and other chatbot systems are limited to the knowledge that has been crafted in their
files. [13] As a result, ALICE continues to adopt a rule-based technique to acquire the
response, repeatedly invoking a pattern-matcher. In 1972, PARRY was written at Stan-
ford University by psychiatrist Kenneth Colby. It attempted to imitate a person suffering
from paranoid schizophrenia. It was dubbed “ELIZA with attitude” because it used a
more advanced conversational style than ELIZA. In the 1970s, a group of psychiatrists
used a form of the Turing test to examine a combination of actual patients and com-
puters running PARRY. The test transcripts were sent to another group of psychiatrists
 Intelligent Conversational Chatbots 33

to distinguish between human and computer responses, and they were able to guess 48
percent of the time correctly [14].
In 1983, Racter Chatbot was built by William Chamberlain and Thomas Etter for the
Inc Corporation. Z80 microprocessor having 64K of Main memory was used to build
the Racter Chatbot, which is hardware that is incomparable to today’s technology [15].
Jabberwacky is a 1988 AI software that was one of the first attempts to create an
AI program that could simulate human interaction and keeps on a discussion between
users. It was primarily intended as a kind of amusement. Its goal was to changeover
from a text-based system to one that was entirely voice-controlled [14]. CLEVERBOT
or JABBERWACKY are two effective conversational chatter handlers [9].
SHRDLU was another version of chatbot developed in the year 1971. Parsing, gram-
matical detection, and semantic analysis are the techniques used to create a chatbot [16].
A graduate from MIT named Terry Winograd worked at SHRDLU from 1968 to 1970. It
was a computer system that could recognize basic English phrases and carry out orders
in English talks. People could send commands to handle items, and the system would
ask for explanations if its heuristics programs couldn’t decipher a language based on
context and physical knowledge (Winograd 1971) [17].
Dr. Sabaitso (“Sound Blaster Artificial Intelligent Text to Speech Operator”) in 1991,
a chatbot took advantage of a technologically revolutionary innovation – Creative Labs’
Sound Cards. It grew added humanistic than just its predecessors in some ways because
it can imitate dialog – it communicated vocally [15].
ALICE (“Artificial Linguistic Internet Computer Entity”) was a chatbot based on
ELIZA that existed from 1995 until 2000. [18] The Artificial Intelligence Markup Lan-
guage was also used to create ALICE (AIML), which has a structure that is quite close to
that of today’s current solutions [1, 19–21]. It’s an NLP chatterbot that uses systematic
pattern-matching techniques to imitate conversation, however, it failed the Turing test.
ALICE is built on the foundation of XML knowledge bases [14].
SmarterChild is the next chatbot in its evolution and it was made available in messag-
ing applications [22]. ActiveBuddy was created in 2001, and it was extensively diffused
over worldwide instant messengers and SMS networks. It was a forerunner of Apple’s
SIRI and Samsung’s S VOICE since it provided a lively, personalized dialogue [14].
StudyBuddy and SmarterChild were launched in the AIM context, both of which
can communicate with people in non-formal training to assist them. The app offered
discussion boards where members of the online community may speak in addition to
effective engagement [18].
Watson is a question-answering (QA) computing system that applies modern NLP,
retrieval of information, expert systems, task-specific algorithms, and machine learning
technologies to the domain of open and distributed query answering. Watson was created
by IBM in 2006. Watson uses IBM’s Deep Quality Assurance software as well as the
Apache UIMA framework (Unstructured Information Management Architecture). It runs
on the SUSE Linux Enterprise Server operating system and leverages the Apache Hadoop
framework to facilitate distributed computation [14].
34 J. Gnanaprakasam and R. Lourdusamy

The development of virtual personal assistants, such as Apple Siri [1] also called
Cortana [20], was the next stage. It is an example of voice-based task-oriented chat-
bots/conversational agents that try to respond to the task they have been given. In confined
areas, task-oriented chatbots operate effectively [23].
Google Assistant/ Google Now (IPAs – Intelligent Personal Assistants) [24] is a type
of virtual personal assistant. [1] It makes suggestions for locations people might want to
visit, and Tesla can drive us there [25]. It was invented by Google for Google’s mobile
search applications. Android 4.1 (“Jelly Bean”), which was released in 2012, was the
first to have it. By delivering queries to a series of web services, it used a natural language
interface to answer inquiries, offer recommendations, and forecast actions. Google also
created Google Assistant, a smart extension of Google Now that participated in two-way
communication with the user [14].
Mitsuku is an AIML-based chatbot [4]. Steve Worswick built it using AIML to
comprehend and respond to people. Her capacity to reason with certain items is part of
her intellect. She has won the Loebner Prize twice, in 2013 and 2016, and was runner-up
in 2015 [14]. It provides web-based services [26].
Microsoft developed Cortana Chatbot [27]. Microsoft Cortana assists us in resolving
everyday issues such as finding restaurants, places, and more. It focuses on a single task
and does it well [28].
Amazon Alexa [1] is a digital tool existing as hardware. [3] It is a voice-based task-
oriented chatbot [23]. Because of its natural language contact with people, it strives
to assist humans by delivering information or doing specified jobs for them. These
systems are designed to address two distinct but connected issues: (a) natural language
comprehension and (b) conversation flow management (dialogue). Different frameworks
have been built based on the two mentioned difficulties. These platforms offer researchers
methods capable of dealing with the issues mentioned, allowing them to create Virtual
Assistants and chatbots for specialized areas [24].
Bots for Messengers – Facebook developed a chat framework in 2016 that allows
programmers to construct bots that could communicate with Facebook users. M bots
were ready by the end of 2016, spanning a wide variety of use cases [14].
Microsoft released Tay, an experimental artificial intelligence chatbot, in 2016. Tay
was terminated after a day due to their profane and offensive tweets, contacts, and
interactions with people on Twitter [5]. Microsoft introduced Tay to the public on March
23, 2016. During Tay’s brief public existence, the bot sent out over 93,000 tweets [5].
‘Thinking about you is abbreviated as Tay [15].

2.5 The Taxonomy of Chatbots

Chatbots are often classified as, task-oriented and non-task-oriented chatbots. Task-
focused chatbots are meant to do certain tasks depending on the user’s instructions,
whereas non-task-oriented chatbots have several functions as in Figs. 2 and 3.

2.6 Applications/Use of Chatbots

An intelligent Agent is a computer conversation system that uses natural language to
connect with humans [29]. Chatbot mechanisms have driven new possibilities for a
 Intelligent Conversational Chatbots 35

Fig. 2. Taxonomy of Chatbots.

Fig. 3. Chatbot Classification.

wide range of industries. Chatbots are implemented in educational domains, not just to
improve the interactive skills of the students but also to aid instructors, in such a way
by providing computerization. Virtual assistants in education improve connectedness,
efficiency, and interactions by reducing uncertainty. It can easily provide a virtual learn-
ing atmosphere that is focused, personalized, and outcome-based, which is exactly what
today’s modern academic institutions want. If an institution of higher learning utilizes
a chatbot to connect with students, the application’s first mistake rate is quite high [18].
Education Environments, Customer Service/Support, Health, Robotics, Industrial Use
Cases, Ticket Booking, Instant Query Solutions, Gathering Feedback, Email Listing,
Friendly Conversation, Marketing, and sales.

3 Related Works

An intelligent chatbot tends to improve its efficiency through conversations all the time.
The chatbot modules which include user simulation modules and the module for under-
standing the natural language will work better through continuous interaction and learn-
ing processes. Algorithms of machine learning (ML) and human supervisors enable the
chatbot to be trained properly. To ensure that the AI chatbot becomes a successful learner,
ML methods such as supervised reinforcement learning and unsupervised methods can
be leveraged. Chatbots can become successful learners with neural networks and deep
36 J. Gnanaprakasam and R. Lourdusamy

learning. This study aims to analyze and investigate the earlier chatbot surveys that were
conducted previously.
A systematic review on “Chatbot: History, Technology and Applications” by
Adamopoulou, Eleni, and Moussiades, Lefteris [2] explains the history of chatbots in
detail with the technology used. The key objective of the survey was to present the
existing chatbot types and the approaches for developing a chatbot. A chatbot may be
created using computer languages such as Java and Python, as well as a commercial or
open-source chatbot creation environment. Chatbot usage comes with several risks, the
most significant of which is concerning the security of personal data. Protective tech-
niques are now being studied in this area. The most severe problem they now have with
chatbots is their incapacity to understand and generate natural conversations. They may
occasionally become unable to interpret a term, resulting in communication breakdowns
and tense interactions with their interlocutor. The most essential stage in the ongoing
development of chatbots is to improve language understanding and output.
According to Eleni Adamopoulou and Lefteris Moussiades, the quantity of chatbot-
related technologies is out of control, and it is getting worse every day. Chatbots may
be created using a programming language like Java, Clojure, Python, C++, PHP, Ruby,
Lisp, or cutting-edge platforms. The knowledge area, service, as well as objectives, input
processing, response production technique, human aid, and chatbots can be classified
based on a variety of variables, including their creation process. The study has listed and
explained the Concepts associated with chatbot technologies: Pattern Matching, AIML,
LSA, Chat script, RiveScript, NLP, NLU, entity, and finally contexts [1].
The Survey “Development of Conversational Agent to Enhance Learning Experi-
ence” by Nor Hayati Jaya et al. [31], aims to evaluate the use of chatbots and explained
the importance of NLP Techniques that will be used in designing a chatbot. The study
highlighted the important components: Basic input processing; Input Understanding;
Dialogue Manager; Knowledge Base; Response Generator and output. The author
opined that the majority of students use chatbots for learning purposes. A chatbot helps
immensely in the teaching-learning process especially for students because it gives
instant messages and responses to the user. Besides this, based on the student genders,
there is no significant difference towards the conversational agent. The study focuses
largely on the usage of Chatbots in higher education institutions for teaching and learn-
ing, with less attention paid to other forms of education: such as healthcare, sociology,
and so on.
Jan Deriu [32] proposed the classification of characteristics of the various types
of intelligent conversation systems based on Task-Oriented, Conversational Agents,
and Interactive questioning and answering. This article also concentrated on efforts
to automate the conversation system’s evaluation procedure.
Divya S et al. [33] describe Chatbots as Software that communicates with people in
natural language. Even though every chatbot has a unique field of specialty, this approach
is the same as other kinds of chatbots. One input from a human is matched against the
chatbot’s knowledge base. Chatbots rely on artificial intelligence to function.
A study conducted by Erika Bonnevie et al. [34], elaborates on the requirement for
considerable finance to continue traditional media buys, community-based programs that
 Intelligent Conversational Chatbots 37

leverage expensive mass media techniques as a core tactic may face hurdles in ensuring
long-term sustainability.
The authors of “AI-Powered Health Chatbots: Toward a General Architecture”
present a robust Structure of an Intelligence Medical Bot with four factors to accomplish
two objectives that integrate conversation and interaction aspects in natural language
understanding (NLU) and natural language generation (NLG), as well as a supervised
neural expert portion that provides adequate answers from pre-formatted information
[35].
A study from [36] “Conversation Technology With Micro-Learning: The Impact of
Chatbot-Based Learning on Students’ Learning Motivation and Performance” by Jiaqi
Yin et al. expressed that research has a lot of potential for the future development of
chatbot-based micro-learning systems with varying levels of interactivity. The influ-
ence of a chatbot-based micro-learning system on students’ learning motivation and
performance was explored in this study.
Adam et al. conducted an online study to prove how verbal humanities promote
better and the finger technique both increase user participation when a chatbot requests
customer comments [30].
Florian Brachten et al. created a model, which is based on the deconstructed ‘Theory
of Planned Behavior. Employee intrinsic motivation has a substantial positive influence
on the intention to utilize Enterprise Bots (EB), according to the results of a structural
equation model, but external variables have a lower impact. One of the most intriguing
discoveries is that one’s attitude about utilizing electronic books is the most important
driver of real EB usage intention. It demonstrates that internal impacts (e.g., subjective
standards or perceived behavioral control) are more relevant than external factors (e.g.,
external influences), allowing change managers to include these components in their
efforts [37].
Vrushil Gajra et al. [38] have proposed that Robotic Process Automation (RPA)
and Chatbot Technology may be used in a variety of manufacturing and organization
procedures. Chatbots may be used successfully and efficiently in the retail and consumer
products businesses, as well as in the education, manufacturing, and airline industries.
Finance, government agencies, manufacturing, human resources, and education are just
a few of the industries that might benefit from automation. Education is one of the
industries where both chatbots and RPA may be used at the same time.
Santosh Maher et al. [39] described that, due to the widespread usage of messaging
services and the advent of NLU, chatbots have lately gained popularity. Rule, Retrieval,
and Generative-based are the three types of chatbots. A specified collection of phrases
is grouped in a question-answer system in which each question is defined as a response
in the form pair in a rule-based chatbot. AIML, an XML-based language for designing
bots, was released in 2001.
Agarwal et al. [40] clarify that an open-domain dialog agent is supposed to handle
several domains, whereas a task-oriented chatbot is domain-specific. When it comes to
multi-turn conversations, it is critical that the chatbot can discuss in detail a topic, very
much like a human.
Biduski et al. [41] elucidate that Support Systems frequently feature virtual agents.
Zhou et al. [42] consider both intelligent and emotional quotients in system design,
38 J. Gnanaprakasam and R. Lourdusamy

casting human-machine social conversation as decision-making over Markov Decision

Processes; XiaoIce is optimized for lengthy user participation as assessed by anticipated
Discussion Each session.
William Villegas-Ch has proposed an “Architecture for the Integration of a Chatbot
with Artificial Intelligence in a Smart Campus for the Improvement of Learning” [43]
the author elaborated on the use of technologies in institutions and he put forth that it
motivates students to learn, and the student’s enthusiasm increases, encouraging them to
participate actively in their study. However, given the number of factors involved in this
goal, creating an atmosphere conducive to active learning will take a significant amount
of time and work.
Yiping Song et al. have described [44] that Retrieval-based and generation-based
systems are the two broad groups of conversation systems. Retrieval scans a huge con-
versation resource for a user-issued utterance such as a query and delivers a response
that best matches the inquiry. New responses are synthesized using creative ways [45].

3.1 Comparison of Chatbot

Table 1 depicts the evolution of Chatbots from its beginning. It presents a comparison
of conversational systems based on a set of key factors. From the comparison table, the
year of publication, Techniques, Scheme, Drawback, and developers was listed from the
research articles.

Table 1. Evolution of Chatbots

Year Chatbot Technique Scheme Drawback Authors Application Domain

1950 Turing Test Natural Language An interrogator It cannot measure Alan Turing -
[10] had to determine intelligence that
which player is beyond the
was a human ability of humans
1966 ELIZA [5] Rule-Based Response Knowledge Artificial Hospital
Approach selection scheme Limited Intelligence
Pattern Matching (rephrasing the Particular Laboratory in
patient’s Domain of Topic MIT LAB
statement) Cannot Keep long
Conversations or
discover the same
1972 PARRY [11] Rule-Based Interviewing Low Speed of Kenneth Mark Hospital
Approach with his therapist Responding Colby,
Cannot Learn Stanford’s
from the Psychiatry
Conversation Department
1983 Racter [15] Knowledge-Based Emotional Low capabilities William Education
Responses concerning Chamberlain
language and Thomas
understanding Etter, Inrac
Corporation
(continued)
 Intelligent Conversational Chatbots 39

Table 1. (continued)

Year Chatbot Technique Scheme Drawback Authors Application Domain

1988 Jabberwacky Contextual Pattern Mimic human Still unable to Rollo Entertainment,
(written in Matching interaction and respond at a rapid Carpenter Marketing, Robots &
CleverScript) carry out rate Robotics, Digital Pets,
[9] conversation Gadgets & Games
among users
1971 SHRDLU Clarifications, Understood Could not MIT, Terry Education (English
[16] [17] grammatical basic English understand a Winograd Dialogue System)
detection, statements and sentence through
Semantic executed the use of context
Analysis, Parsing commands in and physical
English knowledge
dialogues
1992 Dr. Sabaitso Text-to-Speech First bot to Knowledge Creative Labs Education
(speech utilize Limited for MS-Dos
synthesis) text-to-speech
[15] functioning
1995 ALICE Heuristic pattern Domain-Specific Alice lacked Richard Entertainment
(Alicebot) matching clever traits. Wallace
XML knowledge Couldn’t come up
bases with human-like
responses,
expressions, or
attitudes
2001 SmarterChild Processing Throughout the Knowledge ActiveBuddy Education
(commercial queries asked in world’s instant Limited
instant natural language chat and SMS
messaging networks
bot) [18]
2010 Apple Siri Natural Language Responds to user It requires an Apple for iOS Social Media(Post
(intelligent UI requests using internet statuses or tweets
voice Natural Language various internet connection. You to Facebook or Twitter
assistant) Processing services can’t use it when reminders, driving
[23] you are offline directions and playing
It is multilingual, music)
but instructions
are only in
English. Doesn’t
understand all
spelling
variations
It also has
difficulties in
hearing
2011 Watson[14] Natural Language AI-powered It supports only IBM Weather forecast
Knowledge Base virtual agent English (Building codes,
designed to Slow integration Weather forecasting,
provide Fashion, Advertising,
customers with Healthcare,…etc.)
fast, consistent,
and accurate
answers across
any messaging
platform,
application,
device, or
channel
(continued)
40 J. Gnanaprakasam and R. Lourdusamy

Table 1. (continued)

Year Chatbot Technique Scheme Drawback Authors Application Domain

2012 Google Natural Language predict actions Smart home Google Inch Voice Assistant
Now/Google by passing capability is
Assistant requests to a set limited
[45] [24] of web services
two-way
dialogue with
the user
2005 Mitsuku AIML technology Intelligent Ability to reason Steve
(Kuki or with specific Worswick
Pandorabot) objects
[4]
2014 Cortana [27] Goal-Based Perform tasks It can launch Microsoft for Use information from
like reminders software that windows-based the Bing web search to
Available in distributes devices make reminders,
different malware identify natural
languages speech, and answer
queries
2014 Alexa [1] Knowledge-based Which is built-in Sometimes, Amazon Google Devices
& virtual assistant home users’ queries can
technology automation (IoT) be
misunderstanding
occur
Cannot
understand
particular
language (oral
speech)
2016 Bots for Rule-Based Understand Have limited Facebook’ Social Media
Messengers questions, responses and are
[14] provide answers, not often able to
and execute answer multi-part
tasks questions or
questions that
require decisions
2016 TAY [15] Experimental Twitter to Had to be shut Microsoft Social Media
artificial mimic the down just 16 h
intelligence speech and after launch
habits of a
teenage girl

4 Methodology
PRISMA is an acronym for Preferred Reporting Items for Systematic Reviews and Meta-
Analyses. It is a collection of evidence-based learned elements designed to assist narrates
in reporting a wide range of systematic reviews and meta-analyses, which are primarily
used to explore the benefits. PRISMA focuses on how writers can transparently present
their research.
The criteria for inclusion were as shown in Fig. 4 shows how information passes
across the several phases of systematic reviews using the PRISMA flow diagram. The
PRISMA flow diagram shows the number of records identified, including rejected ones,
 Intelligent Conversational Chatbots 41

and the explanations for exclusions. The following processes of the systematic literature
review were presented to survey the present study in the field of chatbots.

Fig. 4. PRISMA – Flow Diagram for the selected study. (PRISMA 2020 flow diagram for new
systematic reviews which included searches of databases and registers only)

The selected research articles published in Books, Conferences, Generic, and

Journals of the study were presented in Fig. 5.
As shown in Fig. 6 the diagrammatic representation of the previously published
research articles that were analyzed for this study. The selected research articles
are classified into Book Publications, Conference Publications, Generic, and Journal
Publications year-wise, as four categories.

4.1 Identification

Using the PRISMA model, exclusively in line with the Identification step, eighty-nine
research articles have been selected for the research study, of which two articles were
removed for duplication and eighteen articles were eliminated because it is not fit for the
42 J. Gnanaprakasam and R. Lourdusamy

Fig. 5. Classification of Research Work for the study.

Fig. 6. Selected Research Work for the study based on the Year with type.

problem definition but all the eliminated articles can be used for the chatbot development
process.

4.2 Screening
It initially records the number of articles found and then makes the selection process. In
the initial screening step, 69 articles were screened of which 22 articles were excluded
because the titles and abstracts were not relevant to the research or study. Chatbot and
its design technique-related research articles are taken for further deeper study for this
review.

4.3 Included
In the next process of screening, 47 articles were included for the final study based on
the various eligibility for inclusion in the systematic review. Based on the study, Table 1
presents the comparison of chatbot generations with their techniques, schemes, draw-
backs, and used domains. Selected research articles are tabulated and the categorization
is also there for future researchers to look into in detail for further study.

5 Discussion/Results
The study of the literature review was primarily based on a PRISMA investigation
model. This study looks at the history of chatbots and their use in many fields. In the last
several years, there has been substantial development in the invention and the usage of
 Intelligent Conversational Chatbots 43

the chatbot, with numerous advantages in a variety of fields such as education, industry,
health care, research labs, E-Marketing platforms, Group Chat Apps, and advertising.

5.1 Research Gaps

Presently Intelligent Conversational Systems have a verbal intelligence that is con-

strained. The most significant problem that chatbots are now facing is maintaining track
of the situation of the conversation and understanding the human inputs and the respec-
tive responses [4]. Another key issue of a chatbot is to examine the various aspects of
emotional difficulties that instructors experience and their origins. The most essential
step in the development plans of chatbots is to strengthen language understanding and
interpretation [2]. Chatbots are intelligent systems but they cannot think by themselves
based on learning, which is a drawback of the present technology in building chat-
bots. For an effective intelligent conversational system, chatbot development needs to
be planned and the platforms used to implement chatbot needs to be chosen properly
[18].
Chatbots are largely employed in the education domain – teaching and learning pro-
cess. As a dialogue system, AI systems might be used to deliver course learning content
via a digital platform, capable of providing appropriate information to the users. Most
systems focus on pattern-matching techniques but semantic approaches and computer
reasoning systems need to be incorporated for the future development part of an intel-
ligent system. Moreover, inclusive knowledge base designing needs to be improved in
the development of general-purpose chatbots [46].
For Teaching-Learning Process, in LMS new ad-hoc can be added to help stu-
dents and the institution with a personalized way of dealing with [47]. The majority
of chatbot developments are done only for goal-oriented [16] and AI-oriented tasks.
NLP technologies need to be used for building AI-Chatbots in the future.

6 Conclusion
Chatbots have a higher success rate than humans in using messaging apps to reach out to
a large audience and generate substantial cost savings. This study presents an overview
of the fundamental concepts of chatbots. The current state of technology is in a rapid
transition phase, intending to have equipment that can deliver solutions with minimum
human intervention. This study examines several techniques to develop a chatbot, as
well as the categorization and structure of these conversational bots. It also presents the
design principles used to create many chatbots in chronological sequence.
The inability of today’s modern Bots to recognize and produce actual discussions
is the most serious issue they confront. At times, they would be unable to interpret a
phrase, leading to inconsistency in communication and uncomfortable interaction with
their listener. The most crucial phase in Chatbot development is to increase language
understanding and output.
44 J. Gnanaprakasam and R. Lourdusamy

References
1. Adamopoulou, E., Moussiades, L.: An Overview of Chatbot Technology. In: Maglogiannis,
I., Iliadis, L., Pimenidis, E. (eds.) Artificial Intelligence Applications and Innovations: 16th
IFIP WG 12.5 International Conference, AIAI 2020, Neos Marmaras, Greece, June 5–7, 2020,
Proceedings, Part II, pp. 373–383. Springer International Publishing, Cham (2020). https://
doi.org/10.1007/978-3-030-49186-4_31
2. Adamopoulou, E., Moussiades, L.: Chatbots: History, technology, and applications. Mach.
Learn. Appl. 2, 100006 (2020). https://doi.org/10.1016/j.mlwa.2020.100006
3. Vaidyam, A.N., Wisniewski, H., Halamka, J.D., Kashavan, M.S., Torous, J.B.: Chatbots
and conversational agents in mental health: a review of the psychiatric landscape. Can. J.
Psychiatry 64(7), 456–464 (2019). https://doi.org/10.1177/0706743719828977
4. Ramesh, K., Ravishankaran, S., Joshi, A., Chandrasekaran, K.: A survey of design techniques
for conversational agents. Commun. Comput. Inf. Sci. 750, 336–350 (2017). https://doi.org/
10.1007/978-981-10-6544-6_31
5. Neff, G., Nagy, P.: Automation, algorithms, and politics| talking to bots: symbiotic agency
and the case of tay. Int. J. Commun. 10, 17 (2016)
6. Akma, N., Hafiz, M., Zainal, A., Fairuz, M., Adnan, Z.: Review of chatbots design techniques.
Int. J. Comput. Appl. 181(8), 7 (2018). https://doi.org/10.5120/ijca2018917606
7. Ciechanowski, L., Przegalinska, A., Magnuski, M., Gloor, P.: In the shades of the uncanny
valley: an experimental study of human–chatbot interaction. Futur. Gener. Comput. Syst. 92,
539–548 (2019). https://doi.org/10.1016/j.future.2018.01.055
8. Hill, J., Ford, W.R., Farreras, I.G.: Real conversations with artificial intelligence: A compari-
son between human–human online conversations and human–chatbot conversations. Compu.
Hum. Behav. 49, 245–250 (2015). https://doi.org/10.1016/j.chb.2015.02.026
9. Sameera, A., John, D.: Survey on chatbot design techniques in speech conversation systems.
Int. J. Adv. Comput. Sci. App. 6(7), 72–80 (2015). https://doi.org/10.14569/IJACSA.2015.
060712
10. Review, A.Q.: MIND – A Quarterly Review of Psychology and Philosophy, pp. 433–460
(1950)
11. Pamungkas, E.W.: Emotionally-Aware Chatbots: A Survey. (2019). http://arxiv.org/abs/1906.
09774
12. Weizenbaum, J.: ELIZA—a computer program for the study of natural language communi-
cation between man and machine. Commun. ACM 26(1), 23–28 (1983). https://doi.org/10.
1145/357980.357991
13. Shawar, B.A., Atwell, E.: Accessing an information system by chatting. In: Meziane, F.,
Métais, E. (eds.) Natural Language Processing and Information Systems, pp. 407–412.
Springer Berlin Heidelberg, Berlin, Heidelberg (2004). https://doi.org/10.1007/978-3-540-
27779-8_39
14. Swathi, N.S., Tejas, R.: Intelligent bot for novice reciprocal action: a survey. Int. J. Sci. Res.
Comput. Sci. Eng. Inf. Technol. 4(6), 149–155 (2018)
15. Zemčík, M.T.: A Brief History of Chatbots. DEStech Trans. Comput. Sci. Eng. (2019). https://
doi.org/10.12783/dtcse/aicae2019/31439
16. Masche, J., Le, N.T.: A review of technologies for conversational systems. Adv. Intell. Syst.
Comput. 629, 212–225 (2018). https://doi.org/10.1007/978-3-319-61911-8_19
17. Candello, H., Pinhanez, C.: Designing Conversational Interfaces. no. May (2019)
18. Molnar, G., Szuts, Z.: The role of chatbots in formal education. In: SISY 2018 – IEEE 16th
International Symposium Intelligent System Informatics, Proceedings, pp. 197–201 (2018).
https://doi.org/10.1109/SISY.2018.8524609
 Intelligent Conversational Chatbots 45

19. Almansor, E.H., Hussain, F.K.: Survey on intelligent chatbots: state-of-the-art and future
research directions. Adv. Intell. Syst. Comput. 993, 534–543 (2020). https://doi.org/10.1007/
978-3-030-22354-0_47
20. Lin, L., D’Haro, L.F., Banchs, R.E.: A web-based platform for collection of human-chatbot
interactions. In: HAI 2016 – Proceedings of the 4th International Conference on Human Agent
Interaction, pp. 363–366 (2016). https://doi.org/10.1145/2974804.2980500
21. Marietto, M.G.B., et al.: Artificial intelligence markup language: a brief tutorial. Int. J.
Comput. Sci. Eng. Surv. 4(3), 1–20 (2013). https://doi.org/10.5121/ijcses.2013.4301
22. Adamopoulou, E., Moussiades, L.: An overview of chatbot technology. In: Maglogiannis, I.,
Iliadis, L., Pimenidis, E. (eds.) AIAI 2020. IAICT, vol. 584, pp. 373–383. Springer, Cham
(2020). https://doi.org/10.1007/978-3-030-49186-4_31
23. Hussain, S., Sianaki, O.A., Ababneh, N.: A survey on conversational agents/chatbots clas-
sification and design techniques. In: Barolli, L., Takizawa, M., Xhafa, F., Enokido, T. (eds.)
WAINA 2019. AISC, vol. 927, pp. 946–956. Springer, Cham (2019). https://doi.org/10.1007/
978-3-030-15035-8_93
24. Alexiadis, A., Nizamis, A., Koskinas, I., Ioannidis, D., Votis, K., Tzovaras, D.: Applying
an intelligent personal agent on a smart home using a novel dialogue generator, vol. 584.
Springer International Publishing, IFIP (2020)
25. Nimavat, K., Champaneria, T.: Chatbots: an overview types, architecture, tools and future
possibilities. Int. J. Sci. Res. Dev. 5(7), 1019–1026 (2017)
26. Smutny, P., Schreiberova, P.: Chatbots for learning: a review of educational chatbots for the
Facebook Messenger. Comput. Educ. 151, 103862 (2020). https://doi.org/10.1016/j.compedu.
2020.103862
27. Eisman, E.M., Navarro, M., Castro, J.L.: A multi-agent conversational system with heteroge-
neous data sources access. Expert Syst. Appl. 53, 172–191 (2016). https://doi.org/10.1016/j.
eswa.2016.01.033
28. Haristiani, N.: Artificial intelligence (ai) chatbot as language learning medium: an inquiry. J.
Phys. Conf. Ser. (2019). https://doi.org/10.1088/17426596/1387/1/012020
29. Shawar, B.A., Atwell, E.S.: Using corpora in machine-learning chatbot systems. Int. J. Corpus
Linguist. 10(4), 489–516 (2005). https://doi.org/10.1075/ijcl.10.4.06sha
30. Adam, M., Wessel, M., Benlian, A.: AI-based chatbots in customer service and their effects
on user compliance. Electron. Mark. 31(2), 427–445 (2021). https://doi.org/10.1007/s12525-
020-00414-7
31. Jaya, N.H., Mahyan, N.R., Suhaili, S.M., Jambli, M.N., Solehah, W., Ahmad, W.: Develop-
ment of conversational agent to enhance learning experience: case study In Pre (02), 124–130
(2021)
32. Deriu, J., et al.: Survey on evaluation methods for dialogue systems. Artif. Intell. Rev. 54(1),
755–810 (2020). https://doi.org/10.1007/s10462-020-09866-x
33. Habib, F.A., Shakil, G.S., Iqbal, S.S.M., Sajid, S.T.A.: Self-diagnosis medical chatbot using
artificial intelligence. In: Goyal, D., Chaturvedi, P., Nagar, A.K., Purohit, S.D. (eds.) Pro-
ceedings of Second International Conference on Smart Energy and Communication: ICSEC
2020, pp. 587–593. Springer Singapore, Singapore (2021). https://doi.org/10.1007/978-981-
15-6707-0_57
34. Bonnevie, E., Lloyd, T.D., Rosenberg, S.D., Williams, K., Goldbarg, J., Smyser, J.: Layla’s
Got You: developing a tailored contraception chatbot for Black and Hispanic young women.
Health Educ. J. 80(4), 413–424 (2021). https://doi.org/10.1177/0017896920981122
35. Khadija, A., Zahra, F.F., Naceur, A.: AI-powered health chatbots: toward a general archi-
tecture. Procedia Comput. Sci. 191, 355–360 (2021). https://doi.org/10.1016/j.procs.2021.
07.048
46 J. Gnanaprakasam and R. Lourdusamy

36. Yin, J., Goh, T.T., Yang, B., Xiaobin, Y.: Conversation technology with micro-learning: the
impact of chatbot-based learning on students’ learning motivation and performance. J. Educ.
Comput. Res. 59(1), 154–177 (2021). https://doi.org/10.1177/0735633120952067
37. Brachten, F., Kissmer, T., Stieglitz, S.: The acceptance of chatbots in an enterprise context – A
survey study. Int. J. Inform. Manag. 60, 102375 (2021). https://doi.org/10.1016/j.ijinfomgt.
2021.102375
38. Gajra, V., Lakdawala, K., Bhanushali, R., Patil, S.: Automating student management system
using ChatBot and RPA technology. SSRN Electron. J. (2020). https://doi.org/10.2139/ssrn.
3565321
39. Maher, S., Kayte, S., Nimbhore, S.: EasyChair Preprint Chatbots & Its Techniques Using AI:
an Review (2020)
40. Agarwal, R., Wadhwa, M.: Review of state-of-the-art design techniques for chatbots. SN
Comput. Sci. 1(5), 1–12 (2020). https://doi.org/10.1007/s42979-020-00255-3
41. Biduski, D., Bellei, E.A., Rodriguez, J.P.M., Zaina, L.A.M., De Marchi, A.C.B.: Assess-
ing long-term user experience on a mobile health application through an in-app embedded
conversation-based questionnaire. Comput. Hum. Behav. 104, 106169 (2020). https://doi.org/
10.1016/j.chb.2019.106169
42. Zhou, L., Gao, J., Li, D., Shum, H.Y.: The design and implementation of xiaoice, an empathetic
social chatbot. Comput. Linguist. 46(1), 53–93 (2020). https://doi.org/10.1162/COLI_a_
00368
43. Villegas-Ch, W., Arias-Navarrete, A., Palacios-Pacheco, X.: Proposal of an architecture for
the integration of a chatbot with artificial intelligence in a smart campus for the improvement
of learning. Sustainability 12(4), 1500 (2020). https://doi.org/10.3390/su12041500
44. Song, Y., Te Li, C., Nie, J.Y., Zhang, M., Zhao, D., Yan, R.: An ensemble of retrieval-
based and generation-based human-computer conversation systems. Int. Jt. Conf. Artif. Intell.
2018-July, 4382–4388 (2018). https://doi.org/10.24963/ijcai.2018/609
45. Luke MacNeill, A., MacNeill, L., Doucet, S., Luke, A.: Professional representation of con-
versational agents for health care: a scoping review protocol. JBI Evid. Synth. 20(2), 666–673
(2021). https://doi.org/10.11124/JBIES-20-00589
46. Okonkwo, C.W., Ade-Ibijola, A.: Chatbots applications in education: a systematic review.
Comput. Educ. Artif. Intell. 2, 100033 (2021). https://doi.org/10.1016/j.caeai.2021.100033
47. Shukla, V.K., Verma, A.: Enhancing LMS experience through AIML base and retrieval base
Chatbot using R language. In: 2019 International Conference on Automation, Computational
and Technology Management, ICACTM 2019, pp. 561–567 (2019). https://doi.org/10.1109/
ICACTM.2019.8776684
 Unleashing the Power of Dynamic Mode
Decomposition and Deep Learning
for Rainfall Prediction in North-East
India

Paleti Nikhil Chowdary1 , Sathvika Pingali1 , Pranav Unnikrishnan1 ,

Rohan Sanjeev1 , V. Sowmya1(B) , E. A. Gopalakrishnan1,2 ,
and M. Dhanya3
1
Amrita School of Artificial Intelligence, Amrita Vishwa Vidyapeetham,
Coimbatore, India
v [email protected]
2
Department of Computer Science and Engineering, Amrita School of Computing,
Amrita Vishwa Vidyapeetham, Bangalore, India
3
Center for Wireless Networks and Applications (WNA),
Amrita Vishwa Vidyapeetham, Amritapuri, India

Abstract. Accurate rainfall forecasting is crucial for effective disaster

preparedness and mitigation in the North-East region of India, which is
prone to extreme weather events such as floods and landslides. In this
study, we investigated the use of two data-driven methods, dynamic mode
decomposition (DMD) and long-short-term memory (LSTM), for rain-
fall forecasting using daily rainfall data collected from the India Mete-
orological Department in northeast region over a period of 122 years.
We conducted a comparative analysis of these methods to determine
their relative effectiveness in predicting rainfall patterns. Using histori-
cal rainfall data from multiple weather stations, we trained and validated
our models to forecast future rainfall patterns. Our results indicate that
both DMD and LSTM are effective in forecasting rainfall, with LSTM
outperforming DMD in terms of accuracy, revealing that LSTM has the
ability to capture complex nonlinear relationships in the data, making
it a powerful tool for rainfall forecasting. The study reveals that the
DMD method achieved Mean Squared Error (MSE) values ranging from
150.44 mm to 263.34 mm and Mean Absolute Error (MAE) values from
91.34 mm to 154.61 mm. In contrast, the Deep Learning (DL) approach,
utilizing LSTM, demonstrated a normalized MAE value of 0.35 and a
normalized RMSE value of 0.534. Our findings suggest that data-driven
methods such as DMD and deep learning approaches like LSTM can sig-
nificantly improve rainfall forecasting accuracy in the North-East region
of India, helping to mitigate the impact of extreme weather events and
enhance the region’s resilience to climate change.

Keywords: DMD · LSTM · rainfall · forecasting · Data driven

methods
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 47–62, 2025.
https://doi.org/10.1007/978-3-031-71481-8_4
48 P. N. Chowdary et al.

1 Introduction
List of Abbreviations

Abbreviation Definition
DMD Dynamic Mode Decomposition
LSTM Long Short-Term Memory
DL Deep Learning
RNN Recurrent Neural Network
MAE Mean Absolute Error
RMSE Root Mean Square Error
IMD India Meteorological Department
CNN Convolutional Neural Network
SVD Singular Value Decomposition

Rainfall is one of the most important climatic variables that affects various
aspects of human life and natural ecosystems. Accurate rainfall forecasting is
crucial for effective disaster preparedness and mitigation. The North-East region
of India, commonly known as the “Seven Sisters” is a topographically diverse
region with a unique mix of flora and fauna. However, it is also one of the
most vulnerable regions in the world, with a high incidence of natural disasters,
particularly floods and landslides. In this region, accurate rainfall forecasting is
essential for effective disaster preparedness and mitigation, particularly in the
face of increasing occurrences of extreme weather events.
The North-East region receives the highest annual rainfall in India [3], with
its hilly terrain increasing its susceptibility to landslides and flash floods. The
region also experiences cyclones and thunderstorms, which have the potential
to cause extensive damage to infrastructure and agriculture [5]. Climate change
has exacerbated the frequency and intensity of these weather events, leading
to prolonged dry spells and erratic rainfall patterns, further exacerbating the
challenges faced by the region.
Rainfall prediction methods in the past relied on empirical relationships
between atmospheric variables (temperature, humidity, wind, and pressure)
using statistical [17] and dynamical approaches involving computer simulations
[12]. However, their effectiveness was limited due to complex interactions and
uncertainty. Recently, there is increasing interest in using machine learning (ML)
and data-driven techniques to enhance rainfall predictions [7].
The present work aims to build data driven models to predict rainfall in
mm using monthly average rainfall data. The work involved implementing Deep
Learning and Dynamic Mode Decomposition techniques and performing various
experiments to determine the hyperparameters that yield the best results.
 Rainfall Prediction 49

The study leverages a comprehensive dataset from the India Meteorological

Department (IMD), spanning over a century, to train and validate these models.
By comparing the performance of DMD and LSTM models in forecasting rainfall,
we aim to establish a more reliable method for predicting weather patterns in
this ecologically sensitive and disaster-prone region.
Our results demonstrate the effectiveness of both DMD and LSTM in rain-
fall forecasting, with LSTM showing superior performance due to its ability to
discern complex nonlinear dependencies in the data. This research not only con-
tributes to the field of meteorological science but also offers practical applications
in disaster management and climate change mitigation for the North-East region
of India.
The paper is structured as follows: Section 2 provides an overview of the
existing literature, Sect. 3 outlines the proposed methodology, Sect. 4 presents
the results and corresponding discussion, and Sect. 5 provides the conclusion.

2 Literature Review

The field of rainfall forecasting has witnessed significant advancements with the
advent of deep learning-based approaches. A notable development in this realm
is the creation of Tiny-RainNet by Zhang et al. [18], a hybrid network that
melds Convolutional Neural Networks (CNNs) with Bi-directional Long Short-
Term Memory (LSTM) networks, tailored for short-term rainfall prediction from
radar images. While this model demonstrates efficacy, its primary limitation lies
in its forecasting capability, which extends only one or two hours into the future.
Expanding upon this, recent studies have underscored the importance of
incorporating geospatial information into deep learning models for rainfall fore-
casting. Men et al. [9], in their paper titled “Spatio-temporal Analysis of Precip-
itation and Temperature: A Case Study Over the Beijing-Tianjin-Hebei Region,
China,” introduced a deep learning framework that harmonizes both spatial
and temporal data. This approach employs CNNs for extracting spatial fea-
tures from rainfall data, complemented by LSTM networks for temporal analy-
sis. Their model, when tested on an extensive dataset from the Beijing-Tianjin-
Hebei region, exhibited superior performance compared to traditional forecasting
methods.
In another stride forward, Luo et al. [8] presented PredRANN in their study
“PredRANN: The Spatiotemporal Attention Convolution Recurrent Neural Net-
work for Precipitation Nowcasting.” This innovative model integrates CNNs and
LSTM networks with a spatial-temporal attention mechanism, enhancing the
focus on salient spatiotemporal features within rainfall data. The model’s appli-
cation to real-world rainfall datasets revealed that PredRANN surpasses existing
deep learning models in accuracy and predictive capability.
These developments collectively underscore the burgeoning potential of deep
learning and data-driven methods in enhancing rainfall prediction accuracy. By
adeptly incorporating spatiotemporal features and location-specific data, these
advanced methodologies outshine conventional forecasting techniques. The focus
50 P. N. Chowdary et al.

of this paper is to further explore and refine DL and DMD techniques for rain-
fall forecasting. Emphasis will be placed on fine-tuning the parameters within
these models to optimize accuracy and reliability, particularly for applications
in regions susceptible to extreme weather conditions.

3 Methodology

This section outlines the methodology adopted for the study, including the data
acquisition process, model development, and training procedures.

3.1 Dataset

The dataset and training files utilized in this study are publicly available and
can be accessed at the following GitHub repository: https://github.com/Nikhil-
Paleti/rainfall-prediction. This repository contains data and resources essential
for replicating the research findings or for further exploration and analysis.

Fig. 1. Selected Grid Points in North-East India.

The dataset for this study was sourced from the India Meteorological Depart-
ment (IMD) [https://imdpune.gov.in/lrfindex.php], which provides gridded rain-
fall data across India at a spatial resolution of 0.25 × 0.25◦ , covering 122 years
from 1901 to 2022 [11]. For the Dynamic Mode Decomposition (DMD) analysis,
we utilized data from 1901 to 2018, focusing on the North-east region of India,
defined by longitudes 89.750 E to 980 E and latitudes 21.750 N to 300 N (Fig. 1).
This data was aggregated into monthly averages for 435 grid points, from which
a subset covering a period of 10 years (120 months) was used to predict the
average rainfall for the following 12 months.
For the Deep Learning (DL) analysis, data from 1951 to 2021 was used,
specifically focusing on four key locations in the North-east region: Agartala,
Guwahati, Imphal, and Itanagar. The monthly average rainfall data for these
 Rainfall Prediction 51

Table 1. Geographical Coordinates of Locations for DL Analysis

Location Latitude (N) Longitude (E)

Agartala 23.8315 91.2868
Guwahati 26.1158 91.7086
Imphal 24.8170 93.9368
Itanagar 27.0844 93.6053

locations were extracted from the corresponding grid points in the IMD dataset
(refer to Table 1 for geographical coordinates). The DL models were trained
using a dataset divided into a training set of 681 months and a test set of 171
months, with each model predicting the rainfall for the next month based on
varying input window lengths.

3.2 Data Driven Modelling

Data-driven modeling leverages data to develop precise predictions and insights

in various fields, such as engineering and computer science. Unlike traditional
methods, which rely heavily on theoretical assumptions, data-driven approaches
are inherently flexible, adapting to diverse and complex scenarios. These mod-
els efficiently handle large datasets, aiding researchers in deciphering complex
systems and enabling accurate predictions. In this study, we examine two spe-
cific data-driven techniques, Dynamic Mode Decomposition (DMD) and Deep
Learning (DL), for forecasting time series rainfall data.

Dynamic Mode Decomposition (DMD). DMD is a data-driven technique

employed to discern the underlying dynamic structures and patterns within com-
plex, high-dimensional systems. The approach involves decomposing the dataset
into a series of modes, each representing a specific spatial-temporal pattern of
behavior. These modes are defined by the eigenvectors of a data-constructed
matrix, with their respective eigenvalues indicating the temporal dynamics of
each mode (Tu et al.) [16]. DMD’s versatility is evidenced by its applications in
diverse fields, ranging from identifying rice leaf diseases [6] to defect detection in
cantilever beams [10]. The growing popularity of DMD across various domains
highlights its potential to significantly enhance our understanding of a myriad
of complex phenomena. As such, DMD stands out as an increasingly valuable
method in the realm of data-driven analysis, offering insights and predictions
that were previously unattainable with traditional modeling approaches.
Our workflow for DMD analysis is presented in Fig. 2.
52 P. N. Chowdary et al.

Algorithm 1. DMD Procedure

1: Collect time series data from the system to be analyzed, comprising measurements
of the system’s state variables.
2: Construct the data matrix X = [x1 , x2 , . . . , xN ], where each xi ∈ Cn represents a
snapshot of the system’s state at a specific time.
3: Perform SVD on X:
– X = UΣV∗ , where U is a matrix of left singular vectors, Σ is a diagonal
matrix of singular values, and V is a matrix of right singular vectors.
4: Construct the DMD modes Φ using SVD results:
– Φ = X VΣ −1 U∗ , where X is X with the last snapshot removed.
5: Calculate the eigenvalues Λ of the DMD modes:
– Λ = VΣ −1 U∗ X X, representing frequencies and growth rates.
6: Reconstructsystem dynamics using DMD modes and eigenvalues:
– x(t) ≈ rk=1 φk eωk t bk , where φk and ωk are the k-th DMD mode and eigen-
value, r is the number of significant modes, and bk are coefficients from initial
conditions.

Fig. 2. Applying DMD to Rainfall data

Deep Learning. DL, a branch of machine learning, leverages artificial neural

networks to process and learn from large datasets. Its key strengths include auto-
matic feature extraction and proficiency in handling complex data types such
as images, speech, and language, thereby eliminating the need for manual inter-
vention. DL has shown remarkable performance in areas like image recognition,
speech processing, natural language processing (NLP), and autonomous driving
systems.
 Rainfall Prediction 53

A prominent type of neural network used in DL is the RNN. RNNs are spe-
cially designed to process sequential data by maintaining a memory of previous
inputs. This feature makes RNNs ideal for applications in speech recognition,
NLP, and time series forecasting. For instance, RNN-based LSTM models have
been employed to predict COVID-19 trends across various Indian states [13], and
combined RNN and CNN models have been utilized for stock price prediction
[15].
However, RNNs have limitations, such as the vanishing gradient problem.
This issue arises when the gradients, essential for updating the network’s weights,
diminish through layers, impacting the network’s ability to learn long-term
dependencies. Additionally, RNNs are prone to overfitting, where they mem-
orize the training data rather than learning to generalize.
To address these challenges, researchers have developed advanced techniques
like LSTM (Hochreiter & Schmidhuber) [4] and GRU (Chung et al.) [2] net-
works, which incorporate gating mechanisms to better retain information over
extended sequences. Furthermore, strategies like gradient clipping and weight
regularization (Pascanu et al.) [14] have been introduced to improve training
stability and performance.

Fig. 3. Model Architecture

LSTM networks, a variant of RNN, are designed to overcome the vanishing

gradient problem. LSTMs comprise memory cells that selectively store or discard
information over time and are regulated by input, forget, and output gates.
These gates control the flow of information, enabling the network to make more
nuanced decisions about what to retain or forget.
The proposed LSTM model architecture, depicted in Fig. 3, comprises several
layers:

– An LSTM layer with 64 units, designed to capture long-term dependencies

in sequential data.
– A Dropout layer with a rate of 0.2, to mitigate overfitting.
– A Dense layer with a single unit, initialized with zeros.
– A Reshape layer to format the output shape to [1, 1].

This architecture is optimized for processing sequential data and includes

regularization techniques to enhance generalization.
As illustrated in Fig. 4, the LSTM model’s workflow begins with normalizing
the monthly average rainfall data and dividing it into time-windowed sequences.
These sequences are then used to train the LSTM network. Post-training, the
model can predict future rainfall using historical data windows.
54 P. N. Chowdary et al.

Fig. 4. Deep Learning Approach to Rainfall Data

3.3 Model Training

Dynamic Mode Decomposition (DMD). Our DMD analysis focused on

rainfall data from the India Meteorological Department, spanning 1901 to 2018,
with a specific emphasis on North-east India. The methodology involved the
following key steps:

1. Data Collection and Preprocessing: We segmented the IMD dataset into

monthly intervals using Python scripts, targeting the North-east region of
India based on geographical coordinates.
2. Matrix Construction: A data matrix X was constructed, representing
monthly rainfall over a 10-year period (120 months). This matrix encapsu-
lated the temporal dynamics of rainfall for the entire region.
3. Singular Value Decomposition (SVD): SVD was applied to X to identify
dominant rainfall patterns, a crucial step for distilling key dynamics from the
complex data.
4. Forecasting Analysis: Using the DMD algorithm, we forecasted future rain-
fall based on the historical 10-year dataset, calculating DMD modes, eigen-
values, and a dynamic matrix for future state prediction.
5. Performance Evaluation: The model’s accuracy was assessed using RMSE
and MAE.
6. Visualization: Actual vs. predicted rainfall data were visualized for key
locations to assess the model’s accuracy and understand temporal rainfall
patterns.

This methodology enabled effective application of DMD to rainfall prediction,

offering insights into the meteorological dynamics of North-east India.

Deep Learning (DL) Analysis. For the Deep Learning (DL) analysis, we uti-
lized Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory
(LSTM) networks, to model and predict rainfall data. This analysis focused on
key locations in North-east India, with the data spanning from 1951 to 2021.
The LSTM model was chosen for its ability to handle sequential data and its
robustness in capturing long-term dependencies.
 Rainfall Prediction 55

Data Preprocessing: The data was first standardized to form monthly aver-
ages. The dataset was then split into training and testing sets, with the training
set comprising 80% of the data.
Model Training: Our LSTM model training involved a series of experimen-
tal setups with varying parameters. The core architecture consisted of multiple
layers, including an LSTM layer and a Dropout layer. To optimize the model’s
performance, we experimented with different configurations, varying the num-
ber of hidden units in the LSTM layer, the dropout rate, the input and output
window sizes, and the choice of optimizer. The training involved adjusting these
parameters to find the most effective combination for our specific dataset and
objectives.
Model Evaluation: The performance of the model was assessed using metrics
such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).
Additionally, various plots were generated to visualize the training process and
compare the predicted rainfall against actual data.
Results Visualization: The model’s predictions were visualized for individual
locations, providing an intuitive understanding of the model’s performance over
time. This included a comparison of the predicted and actual rainfall data.
This DL approach, centered around LSTM networks, offered a powerful tool
for forecasting rainfall, demonstrating the potential of advanced neural networks
in meteorological data analysis.

Evaluation Metrics. For model evaluation, we used two primary metrics:

1. Root Mean Square Error (RMSE): A key metric for regression models,
RMSE measures the average magnitude of the error. It is calculated as the
square root of the mean squared differences between predicted and actual
values. Lower RMSE values indicate better model performance. The formula
is given by:

 n
1 
RMSE =  (yi − yˆi )2
n i=1

where n is the number of observations, yi is the actual value, and yˆi is the
predicted value for the i-th observation.
2. Mean Absolute Error (MAE): MAE calculates the average absolute error
between predicted and actual values. It’s particularly useful in datasets with
high variability or outliers. The formula for MAE is:
n
1
MAE = |yi − yˆi |
n i=1
Again, n represents the number of observations, with yi and yˆi as the actual
and predicted values, respectively.
56 P. N. Chowdary et al.

4 Results and Discussion

This section discusses the results obtained using Deep Learning and Dynamic
Mode Decomposition (DMD) techniques to predict rainfall. Our results showed
that both methods were able to accurately predict rainfall demonstrating the
potential for using advanced computational methods to improve weather fore-
casting and better prepare for extreme weather events.

4.1 DMD

Table 2 shows results from Dynamic Mode Decomposition (DMD) experiments

on a rainfall data matrix, constructed from 10 years of data, with predictions
made for a single year. DMD was conducted at various projection ranks to
obtain a low-dimensional representation of the data. The RMSE and MAE val-
ues indicate reasonably accurate rainfall predictions, ranging from 150.44 mm to
263.34 mm and 91.34 mm to 154.61 mm, respectively. Notably, the best perfor-
mance was observed for the data from 1995–2005.

Table 2. DMD Results for forecasting 1 year

Start Year Stop Year Predicted Duration Rank RMSE MAE

1929 1939 1 year 106 263.3423 154.6123
1941 1951 1 year 123 260.2758 144.6304
1954 1964 1 year 127 177.3236 107.8259
1973 1983 1 year 128 170.6351 109.3051
1995 2005 1 year 118 150.4379 91.3362
2000 2010 1 year 100 236.3857 124.9998
2005 2015 1 year 123 158.5830 97.9195

The rainfall prediction results for the year 2016, for Agartala, Guwahati,
Imphal, and Itanagar, were obtained by constructing a matrix X from rainfall
data between 2005 and 2015, using a projection rank of 123. The forecasted
precipitation for the entire year of 2016 is shown in Fig. 5. The prediction per-
formance was evaluated using the Mean Squared Error (MSE) and Root Mean
Square Error (RMSE), which were calculated to be 97.9195 and 158.5830, respec-
tively.
 Rainfall Prediction 57

Fig. 5. Predictions for regions of North East India in 2016 using DMD

4.2 Deep Learning

This section presents the results of our time series prediction model for rainfall in
four Indian cities: Itanagar, Guwahati, Agartala, and Imphal. The performance
of our model for each city was assessed based on various experimental setups
involving different optimizers, input/output window sizes, and dropout rates.

Itanagar. For Itanagar, the best performing optimizer and parameter combi-
nation for MAE was Nadam with an input window size of 13, an output window
size of 1, and a dropout of 0.2, yielding an MAE value of 0.3707. For RMSE,
the best performance was achieved by AdamW with an input window size of
14, an output window size of 1, and a dropout of 0.2, resulting in an RMSE
of 0.4527 (Table 3). Figure 6 illustrates the model’s predictions compared to the
actual data on the test set.

Imphal. The analysis for Imphal indicated that the Nadam optimizer with an
input window size of 14, an output window size of 1, and a dropout rate of
0.2 outperformed other settings, achieving the lowest MAE and RMSE values
(Table 4). The model’s predictive performance on the test set is depicted in Fig. 7.

Guwahati. For Guwahati, the optimal results were obtained using the Nadam
optimizer, with an input window size of 13, an output window size of 1, and a
58 P. N. Chowdary et al.

Fig. 6. Prediction Visualisation for Itanagar

Table 3. Performance Analysis of the proposed method for Itanagar

Optimiser Input Window Output Window Dropout Metrics

MAE RMSE
AdamW 13 1 0 0.3637 0.5285
0.2 0.3570 0.5249
14 1 0 0.3650 0.5356
0.2 0.3612 0.4527
15 1 0 0.3792 0.5505
0.2 0.3812 0.5459
Nadam 13 1 0 0.3526 0.5260
0.2 0.3550 0.5258
14 1 0 0.3569 0.5344
0.2 0.3695 0.5372
15 1 0 0.3670 0.5412
0.2 0.3707 0.5467

Fig. 7. Prediction Visualisation for Imphal

dropout rate of 0.2 (Table 5). The performance visualization of the model on the
test dataset is shown in Fig. 8.

Agartala. In the case of Agartala, the combination of Nadam optimizer, an

input window size of 14, an output window size of 1, and a dropout rate of 0
 Rainfall Prediction 59

Table 4. Performance Analysis of the proposed method for Imphal

Optimiser Input Window Output Window Dropout Metrics

MAE RMSE
AdamW 13 1 0 0.4422 0.6505
0.2 0.4421 0.6508
14 1 0 0.4448 0.6533
0.2 0.4477 0.6561
15 1 0 0.4513 0.6593
0.2 0.4519 0.6583
Nadam 13 1 0 0.4434 0.6534
0.2 0.4378 0.6449
14 1 0 0.4424 0.6490
0.2 0.4357 0.6419
15 1 0 0.4381 0.6446
0.2 0.4423 0.6502

Fig. 8. Prediction Visualisation for Guwahati

Table 5. Performance Analysis of the proposed method for Guhawati

Optimiser Input Window Output Window Dropout Metrics

MAE RMSE
AdamW 13 1 0 0.3402 0.4763
0.2 0.3368 0.4827
14 1 0 0.3341 0.4707
0.2 0.3374 0.4668
15 1 0 0.3351 0.4746
0.2 0.3371 0.4757
Nadam 13 1 0 0.3264 0.4665
0.2 0.2990 0.4410
14 1 0 0.3115 0.4555
0.2 0.3091 0.4559
15 1 0 0.3126 0.4695
0.2 0.3126 0.4525

was found to be most effective, as per Table 6. Figure 9 showcases the model’s
predictive accuracy on the test data for Agartala.
60 P. N. Chowdary et al.

Fig. 9. Prediction Visualisation for Agartala

Each city’s data analysis provides unique insights into the model’s perfor-
mance and helps refine the prediction strategy for different geographical loca-
tions.

Table 6. Performance Analysis of the proposed method for Agartala

Optimiser Input Window Output Window Dropout Metrics

MAE RMSE
AdamW 13 1 0 0.3284 0.5322
0.2 0.3887 0.5436
14 1 0 0.3859 0.5394
0.2 0.3839 0.5388
15 1 0 0.3901 0.5451
0.2 0.3820 0.5324
Nadam 13 1 0 0.3786 0.5314
0.2 0.3709 0.5227
14 1 0 0.3698 0.5271
0.2 0.3770 0.5348
15 1 0 0.3804 0.5388
0.2 0.3840 0.5396

5 Conclusion and Future Scope

In conclusion, this study highlights the efficacy of Dynamic Mode Decomposition
(DMD) and Long Short-Term Memory (LSTM) neural networks in predicting
rainfall. The LSTM model particularly excels due to its proficiency in captur-
ing long-term patterns and spatial-temporal dependencies, outshining the DMD
approach in forecasting critical rainfall events. This capability is vital for provid-
ing early flood warnings, crucial for disaster preparedness and mitigation efforts,
especially in vulnerable regions.
 Rainfall Prediction 61

5.1 Future Scope

Looking ahead, two main avenues present themselves for further research:
– Daily Forecasting: Future endeavors could focus on refining the models for
daily rainfall forecasting. This would involve tackling more detailed data and
complex patterns, but could lead to more precise and actionable predictions,
essential for effective short-term emergency response and planning.
– Conformal Prediction: Implementing conformal prediction methodologies
may enhance the utility of the rainfall forecasts. This approach would estimate
the confidence levels of the model’s predictions, adding a layer of reliability
that is particularly important in risk management and decision-making within
meteorological contexts.

To summarize, the study offers significant advancements in rainfall prediction

methods, with LSTM models showing particular promise in early warning sys-
tems. Future research in daily forecasting and conformal prediction is expected to
refine these models further, increasing their applicability in weather forecasting
and disaster management.

Acknowledgments. An earlier version of this work is available as a preprint on arXiv

(arXiv:2309.09336) [1].

References
1. Chowdary, P.N., et al.: Unleashing the power of dynamic mode decomposition and
deep learning for rainfall prediction in north-east India (2023)
2. Chung, J., Gülçehre, Ç., Cho, K., Bengio, Y.: Empirical evaluation of gated recur-
rent neural networks on sequence modeling. CoRR abs/1412.3555 (2014)
3. Dikshit, K.R., Dikshit, J.K.: Weather and climate of north-east India (2014).
https://api.semanticscholar.org/CorpusID:127226018
4. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
5. Jain, S., Kumar, V., Saharia, M.: Analysis of rainfall and temperature trends in
northeast India. Int. J. Climatol. 33 (2013). https://doi.org/10.1002/joc.3483
6. Sudhesh, K.M., Sowmya, V., Kurian, S., Sikha, O.K.: Ai based rice leaf
disease identification enhanced by dynamic mode decomposition. Eng. Appl.
Artif. Intell. 120, 105836 (2023). https://doi.org/10.1016/j.engappai.2023.105836,
https://www.sciencedirect.com/science/article/pii/S0952197623000209
7. Liyew, C.M., Melese, H.A.: Machine learning techniques to predict daily rainfall
amount. J. Big Data 8, 1–11 (2021)
8. Luo, C., Zhao, X., Sun, Y., Li, X., Ye, Y.: PredRANN: the spatiotemporal attention
convolution recurrent neural network for precipitation nowcasting. Knowl.-Based
Syst. 239, 107900 (2022)
9. Men, B., Wu, Z., Liu, H., Tian, W., Zhao, Y.: Spatio-temporal analysis of precipi-
tation and temperature: a case study over the Beijing-Tianjin-Hebei region, China.
Pure Appl. Geophys. 177, 3527–3541 (2020)
62 P. N. Chowdary et al.

10. Nagarajan, K., Ananthu, J., Menon, V.K., Soman, K.P., Gopalakrishnan, E.A.,
Ramesh, A.: An approach to detect and classify defects in cantilever beams using
dynamic mode decomposition and machine learning. In: Reddy, A.N.R., Marla, D.,
Simic, M., Favorskaya, M.N., Satapathy, S.C. (eds.) Intelligent Manufacturing and
Energy Sustainability. SIST, vol. 169, pp. 731–738. Springer, Singapore (2020).
https://doi.org/10.1007/978-981-15-1616-0 71
11. Pai, D., Rajeevan, M., Sreejith, O., Mukhopadhyay, B., Satbha, N.: Development
of a new high spatial resolution (0.25◦ × 0.25◦ ) long period (1901-2010) daily grid-
ded rainfall data set over India and its comparison with existing data sets over
the region. MAUSAM 65(1), 1-18 (2014). https://doi.org/10.54302/mausam.v65i1.
851, https://mausamjournal.imd.gov.in/index.php/MAUSAM/article/view/851
12. Palmer, T.N.: Towards the probabilistic earth-system simulator: a vision for the
future of climate and weather prediction. Q. J. R. Meteorol. Soc. 138(665), 841–
861 (2012). https://doi.org/10.1002/qj.1923, https://rmets.onlinelibrary.wiley.
com/doi/abs/10.1002/qj.1923
13. Pandianchery, M.S., Sowmya, V., Gopalakrishnan, E., Soman, K.: Long short-term
memory-based recurrent neural network model for covid-19 prediction in different
states of india. AI, IoMT, and Analytics, Emerging Technologies for Combatting
Pandemics (2022)
14. Pascanu, R., Mikolov, T., Bengio, Y.: On the difficulty of training recurrent neu-
ral networks. In: Dasgupta, S., McAllester, D. (eds.) Proceedings of the 30th
International Conference on Machine Learning. Proceedings of Machine Learning
Research, vol. 28, pp. 1310–1318. PMLR, Atlanta, Georgia, USA (2013)
15. Selvin, S., Vinayakumar, R., Gopalakrishnan, E., Menon, V.K., Soman, K.: Stock
price prediction using LSTM, RNN and CNN-sliding window model. In: 2017 Inter-
national Conference on Advances in Computing, Communications and Informatics
(ICACCI), pp. 1643–1647. IEEE (2017)
16. Tu, J.H., Rowley, C.W., Luchtenburg, D.M., Brunton, S.L., Kutz, J.N.: On
dynamic mode decomposition: theory and applications (2014)
17. Zaw, W.T., Naing, T.T.: Empirical statistical modeling of rainfall prediction over
Myanmar. World Acad. Sci. Eng. Technol. 2(10), 500–504 (2008)
18. Zhang, C.J., Wang, H.Y., Zeng, J., Ma, L.M., Guan, L.: Tiny-rainnet: a deep
convolutional neural network with bi-directional long short-term memory model
for short-term rainfall prediction. Meteorol. Appl. 27(5), e1956 (2020)
 Patient’s Condition Categorization Using
Drug Reviews

Akshit Kamboj, Shikha Mundra(B) , and Ankit Mundra

Manipal University Jaipur, Jaipur, Rajasthan, India

[email protected]

Abstract. In the modern era, there has been a substantial increase

in health-related concerns among individuals. The prevalence of various
diseases has surged significantly, underscoring the crucial necessity for
timely disease detection. A strategy employed for this purpose involves
the examination of drug evaluations from diverse patient groups. Within
the realm of healthcare, reviews assume a pivotal role. They offer insights
into the efficacy of different drugs and their potential adverse effects.
The primary objective is to categorize a patient’s health condition based
on their feedback pertaining to a specific medication. In this endeavor,
a system for classifying medical conditions has been developed, utiliz-
ing patient reviews related to diverse medications. This system has the
ability to forecast the specific ailment afflicting a patient. To achieve
this objective, the Drug Review Dataset from UCI’s Machine Learn-
ing Repository has been employed. To extract relevant features, multiple
vectorization techniques have been applied, including Bag of Words, TF-
IDF, TF-IDF (Bigrams), and TF-IDF (Trigrams). To execute the neces-
sary categorization, a range of machine learning classification algorithms
have been harnessed, such as Decision Tree, Multinomial Naive Bayes,
Logistic Regression, and Passive Aggressive Classifier.

Keywords: Statistical Features · Depression · BoW · TF-IDF ·

Patient Review

1 Introduction

Over the passage of time, the internet has undergone an astounding transfor-
mation. In its earlier days, it merely served as a platform for individuals to
promote their businesses. However, the landscape of the internet has since meta-
morphosed into a dynamic arena where a plethora of services and products are
assessed by consumers based on the feedback they glean from fellow consumers.
This practice of utilizing reviews and feedback to gauge the caliber of a product
or service is an approach that can also find application within the realm of health-
care. In recent times, the proliferation of online reviews has given rise to novel
domains within the realms of communication and marketing. This, in turn, has
effectively bridged the chasm between feedback that attains a viral status and
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 63–73, 2025.
https://doi.org/10.1007/978-3-031-71481-8_5
64 A. Kamboj et al.

traditional word-of-mouth dissemination. Such a convergence has proven to be

a substantial influence on consumer perceptions. The internet’s evolution from
a simple promotional space to a multifaceted platform for consumer evaluation
has engendered a transformative impact, transcending industries and sectors,
including healthcare, thereby underscoring the formidable role of feedback and
reviews in shaping contemporary perspectives.

1.1 Predicting Medical Conditions Using Patient Feedback

and Machine Learning

Within the realm of Within the realm of healthcare, drug evaluations across
various types of medications can assume heightened significance. These assess-
ments offer assistance through various avenues. To illustrate, they enable the
monitoring of unfavorable responses linked to diverse drugs, identification of
prevailing sentiments surrounding specific medications, and acknowledgment of
the range of ailments addressable by a given drug. In online patient communi-
ties, articulating the medical context for which a medication is being employed
holds paramount importance [1]. This stems from the fact that numerous drugs
possess the potential to alleviate multiple conditions. This research aims to con-
duct a comprehensive analysis of drug reviews for the purpose of predicting
specific medical conditions in individual patients. These reviews originate from
diverse patients discussing various medications on different platforms dedicated
to medicine feedback. The core objective of this project is to construct a clas-
sification system for medical conditions, utilizing patients’ reviews associated
with different drugs [2]. This system will possess the capability to anticipate
the underlying medical condition afflicting a given patient. To accomplish this
goal, the Drug Review Dataset sourced from UCI’s Machine Learning Repos-
itory will serve as the primary dataset. The process of feature extraction will
involve employing several vectorization techniques, including Bag of Words, TF-
IDF, GloVe, and Word2Vec as discussed in [8] and [9]. The subsequent step
entails employing a range of machine learning classification algorithms to fulfill
the necessary categorization tasks. Notably, algorithms such as Decision Tree,
Multinomial Naive Bayes, Logistic Regression, and Passive Aggressive Classifiers
will be leveraged for this purpose.

2 Literature Review

In most of the prior study, machine learning models are experimented based
on the symptoms of the patient. Shatha Melhem’s et al. [1] research focuses on
harnessing the capabilities of machine-learning models to effectively categorize
patient care into two distinct classes: inpatient and outpatient. This innovative
approach has yielded remarkable reductions in time and effort required for such
categorization, thereby facilitating swift and accurate determination of the most
suitable services for individual patients. The consequential decrease in the poten-
tial risk to patients’ lives arises from the notable mitigation of human errors. To
 Patient’s Condition Categorization Using Drug Reviews 65

achieve these goals, the study employs four distinct machine learning models:
Decision Tree, Support Vector Machine, K-Nearest Neighbours (KNN), and Ran-
dom Forest. These models facilitate the meticulous analysis of a patient’s unique
combination of lab test results and overall medical condition. The comparison of
these models is based on a predefined set of criteria, encompassing parameters
like accuracy, specificity, sensitivity, and precision score. The datasets used for
training these models originate from a hospital in Indonesia and encompass a
variety of patient cases, featuring their corresponding laboratory test outcomes.
Among these models, the Random Forest Model emerges as the most effective
performer. Demonstrating its superiority, this model achieves an accuracy rate of
77%, a sensitivity rate of 65%, and a precision rate of 72 Samiullah Shaikh’s et al.
[2] paper endeavors to introduce a comprehensive dataset encompassing intricate
profiles of approximately 12,000 patients. These profiles are intricately linked to
the 20 most frequently encountered diseases. A significant aspect of this paper
involves the facilitation of disease classification. This categorization process is
facilitated through the application of diverse machine learning models, which pre-
dominantly follow a supervised learning approach. Incorporating the principles of
Evidence-Based Medicine (EBM), the paper seeks to embody a decision-making
framework rooted in patient records, clinical expertise, and available evidence.
This approach entails leveraging previous patient records to inform medical pro-
fessionals’ prescription decisions. The rigorous experimentation within this study
relies on the Monte Carlo method. Among the array of methods employed, the
Random Forest Trees (RFT) technique emerges as the optimal performer. This
model attains an impressive accuracy rate of approximately 83%, accentuating
its efficacy in disease classification. Angelower Santana-Velásquez et al. [3] study
endeavors to supplant conventional Diagnosis-Related Groups (DRGs) classifi-
cation techniques with advanced machine learning methodologies. The primary
objective of this research is to assess the viability of machine learning mod-
els in categorizing patients according to the established DRGs standards. This
assessment will hinge on utilizing available patient-related data. The insights
garnered from this investigation will serve as a foundational reference for future
analyses. Specifically, these insights will be juxtaposed with the early-stage pre-
dictions rendered by DRGs during patients’ hospitalization. The outcomes of this
research highlight that by employing Ensemble methods and Artificial Neural
Networks for DRGs classification, remarkable accuracy levels of up to 96% can be
attained. This achievement has significantly bolstered the efficiency of medical
treatments and procedures for patients. Importantly, these advancements have
yielded substantial cost reductions without compromising the quality or duration
of the patients’ hospital stays. The primary objective of the research conducted
by Pushpak Bhattacharyya et al. [6] is to provide comprehensive insights into
the diverse dimensions of sentiment analysis. This encompasses an exploration
of both machine learning-based and knowledge-based methodologies. An addi-
tional facet of this research pertains to the extension of sentiment analysis to
Indian languages, thereby catering to linguistic diversity. Through this inves-
tigation, a significant observation emerged: working with senses, as opposed
66 A. Kamboj et al.

to isolated words, proves notably advantageous in terms of accuracy, particu-

larly in the context of multilingual scenarios. The research also presents note-
worthy findings concerning sarcasm and thwarting, providing valuable insights
into the challenges posed by these linguistic nuances. Yogesh Chandra’s et al.
[4] paper is directed towards leveraging a diverse array of methods to facili-
tate sentiment analysis. The focal point of this research involves the utilization
of machine learning classifiers to achieve the task of sentiment analysis. This
entails the incorporation of Deep Learning models and polarity-based sentiment
analysis to discern whether a given user’s tweet carries a positive or negative
sentiment. To accomplish this categorization, a range of model architectures
have been harnessed, thereby accommodating the variations in thoughts and
opinions prevalent across distinct social media platforms. The outcomes derived
from this study bear substantial relevance for Government and Defense Orga-
nizations. By harnessing the insights garnered, these entities gain the capacity
to assume control and make informed decisions about their subsequent course
of action. Ultimately, this research contributes to enabling more informed and
strategic decision-making based on sentiment analysis across social media con-
tent. The study conducted by Hao Yang et al. [7] focuses on the detection of
malware through the utilization of the TF-IDF model. A notable advantage
of employing Natural Language Processing (NLP) techniques over alternative
machine learning algorithms is the heightened accuracy that NLP can achieve
post-data processing. The study proceeds by constructing subsequent models,
incorporating classifiers such as the Random Forest classifier, Gradient Boosting
classifier, AdaBoost classifier, and ensemble models. In parallel, the study lever-
ages a Convolutional Neural Network (CNN) for training, drawn to its efficiency
in extracting valuable data insights. The TF-IDF model serves a central role,
determining the TF-IDF value attributed to each keyword within a given traffic
packet, without necessitating the packet’s division. The dataset encompasses a
total of 3000 data packets, equally split between black and white data (1500
each). After keyword extraction and the reconstruction of the dataset using the
TF-IDF model, diverse classifiers are trained. Encrypted traffic data undergoes
One-Hot Encoding before being fed into the classifiers and the CNN network
for training and subsequent detection. The study underscores the potential of
combining machine learning and deep learning techniques as a viable solution for
identifying encrypted malware traffic data. Demonstrating a detection accuracy
of 93% using the TF-IDF-based model for encrypted traffic malware detection,
the research presents promising results. This accuracy is anticipated to advance
further through continued research and development in this field.

3 Proposed Methodology

In this work, Bag of Words (BoW), n-gram along with Tf-idf are employed
to identify the patient category from the drug reviews. In BoW, within this
model, the frequency of distinct words within a document is tallied. However,
this methodology disregards the sequence in which the words are presented in
 Patient’s Condition Categorization Using Drug Reviews 67

the document. Consequently, the bag of words approach is characterized as con-

textually agnostic. In this scheme, the value “1” signifies the presence of a word in
the document, while “0” indicates its absence. This model works in the following
manner:
1. Vocabulary Creation: The first step is to create a vocabulary, which is a
unique set of words present in the entire corpus of documents (collection of
texts) you’re working with. Each word is treated as a unique token, and no
consideration is given to the grammatical or semantic meaning of the words
at this stage.
2. Tokenization: In this step, each document in the corpus is divided into indi-
vidual words or tokens. Punctuation and capitalization are usually removed,
and words are typically converted to lowercase to ensure consistent represen-
tation.
3. Word Counting: For each document, the BoW model counts the occurrences
of each word in the vocabulary. This information is then organized into a
numerical vector, where each entry corresponds to a specific word in the
vocabulary, and the value represents the frequency of that word in the docu-
ment. If a word doesn’t appear in the document, its corresponding entry in
the vector will have a value of zero.
4. Vectorization: Once the word counts are obtained for each document, the
documents are represented as vectors in a high-dimensional space. The length
of the vector is equal to the size of the vocabulary, and each dimension of the
vector corresponds to a word in the vocabulary. The value in each dimension
represents how many times the corresponding word appears in the document.
5. Matrix Representation: The collection of these vectors for all documents forms
a matrix, known as the BoW matrix. Rows of this matrix represent individual
documents, while columns represent the words in the vocabulary. Since the
vocabulary is typically large and most documents will only contain a small
subset of the words, the BoW matrix is often sparse (contains mostly zeros)
(Fig. 1).

3.1 Term Frequency - Inverse Document Frequency (TF - IDF)

This is an alternative algorithm that can be employed for feature extraction,
offering greater efficiency in comparison to the bag of words approach. This
technique consists of two distinct components. This model works in the following
manner:

1. Term Frequency (TF): The term frequency of a word in a document is the

ratio of the number of times that word appears in the document to the total
number of words in the document. It gives you an idea of how frequently a
word occurs within a specific document.
2. Inverse Document Frequency (IDF): The inverse document frequency of a
word is a measure of how unique or important that word is across all the
documents in the corpus. It’s calculated as the logarithm of the ratio of the
68 A. Kamboj et al.

Fig. 1. Workflow of Proposed Method

total number of documents to the number of documents containing the word.

Words that appear in many documents have lower IDF values, while words
that appear in fewer documents have higher IDF values, thus emphasizing
the importance of rarer words.
3. TF-IDF Calculation: The TF-IDF score for a word in a document combines
the term frequency and inverse document frequency. It gives a high weight to
words that appear frequently in the document (high TF) but are rare across
the entire corpus (high IDF). Conversely, it assigns a low weight to words
that are either too common or too rare.
4. Vectorization: Each document is represented as a numerical vector, where
each dimension of the vector corresponds to a unique word in the vocabulary.
The value in each dimension is the TF-IDF score of the corresponding word in
the document. This vectorization process converts the text data into a format
suitable for various machine learning algorithms.

4 Experimental Procedure
The Dataset for this work is a Drug Reviews dataset. This dataset has been
obtained from the UCI Machine Learning Repository. The dataset provides infor-
mation about the different reviews related to various types of drugs used by the
patients. There are over 210000 rows (data points) and 6 feature attribute in the
given dataset such as Drug Name, Condition, Review, Rating, Date and Useful
Count. The exploratory analysis of this data is mentioned in next step.

4.1 Exploratory Data Analysis

This stage encompassed achieving a comprehensive understanding of the dataset.
Diverse facets of the dataset were thoroughly examined. During this phase, the
subsequent tasks were undertaken:
 Patient’s Condition Categorization Using Drug Reviews 69

1. Exploration of Selected Features to Identify Trends and Patterns: This

involved a thorough examination of specific attributes within the dataset to
uncover trends, patterns, and potential relationships. The aim was to extract
meaningful insights that could contribute to the subsequent analysis [4].
2. Segregating the Dataframe for Analyzing Individual Conditions: The dataset
was partitioned into distinct segments to facilitate focused analysis of indi-
vidual conditions or categories. This segmentation aimed to reveal condition-
specific nuances and behaviors.
3. Creating WordCloud for Different Classes of the Output Variable: Word-
Clouds were generated for various classes or categories of the output variable
as shown in Fig. 2, 3, 4 and 5. WordClouds visually depicted the frequency
distribution of words within these classes, providing an intuitive glimpse into
the prevalent terms.

Fig. 2. WordCloud for ‘Birth Control’ Class

Fig. 3. WordCloud for ‘Blood Pressure’ Class

70 A. Kamboj et al.

Fig. 4. WordCloud for ‘Depression’ Class

Fig. 5. WordCloud for ‘Diabetes Type 2’ Class

4. Removing Double Quotes from All the Reviews Present in the Dataset: Dou-
ble quotation marks were systematically eliminated from the reviews con-
tained in the dataset. This preprocessing step aimed to ensure uniformity
and eliminate any potential interference with subsequent analyses.
5. Removing the StopWords: StopWords, which are common words with little
contextual significance, were removed from the reviews. This process aimed
to enhance the quality of subsequent analyses by eliminating noise.
6. Converting the Reviews to Individual Important Words: The reviews were
transformed into sets of significant individual words. This transformation was
instrumental in simplifying the data while retaining essential content.
 Patient’s Condition Categorization Using Drug Reviews 71

7. Creating Features and Target Variables: The dataset was organized into fea-
tures (input variables) and a target variable (output variable). This arrange-
ment was vital for subsequent machine learning tasks.
8. Performing Train Test Split: The dataset was divided into training and testing
subsets. The training set was used for model training, while the testing set
was employed to assess model performance [5].
Creating a Function for Plotting Confusion Matrix: A specialized function
was developed to create and visualize confusion matrices. Confusion matrices
are crucial tools for evaluating the performance of classification models.

5 Result Analysis, Conclusion and Future Work

This project employed a trio of diverse feature extraction approaches: Bag
of Words, TF-IDF, TF-IDF (Bigrams), and TF-IDF (Trigrams) as mentioned
Table 1, 2, 3 and 4. In conjunction with each of these techniques, a set of distinct
machine learning models was applied. These models encompassed the Logistic
Regression Model, Decision Tree Classifier, Naive Bayes Classifier, and Passive
Aggressive Classifier.

Table 1. Result analysis using Bag Of Words

Method Accuracy Precision Recall F1-Score

Logistic Regression 97.99 96.33 94.52 0.9539
Decision Tree Classifier 95.98 92.37 90.99 0.9166
Naive Bayes Classifier 97.02 94.17 92.79 0.9346
Passive Aggressive Classifier 97.77 95.28 94.52 0.9489

Table 2. Result analysis using TF-IDF (Unigram)

Method Accuracy Precision Recall F1-Score

Logistic Regression 97.73 96.32 93.38 0.9476
Decision Tree Classifier 95.52 91.49 90.18 0.9081
Naive Bayes Classifier 92.14 95.53 73.06 0.8044
Passive Aggressive Classifier 98.27 96.25 95.71 0.9598

Numerous tasks were undertaken throughout the course of this research

project. The initial step involved formulating a solvable problem statement. Sub-
sequently, an extensive array of research papers closely aligned with the prob-
lem statement were meticulously analyzed. The subsequent phase encompassed
the identification of a suitable dataset, sourced from the UCI machine learning
72 A. Kamboj et al.

Table 3. Result analysis using TF-IDF (Bigrams Approach)

Method Accuracy Precision Recall F1-Score

Logistic Regression 97.50 96.93 91.82 0.9414
Decision Tree Classifier 95.55 90.76 90.14 0.9044
Passive Aggressive Classifier 98.21 96.23 95.55 0.9588

Table 4. Result analysis using TF-IDF (Trigram Approach)

Method Accuracy Precision Recall F1-Score

Logistic Regression 97.09 96.65 90.26 0.9312
Decision Tree Classifier 95.17 90.40 89.57 0.8998
Passive Aggressive Classifier 98.61 97.73 95.89 0.9678

repository. This acquired dataset comprised pivotal features wielding significant

influence in deducing the specific medical condition afflicting a given patient.
Upon dataset preparation, a thorough exploration was conducted, resulting in
the integration of various refinements. Following this, the dataset underwent
Train-Test Splits to facilitate robust model evaluation. Three distinct feature
extraction methodologies were implemented to uncover salient terms within indi-
vidual reviews. This, in turn, streamlined the process of categorizing patients’
medical conditions in a precise and efficient manner. For each of these fea-
ture extraction methods, five diverse machine-learning models were employed.
Notably, the passive-aggressive classifier exhibited remarkable accuracy, emerg-
ing as the most precise model across all feature extraction techniques utilized
within the project. This observation held true for all the employed methods.
The pinnacle of accuracy, approximately 98.57%, was attained through the TF-
IDF: Bigrams Approach. Beyond the passive-aggressive classifier, the logistic
regression model showcased notable accuracy, achieving approximately 97.99%
accuracy in the Bag of Words scenarios. In the context of our work, there
exists a plethora of avenues for implementation, both in the near future and the
long term. In the immediate horizon, the foremost objective involves continued
experimentation with various Machine Learning Models to pinpoint the optimal
model aligned with specific requirements. Concurrently, it becomes imperative
to enhance the predictive accuracy of each attempted model. Attaining these
objectives necessitates addressing various considerations, including the mitiga-
tion of multicollinearity, the prevention of model overfitting, and other relevant
aspects. These concerns are particularly significant when dealing with Decision
Tree and Random Forest Models [6]. Throughout these iterative processes, the
overarching emphasis remains steadfastly fixed on achieving the utmost accuracy
while simultaneously safeguarding against challenges like multicollinearity and
overfitting. Looking towards the long-term trajectory, the approach of employ-
ing diverse feature extraction methods for the purpose of classification holds the
 Patient’s Condition Categorization Using Drug Reviews 73

potential for extension into various other domains. Noteworthy domains ripe
for exploration include e-commerce, banking, and social media sentiment analy-
sis. The adaptability of this methodology underscores its potential applicability
across an array of fields, ultimately contributing to a more nuanced understand-
ing and insightful categorization within these domains.

References
1. Melhem, S., Al-Aiad, A., Al-Ayyad, M.S.: Patient care classification using machine
learning techniques, in 2021 12th International Conference on Information and Com-
munication Systems (ICICS), pp. 57–62 (2021)
2. Shaikh, S., Khan, M.Y., Nizami, M.S.: Using patient descriptions of 20 most common
diseases in text classification for evidence-based medicine. In: Mohammad Ali Jinnah
University International Conference on Computing (MAJICC), vol. 2021, pp. 1–8
(2021)
3. Santana-Vel asquez, A., John Freddy Duitama, M., Arias-Londo no, J.D.: Classifi-
cation of diagnosis-related groups using computational intelligence techniques. In:
2020 IEEE Colombian Conference on Applications of Computational Intelligence
(IEEE ColCACI 2020), pp. 1–6 (2020)
4. Chandra, Y., Jana, A.: Sentiment analysis using machine learning and deep learn-
ing. In: 2020 7th International Conference on Computing for Sustainable Global
Development (INDIACom), pp. 1–2 (2020)
5. Kamboj, A., Kumar, P., Bairwa, A.K., Joshi, S.: Detection of malware in down-
loaded files using various machine learning models. Egypt. Inform. J. 24(1), 81–94
(2023)
6. Bhattacharyya, P.: Sentiment analysis. In: 2013 1st International Conference on
Emerging Trends and Applications in Computer Science (2013)
7. Yang, H., He, Q., Liu, Z., Zhang, Q.: Malicious encryption traffic detection based
on NLP. In: Security and Communication Networks, vol. 2021, Article ID 9960822,
p. 10 (2021)
8. Mundra, S., Mittal, N.: CMHE-AN: code mixed hybrid embedding based attention
network for aggression identification in Hindi English code-mixed text. Multimedia
Tools Appl. 82(8), 11337–11364 (2023)
9. Mundra, S., Mittal, N.: FA-Net: fused attention-based network for Hindi English
code-mixed offensive text classification. Soc. Netw. Anal. Min. 12(1), 100 (2022)
 Systematic Review on Sustainable Healthcare
Practices: A Bibliometric Approach

Shubhangi V. Urkude1(B) and Debajani Sahoo2

1 Department of Operations and IT, IBS, IFHE University, Hyderabad, Telangana, India
[email protected]
2 Department of Marketing, IBS, IFHE University, Hyderabad, Telangana, India

[email protected]

Abstract. In order to map the research trends in this field, the goal of this paper is
to analyse the body of literature on sustainable healthcare practices using biblio-
metric analysis and network visualization. In this work, research literature obtained
from the Scopus and Web of Science databases between 2003 and 2023 is sub-
jected to bibliometric analysis methods. Indicators like changes in research pro-
duction and citations, top contributing authors, countries, journals, keyword co-
occurrences, and co-authorship ties among contributing countries are all evaluated
in the study. The findings shows, the USA and UK are now leading researcher in
this area, with Australia and Canada also playing a significant role. The field of sus-
tainable healthcare practices encompasses a number of different fields, including
public health, healthcare delivery, major clinical trials, questionnaires, middle-
aged adult male and female demographics, clinical practices, psychology, and
public health education. There is a tone of opportunity for research in the future
that connects sustainable healthcare to these subjects in terms of authors, jour-
nals, and organizations that may offer knowledge and possible direction for future
exploration by detailing the research’s organizational framework.

Keywords: Sustainable Healthcare · public health · healthcare delivery ·

Bibliometric analysis · Bibliometrix · Biblioshiny · Citation analysis

1 Introduction

Natural resources are essential to daily life, daily business, and human existence. The
depletion of natural resources could result from disregarding sustainability. Sustainabil-
ity is a crucial area of study. Business students should study sustainability since it makes
companies more consumer-friendly and helps them meet their CSR (Corporate Social
Responsibility) commitments. According to the World Health Organization (WHO), a
sustainable health care system improves, maintains, or restores health while minimizing
negative environmental effects and maximizing opportunities to improve and transform
it to benefit the health and well-being of both present and future generations. Several
projects and studies have offered advice and discussed personal experiences with hospi-
tal energy management, demonstrating improvements and some encouraging outcomes.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 74–85, 2025.
https://doi.org/10.1007/978-3-031-71481-8_6
 Systematic Review on Sustainable Healthcare Practices 75

However, most hospitals continue to struggle with developing solutions for the sustain-
able use of energy [1]. The assessment of environmental influences on healthcare is still
in its early stages and does not yet give enough thorough information to enable healthcare
decision-making, both at the level of particular hospitals and health systems as a whole
[2]. The goal of this present study is to review the research on healthcare management
and practice health management sustainably. The following research questions (RQs)
are what we plan to investigate in the context of this project. To review number of articles
published during 2003–2023 in the era of sustainable healthcare practices.
• To find the trends and frequency of articles in sustainable healthcare across the globe.
• To find the most contributing authors and documents over last 20 years.
Due to the aforementioned goals, this study aims to offer (1) an overview of the bib-
liometric technique and (2) thorough guidelines for carrying out bibliometric analysis
of healthcare research. This paper introduces business scholars to bibliometric analy-
sis by explaining its ideas, methodologies, and practices and providing examples and
arguments. The advantages of this work include, first, helping business academics com-
prehend the bibliometric approach so they can analyze certain fields in the body of
literature utilizing a large corpus and bibliometric data. The document provides a sum-
mary of bibliometric analysis and the rules that govern it. Second, by providing several
recommendations regarding the various techniques for bibliometric analysis in sustain-
able healthcare practices and when to use them, the study can broaden the perspective
of medical professionals and academics in the business world on the alternatives and
justifications for using the various variants of bibliometric analysis.
Rest of the article is organized as follows. Section 2, discussed the research method-
ology with database selection and keywords used to download the paper in detail. Result
and Analysiswas elaborated in Sect. 3 followed by discussion in Sect. 4. Section 5
elaborates conclusion and future scope for the next researchers.

2 Research Methodology

Notably, the development of scientific databases like Scopusand Web of Science has
made it easier to obtain substantial amounts of bibliometric data.Gephi, Leximancer,
VOSviewer and biblioshiny are the bibliometric software available for systematic lit-
erature review [24, 25]. There are many databases available like IEEE, science direct,
Google scholar to download the paper but biblioshiny package in R supports both .csv
and .bibtext file format to input the data, which was easier to get in Scopus and web
of science databases. A recent rise in scholarly interest in bibliometric analysis can be
attributed to this software. In several fields of business research, such as business strategy
[3], electronic commerce [4], and finance [5–7], a bibliometric technique is thus in use. In
the domains of human resources, management, and marketing [8, 9], bibliometric is used
to analyze the intellectual structure in addition to researching collaboration patterns and
publications [10–12]. This work used a standard complete bibliometric analysis, which
starts with a keyword search in the Scopus search syntax, as shown in Fig. 1. Scopus
and web of science databases was used to download the papers in the area of sustainable
healthcare practices with the help of relevant keywords drawn through the literature [26,
76 S. V. Urkude and D. Sahoo

27]. To perform the analysis biblioshiny interface was selected from bibliometric pack-
age in R. Performance analysis without scientific mapping was not advisable [13]. The
keywords relevant to the sustainable healthcare practices were extracted from the past
literature which was published in reputed journal. All the keywords were taken as com-
bination of AND, OR logical format. The relevant papers which satisfy above criteria
was selected and downloaded as .bibtex file. The keywords suggested that papers and
publications with “Sustainable Practices,” OR “Green Practices,” OR “Ecology,” etc.
AND “Healthcare services” OR “Medical services” OR “Health services” in their titles,
keywords, or abstracts are relevant. During the search process, a total of 1755 articles
were selected between the years 2003 and 2023 without any filtration. The search results
were checked for redundant information. After that, only journal articles, review articles,
and articles in English and within the keywords of medicine, healthcare, health policy,
sustainable science, environmental science, and supply chain management are included
for further analysis. A total of 869 articles from the following journal categories: “Med-
ical,” “Social Science,” “Business Management,” “Health Profession,” “Psychology,”
“Environmental Science,” and “Multidisciplinary” were taken into consideration for the
study. The file was saved in “bibtex” format. A similar procedure was followed on the
web of science (WOS) platform to download data to retrieve 22 articles in bibtex format
after purification. Both Scopus and WOS databases have been used to collect a maximum
number of papers in the discipline of sustainability practices in the healthcare industry.

Fig. 1. Article selection criteria

Both .bibtex files are merged using the bibliometric function in R. After removing 4
duplicates, 887 articles were exported for Bibliometric analysis using Biblioshiny. The
biblioshiny platform is the best choice for readily displaying bibliometric maps and is
 Systematic Review on Sustainable Healthcare Practices 77

intelligible for any audience. Biblioshiny was used to build and graphically portray the
network of authors and countries using the bibliometric R program. This study used
Biblioshiny to create a notable keyword cloud that helps to visualize database clusters.

3 Results and Analysis

3.1 Keywords Co-occurrence Analysis

Fig. 2. Co-occurrence network of High-frequency keyword (Color figure online)

Bibliometric typically makes use of the co-occurrence analysis keywords extracted

from the Pub Med database-A specific co-occurrence matrix filtering process must be
performed to eliminate the low-frequency items for the more understandable result due
to poor representativeness. Figure 2 shows the co-occurrence network analysis. This
network showcased terms such as healthcare delivery, organization and management,
healthcare personnel, healthcare quality, and healthcare policy from all demographic
aspects such as male, female, adult, etc.
Two distinct color clusters, red and blue, are displayed in the picture. The circle’s
diameter reflects the keyword’s frequency of use. The interaction between the keywords
in the same cluster shows the dense network with more interaction than the interaction
among the cluster elements.
78 S. V. Urkude and D. Sahoo

3.2 Thematic Analysis

Fig. 3. Thematic map

Figure 3 shows the thematic analysis for the last 20 years (2003–2023) showing four
quadrant and six themes. After investigation, the study received 6 themes based on two
dimensions such as development and relevance degree, which are categorized in 4 quad-
rants. Out of all these 4 quadrants, the first quadrant represents motor themes, which was
study about male, female, human and adult health covering 2 themes. The second quad-
rant represents a space of basic and transversal themes of healthcare delivery and health
care facilities. The third quadrant, niche, represents the study about health care personnel
and questionnaires used in the data collection process. The fourth quadrant represents
sustainability, sustainable development, environmental sustainability, education, human
experiment and organization. It was named as emerging and declining theme as most of
the research was started towards sustainability, which include 2 themes.

3.3 Global Cited Documents in Sustainable Healthcare

Table 1 represents the top 10 globally cited authors with the journal they published. The
analysis shows that the highest number of citations received by Simons 2018(819), with
an average citation per year (of 136.50), followed by Chambers, 2013 having a total
citation (of 749) with an average citation of (68.09). It is interesting to notice that the
citation of sustainability in healthcare service is very striking, and the latest publication,
like Simon, 2018; Kouhizadeh, 2021; Kaika, 2017, have higher citations than older
papers. Normalized TC indicated the average citation for the particular author in the
sustainability area during the given time.

3.4 Publications Per year

The graph of the annual increase in publications is shown in Fig. 4. As per the figure,
although publication on sustainable healthcare practices started in 2004, it picked up from
2015 onwards. During 2020 more focus was given to sustainable healthcare practices
due to the worldwide pandemic.
 Systematic Review on Sustainable Healthcare Practices 79

Table 1. Top 10 authors and Journal names for article published

S. No Author and Total Citations TC per Year Normalized TC DOI

Journal
1 Simons mp, 2018, 819 136.50 30.92 https://doi.org/
Hernia 10.1007/s10029-
017-1668-x
2 Chambers da, 749 68.09 17.88 https://doi.org/
2013, Implement 10.1186/1748-
Sci 5908-8-117
3 Kouhizadeh M, 325 108.33 43.41 https://doi.org/
2021, Int J 10.1016/j.ijpe.
Production 2020.107831
Economics
4 Kaika M, 2017, 260 37.14 13.52 https://doi.org/
Environ Urban 10.1177/095624
7816684763
5 Boissy A, 2016, J 247 30.88 8.69 https://doi.org/
Gen Intern Med 10.1007/s11606-
016-3597-2
6 Armour C, 2007, 211 12.41 3.86 https://doi.org/
Thorax 10.1136/thx.
2006.064709
7 Carling Pc, 2008, 197 12.31 4.51 https://doi.org/
Infect Control 10.1086/591940
Hosp
Epidemiology
8 Pinzone M, 2016, 177 22.13 6.23 https://doi.org/
Journal of 10.1016/j.jcl
Cleaner epro.2016.
Production 02.031
9 Ulrich Rs, 2011, 171 13.15 6.78 https://doi.org/
Health Environ 10.1177/193758
Res Des J 671000400107
10 Proctor E, 2015, 167 18.56 7.25 https://doi.org/
Implement 10.1186/s13012-
Science 015-0274-5

3.5 Most Cited Country with a Total Citation

The total number of citations by country is highlighted in Fig. 5, which reveals that the
United States, the United Kingdom, and Australia have more citations than the other
nations, with respective total of 4712, 1536, and 1104. The United States of America’s
average citations were almost three times higher than its counterpart countries like United
Kingdom and Australia. The reason might be the USA research funding and ideas were
80 S. V. Urkude and D. Sahoo

Fig. 4. Number of publications over the year

Fig. 5. A most cited country with a total citation

far ahead of its counterparts. Conversely, countries like France, Germany, Portugal, and
Spain have very low average citations. Researchers may have become interested in this
sustainability field as a result of the significant growth that these nations’ healthcare
systems have experienced. Countries such as the Netherlands, Italy, and Switzerland
stood in the middle position as per their average citation. They might grow and compete
with the developed country in the upcoming era.

3.6 Bradford Law for Top 10 Journals

A criterion for weakening the search result in the scientific publication was put forth by
Bradford (1934) [14]. Bradford law states that if the number of articles in the particular
field were arranged in a order within three groups then number of articles in each group
will in the ratio of 1:n:n2, that means first group will end with 1 article, second group
with n number of articles and third group with twice the n articles, where n represents
 Systematic Review on Sustainable Healthcare Practices 81

Fig. 6. Bradford law—Top 10 journals in the scope of the study

the sources of downloading the articles. This also indicate even if increase of sources
to find the articles were increase exponentially there was same number of articles will
be available in search which is also stated as exponential diminishing returns. Fig. 6
lists the top 10 journals that contribute to the body of knowledge on sustainable medical
procedures. Over 45% of the research literature was contributed by the top 10 journals,
showing a considerable concentration of study. The journal with the most of the articles
(44), accounting for over 21% of all the papers in the sample set, is BMC Health Service
Research. It is followed by BMJ Open (34 articles) and sustainability, a journal with
a Swiss base, which has 14% of the sample set’s articles (30) out of 210 total articles.
Journals, including Health Management Forum, Rural and Remote Health, and Jour-
nal of Cleaner Productions, simultaneously publish fewer publications on sustainable
healthcare practices.

3.7 Evolution of Dendrogram in Sustainable Healthcare Practices

Fig. 7. Topic dendrogram for articles.

Figure 7 shows the topic dendrogram concerning healthcare sustainability practices.

Dendogram is a hierarchical tree like structure showing how the clusters are merged.
Each cluster is a group of similar objects and one cluster is merged with another cluster
82 S. V. Urkude and D. Sahoo

based on the centroid. Grouping is carried out based on the keywords appeared in the
articles and cited by other articles. Dendogram has clusters with two colors. Blue color
indicate cluster from existing literature on which various studies done and red color indi-
cate very few research has been done on healthcare, sustainability. However, researchers
can focus on this area and get new insights. There are three primary strands to the den-
drogram. The first strand focuses on an extensive clinical and controlled investigation
of diverse demographic groups such as male, female, adult, and middle-aged profile and
their psychology on those sustainable healthcare methods. The second strand focuses
on healthcare delivery, healthcare policy, healthcare planning and practices, and orga-
nizational leadership position on sustainability. Suppose we look at the second block,
which likewise has the most divisions, as an organizational development on sustainabil-
ity. The third strand emphasizes primary healthcare, practice guidelines, patient care,
patient education, etc. There was a correlation between the second and third strands on
the ground of healthcare delivery and healthcare service procedure, so the second stage
of clubbing in the dendrogram occurred again between the second and third strands.
New emerging strand focus on healthcare and sustainability, which was highlighted in
red color. Dendogram indicate that there is a scope for research in sustainability in
healthcare sector.

4 Discussion

Earlier research on green healthcare has examined a number of relevant practical chal-
lenges. Several of these studies have examined leadership [15], hospital management
of healthcare waste challenges, and solutions [16]. The main element influencing the
implementation of environmental sustainability in healthcare institutions is energy con-
sumption [17, 18]. The management of green healthcare supply chains has been the
subject of several studies, including ones that evaluated the environmental performance
of healthcare suppliers [19] and looked at the challenges associated with adopting green
public procurement in the public health sector. In a pharmaceutical supply chain, a novel
multi-objective mathematical model is being proposed [20] for responding to natural dis-
asters while taking quality and cost into account. In the present study, the most productive
authors, institutions, and nations, the annual citation structure, the most cited articles,
the most popular keywords, and the most cited journals are only a few of the bibliometric
indicators employed. The present study holds a significant position among publications
on hospitals and sustainability, making it a reference among the best publications in the
hospital and sustainable sector.
Yet, designing, constructing, and operating hospitals are among the most challeng-
ing tasks [21]. Hospital energy management takes into account the unique needs and
requirements of each department. Intense surgery, for instance, uses equipment with
higher energy needs and stricter constraints on environmental elements like ventilation,
heating, and lighting. Administrative sectors, however, use technology with lower energy
needs [22]. The unique characteristics of these regions add another factor that needs to
be considered. They make it difficult to design electrical installations and manage energy
efficiently in hospitals [23]. This research demonstrates a substantial evolution of the
journal concerning articles, analyzed themes, and research methodologies. This study
 Systematic Review on Sustainable Healthcare Practices 83

will aid to direct academics in the healthcare area toward new subjects and statistical
research methodologies. Also, it accurately describes the study on sustainability and
healthcare management done by this esteemed magazine. The examination of trending
subjects may be able to predict the direction of future research in the hospital and health-
care sectors. On the basis of this research, future studies could investigate the connection
between research published in recognized publications and its usefulness in healthcare
facilities.

5 Conclusions and Future Scope

Energy efficiency is a problem that become more difficult to solve as a result of the
growth in the world’s population, economic activity, and high consumption rates that
are not long-term sustainable. As a result, three crucial issues of energy efficiency in
healthcare should be the focus of healthcare authorities in various regions. (i) Making
use of cutting-edge innovative materials and technologies to improve building efficiency
and minimize energy consumption in healthcare planning. (i) Integrated approaches that
bring together all elements into coherent energy-efficiency programs; (ii) designs to
optimize energy usage in buildings; and (iii) boosting output through more sustainable
sources. An issue with the study could be that citation and co-citation data and patterns
are frequently dynamic and are expected to fluctuate over time. The research field of
sustainable healthcare is garnering a lot of attention from academics and researchers.
This paper summarizes the significant developments in sustainable healthcare research
from 2003 to 2023. Several directions for additional research have been identified after
thoroughly analyzing the sample set of 887 research papers.
Future studies might look further into which specific facets of sustainable health-
care—such as the medical business model, customer value propositions, consumer edu-
cation on sustainable healthcare, and delivery methods—produce measurable gains in
increased customer value. Only a few studies have examined the impact of sustainable
healthcare on consumer behavior, but future research may look at the direct correlation
between sustainable healthcare and certain outcomes including customer satisfaction,
customer loyalty, and future purchase intentions. Exploring these connections among
a diverse group of consumers across various segmentations would be another topic of
inquiry (Male, Female, and Adult). This might aid healthcare service providers in decid-
ing which clientele to target, what kind of investments to make, and how best to streamline
their service delivery. Western nations have primarily dominated research in the area of
sustainable healthcare. Yet, other emerging countries are also increasingly catching on to
the sustainable healthcare trend. Studying these effects on consumer behavior in various
cultural contexts is necessary.

References
1. Campisi, T., Acampa, G., Marino, G., Tesoriere, G.: Cycling master plans in Italy: the I-BIM
feasibility tool for cost and safety assessments. Sustainability 12(11), 4723 (2020)
2. McGain, F., Naylor, C.: Environmental sustainability in hospitals–a systematic review and
research agenda. J. Health Serv. Res. Policy 19(4), 245–252 (2014)
84 S. V. Urkude and D. Sahoo

3. Kumar, S., Lim, W.M., Pandey, N., Westland, J.C.: 20 years of electronic commerce research.
Electron. Commer. Res. 21(1), 1–40 (2021). https://doi.org/10.1007/s10660-021-09464-1
4. Kumar, S., Sureka, R., Lim, W.M., Kumar Mangla, S., Goyal, N.: What do we know
about business strategy and environmental research? Insights from business strategy and
the environment. Bus. Strateg. Environ. 30(8), 3454–3469 (2021)
5. Durisin, B., Puzone, F.: Maturation of corporate governance research, 1993 – 2007: an assess-
ment. In: KITeS, Centre for Knowledge, Internationalization and Technology Studies, vol.
17, No. 3, pp. 266–291. Universita’ Bocconi, Milano, Italy (2009)
6. Linnenluecke, M.K., Chen, X., Ling, X., Smith, T., Zhu, Y.: Research in finance: a review of
influential publications and a research agenda. Pac. Basin Financ. J. 43, 188–199 (2017)
7. Xu, L.D., Xu, E.L., Li, L.: Industry 4.0: state of the art and future trends. Int. J. Prod. Res.
56(8), 2941–2962 (2018)
8. Ellegaard, O., Wallin, J.A.: The bibliometric analysis of scholarly production: how great is
the impact? Scientometrics 105, 1809–1831 (2015)
9. Zupic, I., Čater, T.: Bibliometric methods in management and organization. Organ. Res.
Methods 18(3), 429–472 (2015)
10. Backhaus, K., Lügger, K., Koch, M.: The structure and evolution of business-to-business
marketing: a citation and co-citation analysis. Ind. Mark. Manage. 40(6), 940–951 (2011)
11. Donthu, N., Kumar, S., Pandey, N., Soni, G.: A retrospective overview of Asia Pacific Journal
of Marketing and Logistics using a bibliometric analysis. Asia Pac. J. Mark. Logist. 33(3),
783–806 (2021)
12. Donthu, N., Kumar, S., Pattnaik, D., Lim, W.M.: A bibliometric retrospection of marketing
from the lens of psychology: Insights from Psychology & Marketing. Psychol. Mark. 38(5),
834–865 (2021)
13. Brown, T., Park, A., Pitt, L.: A 60-year bibliographic review of the journal of advertising
research: perspectives on trends in authorship, influences, and research impact. J. Advert.
Res. 60(4), 353–360 (2020)
14. Bradford, S.C.: Sources of information on scientific subjects. Engineering: an illustrated.
Weekly J. 137, 85–86 (1934)
15. Chiarini, A., Vagnoni, E.: Environmental sustainability in European public healthcare: could
it just be a matter of leadership? Leadersh. Health Serv. 29(1), 2–8 (2016). https://doi.org/10.
1108/LHS-10-2015-0035
16. Thakur, V., Anbanandam, R.: Healthcare waste management: an interpretive structural
modeling approach. Int. J. Health Care Qual. Assur. 29(5), 559–581 (2016)
17. Neo, J.R.J., Sagha-Zadeh, R., Vielemeyer, O., Franklin, E.: Evidence-based practices to
increase hand hygiene compliance in health care facilities: an integrated review. Am. J.
Infection Control 44(6), 691–704 (2016)
18. Zadeh, R.S., Xuan, X., Shepley, M.M.: Sustainable healthcare design: existing challenges
and future directions for an environmental, economic, and social approach to sustainability.
Facilities 34(5/6), 264–288 (2016)
19. Malik, A., Budhwar, P., Kandade, K.: Nursing excellence: a knowledge-based view of
developing a healthcare workforce. J. Bus. Res. 144, 472–483 (2022)
20. Balan, S., Conlon, S.: Text analysis of green supply chain practices in healthcare. J. Comput.
Inform. Syst. 58(1), 30–38 (2018)
21. Blass, A.P., da Costa, S.E.G., de Lima, E.P., Borges, L.A.: Measuring environmental
performance in hospitals: a practical approach. J. Clean. Prod. 142, 279–289 (2017)
22. Ryan-Fogarty, Y., O’Regan, B., Moles, R.: Greening healthcare: systematic implementation
of environmental programmes in a university teaching hospital. J. Clean. Prod. 126, 248–259
(2016)
 Systematic Review on Sustainable Healthcare Practices 85

23. Bramer, W.M., Rethlefsen, M.L., Kleijnen, J., Franco, O.H.: Optimal database combinations
for literature searches in systematic reviews: a prospective exploratory study. Syst. Rev. 6,
1–12 (2017)
24. https://www.bibliometrix.org/home/index.php/layout/biblioshiny
25. https://www.bibliometrix.org/
26. https://www.scopus.com/results/results.uri?sort=plf-f&src=s&st1=Sustainable+Practices&
st2=Healthcare+services
27. https://mjl.clarivate.com/
 Violence Detection Through Deep Learning
Model in Surveillance

Anirudh Singh , Satyam Kumar , Abhishek Kumar , and Jayesh Gangrade(B)

Manipal University Jaipur, Jaipur, Rajasthan 303007, India

[email protected]

Abstract. The realm of advanced deep learning architectures for violence detec-
tion is burgeoning and has the potential to revolutionize proposed approaches to
preventing and identifying instances of violence. In this rapidly evolving field,
deep learning architectures exhibit the capability to discern intricate patterns and
make predictions based on vast datasets. Their suitability for violence detection
is particularly pronounced, given their capacity to identify subtle indicators that
may foreshadow aggressive behavior. This research advocates for an autonomous
violence detection system leveraging a combined CNN + LSTM architecture.
Each model adeptly extracts features at the frame level, progressively enhancing
accuracy in the process. The proposed CNN + LSTM model, integrating spa-
tial and temporal analyses, achieves an impressive accuracy rate of 90%. The
synergy of Convolutional Neural Networks (CNNs) for spatial feature extraction
and Long Short-Term Memory (LSTM) networks for temporal aggregation facili-
tates a nuanced analysis of local motion patterns. Rigorous evaluations conducted
on publicly available datasets attest to the models’ efficacy in violence detection.
The amalgamated solution, utilizing the CNN + LSTM architecture, is tailored for
real-time surveillance, ensuring swift identification and notification of potential
violence to relevant authorities for prompt remote intervention.

Keywords: Violence Detection · Deep Learning · CNN + LSTM · Spatial

Analysis · Real-time Surveillance

1 Introduction
In the ever-evolving landscape of monitoring technology, the imperative of deploying
intelligent systems capable of deciphering complex human behaviors during challenging
situations becomes paramount. This study addresses this need by delving into the realm
of cutting-edge machine learning (ML) and artificial intelligence (AI) models to discern
criminal activities. The present state of surveillance systems, while significant, often
falls short in comprehensively capturing the intricacies of real-life violent scenarios.
As part of our methodology, we propose the utilization of a singular, integrated model
that combines Convolutional Neural Network (CNN) and Long Short-Term Memory
(LSTM) architectures [1, 2]. This innovative approach capitalizes on CNN’s proficiency
in spatial feature extraction and the LSTM’s ability to comprehend temporal dynamics
within this amalgamation.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 86–98, 2025.
https://doi.org/10.1007/978-3-031-71481-8_7
 Violence Detection Through Deep Learning Model in Surveillance 87

Proposed methodology represents a departure from traditional studies that adopt a

static perspective. Instead, the CNN + LSTM model embraces a dynamic approach,
allowing for a more sophisticated examination of how motion patterns evolve over time
in video clips. The model achieves a notable accuracy rate of 90%, showcasing its profi-
ciency in violence detection [3–5]. Its simplicity, coupled with its efficacy in capturing
non-linear relationships within data, serves as a benchmark for comparative evaluations,
shedding light on the performance trade-offs between conventional and state-of-the-art
methodologies. As we embark on this exploration of violence detection using a unified
LSTM + CNN model, the research not only seeks to enhance the accuracy of surveil-
lance systems but also aims to contribute nuanced insights into the intricate interplay of
spatial and temporal dynamics in real-world scenarios [6–20]. This endeavor stands at
the intersection of technology and public safety, pushing the boundaries of achievable
outcomes in the creation of intelligent surveillance systems that safeguard society with
precision and foresight.
This research journey is underpinned by the recognition that the landscape of vio-
lence detection demands continuous adaptation to the evolving nature of technology. By
delving into the intricacies of AI and ML models, our pursuit is not merely confined
to the development of a cutting-edge surveillance system. It extends to a broader quest
for a more nuanced understanding of violent behaviors and the creation of intelligent
systems that transcend the limitations of current technologies.
The combination of CNN and LSTM architectures is not only a pragmatic response
to the limitations of existing systems but also a strategic approach to address the mul-
tifaceted nature of violence. CNN’s proficiency in extracting spatial features dovetails
seamlessly with the LSTM’s capacity to decipher temporal dynamics, resulting in a
holistic analysis of video data. This synthesis enables our model to discern subtle pat-
terns that may elude traditional methods, thereby elevating the potential for accurate
violence detection in real-world scenarios. As our exploration unfolds, it is guided by
a commitment to advancing not only the technical aspects of surveillance but also our
collective understanding of the complexities surrounding violence detection. By delv-
ing into the intricate interplay of spatial and temporal dynamics, we aspire to contribute
valuable insights that extend beyond the scope of conventional methodologies.
This study aligns with a broader narrative at the confluence of technological inno-
vation and public safety, aspiring to redefine the possibilities of intelligent surveillance
systems. Through a fusion of cutting-edge techniques and a commitment to real-world
applicability, our endeavor seeks to transcend conventional paradigms.

2 Related Work
Ullah et al. [6] conducted an extensive survey on violence detection in surveillance
videos, providing a comprehensive overview of methodologies and advancements in
the field. Their work serves as a foundational resource for understanding the diverse
approaches employed in the pursuit of effective violence detection. Omarov et al. [7]
contributed significantly to the field through a systematic review of state-of-the-art vio-
lence detection techniques in video surveillance security systems. Their comprehensive
analysis aids in discerning existing methodologies and their respective effectiveness, pro-
viding valuable insights for researchers and practitioners alike. Choque Luque-Roman
88 A. Singh et al.

and Camara-Chavez [8] explored the domain of weakly supervised violence detection
in surveillance video, introducing innovative methods to enhance detection capabilities
under limited supervision. Their work contributes to the ongoing efforts to improve
violence detection systems’ adaptability to various surveillance scenarios. Irfanullah
et al. [9] focused on real-time violence detection using Convolutional Neural Networks
(CNNs), marking a significant contribution to the practical application of deep learning
in surveillance scenarios. Their emphasis on real-time applications addresses a crucial
aspect of violence detection systems. Mahmoodi et al. [10] delved into violence detec-
tion in videos using interest frame extraction and 3D Convolutional Neural Networks.
Their research underscores the importance of capturing temporal features for accurate
detection, providing a nuanced perspective on the temporal dynamics of violent events.
Freire-Obregón et al. [11] proposed an inflated 3D ConvNet context analysis for violence
detection, advancing the field by incorporating context-aware approaches into surveil-
lance technologies. Their work highlights the significance of contextual information in
improving the overall performance of violence detection systems. Sharma et al. [12]
presented a fully integrated violence detection system using CNN and LSTM, show-
casing the combined strengths of these architectures for enhanced detection accuracy.
Their integrated approach provides a holistic understanding of the synergies between
CNN and LSTM in violence detection. Hung et al. [13] explored violent video detection
through a pre-trained model and CNN-LSTM approach, contributing to the develop-
ment of hybrid architectures in violence detection systems. Their work exemplifies the
ongoing exploration of diverse model combinations for improved accuracy. Dandage
et al. [14] provided a comprehensive review of violence detection systems using deep
learning, summarizing the advancements and challenges in the field up to 2019. Their
review serves as a valuable resource for understanding the historical progression of
violence detection research. Ramzan et al. [15] conducted a review on state-of-the-art
violence detection techniques, offering a thorough analysis of the landscape and high-
lighting key trends in violence detection research. Their work contributes to a nuanced
understanding of the evolving methodologies employed in violence detection. Patel [16]
proposed a real-time violence detection system using CNN-LSTM, contributing to the
exploration of integrated deep learning architectures for real-time surveillance appli-
cations. The real-time aspect of their work addresses the critical need for timely and
accurate violence detection. Sumon et al. [17] investigated violence detection using pre-
trained modules and different deep learning approaches, expanding the understanding
of model-based techniques in surveillance. Their research adds to the diverse array of
methodologies available for violence detection. Mumtaz et al. [18] introduced a fast-
learning approach through a deep multi-net CNN model for violence recognition in video
surveillance, emphasizing efficient model training. Their work contributes to the ongo-
ing efforts to enhance the efficiency of violence detection models. Vosta and Yow [19]
proposed a CNN-RNN combined structure for real-world violence detection in surveil-
lance cameras, contributing to the exploration of combined architectures for improved
accuracy. Their research addresses the complexities of real-world surveillance scenar-
ios. Vieira et al. [20] presented a low-cost CNN for automatic violence recognition on
embedded systems, emphasizing practical implementation and efficiency in surveillance
 Violence Detection Through Deep Learning Model in Surveillance 89

applications. Their work aligns with the growing interest in deploying violence detection
systems in resource-constrained environments.
Akash et al. [21] explored human violence detection using various deep learning
techniques, contributing to the understanding of human-centric violence recognition.
Their research sheds light on the unique challenges and opportunities associated with
recognizing violence involving human subjects. Tripathi et al. [22] investigated criminals
and crime detection using Machine Learning & OpenCV, extending the scope of video
surveillance beyond violence recognition. Wang et al. [23] provided baseline results
for violence detection in still images, contributing to the foundational understanding of
violence detection in static contexts. Their research lays the groundwork for exploring
violence detection in scenarios where motion dynamics are limited. Li et al. [24] pro-
posed an end-to-end multiplayer violence detection based on deep 3D CNN, contribut-
ing to the exploration of 3D convolutional architectures for violence detection in videos.
Calero et al. [25] presented “Bindi,” an affective Internet of Things system combat-
ing gender-based violence, offering a unique perspective on violence detection beyond
traditional surveillance methods. Their research introduces innovative approaches to
addressing violence in diverse contexts. Selvi et al. [25] developed a suspicious actions
detection system using enhanced CNN and surveillance video, expanding the scope
of video analytics for improved security. Their work explores the broader applications
of video analytics beyond explicit violence detection. Manikandan and Rahamathun-
nisa [26] proposed a neural network-aided attuned scheme for gun detection in video
surveillance images, addressing specific challenges related to firearm detection. Their
research addresses a niche aspect of violence detection, focusing on the identification
of potentially dangerous objects. Sarkar et al. [27] designed a Weapon Detection Sys-
tem, contributing to the development of specialized systems for identifying potentially
dangerous objects in surveillance scenarios. Their work adds to the growing repertoire
of specialized applications within the broader field of violence detection. Soliman et al.
[28] explored violence recognition from videos using deep learning techniques, adding
to the diverse range of methodologies for automated violence detection. Their research
expands the toolkit of approaches available for violence detection in dynamic video sce-
narios. Taha and Shoufan [29] presented a comprehensive review of machine learning-
based drone detection and classification, offering insights into the evolving landscape
of research in this domain. While not directly related to violence between individuals,
their work contributes to the broader understanding of machine learning applications in
surveillance. This synthesis of related work reflects the depth and breadth of research
in violence detection, showcasing diverse methodologies, approaches, and applications
that collectively contribute to the advancement of this critical field.

3 Proposed Methodology
The workflow diagram of the proposed working model is shown in Fig. 1.
Our research methodology begins with the utilization of a violence dataset compris-
ing 11,063 images extracted from video frames, categorized into Violence (5,832 images)
and Non-Violence (5,231 images). The dataset undergoes data augmentation to enhance
diversity, involving techniques such as rotation, flipping, and zooming. Subsequently,
90 A. Singh et al.

Violence
Data Augmentation
Dataset
and Vectorization

CNN+LSTM
Model

Evaluation
Matrices

Fig. 1. Proposed Architecture

the images are vectorized for optimal compatibility with machine learning models. The
model architecture integrates a Convolutional Neural Network (CNN) to extract spatial
features and a Long Short-Term Memory (LSTM) network to capture temporal depen-
dencies in sequential data, offering a comprehensive understanding of violent events.
The dataset is partitioned into training, validation, and test sets for effective model train-
ing, hyperparameter tuning, and final evaluation. The iterative process of fine-tuning and
optimization involves adjusting model parameters and exploring alternative data aug-
mentation strategies. Thorough documentation covers dataset specifics, preprocessing
steps, model architecture, hyperparameters, and a detailed report on results, emphasizing
both the strengths and limitations of our proposed approach.

3.1 Dataset Description

Proposed violence detection research draws upon a comprehensive dataset obtained from
Kaggle, originating from YouTube videos and curated to include 1000 violence and 1000
non-violence videos. Notably, the violence videos in our dataset encompass real street
fight situations, offering a nuanced representation across diverse environments and con-
ditions as shown in Figs. 2 and 3. To further enrich the dataset, we extracted 11,063
images from the video frames, meticulously dividing them into two distinct classes:
Violence and Non-Violence. The Violence class comprises 5,832 images capturing var-
ious real-life street fight scenarios, providing a rich and authentic set of instances for
model training and evaluation. The Non-Violence class encompasses 5,231 images, con-
tributing to the dataset’s balance and diversity. This dataset serves as a robust foundation
Violence Detection Through Deep Learning Model in Surveillance 91

Fig. 2. Sample image from dataset

Fig. 3. Sample image from dataset

92 A. Singh et al.

for our violence detection research, and its detailed nature ensures the authenticity and
relevance of the scenarios explored. The reference for the dataset is [30].

3.2 Data Vectorization

For video-based violence detection, the initial step involves extracting frames from
videos. This process transforms the temporal data into a series of static images, which
can be fed into various deep learning models. To enhance model generalization and
mitigate overfitting, data augmentation techniques like rotation, flip, scaling, and bright-
ness adjustment are commonly applied to increase the diversity of training samples. For
image-based models (CNN + LSTM), each frame is treated as an individual image.
Frames are resized to a common resolution (224 × 224) to ensure uniform input size.
Pixel values are normalized to a predefined range ([0, 1] or [−1, 1]). Deep pre-trained
models like CNN and LSTM are used to extract high-level features from frames. These
features serve as inputs to the final classification layers. For video sequences, data
vectorization involves encoding temporal information. Video frames are grouped into
sequences of fixed length. Padding may be applied to ensure uniform sequence length.
Frames within a sequence are processed through convolutional layers to extract spa-
tial features. The resulting features are then fed into LSTM layer to capture temporal
dependencies.

3.3 Feature Selection

CNN layers automatically learn to extract relevant features from the input frames. Filters
in these layers identify patterns, textures, and local spatial information. The weights of
these filters adapt during training to emphasize important features for violence detec-
tion. As the model trains, the filters become specialized in capturing violence-related
patterns. In temporal models like LSTM and RNN, the network inherently performs fea-
ture selection. These layers capture temporal dependencies and focus on the sequences
of features that are most informative for violence detection. By adjusting the weights
of the recurrent units during training, the model emphasizes relevant information over
time, effectively selecting features that contribute to the classification task. For deep
convolutional neural networks models, these models are designed with multiple layers
of convolutions that learn progressively more abstract and relevant features. The early
layers capture low-level details, while deeper layers extract high-level features. During
training, the network learns to assign higher weights to features that are more discrim-
inative for the violence detection task. Features that do not contribute significantly are
effectively downweighed in some cases, researchers may fine-tune these pre-trained
models on violence detection datasets. During fine-tuning, the models adapt to the spe-
cific task by adjusting the weights of the final classification layers. This process ensures
that the learned features align more closely with violence-related patterns.

3.4 Deep Learning Model

The Deep Learning Convolutional Neural Network (CNN) + Long Short-Term Memory
(LSTM) model represents a sophisticated architecture designed for violence detection in
 Violence Detection Through Deep Learning Model in Surveillance 93

video data. This model amalgamates the strengths of two distinct neural network types,
namely CNN and LSTM, to capture and analyze both spatial and temporal aspects of
visual information.

3.4.1 Convolutional Neural Networks (CNN)

CNNs excel at extracting spatial features from images. In the context of violence detec-
tion, these spatial features may include patterns, shapes, and textures within individual
frames of a video.

3.4.2 Long Short-Term Memory Networks (LSTM)

LSTMs are adept at understanding temporal dependencies within sequential data. In the
case of video analysis, frames unfold over time. They capture the sequential nature
of video frames, allowing the model to comprehend the evolution of patterns and
movements over time.

3.4.3 Combined Architecture

By combining CNN for spatial feature extraction and LSTM for temporal feature aggre-
gation, the model achieves a comprehensive analysis of both spatial and temporal pat-
terns. This synergy is particularly advantageous for violence detection, as it enables the
system to not only recognize spatial cues within individual frames but also understand
how these cues evolve over the course of the video.
The model summary provides a concise overview of the architecture’s key compo-
nents and their roles in violence detection. Table 1 likely includes details such as the
layers in the CNN and LSTM, the number of parameters, and possibly other relevant
metrics. This table serves as a reference point for understanding the structural compo-
sition of the CNN + LSTM model and its capabilities in discerning violence-related
patterns in video data.In essence, the CNN + LSTM model leverages the complemen-
tary strengths of spatial and temporal analysis to create a robust framework for violence
detection, making it well-suited for capturing the complexities inherent in real-world
scenarios.

3.5 Evaluation Parameters

• Accuracy:
Accuracy assesses the correctness of predictions by comparing the sum of true
positives (TP) and true negatives (TN) to the total instances.

Accuracy = (TP + TN)/(TP + TN + FP + FN) (1)

• Precision:
Precision gauges the precision of positive predictions, determined by the ratio of
true positives (TP) to the sum of true positives (TP) and false positives (FP).

Precision = TP/(TP + FP) (2)

94 A. Singh et al.

Table 1. Model summary of CNN +LSTM Model

Layer (type) Output Shape Param #

time_distributed (TimeDistributed) (None, 16, 2, 2, 1280) 2257984
dropout (Dropout) (None, 16, 2, 2, 1280) 0
time_distributed_1 (TimeDistributed) (None, 16, 5120) 0
bidirectional (Bidirectional) (None, 64) 1319168
dropout_1 (Dropout) (None, 64) 0
dense (Dense) (None, 256) 16640
dropout_2 (Dropout) (None, 256) 0
dense_1 (Dense) (None, 128) 32896
dropout_3 (Dropout) (None, 128) 0
dense_2 (Dense) (None, 64) 8256
dropout_4 (Dropout) (None, 64) 0
dense_3 (Dense) (None, 32) 2080
dropout_5 (Dropout) (None, 32) 0
dense_4 (Dense) (None, 2) 66

• Recall:
Recall measures the ability to capture positive instances, calculated by dividing
true positives (TP) by the sum of true positives (TP) and false negatives (FN).

Recall = TP/(TP + FN) (3)

• F1-Score:
The F1-score offers a balanced assessment, combining precision and recall using
the harmonic mean formula:

F1-Score = 2 ∗ (Precision ∗ Recall)/(Precision + Recall) (4)

4 Results Analysis
The outcomes derived from our extensive investigation into violence detection, utilizing
the CNN + LSTM model, illuminate compelling insights into the model efficacy in
discerning and identifying violent actions within intricate video sequences. This research
not only sheds light on the model’s discernment capabilities but also endeavors to unravel
the multifaceted dimensions of violence detection, delving into the intricacies that define
the model’s effectiveness in real-world scenarios.
The evaluation of the CNN + LSTM model is meticulously examined through an
array of performance metrics, transcending the conventional boundaries of assessment.
This comprehensive evaluation collectively provides a nuanced and holistic perspective
on the model’s multifaceted capabilities in the realm of violence detection.
 Violence Detection Through Deep Learning Model in Surveillance 95

The accuracy metric serves as a fundamental benchmark, gauging the model’s overall
correctness in predicting violent actions. Precision delves into the model’s precision in
correctly identifying violent instances, while recall assesses its ability to capture the
entirety of violent actions within the video sequences. The F1-score, a harmonized
metric of precision and recall, offers a balanced measure of the model’s effectiveness in
violence detection (Table 2).

Table 2. Performance Evaluation using various deep learning models.

Classifier Accuracy (%) Precision (%) Recall (%) F1-Score (%)

CNN + LSTM 90 84 96 90

The CNN + LSTM model stands out with an impressive accuracy of 90%, show-
casing a well-balanced performance with 84% precision, 96% recall, and an F1-score of
90.00%. This combination of spatial and temporal analysis proves effective in capturing
local motion patterns indicative of violent actions. Confusion matrix and ROC Curve
are shown in Figs. 4 and 5.

Fig. 4. Confusion Matrix of CNN + LSTM Model

96 A. Singh et al.

Fig. 5. ROC curve of CNN + LSTM Model

5 Conclusion and Future Work

In conclusion, our research on violence detection using the CNN + LSTM model has
yielded valuable insights into the effectiveness of deep learning architectures in dis-
cerning violent actions within video sequences. By leveraging a model like CNN +
LSTM, we conducted a thorough exploration of spatial and temporal features, signif-
icantly advancing the landscape of surveillance technology. The nuanced strengths of
each model were evident in our findings, with the CNN + LSTM model demonstrating
exceptional accuracy, emphasizing the pivotal role of combining spatial and temporal
analyses for robust violence detection. The inclusion of diverse content, spanning real
street fight scenarios and various non-violent activities, enhanced the robustness and
applicability of our research outcomes. The comprehensive evaluation results provided
a thorough assessment of model performance. This analysis not only aids in model
selection for specific applications but also paves the way for future research endeavors
to refine and optimize violence detection systems. In essence, our research contributes
significantly to the evolving landscape of violence detection, offering valuable insights
into the capabilities of deep learning models and establishing a foundation for continued
advancements in surveillance technology. By addressing the challenges posed by real-
world violence scenarios, our aim is to enhance the practical applicability of our models,
ultimately contributing to the development of safer and more secure public spaces.
Building upon the foundation laid by our current research on violence detection
using CNN + LSTM, there are several promising avenues for future exploration and
enhancement: Investigate advanced temporal fusion techniques to further enhance the
integration of temporal features extracted by the CNN + LSTM. Explore methods to
dynamically adjust the weightage of temporal information based on the evolving nature
 Violence Detection Through Deep Learning Model in Surveillance 97

of violent actions in video sequences. Assess the synergies between visual and non-
visual modalities for enhanced violence detection capabilities. Explore techniques for
generating visual explanations for model predictions, fostering trust, and understanding
in real-world deployment scenarios. Investigate the development of an adaptive model
architecture that dynamically adjusts its complexity based on the complexity of the
observed scene, optimizing resource utilization. Explore techniques to automatically
select the most relevant model architecture for a given surveillance environment. By
pursuing these future research directions, we aim to not only enhance the performance
of our violence detection system but also contribute to the broader landscape of machine
learning applications in surveillance and security.

Availability of Data and Materials. The data set used in the experiments can be found on:
https://www.kaggle.com/datasets/karandeep98/real-life-violence-and-nonviolence-data.

References
1. Marszałek, M., Laptev, I., Schmid, C.: Actions in context, In: IEEE Computer Society
Conference on Computer Vision and Pattern Recognition Workshops (2009)
2. Gracia, I.S., Suarez, O.D., Garcia, G.B., Kim, T.-K.: Fast fight detection. PLoS ONE 10(4),
e0120448 (2015)
3. Dhiman, C., Vishwakarma, D.K.: A review of state-of-the-art techniques for abnormal human
activity recognition. Eng. Appl. Artif. Intell. 77, 21–45 (2019)
4. Ben Mabrouk, A., Zagrouba, E.: Abnormal behavior recognition for intelligent video surveil-
lance systems: a review. Expert Syst. Appl. 91, 480–491 (2018). https://doi.org/10.1016/j.
eswa.2017.09.029
5. Li, Q., Li, W.: A novel framework for anomaly detection in video surveillance using multi-
feature extraction. In: 9th International Symposium on Computational Intelligence and Design
(ISCID), vol. 1, pp. 455–459 (2016)
6. Ullah, F.U.M., et al.: A comprehensive review on vision-based violence detection in
surveillance videos. ACM Comput. Surv. 55(10), 1–44 (2023). https://doi.org/10.1145/356
1971
7. Omarov, B., et al.: State-of-the-art violence detection techniques in video surveillance security
systems: a systematic review. PeerJ Comput. Sci. 8, e920 (2022)
8. Choqueluque-Roman, D., Camara-Chavez, G.: Weakly supervised violence detection in
surveillance video. Sensors 22(12), 4502 (2022)
9. Irfanullah, et al.: Real time violence detection in surveillance videos using Convolutional
Neural Networks. Multimed. Tools Appl. 81(26), 38151–38173 (2022)
10. Mahmoodi, J., Nezamabadi-pour, H., Abbasi-Moghadam, D.: Violence detection in videos
using interest frame extraction and 3D convolutional neural network. Multimed. Tools Appl.
81(15), 20945–20961 (2022)
11. Freire-Obregón, D., et al.: Inflated 3D ConvNet context analysis for violence detection. Mach.
Vis. Appl. 33, I15 (2022)
12. Sharma, S., et al.: A fully integrated violence detection system using CNN and LSTM. Int. J.
Electr. Comput. Eng. 11(4), 3374 (2021)
13. Hung, B.T., et al.: Violent video detection by pre-trained model and CNN-LSTM approach. In:
Singh Mer, K.K., Semwal, V.B., Bijalwan, V., Crespo, R.G. (eds.) Proceedings of Integrated
Intelligence Enable Networks and Computing: IIENC 2020, pp. 979–989. Springer Singapore,
Singapore (2021)
98 A. Singh et al.

14. Dandage, V., et al.: Review of violence detection system using deep learning. Int. Res. J. Eng.
Technol. 6(12), 1899–1902 (2019)
15. Ramzan, M., et al.: A review on state-of-the-art violence detection techniques. IEEE Access
7, 107560–107575 (2019)
16. Patel, M.: Real-Time Violence Detection Using CNN-LSTM.“ arXiv preprint arXiv:2107.
07578 (2021)
17. Sumon, S.A., et al.: Violence detection by pretrained modules with different deep learning
approaches. Vietnam J. Comput. Sci. 07(01), 19–40 (2020). https://doi.org/10.1142/S21968
88820500013
18. Mumtaz, A., Sargano, A.B., Habib, Z.: Fast learning through deep multi-net CNN model for
violence recognition in video surveillance. Comput. J. 65(3), 457–472 (2022)
19. Vosta, S., Yow, K.-C.: A cnn-rnn combined structure for real-world violence detection in
surveillance cameras. Appl. Sci. 12(3), 1021 (2022)
20. Vieira, J.C., et al.: Low-cost CNN for automatic violence recognition on embedded system.
IEEE Access 10, 25190–25202 (2022)
21. Akash, S.A.A., et al.: Human violence detection using deep learning techniques. J. Phys.
Conf. Ser. 2318(1), 012003 (2022)
22. Tripathi, A., et al.: Criminals as well as crime detection using Machine Learning & OpenCV.
Int. Res. J. Modernization Eng. Technol. Sci. 3(4), 2135–2141 (2021)
23. Wang, D., et al.: Baseline results for violence detection in still images. In: 2012 IEEE Ninth
International Conference on Advanced Video and Signal-Based Surveillance. IEEE (2012)
24. Li, C., et al.: End-to-end multiplayer violence detection based on deep 3DCNN. In: Proceed-
ings of the 2018 VII International Conference on Network, Communication, and Computing
(2018)
25. Miranda Calero, J.A., et al.: Bindi: affective internet of things to combat gender-based vio-
lence. IEEE Internet of Things J. 9(21), 21174–21193 (2022). https://doi.org/10.1109/JIOT.
2022.3177256
26. Selvi, E., et al.: Suspicious actions detection system using enhanced CNN and surveillance
video. Electronics 11(24), 4210 (2022)
27. Manikandan, V.P., Rahamathunnisa, U.: A neural network aided attuned scheme for gun
detection in video surveillance images. Image Vis. Comput. 120, 104406 (2022)
28. Sarkar, S., et al.: Design of Weapon Detection System. In: 2022 3rd International Conference
on Electronics and Sustainable Communication Systems (ICESC). IEEE (2022)
29. Soliman, M.M., et al.: Violence recognition from videos using deep learning techniques.
In: 2019 Ninth International Conference on Intelligent Computing and Information Systems
(ICICIS). IEEE (2019)
30. Soliman, M., Kamal, M., Nashed, M., Mostafa, Y., Chawky, B., Khattab, D.: Violence recog-
nition from videos using deep learning techniques. In: Proceedings of the 9th International
Conference on Intelligent Computing and Information Systems (ICICIS’19), pp. 79–84. Cairo
(2019)
 Role of ChatGPT in Decision Making Across
Industries: An Indian Perspective

Debajani Sahoo1 and Shubhangi V. Urkude2(B)

1 Department of Marketing, IBS, IFHE University, Hyderabad, Telangana, India
[email protected]
2 Department of Operations and IT, IBS, IFHE University, Hyderabad, Telangana, India

[email protected]

Abstract. This study delves deeply into the functions of ChatGPT, an AI model
built on the GPT-3.5 architecture, across multiple industries. Through ChatGPT,
artificial intelligence (AI) and natural language processing (NLP) are combined
to create new opportunities for a variety of sectors. This summary highlights
the report’s analysis of ChatGPT’s applications in marketing, education, health-
care, investment decision-making, and agriculture sector. The survey methodol-
ogy involves online distribution of questionnaires to gather insights from respon-
dents in each sector. The data was collected using Google Forms, analyzed and
interpreted to provide nuanced perspectives on ChatGPT’s role in these sectors.
While ChatGPT demonstrates potential to enhance efficiency and productivity,
challenges such as data privacy, biases, and overreliance on technology must be
addressed. The report concludes that responsible integration of ChatGPT in vari-
ous sectors necessitates a balance between AI capabilities and human expertise. As
technology evolves, the transformative potential of ChatGPT continues to shape
these industries and contribute to responsible innovation.

Keywords: ChatGPT · decision making · AI-driven technologies · Industry

application

1 Introduction

In the modern era, industries worldwide are leveraging AI-driven technologies to opti-
mize operations and decision-making processes. ChatGPT’s natural language capabili-
ties are at the forefront of this transformation, enabling innovative communication, infor-
mation retrieval, and problem-solving. The report focused into how ChatGPT addresses
distinct challenges and opportunities in different sectors [1].
ChatGPT revolutionizes marketing strategies by enabling personalized engagement,
content creation, and market analysis [2]. ChatGPT services in education industry act as a
virtual assistant, enriching learning environments by providing valuable information and
resources to students that, improves student productivity. ChatGPT in healthcare industry
aids in patient interactions, diagnoses, and medical research. Survey data portrays a
positive sentiment toward ChatGPT’s potential to enhance communication in the field

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 99–112, 2025.
https://doi.org/10.1007/978-3-031-71481-8_8
100 D. Sahoo and S. V. Urkude

of spinal surgical practice. Similarly, ChatGPT’s rapid data processing abilities in finance
investment industry shows promise in aiding investment decisions. Respondents express
varying degrees of agreement on its potential to mitigate biases and emotional influences.
The integration of ChatGPT with IoT devices in agriculture is met with skepticism,
to optimize irrigation, nutrient application, and pest management remains an area for
exploration.
The role of ChatGPT is undoubtedly worth addressing due to its profound impact on
various sectors and its potential to reshape industries and human interactions. ChatGPT
role spans from improving customer interactions and personalizing marketing strategies
to enhancing education, healthcare, investment, and agriculture sectors needs to explore
further [3] from emerging market prospective.

1.1 Marketing Sector: Personalization and Customer Engagement

In the realm of marketing, ChatGPT has revolutionized the way brands interact with their
audience. The capacity of technology to produce replies that resemble those of a human
being and to have lively discussions has created new opportunities for tailored consumer
engagements. ChatGPT can craft compelling marketing content, draft engaging social
media posts, and even provide real-time customer support through chatbots [4]. This
enhances brand loyalty and customer satisfaction by delivering prompt and accurate
responses.
By analyzing consumer preferences and behaviors, ChatGPT aids in developing data-
driven marketing strategies [5]. Its proficiency in natural language processing enables it
to sift through vast volumes of consumer data to extract valuable insights. Marketers can
tell their campaigns to align with consumer sentiments, resulting in higher conversion
rates and improved ROI.

1.2 Education Sector: Interactive Learning and Knowledge Dissemination

The education sector has witnessed a paradigm shift with the integration of ChatGPT.
Traditional classrooms are no longer the sole venues for learning. ChatGPT facilitates
interactive and personalized learning experiences through virtual tutors and educational
chatbots [6]. Students can seek clarifications, receive explanations, and engage in discus-
sion, transcending geographical barriers. Furthermore, ChatGPT serves as a knowledge
repository, enabling seamless access to information [7]. ChatGPT assists educators in
preparing course materials, generating assessment questions, and providing real-time
feedback to students [8]. Language barriers are diminished as ChatGPT can facilitate
translation, making educational resources accessible to a global audience.

1.3 Healthcare Sector: Diagnostics, Consultations, and Medical Research

In healthcare, ChatGPT’s potential is harnessed for diagnostics and patient interactions.

Chatbots powered by the model can engage patients in preliminary symptom assess-
ments, audience healthcare providers in offering informed guidance [9]. While these
chatbots do not replace medical professionals, they expedite the triage process; ensuring
 Role of ChatGPT in Decision Making Across Industries 101

patients receive appropriate care more efficiently. ChatGPT’s impact extends to medical
research as well. It assists researchers in analyzing vast medical literature, identify-
ing potential drug interactions, and even predicting disease outbreaks [10]. Its rapid
information processing capabilities expedite the development of medical solutions and
contribute to evidence-based decision-making.

1.4 Investment Sector: Data Analysis and Financial Insights

Data drives investment decisions, and ChatGPT excels in data analysis. It may sort
through market patterns, economic statistics, and financial reports to give investors infor-
mation about possible dangers and possibilities. ChatGPT’s ability to generate narra-
tives from raw data simplifies complex financial information, aiding investors in making
informed choices. Furthermore, financial advisory services are enhanced through the
integration of ChatGPT [11]. It provides tailored advice according to each user’s finan-
cial objectives and risk tolerance. This democratization of financial advice empowers a
broader demographic to navigate the intricacies of investment.

1.5 Agriculture Sector: Precision Farming and Crop Management

In agriculture, ChatGPT contributes to the burgeoning field of precision farming. Ana-
lyzing data from sensors, drones, and satellites assists farmers in optimizing irrigation,
fertilization, and pest control. The model generates insights on crop health, disease
detection, and yield estimation, enabling proactive decision-making to enhance produc-
tivity [12]. Additionally, ChatGPT serves as a knowledge hub for farmers. It answers
questions about crop diseases, soil management, and sustainable practices. Farmers can
access information in real time, aiding them in adapting to changing conditions and
adopting innovative approaches.

1.6 ChatGPT Tools

A new multimodal GPT-4 version from ChatGPT enables users to upload and examine a
variety of document types. DALL-E 3, integrated browsing, and advanced data analysis
are all included in GPT-4 all Tools. It’s possible that some new features will render many
third-party ChatGPT plugins. In response to this difficulty, some AI systems have created
custom solutions that may be a better fit for specific requirements. Many popular and
up-to-date ChatGPT tools are available, including GitHub Copilot, OpenAI Playground,
Bard, Jasper Chat, ChatSonic, and many more. Various ChatGPT programs examine
financial data to offer insights on potential investments, including stock prices and market
patterns. ChatGPT can also be used to automate customer care for financial services by
responding to consumer inquiries and offering real-time investment recommendations.

2 Literature Review
The growth of ChatGPT across these diverse sectors is poised to continue it’s upward
trajectory. ChatGPT’s capabilities are expected to become even more refined and ver-
satile as AI technologies advance. Ethical considerations and data privacy will remain
102 D. Sahoo and S. V. Urkude

paramount, necessitating responsible AI deployment. Collaborations between AI devel-

opers and sector-specific experts will drive innovation and yield more tailored solutions
[13]. In order to ensure that the advantages of AI technologies are maximized and the
risks are minimized, governments and regulatory agencies will be essential in guid-
ing the appropriate implementation of this technology [14]. The part of ChatGPT is
worth addressing due to its transformative potential across sectors. India stands at a
juncture where it can leverage ChatGPT’s capabilities to drive innovation, improve ser-
vices, and contribute to the global AI landscape., that enhance its competitiveness and
socioeconomic growth globally [15].
The use of ChatGPT in marketing also entails certain risks and challenges. It is
crucial to consider these factors to ensure responsible and ethical usage such as; data
from inaccurate sources, reliance on outdated data, loss of brand identity due to similar
responses generated for different marketers, overreliance on AI tools, job displacement,
perpetuation of biases if not properly designed and tested, and the need for profession-
alism and expertise in using ChatGPT effectively [5]. Marketers can leverage ChatGPT
in market research to collect and condense consumer feedback and social media con-
versations, providing valuable insights into consumers’ vocabulary, perceptions, and
attitudes towards products and marketing campaigns that facilitating better-informed
decision-making, assists in understanding customer questions, suggesting solutions, and
responding quicker.
It can also assist marketers in generating ideas for marketing campaigns, refining
ad concepts, personalizing campaigns, and analyzing consumer behavior to improve
product development and marketing strategies [16, 17].

3 Research Objective

This study aims to do a thorough analysis of ChatGPT’s function in several domains,

including marketing, education, healthcare, investment decision-making, and agricul-
ture. The report aims to achieve the following research objectives:
• How ChatGPT is utilized in various sectors and the specific applications it brings to
each domain.
• Assess the perceived benefits and challenges associated with integrating ChatGPT
within each sector, highlighting both the potential advantages and limitations.
• Analyze the opinions and perceptions of stake holders; including professionals, con-
sumers, educators, and students, towards ChatGPT’s impact on the sectors under
study.
• Evaluate user acceptance, comfort, and willingness to adopt AI-driven technologies
like ChatGPT within each sector, identifying adoption factors.
• Examine the sustainability and long-term viability of ChatGPT’s integration within
sectors, considering potential shifts in industry practices and evolving technology.
 Role of ChatGPT in Decision Making Across Industries 103

4 Research Methodology

4.1 Data collection

To perform this study sample data was collected by circulating the google form on a
likert scale of 1 to 5 having 1 as strongly disagree, 2 as disagree, 3 as neutral, 4 as agree
and 5 as strongly agree. The questionnaire was distributed among relevant population.
Data consist of 85 samples with demographic details and some questions related to use
of ChatGPT in education sector, investment banking, agriculture, marketing strategies
and medical sector. Collected data was analyzed to get the understanding using excel.
The questions are included below:
• Chat GPT improves student’s productivity by providing valuable information and
resources
• Chat GPT should be used as a complementary tool to human interaction and students’
efforts in learning.
• Chat GPT can significantly enhance marketing strategies by generating content
quickly and with potentially higher quality than human content creators.
• Chat GPT can improve customer interactions and provide valuable insights into con-
sumer feedback and attitudes, thereby facilitating more efficient and effective market
research in the marketing industry.
• Chat GPT’s ability to process large volumes of information quickly can be
advantageous for investment professionals in making informed decisions.
• Use of Chat GPT in investment decision making can help reduce cognitive biases and
emotional influences on investment decisions, providing a more rational perspective
• How positively do you perceive the integration of ChatGPT with IoT devices
in agriculture for optimizing irrigation schedules, nutrient application, and pest
management?
• Chat GPT has the potential to drive innovation and enhance productivity in the
agricultural sector.
• Chat GPT/GPT-4 can enhance communication between medical staff, patients, and
their relatives in the field of spinal surgical
• Chat GPT/GPT-4 can be a valuable tool in accelerating the learning curve and
supporting spine surgeons in mastering

5 Result and Discussion

5.1 Education Sector

Figures 1 and 2 shows, the application of ChatGPT in the education sector. In Fig. 1,
53% respondents are saying that the ChatGPT was helping the students in improving
the productivity, whereas 21% respondents having no idea about the use of ChatGPT in
education sector. 14%, 6%, 6% a respondent was strongly agree, strongly disagree and
disagree on the use in education sector respectively. Overall, the survey results suggest
that ChatGPT is a valuable tool to help students learn and be more productive. However,
a small group of students believe that ChatGPT is not helpful.
104 D. Sahoo and S. V. Urkude

Figure 2, shows ChatGPT can be used as tool between machine and student’s learning
efforts. In this chart 46% of the students are agreed on the statement, whereas 28% are
neutral about the opinion, 11% are strongly agree, 9% disagree and 6% are strongly
disagree on the statement. This result shows that there is positive impact of chat gpt in
the education sector and we will recommend it use for the better future of students who
want to learn new thing, explore new ideas on their own.

Fig. 1. Improve Student’s productivity

Fig. 2. Student’s Efforts in Learning

5.2 Marketing Sector

Figures 3 and 4 shows the use of ChatGPT in enhancing marketing strategies by gen-
erating the quality contents and improve customer interaction by valuable insights into
 Role of ChatGPT in Decision Making Across Industries 105

feedback and attitude. 48% of the respondents agreed on that ChatGPT helps in enhanc-
ing marketing strategies through automatic content generation for marketing the product
using chatboat. But 21% respondents are neutral about their opinion,13% are disagree
and strongly agree respectively. Overall, the pie chart shows that there is a strong belief
in the potential of chatbots to enhance marketing strategies using AI based technol-
ogy. Figure 4 shows that 39% are agree, this suggests that there is a growing belief
in the potential of ChatGPT to improve customer interactions and provide valuable
insights into consumer feedback. It can produce prose of a human calibre, translate
between languages, create a variety of artistic content, and provide insightful answers to
queries.6% strongly disagreed with the statement may be concerned about the accuracy
and reliability of ChatGPT.
36% neutral whereas 15%, 6% and 4% respondents are strongly agree, disagree
and strongly disagree. Overall, the pie chart shows that there is a strong belief in the
potential of ChatGPT to improve customer interactions and provide valuable insights
into consumer feedback. As ChatGPT technology continues to develop, it is likely that
this belief will only grow stronger.

Fig. 3. Enhance marketing strategies

5.3 Investment Banking

Figure 5 shows the application of ChatGPT to process large volume of data at faster
rate and helps the investment professional to make the informed decision. The survey
revealed a diverse range of responses; with 48% of participants agreeing that ChatGPT’s
capacity to rapidly process information could indeed be advantageous for investment
professionals. 16% of participants expressed a strong agreement with this proposition
demonstrates a high level of confidence in the technology’s capability, Similarly, the
relatively low percentage (7%) of strong disagreement points towards a limited number
of detractors who believe that ChatGPT’s rapid information processing might not be
as beneficial as suggested. The presence of 17% neutral responses highlights a critical
aspect of this discussion. This is unsurprising, given that the integration of advanced AI
106 D. Sahoo and S. V. Urkude

Fig. 4. Improve customer Interactions

into traditional investment practices represents a paradigm shift, which can prompt con-
cerns about the technology’s reliability, ethical considerations, and potential unintended
consequences. Additionally, investment decisions often require not just data analysis but
also a nuanced understanding of market dynamics, investor sentiment, and macroeco-
nomic trends – areas where human judgment and expertise still hold considerable value.
In conclusion, the survey responses present a compelling narrative regarding the potential
advantages of ChatGPT’s quick information processing for investment professionals.

Fig. 5. Advantages for investment professional

Figure 6 shows use of ChatGPT in reducing the biased and emotional decision mak-
ing in investment banking. The analysis of the survey responses regarding the belief in
ChatGPT’s potential to mitigate biases and emotional decisions in investment making
provides valuable insights into the perceptions surrounding this topic. A notable 38% of
participants expressed agreement that the utilization of ChatGPT in investment decision-
making could aid in reducing congenital biases and emotional decision-making. Further-
more, 9% of participants strongly agree with the notion, which indicates a smaller yet
 Role of ChatGPT in Decision Making Across Industries 107

significant subset of respondents who hold a more emphatic stance on the matter. The
sizeable percentage of neutral responses, at 33%, reflects a significant segment of par-
ticipants who are undecided or unsure about the extent to which ChatGPT can address
biases and emotional factors in investment decisions. On the contrary, 12% of partici-
pants disagreed that ChatGPT could help reduce biases and emotional decisions, holds
reservations about the technology’s capacity to completely eliminate these innate human
tendencies. Ultimately, while ChatGPT and similar AI models may have the potential to
offer a more rational perspective by reducing biases and emotional decisions, it’s crucial
to acknowledge that AI is a tool that works in conjunction with human expertise. Suc-
cessful integration will likely involve striking a balance between technology and human
judgment, leveraging AI’s strengths while acknowledging its limitations.

Fig. 6. Investment decision making

5.4 Agriculture Sector

Figure 7 shows the results of a survey on how people perceive the integration of ChatGPT
with IoT devices in agriculture for optimizing irrigation schedules, nutrient application,
and pest management. The majority of respondents (42%) agree with the integration of
ChatGPT with IoT devices in agriculture. This suggests that many people are skeptical
about the potential benefits of this technology. The neutral (31%), disagree (7%), strongly
agree (14%) and strongly disagree (6%) responses are distributed respectively. There are
a few possible reasons for the skepticism towards the integration of ChatGPT with IoT
devices in agriculture. One reason is that this technology is still relatively new and
untested. There is not yet enough data to demonstrate its effectiveness in a real-world
setting. For example, ChatGPT could be used to collect and analyze data from IoT
sensors to optimize irrigation schedules and nutrient application, identify and manage
pests.
Overall, the results of this survey suggest that there is still some uncertainty about the
potential benefits of integrating ChatGPT with IoT devices in agriculture. However, there
108 D. Sahoo and S. V. Urkude

are also some potential benefits that could make this technology worth exploring further.
The survey’s findings represent merely a moment in time in the public’s perception.
Should the survey be administered once more in the future, the outcomes can vary. It is
important to note that the survey did not ask respondents to explain their reasons for their
responses. This makes it difficult to understand the underlying reasons for the skepticism
towards the integration of ChatGPT with IoT devices in agriculture.
Figure 8 shows, to what extends ChatGPT have the potential to drive innovation and
enhance productivity in the agricultural sector. The majority of respondents (39%) agree
the potential of chatbots in agriculture. This suggests that many people are skeptical
about the ability of chatbots to drive innovation and enhance productivity in this sector.
The neutral (35%), strongly agree (12%), disagree (9%) and strongly disagree (5%)
responses are distributed respectively. There are a few possible reasons for the skepticism
towards the potential of chatbots in agriculture. One reason is that this technology is still
relatively new and untested. There is not yet enough data to demonstrate its effectiveness
in a real-world setting. For example, chatbots could be used to: Provide farmers with
information and advice on crop management, help farmers to automate tasks, such
as ordering supplies or scheduling deliveries, monitor crops and livestock for signs
of disease or pests, connect farmers with other businesses and organizations in the
agricultural sector.

Fig. 7. Integration of ChatGPT & IoT

Overall, the pie chart provides a useful overview of public opinion on this issue.
However, it is important to consider the limitations of the survey results before drawing
any definitive conclusions. It was believed that these challenges can be overcome, and
that chatbots have the potential to revolutionize the agricultural sector. Chatbots can help
farmers to improve their yields, reduce their costs, and make better decisions about their
businesses. They can also help to connect farmers with other businesses and organizations
in the agricultural sector, and to provide farmers with access to information and advice.
I was believed that this technology has the potential to make a positive impact on the
lives of farmers and the world’s food supply.
 Role of ChatGPT in Decision Making Across Industries 109

Fig. 8. Productivity in Agriculture

5.5 Healthcare Sector

Figure 9 shows the results of how much they agree that ChatGPT/GPT-4 can enhance
communication between medical staff, patients, and their relatives in the field of spinal
surgical practice. The largest group, 39%, agreed that ChatGPT/GPT-4 can enhance com-
munication, 7% strongly disagreed, 35% neutral, 15% disagreed, 4% strongly agreed.
Strongly disagree (7%), respondents believes that ChatGPT/GPT-4 is not capable of
enhancing communication between medical staff, patients, and their relatives in the
field of spinal surgical practice. Agree (39%), respondents believe that ChatGPT/GPT-4
has the potential to significantly enhance communication between medical staff, patients,
and their relatives in the field of spinal surgical practice. They may have concerns about
the accuracy of the language model, its ability to understand complex medical termi-
nology, or its ability to build rapport with patients. Disagree (15%), respondents was
less strongly opposed to the use of ChatGPT/GPT-4, but they still do not believe that it
is the best solution for enhancing communication in spinal surgical practice. They may
have similar concerns to the strongly disagree group, or they may believe that there are
other, more effective ways to improve communication. Neutral (35%), respondents were
not sure whether or not ChatGPT/GPT-4 can enhance communication in spinal surgical
practice. They may need more information about the language model or its capabilities
before forming an opinion.
The strongly agreed (4%), respondents were believed that ChatGPT/GPT-4 has the
potential to surely enhance communication between medical staff, patients, and their
relatives in the field of spinal surgical practice. This suggests that there are significant
concerns about the language model’s capabilities and limitations.
110 D. Sahoo and S. V. Urkude

Figure 10 shows the results how much they believe that ChatGPT/GPT-4 can be
a valuable tool in accelerating the learning curve and supporting them in mastering
endoscopic spinal surgery. Most respondents (34%) agreed that ChatGPT/GPT-4 can
be a valuable tool. 31% were neutral, 19% disagreed, 7% strongly disagreed and 9%
strongly agreed. This suggests that there is no clear consensus among the respondents on
the value of ChatGPT/GPT-4 as a tool for learning endoscopic spinal surgery. However,
it is worth as the majority of respondents believe that it a valuable tool but most of them
were neutral and disagreed with the situation.
Here are some possible reasons for this:
• Surgeons may be skeptical of the ability of a chatbot to provide accurate and reliable
information about a complex medical procedure.
• If they were to utilize a chatbot, they could worry about the safety and privacy of
their data.
• They may prefer to learn endoscopic spinal surgery from a human instructor.
It is significant to remember that a limited sample of persons participated in this
survey. It is plausible that the outcomes could alter if a more extensive and representative
sample of individuals was polled. Overall, the results of this survey suggest that there
is no clear consensus among respondents on the value of ChatGPT/GPT-4 as a tool for
learning endoscopic spinal surgery. However, it is worth noting that most respondents
believe that it is a valuable tool. More research is needed to determine the true value of
ChatGPT/GPT-4 for this purpose.

Fig. 9. Enhance communication between patient and medical staff

 Role of ChatGPT in Decision Making Across Industries 111

Fig. 10. Accelerate learning in medical staff

6 Limitations

The survey sample size of 85 participants may only represent part of the population in
each sector, potentially leading to biased results. Participants who voluntarily took the
survey might have pre-existing opinions or experiences related to ChatGPT, leading to
skewed results. The survey was conducted online, which might exclude individuals who
are less comfortable with technology or have limited internet access. Each sector has
unique challenges and complexities that a single survey might only partially capture.

7 Conclusions and Future Scope

ChatGPT’s integration across various sectors signifies its role as a transformative force
in modern technology. Its application in marketing is revolutionizing customer engage-
ment and content creation. While in education, it’s reshaping learning environments
through virtual assistance and knowledge facilitation. In healthcare, ChatGPT is improv-
ing patient interactions, diagnoses, and medical research. ChatGPT can be used a tool
for decision making in financial sector. Additionally, in agriculture, it offers insights
into crop management, weather forecasts, and market trends, empowering farmers for
sustainable food production. The survey-based analysis provides valuable insights into
perceptions within these sectors, showing a mix of optimism and skepticism regarding
ChatGPT’s capabilities. While it’s potential to enhance communication, learning, and
decision-making is evident, privacy, reliability, and human expertise must be addressed
for successful integration. Ultimately, ChatGPT’s transformative impact depends on
balancing its strengths with human judgment, thus shaping a future where AI augments
various sectors for the better.
In future we can extend this work to collect more real time data and deeper analysis
can be performed by collecting sufficient sample to get a generalized results.
112 D. Sahoo and S. V. Urkude

References
1. George, A.S., George, A.H.: A review of ChatGPT AI’s impact on several business sectors.
Partners Univers. Int. Innov. J. 1(1), 9–23 (2023)
2. Limna, P., Kraiwanit, T., Jangjarat, K., Klayklung, P., Chocksathaporn, P.: The use of ChatGPT
in the digital era: Perspectives on chatbot implementation. J. Appl. Learn. Teachi. 6(1) (2023)
3. Wiredu, J., Kumi, M., Ademola, P.H., MuseshaIyela, P.: An investigation on the characteris-
tics, abilities, constraints, and functions of artificial intelligence (AI): the age of ChatGPT as
an essential ultramodern support tool. Int. J. Dev. Res. 13(05), 62614–62620 (2023)
4. Saputra, R., Nasution, M.I.P., Dharma, B.: The impact of using AI chat GPT on marketing
effectiveness: a case study on instagram marketing. Indonesian J. Econ. Manag. 3(3), 603–617
(2023)
5. Rivas, P., Zhao, L.: Marketing with ChatGPT: navigating the ethical terrain of GPT-based
chatbot technology. AI 4(2), 375–384 (2023). https://doi.org/10.3390/ai4020019
6. Aithal, S., Aithal, P.S.: Effects of AI-based ChatGPT on higher education libraries. Int. J.
Manag. Technol. Soc. Sci. 8(2), 95–108 (2023)
7. Urkude, S., Gupta, K.: Student intervention system using machine learning techniques. Int.
J. Eng. Adv. Technol. 8(6), 21–29 (2019)
8. Grassini, S.: Shaping the future of education: exploring the potential and consequences of AI
and ChatGPT in educational settings. Educ. Sci. 13(7), 692 (2023)
9. Cascella, M., Montomoli, J., Bellini, V., Bignami, E.: Evaluating the feasibility of ChatGPT
in healthcare: an analysis of multiple clinical and research scenarios. J. Med. Syst. 47(1), 33
(2023)
10. Biswas, S.S.: Role of chat GPT in public health. Ann. Biomed. Eng. 51(5), 868–869 (2023)
11. Mhlanga, D.: Digital transformation education, opportunities, and challenges of the applica-
tion of ChatGPT to emerging economies. Educ. Res. Int. 2023, 1–13 (2023). https://doi.org/
10.1155/2023/7605075
12. Biswas, S.: Importance of chat GPT in Agriculture: According to chat GPT, 30 Mar 2023.
SSRN: https://ssrn.com/abstract=4405391 or https://doi.org/10.2139/ssrn.4405391
13. Urkude, S.V., Hasanuzzaman, V.R.U., Kumar, C.S.: Comparative analysis on machine
learning techniques: a case study on Amazon product. Int. J. Mech. Eng. 6(3), 739–744
(2021)
14. Pokkakillath, S., Suleri, J.: ChatGPT and its impact on education. Res. Hospitality Manag
13(1), 31–34 (2023)
15. Haugom, E., Lyocsa, S., Halousková, M.:The financial impact of ChatGPT for the higher
education industry in the US. Available at SSRN 4546522 (2023)
16. Jain, V., Rai, H., Parvathy, Mogaji, E.: The Prospects and Challenges of ChatGPT on
Marketing Research and Practices. SSRN Electron. J. (2023)
17. Tanveer, M., Khan, N., Ahmad, R.: AI support marketing: understanding the customer journey
towards the business development (2021)
 Identification of Splice Image Forgeries
with Enhanced DenseNet201 and VGG19

Satyendra Singh1 , Rajesh Kumar1,2(B) , and Chandrakant Kumar Singh2

1 Department of Electronic and Communication, University of Allahabad, Prayagraj, India
[email protected], [email protected]
2 Uttar Pradesh Rajrshitandon Open University, Prayagraj, India

Abstract. Digital image manipulation has become increasingly sophisticated,

leading to a rise in image forgery cases. In this work proposed model to detect splice
and copy-move manipulations using deep learning model, Enhanced DenseNet201
and VGG19, on the widely used CASIA 2.0 and CASIA 1.0 dataset. The pro-
posed methodology focuses on enhancing the accuracy of forge image classifica-
tion through the utilization of advanced convolutional neural networks (CNNs).
Enhanced DenseNet201 and VGG19 models have fine-tuned and trained on the
CASIA 1.0 and CASIA 2.0 dataset, which comprises authentic and tampered
photos. The models are equipped to identify instances of splice and copy-move
forgeries. By leveraging the deep features extracted by these models, the pro-
posed approach demonstrates remarkable performance in accurately distinguish-
ing between genuine and manipulated images. The evaluation of the models is car-
ried out using standard evaluation metrics accuracy, precision, recall, and auc. The
achieved results underscore the effectiveness of the Enhanced DenseNet201 app-
roach, with accuracy, precision, recall, and auc values of 0.9181, 0.9157, 0.9113,
and 0.9357, respectively. These metrics showcase the models’ robustness and
capability to classify splice and original images with high precision and recall.

Keywords: copy-move · splicing · DensNet201; VGG19 · deep learning ·

Transfer learning

1 Introduction
A digital image is a picture that has been created, captured, or digitized electronically,
and can be viewed, edited, and manipulated using digital devices and software [1]. A
fake image is an image that has been manipulated or generated to misrepresent reality. It
can be created using various digital image manipulation techniques and it can be used for
various purposes, such as creating a false impression or deception, political propaganda.
Fake image deceives viewers into believing that the image is authentic [2]. Manipulated
pictures that changed or modified by software tools such as Adobe Photoshop or GIMP.
For example, someone might manipulate an image to remove or add certain elements,
change colours, or adjust the brightness and contrast to create a false impression [3].
Manipulated images, also known as “photoshopped” images, are digital images that
have been altered or edited in some way to change their appearance or convey a different

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 113–123, 2025.
https://doi.org/10.1007/978-3-031-71481-8_9
114 S. Singh et al.

message. The manipulation can be done using various software tools such as Adobe
Photoshop, GIMP, or other photo editing software. Manipulated images can be created
for many reasons, including artistic expression, advertising, social media, or to deceive
people [4]. In Fig. 1 shows the examples of spliced images and original images.

Fig. 1. Original and spliced images

In some cases, manipulated images can be used to create a humorous effect or to

emphasize a point or message. However, in other cases, manipulated images can be used
to spread false information or to mislead people. Examples of manipulation include
altering the colour, lighting, or contrast of an image, removing or adding objects or
people, or creating composite images by merging multiple images together. Additionally,
other more complex manipulations can include creating an entirely new scene, distorting
the shape or size of objects or people, or creating fake images of events that never occurred
[5].
In brief, this paper highlights our main contributions as follows:
We proposed an approach for detecting image splicing using Enhanced DenseNet201
models. This method comprises a backbone network to extract features and an artificial
neural network-based binary classifier to identify spliced photos.
In the proposed model a combination of convolutional layers from the Enhanced
DenseNet201 model (CNN) for feature extraction and fully connected layers (ANN) for
the classification task.
In the context of classification, we have been utilized a fully connected artificial
neural network as an alternative to conventional classifiers like Support Vector Machines
(SVMs).
Remaining paper has been arranged as follows: In Sect. 2, we provide an overview
of the relevant existing literature. In Sect. 3, introduce and elaborate on our proposed
model, including dataset details and parameter selection. Sect. 4 presents an extensive
and detailed summary of our experiment, results, in-depth discussion. Finally, in Sect. 5,
conclusions of the presented study and provide insights into future research directions.
 Identification of Splice Image Forgeries 115

2 Literature Review

We conducted a comprehensive review of existing research focused on image forgery

detection, thoroughly examining the findings presented in these studies.
Guo et al. [6] describes a study focused on detecting fake colorized images, a rel-
atively new type of image manipulation technique that can be used to deceive object
recognition algorithms. In this paper highlights the importance of image forensics in
detecting image manipulations and the need for research in detecting emerging tech-
niques like colorization. Also describes the limitations of existing techniques and the
contribution of the proposed methods in detecting fake colorized images. Li et al. [7]
described model t detect fake colourized images using neural network.
He et al. [8] proposed a splicing manipulation detection model. Authors used Markov
feature in DWT and DCT and support vector machine for classification. Hamid, et al.
[9] present an image manipulation detection model. The research paper highlights the
drawbacks of current techniques used for identifying fake images and emphasizes the
requirement for more robust methods that can effectively handle images generated by
Generative Adversarial Networks (GANs).
Mo et al. [10] discusses the potential ethical and legal issues that may arise from the
use of high-quality fake face images generated by the Generative Adversarial Network
(GAN) model in image tampering. The paper proposes a CNN based method for detecting
such manipulated face photos and demonstrates through experimental evidence that this
model achieves an overall accuracy of over 99.4%. In this work highlights the need for
reliable methods for detecting fake face images generated by GANs to address potential
ethical and legal concerns.
Wang et al. [11] proposed, an image splicing manipulation becomes more prevalent,
the need for accurate detection methods has grown. The authors propose a new approach
to detect splice images based on a CNN with a weight combination technique. The
presented model uses three types of features like YCbCr, edge, and photo response
non-uniformity to distinguish splicing manipulation, which are combined with varying
weights determined during the CNN training process. Zhao et al. [12] introduced a
chroma space for splicing image forgery detection with SVM classifier.
Abd El-Latif et al. [13] image splicing forgery that is often difficult to detect. The
detection of spliced images is a topic of great interest in the field of multimedia forensics
and security. Various algorithms have been proposed for spliced image detection, but they
often suffer from issues such as low accuracy and high FPR. This paper also performs
additional experiments using discrete cosine transform (DCT) and principal component
analysis. Thakur et al. [14] proposed spliced photo manipulation detection method. In
this model, author used SVM as a classifier. Bird et al. [15] describe a method to classify
machine generated images and achieved 92.98 % classification accuracy.
In Table 1 show the comparation of the existing work related to digital image forgery
detection.
116 S. Singh et al.

Table 1. Comparative analysis of the related work

Author Year Detection Dataset Observation

Technique
Guo, et al. [6] 2018 SVM, Fisher ImageNet Developed two
Vector methods FCID-HIST
and FCID-FE. The
FCID-FE performed
better than other
Wang et al. [11] 2020 YCbCr, CNN, CASIA v1 and This method generates
PRNU CASIA v2 redundant data and not
perform better for all
splice images
Abd El-Latif et al. 2020 CNN, DWT and CASIA v1 and This method works on
[13] SVM CASIA v2 low dimensional
feature vector
Thakur et al. [14] 2018 DWT, SURF and CASIA v1 It is work on scaling
SVM and orientation attacks
Bird et al. [15] 2023 Deep CNN CIFAKE Accuracy 92.98%

3 Proposed Model

In this proposed work, we developed a binary classification model to identify image is

spliced or original. To the detection of spliced image using a deep neural network based
DenseNet201 and VGG19 model. The developed model includes some important steps:
input images from CASIA 1.0 and CASIA 2.0 datasets, apply Error Level Analysis, Data
preparation, DEnseNet201 deep learning model, Model evaluation, and final classifica-
tion steps to classify forged or original. These steps of the proposed model demonstrate
in Fig. 2.

Fig. 2. Block diagram of proposed model

 Identification of Splice Image Forgeries 117

3.1 Dataset Description

Our experiments are conducted on two widely used dataset CASIA 1.0 and CASIA 2.0
for manipulation detection [16]. The CASIA 1.0 has 800 original and 921 manipulated
images and CASIA 2.0 revised dataset poses a greater challenge as it introduces post-
processing on the boundary area of forged regions. This dataset includes 7,491 original
and 5,123 manipulated colour images in various formats such as BMP, JPEG, and TIFF,
with sizes ranging from 240 × 160 to 900 × 600 pixels. The CASIA 2.0 datasets
contain splicing and copy-move tampered images. In Table 1 shows CASIA 2.0 dataset
dimensions.

Table 2. Shows the dataset details which is used in proposed model

Dataset Original Manipulated Total

Casia 1.0 800 921 1721
Casia 2.0 7491 5123 12614

3.2 DenseNet 201

It is a deep convolutional neural network (CNN) architecture and an extension of the orig-
inal DenseNet model [17]. The primary motivation behind DenseNet201 is to address the
vanishing gradient problem and improve information flow in deep CNNs. DenseNet201
is the use of dense connectivity. This dense connectivity allows for maximum informa-
tion flow across the network, enabling each layer to access the features extracted by all
preceding layers. As a result, feature reuse is promoted, and gradients can flow more
easily through the network, addressing the vanishing gradient problem. In the Fig. 3
illustrate the DenseNet 201 architecture, which is used in the proposed model.

Fig. 3. DenseNet 201 with 4 blocks, transition layers between two blocks, which modify the
feature map size using convolutional operation and pooling

In total, the DenseNet201 architecture has 98 convolutional layers, including the

layers in the dense blocks and the transition blocks. The transition blocks are responsible
for down-sampling the spatial dimensions of the feature maps between dense blocks. In
the proposed DenseNet 201, dense block 1 to dense block 4 presents the number of layers
6, 12, 48, and 32 respectively. The architecture also includes a global average pooling
118 S. Singh et al.

layer, which aggregates global information from the entire feature map, converting it into
a one-dimensional vector before the final classification layer. This flattened vector is then
connected to a series of fully connected (FC) layers. The first FC layer has 256 nodes and
applies the Relu activation function, dropout (0.6 rate) to prevent overfitting. Finally, the
output from the last FC layer is fed into a FC layer with 2 nodes, corresponding to the
binary classification result. The sigmoid activation function used in final layer, which
maps the output values to a range (0 and 1), representing the probability of each class.

3.3 Pretrained VGG19 Model

In this section, we propose to leverage the power of pretrained deep learning models
to extract high-level features from images. Specifically, we will use the VGG19 neural
network, a popular and well-performing convolutional neural network (CNN), pretrained
on a large dataset to perform image classification.

Fig. 4. Proposed architecture of VGG19 model

From Fig. 4 the architecture of the VGG19 model, which is a pre-trained model used
for various computer vision tasks such as image classification. It includes 16 conv2D
layers, 5 maxpooling layers, and 3 FC layers, along with a sigmoid classifier. The output
from the last max pooling operation is then passed to a flatten layer, which transforms
the multidimensional data into a 1D vector. This flattened vector is then connected to
a series of fully connected layers. The first FC layer has 256 nodes and applies the
Relu activation function, dropout layer (rate of 0.5) to prevent overfitting. The process
continues with additional fully connected layers, each with decreasing numbers of nodes
(64 and 32, respectively) and Relu activation. Finally, output from the last FC layer is
fed into a final FC layer with 2 nodes, for binary classification. The block diagram shows
the architecture of the VGG19 method, which is a variant of the VGG. This method is
designed to classify input images into one of two classes.

3.4 Details of Implementation

Throughout the training process, we train Enhanced DenseNet201 models using the
Adam optimizer. Enhanced DenseNet201 is trained on 150 epochs. The starting learn-
ing rate for models is set to 0.0001. During training, we use a batch size of 8, which means
the models process 8 samples in parallel during each training iteration. For validation
 Identification of Splice Image Forgeries 119

purposes, we allocate 10% data for training and validation set to assess the models’ per-
formance on unseen data and prevent overfitting. The experiment performed on Google
Colab, which serves as our training framework. Google Colab provides cloud-based
GPU acceleration, making it a suitable platform for training deep learning models with
computational power. The chosen GPU for the experiments is T4 with 25 GB of RAM.

4 Experimental Results
The propose transfer learning model for splice image classification has been done on
casia 1.0 and casia 2.0 datasets. The proposed model results illustrate on Table 2 and
Table 3. In Table 2 show, training results of the metrices accuracy, loss, precision, recall
and AUC.

Table 3. Training results on casia 2.0 and casia 1.0 datasets.

Method Dataset Accuracy Loss Precision Recall AUC

DenseNet201 Casia 1.0 0.8960 0.2607 0.8963 0.8934 0.9588
Casia 2.0 0.9936 0.023 0.9936 0.9931 0.9990
VGG19 Casia 2.0 0.9201 0.2120 0.9248 0.9253 0.9768

In the Table 3 display the validation results of the proposed model for metrices
accuracy, loss, precision, recall and AUC.

Table 4. Validation result on casia 2.0 and casia 1.0 datasets

Method Dataset Accuracy Loss Precision Recall AUC

DenseNet201 Casia 1.0 0.8837 0.317 0.8830 0.8779 0.9417
Casia 2.0 0.9181 0.5461 0.9157 0.9113 0.9357
VGG19 Casia 2.0 0.8824 0.3265 0.8849 0.8855 0.9368

In Table 4 Compare the proposed results from existing studies which is related to
splice image identification (Table 5).

4.1 Analysis of the Results

The experiment has been performed for the spliced image identification with the help of
enhanced DenseNet201 transfer learning-based model. The performance of the model
measure in the follows metrices as accuracy, recall, precision and auc. The performance
of the developed model is better as compared to the results of related studies presented
in Table 4.
In Fig. 5 display the train and validation results of the proposed spliced image
identification model on CASIA 2.0 and Fig. 6 casia 1.0 dataset. The analysis of the
performance of train and validation results of the proposed model shows in Fig. 7.
120 S. Singh et al.

Table 5. Results compare with existing research

Author Dataset classifier Accuracy

Almawas, et al. [18] CASIA 2.0 SVM 86.5%
Naïve Bayes 87.01%
KNN 88.6%
CASIA 1.0 SVM 49.89%
Jaiswal, et al. [19] CASIA 2.0 Multiclass Model 70.26%
KNN 59.91%
Naïve Bayes 59.91%
Proposed CASIA 2.0 DenseNet201 91.81%
VGG19 88.49%

Fig. 5. Performance of the DenseNet201 model on casia 2.0 dataset.

Fig. 6. Performance of DenseNet201 on casia v1.0 dataset.

4.2 Model Performance Evaluation

The important parameter for the performance evaluation of developed binary classifica-
tion model are
 Identification of Splice Image Forgeries 121

Fig. 7. VGG19 model performance on CASIA 2.0

True Positives (Tp): The number of cases that were accurately categorized as positive
(Forge).

True Positives (Tp) = (True Positive Predictions) (1)

False Positives (Tp): The number of cases that were misclassified as positive when they
were negative (Original).

False Positives (Tp) = (False Positive Predictions) (2)

True Negatives (Tn): The cases that were correctly categorized as negative (Original).

True Negatives (Tn) = (True Negative Predictions) (3)

False Negatives (Tn): The number of cases falsely categorized as negative (Original)
even though they are positive (Forge).

False Negatives (Tn) = (False Negative Predictions) (4)

The Eqs. 1, 2, 3, and 4 represent formula for the measurement of performance of

classification model [20, 21].
AUC: it is stand for area under curve and calculated using the area under the ROC curve
[22].

5 Conclusion

This research contributes to the field of digital image forensics by introducing a powerful
approach for identification of splice and copy-move forgeries. Enhanced DenseNet201
and VGG19 models performed on the CASIA v1.0 and CASIA 2.0 dataset. The results of
the proposed model have been impressive than other studies. The deep learning-based
classifier has been better results as compared to conventional classifiers like Support
Vector Machines (SVMs). The proposed model is very useful for the spliced image
identification and future research in this field.
122 S. Singh et al.

References
1. Asghar, K., Habib, Z., Hussain, M.: Copy-move and splicing image forgery detection and
localization techniques: a review. Aust. J. Forensic Sci. 49(3), 281–307 (2017)
2. Qazi, T., et al.: Survey on blind image forgery detection. IET Image Proc. 7(7), 660–670
(2013)
3. Al-Qershi, O.M., Khoo, B.E.: Passive detection of copy-move forgery in digital images:
state-of-the-art. Forensic Sci. Int. 231(1–3), 284–295 (2013)
4. Mahmood, T., Nawaz, T., Irtaza, A., Ashraf, R., Shah, M., Mahmood, M.T.: Copy-move
forgery detection technique for forensic analysis in digital images. Math. Probl. Eng. 2016,
1–13 (2016)
5. Kakar, P., Sudha, N., Ser, W.: Exposing digital image forgeries by detecting discrepancies in
motion blur. IEEE Trans. Multimedia 13(3), 443–452 (2011)
6. Guo, Y., Cao, X., Zhang, W., Wang, R.: Fake colorized image detection. IEEE Trans. Inf.
Forensics Secur. 13(8), 1932–1944 (2018)
7. Li, Y., Zhang, Y., Lu, L., Jia, Y., Liu, J.: Using neural networks for fake colorized image
detection. In: Advances in Digital Forensics XV: 15th IFIP WG 11.9 International Conference,
Orlando, FL, USA, 28–29 Jan 2019, Revised Selected Papers 15, pp. 201–215. Springer
International Publishing (2019)
8. He, Z., Lu, W., Sun, W., Huang, J.: Digital image splicing detection based on Markov features
in DCT and DWT domain. Pattern Recogn. 45(12), 4292–4299 (2012)
9. Hamid, Y., Elyassami, S., Gulzar, Y., Balasaraswathi, V.R., Habuza, T., Wani, S.: An
improvised CNN model for fake image detection. Int. J. Inf. Technol. 15(1), 5–15 (2023)
10. Mo, H., Chen, B., Luo, W.: Fake faces identification via convolutional neural network. In
Proceedings of the 6th ACM Workshop on Information Hiding and Multimedia Security,
pp. 43–47 (2018)
11. Wang, J., Ni, Q., Liu, G., Luo, X., Jha, S.K.: Image splicing detection based on convolutional
neural network with weight combination strategy. J. Inform. Secur. Appl. 54, 102523 (2020)
12. Zhao, X., Li, J., Li, S., Wang, S.: Detecting digital image splicing in chroma spaces. In:
Digital Watermarking: 9th International Workshop, IWDW 2010, Seoul, Korea, 1–3 Oct
2010. Revised Selected Papers 9, pp. 12–22. Springer Berlin Heidelberg (2011)
13. Abd El-Latif, E.I., Taha, A., Zayed, H.H.: A passive approach for detecting image splicing
based on deep learning and wavelet transform. Arab. J. Sci. Eng. 45, 3379–3386 (2020)
14. Thakur, T., Singh, K., Yadav, A.: Blind approach for digital image forgery detection. Int. J.
Comput. Appl. 975, 8887 (2018)
15. Bird, J.J., Lotfi, A.: CIFAKE: Image Classification and Explainable Identification of AI-
Generated Synthetic Images. arXiv preprint arXiv:2303.14126 (2023)
16. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional
networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition,
pp. 4700–4708 (2017)
17. Dong, J., Wang, W., Tan, T.: Casia image tampering detection evaluation database. In 2013
IEEE China Summit and International Conference on Signal and Information Processing,
pp. 422–426. IEEE (2013)
18. Almawas, L., Alotaibi, A., Kurdi, H.: Comparative performance study of classification models
for image-splicing detection. Procedia Comput. Sci. 175, 278–285 (2020)
19. Jaiswal, A.K., Srivastava, R.: Image splicing detection using deep residual network. In: Pro-
ceedings of 2nd International Conference on Advanced Computing and Software Engineering
(ICACSE) (2019)
20. Wu, Y., Abd-Almageed, W., Natarajan, P.: Deep matching and validation network: an end-
to-end solution to constrained image splicing localization and detection. In: Proceedings of
the 25th ACM international conference on Multimedia, pp. 1480–1502 (2017)
 Identification of Splice Image Forgeries 123

21. Yang, B., Sun, X., Chen, X., Zhang, J., Li, X.: Exposing photographic splicing by detecting
the inconsistencies in shadows. Comput. J. 58(4), 588–600 (2015)
22. Nath, S., Naskar, R.: Automated image splicing detection using deep CNN-learned features
and ANN-based classifier. Signal Image Video Process. 15, 1601–1608 (2021)
 Identification of DR (Diabetic Retinopathy)
from Messidor-2 Dataset Images Using Various
Deep and Machine Learning Techniques:
A Comparative Analysis

Piyush Jain1(B) , Deepak Motwani1 , and Pankaj Sharma2

1 Amity University, Gwalior, Madhya Pradesh, India
[email protected], [email protected]
2 Eshan College of Engineering, Mathura, India

Abstract. Diabetic retinopathy (DR) is a complication of diabetes that affects one

in four people with the disease. Retinal degeneration occurs as a result. DR is a lead-
ing cause of preventable blindness, & early detection through automated screening
is important to ensuring the health of a patient’s eyes. Preventing a significant loss
of vision requires prompt medical attention, ideally from an ophthalmologist.
Given the recent advancements in machine learning (ML) applied to healthcare,
intelligent systems may prove effective in the early diagnosis of DR. This research
seeks to identify the best method for screening diabetic retinopathy patients using
datasets such as Messidor-2 datasets by comparing eight widely used classification
algorithms (Inception-V3, DR2Net, ResNet50, IncRes-v2, CNN, SVM, RetNet-
10, and ELM with CNN-SVD). A number of metrics, like F1-score, Precision,
Recall, Accuracy, and AUC-ROC, are utilized to evaluate and contrast different
algorithms. From the obtained results, it is observed that proposed RetNet-10
model outperforms other algorithms, having an accuracy of 99.46% along with a
good AUC-ROC score of 99.67%. In addition, Inception-V3 obtains an accuracy
of 94.59% on Messidor-2 dataset, which is quite low as the dataset is small. The
simulation results show how classification algorithms perform differently, and this
can be a decisive factor in which approach to take.

Keywords: DR · ML · DL · Transfer Learning · Messidor-2 dataset

1 Introduction

According to the WHO, the number of people with diabetes is anticipated to rise from
the current 422 million to 552 million by 2030 [1]. According to the National Diabetes
Statistics Report [2] published by the United States Department of Health & Human
Services, an estimated 30.5%, or 30.5 million Americans (across all age categories),
will have diabetes in 2020. They will go undiagnosed in the number of 7.3 million.
Diabetic eye illnesses, like DR, DME, & Glaucoma, are common among people with
diabetes. Damage to a blood vessel in a retina is at a heart of DR, a most common cause

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 124–135, 2025.
https://doi.org/10.1007/978-3-031-71481-8_10
 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 125

of blindness in Western countries. Disease progression is indicated by hard exudates, the

development of vitreous hemorrhage, retinal detachment & microaneurysms. Images of
a retina with varying degrees of DR are displayed in Fig. 1: (a) normal, (b) mild, (c)
moderate, (d) severe, & (e) proliferative. [3].

Fig. 1. Retina images from the DR dataset with different DR levels: (a) normal, (b) mild, (c)
moderate, (d) severe, and (e) proliferative [3].

Due to a high cost and lengthy nature of DR diagnosis performed manually by medi-
cal professionals, there is a continuing need for research and development of automation
methods for DR identification. Computer-assisted diagnosis (CAD) for DR detection
is a potentially useful application of these methods [4]. Suitable DR detection meth-
ods for CAD frequently employ ML and DL methods, as well as a large variety of
fully pretrained DL models known as TL models. Automating DR detection in colour
fundus retinal pictures has been achieved using TL models for binary and multi-class
classification [5–8]. These models have been effective in an automatic detection of DR
outside of clinical settings, where a small dataset increases the risk of underfitting and
generalization error. As a result, TL is favoured over traditional DL methods in certain
circumstances. Although deep learning techniques, and especially CNNs, are still a rel-
atively novel area of study, they have already been put to use in a variety of contexts,
including DR detection. Neural networks, which are utilized in deep learning, are used to
perform image categorization using hundreds of equations with millions of parameters
[9].
The novelty of this study is found in its detailed evaluation of eight different classifi-
cation methods for diabetic retinopathy identification using the Messidor-2 dataset. For
studies comparing diabetic retinopathy, the Messidor-2 dataset is the way to go because
of its diversity, realism, and extensive use. Being a benchmark dataset for evaluating
the effectiveness of classification methods, it stands out from other accessible datasets,
with 1748 photos representing differences in image quality as well as DR severity. The
research stands out by taking into account other parameters beyond accuracy, showcas-
ing proposed RetNet-10 as the best performance. By being open and honest about the
126 P. Jain et al.

difficulties of preprocessing and analyzing the characteristics of datasets, researchers

might get important insights. The effect of preprocessing on various models’ process-
ing costs is recognized in the study. Image preprocessing before employing them in
different models decreases processing costs, according to the research. Models’ compu-
tational needs are reduced by picture cleaning and scaling, resulting in processing that
is both efficient as well as easier on resources. The following are the main findings of
the suggested investigation:
• The model’s efficacy was evaluated using the extensive publicly available dataset.
Messidor-2 dataset.
• Different models were utilized to separate the retina from the blood vessels after
initial processing and data augmentation.
• Images of the fundus are classified using pre-trained models based on ML, DL, &
transfer learning (TL).
• Precision, accuracy, F1-score, specificity, sensitivity, & AUC-ROC are just few of the
performance indicators used to verify the proposed model’s diagnostic efficacy.
Here is how a rest of a paper is structured: Sect. 2 gives an update on a state of an
investigation. In Sect. 3, we detail the datasets we used and present our recommended
technique. Section 4 explains a findings & the discussion, and Sect. 5 wraps up the study.

2 Literature Review

Recent DR detection literature focuses on alternate methods for refraction, illumination,

and occlusion, which are common issues with typical fundus pictures. Researchers and
data scientists from across the globe attended the meeting to learn about DL-ML-TL DR
detection methods.
The proposed structure is a hybrid of three distinct phases. To begin with [10],
a normalized procedure and an enhanced method are utilized to pre-process a fundus
image. The Indus pictures in the eye will be extracted by a trained method with Beginning
V3 architecture & Resnet50, and the outcome will come with astounding accuracy of
83.79 percentile through the use of various activation functions.
In [11], describes a novel approach to using Grad-CAM with a multi-label classifi-
cation method that is based on deep learning. The results show accurate lesion outlining
and strong DR classification accuracy (93.9%) and specificity (94.4%).
In [12], SEResNeXt32x4d and EfficientNetb3 models have been pretrained. The
standardized 224×224 pictures are improved with AUGMIX and then pooled using
GeM. By reaching a high training accuracy of 0.91, the approach may reduce the amount
of work required by medical staff members who must analyze retinal images.
In [13] presents a new approach to identifying diabetic retinopathy (DR), with a
particular emphasis on the precise identification of aneurysm in colour images of the
fundus. The study used a LSVM to obtain a high level of sensitivity (96%) and precision
(92%), respectively.
In [14], offers a ResNet CNN model for feature extraction, a compressed feature
vector is produced. The overall categorization accuracy is an impressive 86.67% as a
consequence, outperforming current techniques.
 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 127

Table 1. Comparative Study of Related Work

Author & year Model Dataset Results Research gap

[15] VGG19, DR Kaggle The method’s The aim is for
InceptionV3 data classification computers to
accuracy can reach recognise DR with
0.60, the same
retinopathy grade
accuracy as
physicians
[16] Inception V3, Messidor an outcomes of Machine learning
EyeWeS dataset Inception V3 by hybrid research will
94.9% AUC to improve this work
95.8%
[17] CNN DR Kaggle In training, the Even with a limited
data model was 98.6 number of pictures,
percent accurate. Inception Version 3
There was a No DR CNN trained and
accuracy of 86.6% tested well for
for the model, diabetic eye
condition diagnosis
[18] ResNet-50 Diabetic ResNet-50 model A future author may
model Retinopathy network achieved improve this
Kaggle 89.4% a approach and use it
to treat diabetes
[19] VGG-16, APTOS The suggested More ML and DL
DenseNet dataset model had a approaches are
98.32% accuracy required for DR
on the APTOS detection
dataset and a
98.71% accuracy
on the Kaggle
dataset

3 Methodology
Methods for identifying DR in fundus photographs of the retina are compared in this
research. Messidor-2 makes use of freely available datasets. Image scaling, data label-
ing, and data cleaning are just a few examples of preprocessing procedures that improve
quality. An accuracy of a classifications obtained is directly related to a quality of pre-
processed images. Extracting the OD and the blood vessel using preprocessing tech-
niques yields the enhanced image. In the end, after some preprocessing for feature
extraction and selecting the best ones, a number of ML, transfer learning, & DL models
were developed. Finally, an enhanced picture dataset is employed in conjunction with
the proposed transfer learning-based models for DR diagnosis. AUC-ROC, F1 score,
recall, accuracy, and precision are just a few of the performance metrics used to judge
128 P. Jain et al.

the suggested method’s effectiveness. In Fig. 2, we see a flowchart of the proposed

procedure.

3.1 Data Collection

The Messidor-2 dataset was the most often utilized public dataset for testing the perfor-
mance of the proposed system. There are 1748 retinal fundus photos in the Messidor-2
dataset [20]. There is a wide range in image quality, colour accuracy, lighting, and reso-
lution within the database. It includes 1748 photos from 874 different examinations. We
didn’t have a way to categorize tree images, so we left them out. Ten hundred and twelve
photos of funduses without DR lesions and seven hundred and thirty-three images of
funduses with DR lesions of varying severity (0: no DR, 1: mild DR, 2: moderate DR,
3: severe DR, NPDR, 4: PDR) were used in a study’s tests. Images with lesions are rare
compared to those without, and they are spread evenly across the four abnormal grades
(1:4). At the outset of the study, the number of photos in a Messidor-2 database was
doubled by processing a vertical mirroring version of a database, bringing the total to
3490.

Start

Data Collection

Data Pre-processing

Data Splitting

Training Testing /Validation

Detections and classification models (Inception-V3, DR2Net,

ResNet50, IncRes-v2, CNN, SVM, RetNet-10, and ELM with
CNN-SVD)

Evaluate models performance according to accuracy, precision,

recall, f-score and AUC-ROC

End

Fig. 2. Methodology flowchart.

 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 129

3.2 Data Preprocessing

Quality of pre-processed photos affects categorization accuracy. Preprocessing photos
before sending them into a neural network is essential for accuracy. It removes arte-
facts noise and highlights confusing but significant elements. High-quality colour reti-
nal images are needed to classify DR quickly and accurately [21]. Public retinal fundus
image collections may have background noise due to their varying resolutions and com-
pression methods [22]. DR may be difficult to classify without preprocessing because
to the neural network model utilised for classification. After obtaining the preprocessed
image, merely feeding the neural network the same-sized images reduce the model’s
processing cost. We resized the revised photos to meet these requirements. After image
preparation, we get clean, noise-free images.

3.3 Data Splitting

After the dataset has been partitioned, the photos can be used in the suggested model. The
preprocessed dataset was splitted into three equal phases for use in training, validation,
and testing, with a ratio of 80:10:10.

3.4 Classification Methods

This study tested various ML & DL models, including Inception-V3, DR detection
model (DR2Net), ResNet50, IncRes-v2, CNN, SVM, proposed RetNet-10, and ELM
with CNN-SVD, using the Messidor-2 distinct diabetic retinopathy (DR) dataset. To
enable a complete evaluation of the models’ effectiveness, each dataset was carefully
picked.

4 Experiments Results Analysis

In this section, we provide comparative outcomes & discussion of multiple machine
learning, deep learning, transfer learning and ELM models utilizing a Messidor-2 dataset
for DR detection and classification.

4.1 Performance Measures

With the goal of evaluating every type of categorization model currently available, such
as machine learning, transfer learning, DL, and so on. FP, FN, TP, & TN values are used
in general to calculate performance metric values utilizing the corresponding Eqs. (1)–
(4) below [23, 24]. We evaluated an effectiveness of a proposed methodology by looking
at its F-1 score, precision, AUC-ROC, sensitivity, specificity, & accuracy.
Classification Accuracy – calculates a percentage of correct classifications compared
to a total number of forecasts by using the following formula:
TP + FN
Accuarcy = (1)
TP + TF + FP + FN
130 P. Jain et al.

Precision – Percentage representing the proportion of correctly labelled images relative

to the total number of predicted images.
TP
Precision = (2)
TP + FP
Recall – Percentage representing the proportion of images in a specific class that were
successfully identified.
TP
Recall = (3)
TP + FN
F1 Score – Identifies the model’s performance as a whole by averaging its Precision and
Recall ratings. To express the score as a number, we use the following notation:
2(precision × recall)
F1 − score = (4)
precision + recall
Specificity – The specificity of a test is its ability to designate an individual who does not
have a disease as negative. A highly specific test means that there are few false positive
results.
TN
Specificity = (5)
TN + FP
AUC (Area under the ROC Curve)- An AUC measures overall performance at any
given categorization cutoff. The AUC can be thought of as the likelihood that the model
will assign a higher score to a positive example than a negative example at random. The
ROC curve plots the true positive rate (Sensitivity) against the false positive rate (100-
Specificity) as its discrimination threshold is varied. All the points on the ROC curve
can be regarded as a sensitivity/specificity pair corresponding to a particular decision
threshold.

4.2 Comparative Results on Messidor-2 Dataset

The following sections provide the compare analysis of DR detection using the Kaggle
Messidor-2 dataset. Below are the comparative results of different ML models according
to Accuracy, Recall, F1-Score, AUC-ROC, and Precision measure. Table 1 shows the
model performance (Table 2).
The bar graph in Fig. 3 compares the accuracy of several models for DR image
detection and automated categorization. Blue bars make up the horizontal bar graph.
The most accurate is RetNet-10 at 99.46%. With 94.59% accuracy, Inception-V3 has
lowest accuracy.
In Fig. 4, a bar graph compares the accuracy of several models for DR image detection
and automated categorization. The comparison shows that SVM has 100% precision and
Inception-V3 93%.
The comparative Fig. 5 shows the recall measure performance with multiple models,
including ML, DL and TL models. The CNN model obtains highest recall with 100%,
while lower recall is by the ELM with CNN_SVD model.
 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 131

Table 2. Comparison between various models’ performance on the Messidor-2 dataset.

Models Accuracy Precision Recall F1-Score AUC-ROC

Inception-V3 [25] 94.59 95.12 94.81 92.99 96.9
DR2Net [26] 96.85 95.3 97.3 96.29 95.59
ResNet50 [3] 99.1 98.8 98.3 98.7 99.78
IncRes-v2 [27] 96.2 96.9 96.7 89.1 97.9
CNN [28] 98.37 96.2 100 98.08 98.37
SVM [28] 97.23 100 94.6 97.22 97.10
RetNet-10 99.46 98.65 98.66 98.65 99.67
ELM with CNN-SVD [29] 96.26 98 94 96 98.42

Comparaison of Accuracy Measure

100 99.46
99.1
99 98.37
98 97.23
96.85
97 96.2 96.26
in %

96
94.59
95
94
93
92

Models

Fig. 3. Comparative bar graph of accuracy measure with different ML models

Figure 6 compares f1-score performance among models. With 98.65%, ResNet10

has the highest f1-score, whereas IncRes-V2 has 89.1%. These models have f1-scores
of 97.22% for SVM and 96.29 for DR2Net. Another model demonstrates f1-score
performance.
132 P. Jain et al.

Comparaison of Precision Measure

101 100
100 98.8 98.65
99 98
98 96.9
97 96.2
in %

96 95.12 95.3
95
94
93
92

Models

Fig. 4. Comparative bar graph of precision measure with different ML models

Comparaison of Recall Measure

101 100
100
98.3 98.66
99
98 97.3
96.7
97
in %

96 94.81 94.6
95 94
94
93
92
91

Models

Fig. 5. Comparative bar graph of Recall measure with different ML models

Figure 7 compares AUC-ROC measure performance among models. In image, x-

axis shows model AUC % and y-axis shows model values with bars. Inception-V3
has 96.9% AUC-ROC, DR2Net 95.59%, ResNet50 99.78%, IncRes-v2 97.9%, CNN
98.37%, SVM 97.1%, RetNet-10 99.67%, and ELM with CNN-SVD 98.42% on the
y-axis labelled “Models”. The maximum AUC-ROC was 99.78% for RetNet-10, while
the lowest was 95.59%, respectively.
 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 133

Comparaison of F1-Score Measure

100 98.7 98.08 98.65
97.22
98 96.29 96
96
94 92.99
in %

92
90 89.1

88
86
84

Models

Fig. 6. Comparative bar graph of F1-score measure with different ML models

Comparaison of AUC-ROC Measure

101
99.78 99.67
100
99 98.37 98.42
97.9
98 97.1
96.9
in %

97
96 95.59

95
94
93

Models

Fig. 7. Comparative bar graph of AUC measure with different ML models

5 Discussion
Different models for diabetic retinopathy categorization have different performance char-
acteristics, according to the comparison. In particular, the outstanding F1-Score, AUC,
recall, accuracy, and precision values displayed by ResNet50 and RetNet-10 attest to their
strong performance. Inception-V3 shows balanced stats, whereas CNN shows flawless
recall. DR2Net provides an excellent compromise between recall and precision, whereas
134 P. Jain et al.

SVM demonstrates great precision. The figures highlight the exceptional performance
of proposed RetNet-10, which is a top-notch model for detecting diabetic retinopathy. Its
accuracy is 99.46% and its AUC-ROC is 99.67%. These findings highlight how critical it
is to select suitable models for trustworthy categorization in medical image processing.

6 Conclusion and Future Work

This study examines a use of automatic screening for diabetic retinopathy cases and
compares eight popular classifier methods. Each classification method’s parameters were
adjusted to produce the best outcomes. This evaluation was carried out so as to be
used as a criterion for selecting the appropriate classification method (which is highly
application dependent) in future studies. The data shows a pattern in which classification
methods work best and how their accuracy varies depending on a number of various
characteristics. The results suggest that the proposed ResNet10 model has the most
potential for diagnosing cases of diabetic retinopathy, however other methods can be
modified to achieve acceptable performance as well. Using a larger dataset for future
research is an option. But we have to think about how the dataset is skewed. Although
the report’s findings don’t specify which TL model was utilized for optimization, it’s
safe to assume that doing so would have helped get rid of underperforming elements.

References
1. Bourne, R.R.A., et al.: Causes of vision loss worldwide, 1990–2010: a systematic analysis.
Lancet Glob. Heal. (2013). https://doi.org/10.1016/S2214-109X(13)70113-X
2. DHHS: National Diabetes Statistics Report (2020)
3. Hacisoftaoglu, R.E., Karakaya, M., Sallam, A.B.: Deep learning frameworks for diabetic
retinopathy detection with smartphone-based retinal imaging systems. Pattern Recognit. Lett.
(2020). https://doi.org/10.1016/j.patrec.2020.04.009
4. Das, A.: Diabetic retinopathy: battling the global epidemic. Investig. Ophthalmol. Vis. Sci.
(2016). https://doi.org/10.1167/iovs.16-21031
5. Alban, M., Gilligan, T.: Automated detection of diabetic retinopathy using fluorescein
angiography photographs. cs231n课程project (2016)
6. Kermany, D.S., et al.: Identifying medical diagnoses and treatable diseases by image-based
deep learning. Cell (2018). https://doi.org/10.1016/j.cell.2018.02.010
7. Mahmoud, M.H., Alamery, S., Fouad, H., Altinawi, A., Youssef, A.E.: An automatic detection
system of diabetic retinopathy using a hybrid inductive machine learning algorithm. Pers.
Ubiquitous Comput. (2023). https://doi.org/10.1007/s00779-020-01519-8
8. Aswathi, T., Swapna, T.R., Padmavathi, S.: Transfer learning approach for grading of diabetic
retinopathy. J. Phys. Conf. Ser. 1767(1), 012033 (2021). https://doi.org/10.1088/1742-6596/
1767/1/012033
9. Hassan, D., Gill, H.M., Happe, M., Bhatwadekar, A.D., Hajrasouliha, A.R., Janga, S.C.:
Combining transfer learning with retinal lesion features for accurate detection of diabetic
retinopathy. Front. Med. (2022). https://doi.org/10.3389/fmed.2022.1050436
10. Patra, P., Singh, T.: Diabetic retinopathy detection using an improved ResNet50-InceptionV3
structure (2022). https://doi.org/10.1109/ICCCNT54827.2022.9984253
11. Jiang, H., et al.: A multi-label deep learning model with interpretable grad-CAM for diabetic
retinopathy classification. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, vol.
2020-July, pp. 1–4 (2020). https://doi.org/10.1109/EMBC44109.2020.9175884
 Identification of DR (Diabetic Retinopathy) from Messidor-2 Dataset Images 135

12. Ramchandre, S., Patil, B., Pharande, S., Javali, K., Pande, H.: A deep learning approach
for diabetic retinopathy detection using transfer learning. IEEE Int. Conf. Innov. Technol.
INOCON 2020, 1–5 (2020). https://doi.org/10.1109/INOCON50539.2020.9298201
13. Kumar, S., Kumar, B.: Diabetic retinopathy detection by extracting area and number of microa-
neurysm from colour fundus image. In: 5th Int. Conf. Signal Process. Integr. Networks, SPIN
2018, pp. 1–6 (2018). https://doi.org/10.1109/SPIN.2018.8474264
14. Elswah, D.K., Elnakib, A.A., El-Din Moustafa, H.: Automated Diabetic Retinopathy Grading
using Resnet (2020). https://doi.org/10.1109/NRSC49500.2020.9235098
15. Wu, Y., Hu, Z.: Recognition of diabetic retinopathy based on transfer learning. In: IEEE 4th
Int. Conf. Cloud Comput. Big Data Anal. ICCCBDA 2019, pp. 1–4 (2019). https://doi.org/
10.1109/ICCCBDA.2019.8725801
16. Costa, P. et al.: EyeWeS: Weakly supervised pre-trained convolutional neural networks for
diabetic retinopathy detection. Proc. 16th Int. Conf. Mach. Vis. Appl. MVA 2019, pp. 1–6
(2019). https://doi.org/10.23919/MVA.2019.8757991
17. Jayakumari, C., Lavanya, V., Sumesh, E.P.: Automated diabetic retinopathy detection and
classification using imagenet convolution neural network using fundus images. Int. Conf.
Smart Electr. Commun. 2020, 577–582 (2020). https://doi.org/10.1109/ICOSEC49089.2020.
9215270
18. Rajkumar, R.S., Jagathishkumar, T., Ragul, D., Selvarani, A.G.: Transfer learning approach for
diabetic retinopathy detection using a residual network. In: Proc. 6th Int. Conf. Inven. Comput.
Technol. ICICT, pp. 1–5 (2021). https://doi.org/10.1109/ICICT50816.2021.9358468
19. Bhimavarapu, U., Chintalapudi, N., Battineni, G.: Automatic detection and classification of
diabetic retinopathy using the improved pooling function in the convolution neural network.
Diagnostics 13(15), 1–19 (2023). https://doi.org/10.3390/diagnostics13152606
20. Decencière, E., et al.: Feedback on a publicly distributed image database: the Messidor
database. Image Anal. Stereol. (2014). https://doi.org/10.5566/ias.1155
21. Rasta, S.H., Partovi, M.E., Seyedarabi, H., Javadzadeh, A.: A comparative study on prepro-
cessing techniques in diabetic retinopathy retinal images: illumination correction and contrast
enhancement. J. Med. Signals Sens. (2015). https://doi.org/10.4103/2228-7477.150414
22. Pinedo-Diaz, G., et al.: Suitability classification of retinal fundus images for diabetic
retinopathy using deep learning. Electron. (2022). https://doi.org/10.3390/electronics1116
2564
23. Grubbs, F.E.: Errors of measurement, precision, accuracy and the statistical comparison of
measuring instruments. Technometrics (1973). https://doi.org/10.1080/00401706.1973.104
89010
24. Sharma, M., Sharma, S., Singh, G.: Performance analysis of statistical and supervised learning
techniques in stock data mining. Data (2018). https://doi.org/10.3390/data3040054
25. Bilal, A., Zhu, L., Deng, A., Lu, H., Wu, N.: AI-based automatic detection and classification
of diabetic retinopathy using U-Net and deep learning. Symmetry (Basel) (2022). https://doi.
org/10.3390/sym14071427
26. Berbar, M.: Diabetic retinopathy detection and grading using deep learning. Menoufia J.
Electron. Eng. Res. (2022). https://doi.org/10.21608/mjeer.2022.138003.1057
27. Chetoui, M., Akhloufi, M.A.: Explainable end-to-end deep learning for diabetic retinopathy
detection across multiple datasets. J. Med. Imaging (2020). https://doi.org/10.1117/1.jmi.7.
4.044503
28. Berbar, M.A.: Features extraction using encoded local binary pattern for detection and grading
diabetic retinopathy. Heal. Inf. Sci. Syst. (2022). https://doi.org/10.1007/s13755-022-00181-z
29. Nahiduzzaman, M., Islam, M.R., Islam, S.M.R., Goni, M.O.F., Anower, M.S., Kwak, K.S.:
Hybrid CNN-SVD based prominent feature extraction and selection for grading diabetic
retinopathy using extreme learning machine algorithm. IEEE Access (2021). https://doi.org/
10.1109/ACCESS.2021.3125791
 Access-Based Authentication of Healthcare
Record in Blockchain: A Survey from Security
Perspectives

Shashank Saroop1,2(B) , Rajesh Kumar Tyagi2 , Shweta Sinha2 , Shafiqul Abidin3 ,

and Lokesh Kumar1
1 ADGITM, New Delhi, India
[email protected]
2 Amity University, Gurugram, India
3 Aligarh Muslim University, Aligarh, India

Abstract. The rapid digitization of medical records has completely changed how
patient data is shared and kept. However, maintaining these documents’ privacy
and security is still a top priority. A viable method for boosting the security and
integrity of medical data is blockchain technology. Access-based authentication
mechanisms in blockchain, in particular, can offer a solid framework for monitor-
ing and restricting access to private medical records. This study provides a thor-
ough analysis of Blockchain technology used for medical records with an emphasis
on security issues. The survey investigates various strategies and methods used to
safeguard medical records in a blockchain setting.

Keywords: Access-based authentication · Healthcare records · Blockchain

1 Introduction
The digitization of healthcare records has transformed the healthcare industry, enabling
efficient storage, retrieval, and sharing of patient information. However, ensuring the
security and privacy of these records has become a critical concern in this digital era.
Blockchain technology has emerged as a promising solution for enhancing the security
and integrity of healthcare data [1]. By leveraging the decentralized and immutable nature
of blockchain, access-based authentication mechanisms can provide a robust framework
for controlling and managing access to sensitive healthcare records. The adoption of
blockchain in healthcare has gained significant attention due to its potential to address
security and privacy challenges. Blockchain ensures transparency, tamper resistance, and
data integrity through its distributed ledger technology [2]. Access-based authentication
mechanisms built on top of blockchain can strengthen the security of healthcare records,
preventing unauthorized access and maintaining patient confidentiality (Fig. 1).
In this survey, we aim to comprehensively explore the access-based authentication
of healthcare records in blockchain systems from a security perspective. By examin-
ing existing literature and research works, we seek to analyze the various approaches

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 136–149, 2025.
https://doi.org/10.1007/978-3-031-71481-8_11
 Access-Based Authentication of Healthcare Record in Blockchain 137

Fig. 1. Working of Blockchain

and techniques employed in securing healthcare data within blockchain environments.

The survey will provide insights into the different access control models, authentication
mechanisms, and privacy-preserving techniques used in conjunction with blockchain
technology. Moreover, we will discuss the challenges faced in the healthcare domain
concerning security and privacy and explore how blockchain-based access control can
effectively address these challenges. By understanding the strengths and limitations of
access-based authentication methods, we can identify opportunities for improvement and
advancements in securing healthcare records. Also, we aim to contribute to the exist-
ing body of knowledge on access-based authentication in blockchain-based healthcare
record management systems. By highlighting the current state of research, identifying
gaps, and discussing potential future directions, we seek to promote the development of
secure and trustworthy healthcare information systems.

1.1 Background
The digitization of healthcare records has revolutionized the healthcare industry, replac-
ing traditional paper-based systems with electronic storage and management of patient
information. Electronic Health Records (EHRs) provide numerous benefits such as
improved accessibility, streamlined healthcare processes, and enhanced patient care.
However, this digital transformation has also brought forth significant security and pri-
vacy concerns [3]. Healthcare records contain highly sensitive and personal information,
including medical history, diagnoses, treatments, and demographic data. Unauthorized
access to these records can lead to breaches of patient privacy, identity theft, and misuse
of medical information. Ensuring the security and confidentiality of healthcare records
is crucial to maintaining patient trust and complying with privacy regulations such as the
Health Insurance Portability and Accountability Act (HIPAA) in the United States and
the General Data Protection Regulation (GDPR) in the European Union [4]. Traditional
centralized systems used for healthcare record storage and management are susceptible
to security vulnerabilities, as they rely on a single authority to authenticate and authorize
access to records. These centralized systems face risks such as data breaches, insider
threats, and single points of failure. To address these concerns, blockchain technology has
emerged as a potential solution for secure healthcare record management. Blockchain,
originally introduced as the underlying technology for cryptocurrencies like Bitcoin,
is a decentralized and distributed ledger that maintains a tamper-resistant record of
transactions. It offers unique features such as immutability, transparency, and consensus
mechanisms that enhance the security and integrity of data [5]. Blockchain technology
provides a decentralized approach to authentication and access control, eliminating the
need for a central authority and enabling secure and auditable data sharing. By integrating
138 S. Saroop et al.

access-based authentication mechanisms into blockchain systems, healthcare organiza-

tions can establish granular control over who can access, modify, and share patient
records. Access control models such as role-based access control (RBAC) and attribute-
based access control (ABAC) can be implemented within blockchain networks, ensuring
that only authorized individuals or entities have access to specific healthcare information
[6]. Furthermore, authentication mechanisms based on public key infrastructure (PKI),
digital signatures, and smart contracts can be leveraged in blockchain-based healthcare
record systems to verify the identities of users and ensure the authenticity of data (Kuo
et al., 2018). These mechanisms contribute to establishing a trusted environment for
accessing and managing healthcare records.

2 Blockchain Technology in Healthcare

2.1 Overview of Blockchain Technology

Blockchain technology has garnered significant attention in the healthcare industry due
to its potential to address security, privacy, and interoperability challenges associated
with healthcare data management. This section provides an overview of blockchain
technology in healthcare, highlighting its advantages and potential applications (Fig. 2).

Fig. 2. Application of Blockchain in healthcare

Enhanced Security and Data Integrity:

Blockchain’s distributed ledger architecture offers enhanced security and data integrity
for healthcare records. The decentralized nature of blockchain eliminates the need for
a central authority, reducing the risk of data breaches and unauthorized access [2]. The
immutability of blockchain transactions ensures that healthcare data remains tamper-
proof and auditable [7].

Interoperability and Data Sharing:

Blockchain technology provides a promising solution for interoperability challenges in
healthcare by enabling secure and efficient data sharing among different stakeholders.
Blockchain’s shared ledger allows for real-time updates and data synchronization across
multiple entities, improving care coordination and facilitating a seamless exchange of
information [6].
 Access-Based Authentication of Healthcare Record in Blockchain 139

Smart Contracts for Automated Processes:

Smart contracts, self-executing contracts with predefined rules encoded on the
blockchain, offer automation and efficiency in healthcare processes. These contracts can
facilitate automated claims processing, enforce compliance with contractual obligations,
and streamline administrative tasks [4].

Patient-Centric Control of Data:

Blockchain technology empowers patients by giving them control over their health data.
Patients can securely manage their consent preferences and grant selective access to
healthcare providers, ensuring privacy and facilitating patient-centered care [8].
Medical Research and Clinical Trials:
Blockchain technology has the potential to revolutionize medical research and clinical
trials. By securely storing and sharing research data on the blockchain, transparency and
trust can be established, preventing data manipulation and ensuring the reproducibility
of results [9].
Supply Chain and Drug Traceability:
Blockchain can improve supply chain management in the pharmaceutical industry by
enhancing the traceability, authentication, and transparency of drugs. It enables the
tracking of drug provenance, ensuring the integrity of the supply chain, and mitigating
the risks of counterfeit drugs [10].

2.2 Advantages of Blockchain in Healthcare

Advantages of Blockchain in Healthcare Description

Enhanced security and data integrity [2] Blockchain’s distributed ledger architecture
ensures tamper-proof storage and immutability
of healthcare data, enhancing security and data
integrity. (Mettler, 2016)
Improved interoperability and data sharing Blockchain enables secure and efficient data
[11] sharing among different healthcare
stakeholders, promoting interoperability and
facilitating seamless exchange of information.
(Kuo et al., 2017)
Efficient and transparent claims processing Blockchain-based smart contracts automate and
[7] streamline the claims processing in healthcare,
reducing administrative burdens and ensuring
transparency. (Ekblaw et al., 2016)
Patient-centric control of health data [8] Blockchain empowers patients by giving them
control over their health data and enabling
selective sharing of information, fostering
patient-centered care. (Linn et al., 2019)
(continued)
140 S. Saroop et al.

(continued)
Advantages of Blockchain in Healthcare Description
Enhanced research and clinical trial Blockchain provides a transparent and auditable
capabilities [12] platform for medical research and clinical trials,
improving data integrity, reproducibility, and
trust. (Benchoufi et al., 2017)
Reliable drug traceability and supply chain Blockchain enables end-to-end traceability of
management [13] drugs, ensuring the authenticity and integrity of
the supply chain and mitigating the risks of
counterfeit products. (Kamal et al., 2020)

2.3 Security and Privacy Challenges in Healthcare Records

See Fig. 3.

Fig. 3. Challenges in Healthcare Records

• Data Breaches and Unauthorized Access [14]:

Healthcare records are vulnerable to data breaches and unauthorized access, leading
to potential misuse or exposure of sensitive patient information. (Roman et al., 2018).
• Insider Threats [15]:
Insiders with authorized access to healthcare records, such as healthcare providers
or employees, may misuse or exploit patient data for personal gain or malicious
purposes. (Vijayakumar et al., 2018).
• Data Integrity and Trustworthiness [16]:
Ensuring the integrity and trustworthiness of healthcare records is crucial to prevent
unauthorized modifications, tampering, or falsification of patient data. (Fernández-
Alemán et al., 2013).
• Identity and Access Management [17]:
Effective identity and access management solutions are needed to authenticate and
authorize legitimate users while preventing unauthorized access to healthcare records.
(Dinh et al., 2018).
 Access-Based Authentication of Healthcare Record in Blockchain 141

• Interoperability and Data Sharing [18]:

Ensuring secure and privacy-preserving interoperability and data sharing among
different healthcare systems and stakeholders remains a challenge. (Sun et al., 2019).
• Consent Management and Patient Privacy [19]:
Managing patient consent for data sharing and protecting patient privacy in the context
of healthcare records present significant challenges, particularly in an increasingly
interconnected healthcare ecosystem. (Prasser et al., 2017).

2.4 Cryptocurrency

• Streamlining Payments and Transactions [20]:

Cryptocurrency can be used in healthcare to streamline payments and transactions
between different stakeholders, such as patients, healthcare providers, and insurance
companies. By eliminating intermediaries and reducing transaction costs, cryptocur-
rency can facilitate faster and more efficient financial transactions in the healthcare
industry. (Dagher et al., 2018).
• Enhancing Data Security and Privacy [21]:
The underlying technology of cryptocurrency, blockchain, provides a secure and
decentralized platform for storing and managing healthcare data. With cryptographic
techniques and distributed consensus mechanisms, blockchain-based cryptocurrency
can enhance data security and protect patient privacy in healthcare transactions. (Yli-
Huumo et al., 2017).
• Enabling Microtransactions and Incentives [22]:
Cryptocurrency allows for microtransactions in healthcare, enabling the implementa-
tion of innovative incentive mechanisms. For example, patients can be rewarded with
tokens or coins for participating in research studies or adhering to treatment plans,
fostering patient engagement and motivation. (Yli-Huumo et al., 2016).

Improving Supply Chain Management [20]:

Cryptocurrency can enhance supply chain management in healthcare by providing an
immutable and transparent ledger to track the movement of pharmaceuticals, medi-
cal devices, and other healthcare products. This helps prevent counterfeit drugs and
ensures the authenticity and integrity of the supply chain. (Dagher et al., 2018).
Facilitating Medical Research and Data Sharing [21]:
Cryptocurrency and blockchain technology can facilitate secure and transparent
sharing of medical research data. Researchers can use tokens or coins to incentivize
data contributors, ensure data provenance, and create a decentralized marketplace for
sharing research findings and insights. (Yli-Huumo et al., 2017).

3 Literature Review

See Table 1.
142 S. Saroop et al.

Table 1. Analysis of Various Security based approaches to secure healthcare record using
blockchain

Title Authors & Year Citations Description Results Pros Cons

Blockchain Alzahrani et al. & [23] This paper provides The paper Improved Lack of standards
Technology for 2022 a review of the concludes that security and and
Healthcare: A literature on the use blockchain has the privacy of interoperability
Review of of blockchain potential to healthcare Need for security
Applications and technology for significantly records and Privacy
Security Concerns healthcare improve healthcare. Reduced risk of
applications. The However, the paper fraud and abuse
paper discusses the also notes that there
different ways that are still some
blockchain can be security concerns
used to improve that need to be
healthcare, addressed before
including blockchain can be
improving the widely adopted in
security of healthcare
healthcare records,
improving the
efficiency of
healthcare delivery,
and reducing the
cost of healthcare.
The paper also
discusses the
security concerns
associated with
blockchain
technology in
healthcare
A Survey on Kumar et al. & [24] This paper provides The paper Improved Lack of standards
Blockchain 2022 a survey of the concludes that security and and
Technology in literature on the use blockchain has the privacy of interoperability
Healthcare: of blockchain potential to healthcare Potential for
Applications, technology in significantly records cyberattacks
Security, and healthcare. The improve healthcare. Potential for new
Privacy Concerns paper discusses the However, the paper healthcare
different ways that also notes that there applications
blockchain can be are still some
used to improve security and privacy
healthcare, concerns that need
including to be addressed
improving the before blockchain
security of can be widely
healthcare records, adopted in
improving the healthcare
efficiency of
healthcare delivery,
and reducing the
cost of healthcare.
The paper also
discusses the
security and privacy
concerns associated
with blockchain
technology in
healthcare

(continued)
 Access-Based Authentication of Healthcare Record in Blockchain 143

Table 1. (continued)

Title Authors & Year Citations Description Results Pros Cons

Blockchain for Kumar et al. & [25] This paper provides The paper Reduced risk of Lack of standards
Healthcare 2021 a systematic review concludes that fraud and abuse and
Applications: A of the literature on blockchain has the Increased interoperability
Systematic Review the use of potential to efficiency of Need for security
blockchain significantly healthcare data and privacy
technology for improve healthcare. sharing
healthcare However, the paper
applications. The also notes that there
paper discusses the are still some
different ways that challenges that need
blockchain can be to be addressed
used to improve before blockchain
healthcare, can be widely
including adopted in
improving the healthcare
security of
healthcare records,
improving the
efficiency of
healthcare delivery,
and reducing the
cost of healthcare
Security and Alzahrani et al. & [26] This paper surveys The paper Improved Lack of standards
Privacy Issues in 2020 the security and concludes that transparency and and
Blockchain-Based privacy issues of blockchain-based collaboration in interoperability
Healthcare blockchain-based healthcare systems healthcare Potential for
Systems healthcare systems. are a major target Potential for new cyberattacks
The paper discusses for cyberattacks and healthcare
the different types that there is a need applications
of cyberattacks that for new security
can be launched and privacy
against solutions for
blockchain-based blockchain-based
healthcare systems healthcare systems
and the different
ways that data can
be compromised in
blockchain-based
healthcare systems.
The paper also
discusses the
different ways that
blockchain-based
healthcare systems
can be made more
secure and private

(continued)
144 S. Saroop et al.

Table 1. (continued)

Title Authors & Year Citations Description Results Pros Cons

Blockchain for Kumar, S. et al.& [27] This paper provides The paper : Reduced risk of Need for security
Healthcare: A 2020 a comprehensive concludes that fraud and abuse and privacy
Comprehensive overview of the use blockchain has the - Increased Potential for
Overview of blockchain potential to efficiency of cyberattacks
technology in significantly healthcare data
healthcare. The improve healthcare. sharing
paper discusses the However, the paper - Improved
different ways that also notes that there transparency and
blockchain can be are still some collaboration in
used to improve challenges that need healthcare
healthcare, to be addressed
including before blockchain
improving the can be widely
security of adopted in
healthcare records, healthcare
improving the
efficiency of
healthcare delivery,
and reducing the
cost of healthcare
Blockchain in He, D. et al. & [28] This paper reviews The paper Improved Lack of standards
Healthcare: A 2020 the literature on the concludes that security and and
Review of use of blockchain blockchain has the privacy of interoperability
Applications, technology for potential to healthcare Need for security
Challenges, and healthcare significantly records and privacy
Opportunities applications. The improve healthcare. Potential for new
paper discusses the However, the paper healthcare
different ways that also notes that there applications
blockchain can be are still some
used to improve challenges that need
healthcare, to be addressed
including before blockchain
improving the can be widely
security of adopted in
healthcare records, healthcare
improving the
efficiency of
healthcare delivery,
and reducing the
cost of healthcare
Access-Based Jose et al. & 2019 [29] This paper surveys The paper Reduced risk of Potential for
Authentication of the use of concludes that fraud and abuse cyberattacks
Healthcare Record blockchain blockchain Improved Need for security
in Blockchain: A technology for technology has the transparency and and privacy
Survey from access-based potential to collaboration in
Security authentication of significantly
Perspectives healthcare records. improve the
The paper discusses security of
the security healthcare records.
challenges of However, the paper
healthcare records also notes that there
and how blockchain are still some
can be used to challenges that need
address these to be addressed
challenges. The before blockchain
paper also discusses can be widely
the different types adopted in
of blockchain-based healthcare
access control
systems that have
been proposed
 Access-Based Authentication of Healthcare Record in Blockchain 145

4 Open Research Challenges and Future Directions

Open research challenges and future directions in the use of cryptocurrency in healthcare
can include the following:
Scalability and Performance:
One of the key challenges is addressing the scalability and performance limitations
of blockchain networks in healthcare. As the volume of healthcare data and transac-
tions increases, ensuring that the cryptocurrency infrastructure can handle the growing
demands becomes crucial.
Regulatory and Legal Frameworks:
Developing appropriate regulatory and legal frameworks for the use of cryptocurrency in
healthcare is essential. Clear guidelines and regulations need to be established to ensure
compliance with data privacy, security, and financial regulations, while also considering
the unique aspects of blockchain technology.
Interoperability and Standardization:
Achieving interoperability among different blockchain networks and healthcare systems
is a challenge. Establishing common standards and protocols for data exchange, smart
contracts, and identity management will be necessary to enable seamless integration and
collaboration.
Data Privacy and Security:
While blockchain technology provides inherent security features, ensuring robust data
privacy and security in cryptocurrency-based healthcare systems is critical. Privacy-
preserving techniques, encryption methods, and access control mechanisms need to be
developed and implemented to protect sensitive patient data.
User Experience and Adoption:
Enhancing the user experience and promoting the adoption of cryptocurrency-based
healthcare systems is vital. User-friendly interfaces, intuitive wallets, and educational
resources should be developed to make it easier for healthcare professionals and patients
to use and understand the benefits of cryptocurrency in healthcare.

• Ethical Considerations:
Exploring the ethical implications of using cryptocurrency in healthcare is important.
Ethical issues related to consent, data ownership, fairness, and transparency should
be addressed to ensure the responsible and ethical use of cryptocurrency in healthcare
applications.
• Cost and Financial Considerations:
Assessing the cost-effectiveness and financial viability of implementing cryptocur-
rency in healthcare is a future direction. Analyzing the economic benefits, potential
savings, and return on investment will help determine the feasibility and sustainability
of adopting cryptocurrency in healthcare systems.
146 S. Saroop et al.

4.1 Emerging Trends and Technologies

• Artificial Intelligence (AI) in healthcare [30]:
AI is increasingly being applied in healthcare for tasks such as medical imaging
analysis, disease diagnosis, personalized medicine, and drug discovery. It has the
potential to improve healthcare outcomes, enhance efficiency, and enable more precise
and personalized treatments. (Topol, 2019).
• Internet of Medical Things (IoMT) [31]
IoMT refers to the network of connected medical devices and wearable sensors that
collect and transmit patient data for remote monitoring and healthcare management.
It enables real-time patient monitoring, early detection of diseases, and improved
patient outcomes. (Khan, Yaqoob, Hashem, et al., 2014).
• Precision Medicine [32]:
Precision medicine focuses on tailoring medical treatments to individual patients
based on their genetic, environmental, and lifestyle factors. It leverages technologies
such as genomics, proteomics, and bioinformatics to enable targeted therapies and
improve patient outcomes. (Collins & Varmus, 2015).
• Blockchain in Healthcare [21]:
Blockchain technology offers decentralized and secure storage and sharing of health-
care data, enabling enhanced data privacy, interoperability, and data integrity. It has the
potential to revolutionize healthcare data management, clinical trials, and healthcare
supply chain processes. (Yli-Huumo, Ko, Choi, et al., 2017).
• Telemedicine and Remote Patient Monitoring [33]:
Telemedicine allows for remote healthcare consultations and monitoring, enabling
patients to access healthcare services from their homes. With the advancements
in communication technologies and wearable devices, remote patient monitoring
has become increasingly feasible, leading to improved access to care and reduced
healthcare costs. (Bashshur, Shannon, & Bashshur, 2016).

4.2 Recommendations for Future Research

Privacy-Preserving Techniques for Blockchain:
Investigate and develop privacy-preserving techniques that ensure the confidentiality
of sensitive healthcare data stored on blockchain. This includes exploring encryption
methods, zero-knowledge proofs, and differential privacy approaches to protect patient
privacy while maintaining the benefits of blockchain technology.
Interoperability and Data Standardization:
Focus on developing interoperability frameworks and data standards that facilitate seam-
less data exchange and integration across different healthcare systems. This research can
involve exploring interoperable data models, ontologies, and protocols to ensure efficient
and secure data sharing among healthcare providers, patients, and other stakeholders.
Explainable Artificial Intelligence (AI) in Healthcare:
Conduct research to enhance the explainability and interpretability of AI algorithms used
in healthcare. Develop methodologies and techniques that provide transparent explana-
tions for AI-based decisions, allowing healthcare professionals and patients to under-
stand and trust the reasoning behind AI-driven diagnoses, treatment recommendations,
and patient outcomes.
 Access-Based Authentication of Healthcare Record in Blockchain 147

Ethical and Legal Implications of Emerging Technologies:

Investigate the ethical and legal implications of emerging technologies in healthcare,
such as AI, blockchain, and genomics. Research should focus on addressing concerns
related to data privacy, bias in algorithms, informed consent, data ownership, and the
impact of these technologies on healthcare equity and access.
User Experience and Human-Centered Design:
Explore user-centered design approaches to enhance the usability and user experience
of healthcare technologies. Investigate ways to involve end-users, such as patients and
healthcare professionals, in the design and development process to ensure that emerging
technologies meet their needs, are intuitive to use, and facilitate effective healthcare
delivery.
Cybersecurity and Data Protection:
Conduct research on cybersecurity measures and data protection strategies to mitigate the
risks associated with healthcare data breaches, ransomware attacks, and unauthorized
access. Explore advanced encryption techniques, secure data sharing protocols, and
proactive threat detection and response mechanisms to safeguard healthcare systems
and patient data.
Health Equity and Social Determinants of Health:
Investigate the impact of emerging technologies on health equity and address disparities
in healthcare access, delivery, and outcomes. Research should focus on understanding
and addressing the social determinants of health, leveraging technology to improve
health outcomes for marginalized populations, and ensuring equitable distribution of
healthcare resources and services.

5 Conclusion
In conclusion, the use of blockchain technology for access-based authentication of
healthcare records presents significant opportunities for improving security, privacy, and
efficiency in the healthcare industry. This survey from a security perspective has explored
the advantages of blockchain in healthcare, including enhanced data integrity, decen-
tralization, and auditability. However, several challenges and open research directions
have also been identified. The adoption of blockchain in healthcare requires address-
ing scalability issues, developing regulatory frameworks, ensuring interoperability, and
addressing data privacy and security concerns. Additionally, user experience, cost con-
siderations, and ethical implications need to be carefully considered and integrated into
the design and implementation of blockchain-based healthcare systems.
Future research should focus on privacy-preserving techniques for blockchain, inter-
operability standards, explainable AI, ethical and legal implications, user-centered
design, cybersecurity measures, health equity, and the impact of emerging technolo-
gies on healthcare. By addressing these challenges and advancing research in these
areas, we can unlock the full potential of blockchain technology in healthcare, leading
to improved patient care, secure data management, and transformative advancements in
healthcare delivery.
In conclusion, blockchain-based access-based authentication holds great promise
for enhancing healthcare record security, privacy, and interoperability. With continued
148 S. Saroop et al.

research and innovation, we can overcome the current limitations and pave the way for
a more secure and efficient healthcare ecosystem.

References
1. Smith, K., Vazirani, A., Xiong, L.: Blockchain technology in the healthcare sector: a
systematic review. Healthcare (Basel, Switzerland) 5(4), 56 (2017)
2. Mettler, M.: Blockchain technology in healthcare: the revolution starts here. In: Proceedings
of the 2016 IEEE 18th International Conference on e-Health Networking, Applications and
Services (Healthcom), pp. 1–3. IEEE (2016)
3. Liu, M., Liu, X., Song, L., Zhang, X., Xiong, N.N.: Secure and privacy-preserving access
control scheme for electronic health records in blockchain. J. Med. Syst. 43(3), 54 (2019)
4. Azaria, A., Ekblaw, A., Vieira, T., Lippman, A.: MedRec: Using blockchain for medical data
access and permission management. In: Proceedings of the 2016 2nd International Conference
on Open and Big Data (OBD), pp. 25–30. IEEE (2016)
5. Nakamoto, S.: Bitcoin: a peer-to-peer electronic cash system (2008). Retrieved from https://
bitcoin.org/bitcoin.pdf
6. Kuo, T.T., Kim, H.E., Ohno-Machado, L.: Blockchain distributed ledger technologies for
biomedical and health care applications. J. Am. Med. Inform. Assoc. 25(9), 1211–1220 (2018)
7. Ekblaw, A., Azaria, A., Halamka, J.D., Lippman, A.: A case study for blockchain in healthcare:
“MedRec” prototype for electronic health records and medical research data. In: Proceedings
of IEEE Open & Big Data Conference, pp. 1–8. IEEE (2016)
8. Linn, L.A., Koo, M.B., Jani, Y.H.: Blockchain for health data and its potential use in health
IT and health care-related research. Fed. Pract. 36(Suppl 4), S37–S41 (2019)
9. Benchufi, M., Porcher, R., Ravaud, P., Boutron, I.: Blockchain protocols in clinical trials:
transparency and traceability of consent. Lancet Oncol. 20(5), e254–e261 (2019)
10. Li, X., Zhao, Z., Wei, Y., Li, X.: Blockchain-based data integrity service framework for
pharmaceutical supply chain. IEEE Access 8, 96051–96059 (2020)
11. Kuo, T.T., Kim, H.E., Ohno-Machado, L.: Blockchain distributed ledger technologies for
biomedical and health care applications. J. Am. Med. Inform. Assoc. 24(6), 1211–1220 (2017)
12. Benchoufi, M., Ravaud, P., Boutron, I.: Blockchain technology for improving clinical research
quality. Trials 18(1), 335 (2017)
13. Kamal, S., Singh, R., Letaief, K.B.: Blockchain-enabled trustworthy and privacy-aware
healthcare system. IEEE Trans. Serv. Comput. 13(4), 689–702 (2020)
14. Roman, R., Zhou, J., Lopez, J.: On the features and challenges of security and privacy in
distributed Internet of Things. Comput. Netw. 129, 441–458 (2018)
15. Vijayakumar, A.N., Othman, Z.A., Jawawi, D.N.A.: Detection and prevention techniques for
insider threats in healthcare. J. Med. Syst. 42(12), 237 (2018)
16. Fernández-Alemán, J.L., Señor, I.C., Lozoya, P.Á.O., Toval, A.: Security and privacy in
electronic health records: a systematic literature review. J. Biomed. Inform. 46(3), 541–562
(2013)
17. Dinh, H.T., Lee, C., Niyato, D., Wang, P.: A survey of mobile cloud computing: architecture,
applications, and approaches. Wirel. Commun. Mob. Comput. 2018, 1–35 (2018)
18. Sun, J., Zhang, Y., Xiong, H., Xue, Y., Chen, J.: Privacy-preserving medical big data sharing
and analytics framework with blockchain support. J. Med. Syst. 43(6), 169 (2019)
19. Prasser, F., Kohlmayer, F., Spengler, H.: Scalable and privacy-preserving distributed consent
management for electronic health records. J. Biomed. Inform. 72, 25–36 (2017)
20. Dagher, G.G., Mohler, J., Milojkovic, M., Marella, P.B., Marella, W.M.: Ancile: privacy-
preserving framework for access control and interoperability of electronic health records
using blockchain technology. Sustain. Cities Soc. 39, 283–297 (2018)
 Access-Based Authentication of Healthcare Record in Blockchain 149

21. Yli-Huumo, J., Ko, D., Choi, S., Park, S., Kim, K.: Blockchain in healthcare: a systematic
review. Healthc. Inform. Res. 23(3), 181–193 (2017)
22. Yli-Huumo, J., Ko, D., Choi, S., Park, S., Smolander, K.: Where is current research on
blockchain technology?—a systematic review. PLoS ONE 11(10), e0163477 (2016)
23. Alzahrani, S., Jararweh, Y., Aljaloud, A.: Blockchain technology for healthcare: a review of
applications and security concerns. Healthcare 10(2), 310 (2022)
24. Kumar, A., Kumar, S., Garg, S., Singh, M.: A survey on blockchain technology in healthcare:
applications, security, and privacy concerns. IEEE Access 10, 62114–62133 (2022)
25. Kumar, A., Kumar, S., Garg, S., Singh, M.: Blockchain for healthcare applications: a
systematic review. J. Med. Syst. 45(11), 351 (2021)
26. Jose, M., Chacko, M.M., Thomas, S.B., Nadesh, R.K.: Access-based authentication of
healthcare record in blockchain: a survey from security perspectives. J. Med. Syst. 43(1),
1 (2019)
27. Alzahrani, S., Alotaibi, F., Jararweh, Y., Aljaloud, A.: Security and privacy issues in
blockchain-based healthcare systems. In: 13th International Conference on Information
Technology-New Generations (ITNG), pp. 122–129. IEEE (2020)
28. Kumar, S., Dash, S.: Blockchain for healthcare: a comprehensive overview. In: 2nd Inter-
national Conference on Computing, Communication, and Artificial Intelligence (ICCCA),
pp. 1–6. IEEE (2020)
29. He, D., Chen, L., Lin, Z.: Blockchain in healthcare: a review of applications, challenges, and
opportunities. Healthcare 8(3), 125 (2020)
30. Topol, E.J.: High-performance medicine: the convergence of human and artificial intelligence.
Nat. Med. 25(1), 44–56 (2019)
31. Khan, R., Yaqoob, I., Hashem, I.A.T., et al.: Big data: survey, technologies, opportunities,
and challenges. Scientific World J. 2014, 712826 (2014)
32. Collins, F.S., Varmus, H.: A new initiative on precision medicine. N. Engl. J. Med. 372(9),
793–795 (2015)
33. Bashshur, R.L., Shannon, G.W., Bashshur, N.: Telemedicine: an updated assessment of the
scientific literature. Telemed. J. E Health 22(2), 1–47 (2016)
 Smart Entry Management Using IoT
and Number Plate Recognition
for Organizational Security

Love Lakhwani, Vibhanshu Singh, and Sunil Kumar(B)

Department of IoT and Intelligent Systems (School of CIS),

Manipal University Jaipur, Jaipur, Rajasthan, India
[email protected], [email protected],
[email protected]

Abstract. Modern organizations require robust security measures to

ensure controlled access to their premises. The proposed system aims to
enhance security by allowing only registered vehicles to enter the organi-
zation’s premises. Using IoT connectivity, our system will seamlessly inte-
grate with existing infrastructure to capture and analyze vehicle number
plates at entry points. The number plate recognition process will be auto-
mated and efficient, enabling real-time identification and authentication
of vehicles. This technology eliminates the need for manual interventions,
streamlining the access control process and minimizing potential security
breaches.

Keywords: robust security · automated · real-time identification ·

authentication of vehicles

1 Introduction

From residential communities and commercial establishments to parking lots

and public facilities, the need for efficient and reliable entry management has
never been greater. To address this pressing need, we present the project “Smart
Entry Management using IoT and Number Plate.”
The purpose of this project is to develop an intelligent entry management sys-
tem that leverages the power of Internet of Things (IoT) and Number Plate tech-
nology to revolutionize the way entry is managed. By integrating IoT devices and
utilizing the information encoded in number plates, our system aims to provide a
seamless and secure experience for visitors and personnel alike. The significance
of this project lies in its ability to address the challenges faced by traditional
entry management systems. Conventional methods often rely on physical tokens,
such as access cards or keys, which can be lost, stolen, or easily replicated. This
poses a significant security risk and can lead to unauthorized access, compromis-
ing the safety of individuals and assets. Moreover, manual entry management

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 150–163, 2025.
https://doi.org/10.1007/978-3-031-71481-8_12
 Smart Entry Management Using IoT 151

processes are time-consuming and prone to human error, resulting in inefficien-

cies and delays. automation to a realm where it interacts intelligently with its
environment.
By harnessing the power of IoT and Number Plate technology, our system
eliminates the need for physical tokens and introduces a more secure and efficient
approach to entry management. With the integration of IoT devices, such as
smart cameras and sensors, our system can monitor and analyze entry data in
real-time.

1.1 Contribution’s

The implementation of the “Smart Entry Management using IoT and Num-
ber Plate” system holds immense potential for contributing positively to society
on various fronts. First and foremost, the enhanced security measures provided
by the system address a critical need in contemporary urban environments. As
cities and organizations face escalating security challenges, a streamlined and
automated entry management system mitigates risks associated with unautho-
rized access, contributing to the overall safety and well-being of the community.
By minimizing the potential for security breaches, the system creates a more
secure and trustworthy environment for residents, employees, and visitors.
Secondly, the project’s emphasis on leveraging IoT technology aligns with
the broader movement toward smart cities. Integrating intelligent devices and
sensors not only enhances security but also lays the foundation for a more inter-
connected urban infrastructure. The data collected by the system can be utilized
for more comprehensive urban planning, traffic management, and resource opti-
mization. This contributes to the creation of efficient and sustainable urban
spaces, promoting a higher quality of life for the community members.
Lastly, the adoption of the proposed entry management system promotes
efficiency and reduces operational overhead. By eliminating the need for man-
ual interventions, the system streamlines access control processes, minimizing
delays and potential errors associated with traditional entry management meth-
ods. This efficiency gain has a ripple effect, positively impacting productiv-
ity and resource utilization not only within organizations but also within the
broader societal context. Ultimately, the societal benefits extend beyond secu-
rity enhancement to encompass improved urban living, increased operational
efficiency, and a more resilient and connected community.

2 Literature Review
In this literature review, we explore license plate detection technology through
live camera feeds. Focusing on real-time surveillance, we examine the evolving
landscape, methodologies, and innovations that define this intersection of com-
puter vision and security applications.
Sanghyeop Lee et al. (2018) proposed Vehicle License Plate Recognition with
Convolutional Neural Networks in which he discuss about harnessing the power
152 L. Lakhwani et al.

of Convolutional Neural Networks (CNNs), this paper pioneers an integrated

solution that tackles both license plate detection and character recognition. This
approach capitalizes on the deep learning capabilities of CNNs to achieve accu-
rate and efficient recognition of license plates [1].
Ibrahim El Khatib et al. (2015) proposed an Efficient Automatic Car License
Plate Recognition System. This research offers a streamlined license plate recog-
nition system that harmonizes several vital components. Incorporating image
preprocessing, connected component analysis, and template matching, it show-
cases an efficient pipeline to accurately extract license plate information from
vehicle images [2].
Sergio M. Silva et al. (2022) proposed License Plate Detection and Recog-
nition in Unconstrained ScenariosIn the face of challenging real-world scenarios
marked by erratic lighting and occlusions, this paper’s integration of edge detec-
tion and neural networks presents a formidable solution. This dynamic combi-
nation effectively addresses the complexity of identifying license plates across
varying environmental conditions [3].
Layan Hewa Godage et al. (2019) proposed Real-time Vehicle License Plate
Recognition for CCTV Applications.With the imperative of real-time surveil-
lance applications, this study introduces an end-to-end system. Comprising
image preprocessing, connected component analysis, and a probabilistic neural
network, the system demonstrates prowess in swiftly and accurately recognizing
license plates in live CCTV feeds [4].
Prashengit Dhar et al. (2022) proposed License Plate Recognition using Edge
Detection and Neural Networks discuss how we can Skillfully navigating the
intricacies of license plate recognition, this paper orchestrates an approach that
amalgamates edge detection and neural networks. This holistic strategy, supple-
mented by rigorous preprocessing, yields enhanced accuracy and reliability in
deciphering license plates [5].
Nivethika et al. demonstrate a four-wheeled robot that follows people and
has an ultrasonic sensor to avoid collisions. The robot locates the user using
ultrasonic sensor feedback. This robot can help humans carry various accessories,
luggage, and highly sensitive chemicals in a variety of applications. An algorithm
was developed using ultrasonic sensors to track and position a human user while
keeping a constant distance from the robot. The IR sensors are used to uniquely
identify the user in the presence of numerous such obstacles [6].
K. Yilmaz et al. (2011) proposed A Hybrid Approach to License Plate Recog-
nition. Acknowledging the multifaceted nature of license plate recognition, this
paper pioneers a hybrid methodology. This synergy between color-based seg-
mentation and Convolutional Neural Networks (CNNs) showcases the potential
of combining classical computer vision techniques with modern deep learning
advancements [7].
M.R.C.Niluckshini; et al. (2020) License Plate Detection and Recognition
in Uncontrolled Environments “Tackling the intricate task of recognizing license
plates in dynamic and uncontrolled environments, this paper introduces a robust
two-stage approach. By deftly addressing varying lighting, angles, and occlu-
sions, this method exhibits adaptability to the unpredictability of real-world
settings [8].
 Smart Entry Management Using IoT 153

Wen-bin GONG et al. (2013) License Plate Recognition using Haar Cascade
ClassifiersCapitalizing on the strength of Haar-like features and cascade classi-
fiers, this study lays a robust foundation for accurate license plate recognition.
Its sequential classification approach empowers the system to progressively refine
license plate detection with each subsequent stage [9].
Bini Omman et al. (2017) proposed Efficient Car License Plate Recogni-
tion with Support Vector Machines”. Tailored for efficiency, this study embraces
Support Vector Machines (SVMs) to unravel license plate recognition intrica-
cies. Melding feature extraction and classification, this system resonates with
accuracy while demonstrating the efficacy of machine learning techniques [10].
Layan Hewa Godage et al. (2016) proposed “Automatic Number Plate Recog-
nition using Genetic Algorithm and Neural Network. Navigating the intrica-
cies of optimization, this paper pairs genetic algorithms with neural networks.
By meticulously fine-tuning neural network parameters via genetic algorithm
optimization, the system attains enhanced accuracy in automatic number plate
recognition [11].
Nitin Rakesh et al. (2020) proposed “End-to-End Trainable Vehicle License
Plate Recognition via Convolutional Neural Networks”. Simplifying the license
plate recognition process, this paper’s end-to-end trainable system seamlessly
integrates Convolutional Neural Networks (CNNs). In a single holistic frame-
work, license plates are not only detected but recognized, accentuating the effi-
ciency of unified learning [12].
R. Scherer et al. (2020) proposed “Deep Learning Approaches for Number
Plate Detection” This paper explores deep learning techniques for number plate
detection. The authors propose a convolutional neural network (CNN) model
trained on annotated datasets to achieve high accuracy and robustness in detect-
ing number plates [13].
Shivam Singh et al. (2016) proposed Number Plate Detection in Low-Light
Conditions.This paper addresses the challenges of number plate detection in low-
light conditions. The authors propose an adaptive thresholding algorithm that
effectively handles variations in lighting conditions, ensuring accurate detection
even in challenging environments [14].
M V Raghunadh et al. (2018) proposed ”Machine Learning Techniques for
Number Plate Detection. This paper explores the application of machine learning
techniques for number plate detection. The authors compare and analyse various
algorithms, including Support Vector Machines (SVM), Random Forests, and
Gradient Boosting, to identify the most effective approach [15].
The latest trends in IoT-based entry management and security systems high-
light significant technological strides. One key trend involves integrating edge
computing for real-time processing, enhancing system responsiveness and overall
performance. Another notable focus is on leveraging machine learning, especially
deep learning algorithms, to elevate the accuracy of number plate recognition
across varying environmental conditions.
154 L. Lakhwani et al.

The exploration of blockchain technology is a compelling trend in IoT-based

entry management systems, enhancing security through a tamper-proof record of
access events. This reinforces the system’s resilience against cyber threats, pro-
viding a robust foundation for secure access control. Additionally, researchers
are actively exploring the fusion of multiple biometric modalities, such as facial
recognition with number plate recognition, marking a broader shift towards com-
prehensive and multifaceted security solutions in IoT-based entry management.

3 Methodology

Within this section, we meticulously detail the hardware specifications integral to

the execution of the study. Comprehensive insights into key components such as
processing units, memory capacity, storage solutions, and additional peripherals
are provided, aiming to ensure transparency and facilitate the replicability of
the research outcomes.

3.1 Esp 32 Cam (Module) Setup

– Utilize NodeMCU-32S microcontroller with integrated Wi-Fi capabilities.

– Connect a compatible camera module (e.g., OV2640) to the NodeMCU-32S
to enable image capture functionality.

3.2 MD Board

– The ‘md board’ is essential as it facilitates the transfer of frames captured

by the NodeMCU camera module. It accomplishes this by transmitting the
frames through the WiFi module.

3.3 Frame Captured/Transferred

– The frames are received on the server side via the WiFi network.

3.4 Image Processing

– Upon reception of the image on the server side, we apply image processing
algorithms.
– Additionally, machine learning models are utilized to prepare the image for
the subsequent phase of the program.
 Smart Entry Management Using IoT 155

3.5 Number Plate Extraction

– The processed image undergoes edge detection to isolate and extract the
number plate.
– Extraneous details are eliminated from the image during this process.

3.6 Optical Character Recognition

– The OCR library in Python is employed to extract text from the number
plate on the processed image.
– Subsequently, the interpreted text is obtained, revealing the information writ-
ten on the number plate.

3.7 Checking Vehicle

– After extracting the information from the number plate, we proceed to search
our MySQL database.
– The search is conducted to determine if the identified number plate is regis-
tered in our database.

3.8 Vehicle Verifies

– If the vehicle is registered, relevant data is fetched from the database.

– The retrieved information is then displayed to the guard, who can assess and
decide whether to permit entry for the vehicle (Fig. 1).
156 L. Lakhwani et al.

3.9 Flow Chart

Fig. 1. Flow Chart Of Working Model

 Smart Entry Management Using IoT 157

4 Algorithms
4.1 Vehicle Detection Algorithm

This algorithm detects the presence of a vehicle at the gate using IoT sensors
like infrared or ultrasonic sensors.
Function Detectvehicle():
........Repeat:
......Read sensor data
....Until vehicle is detected

4.2 Number Plate Recognition Algorithm

This algorithm processes the image of the vehicle’s number plate and extracts
the alphanumeric characters for recognition.
Function RecognizeNumberPlate(image):
....Preprocess image (e.g., resize, apply filters)
....Apply Optical Character Recognition (OCR) to extract characters
....Return recognized number plate

4.3 Access Control Algorithm

This algorithm checks if the recognized number plate is authorized to access the
premises.
Function AccessControl(numberPlate):
....if numberPlate is in Authorized Vehicles Database:
........Open gate
........Log entry in access control system
....else:
........Display “Access Denied” message

4.4 Database Query Algorithm

This algorithm queries the database of authorized vehicles to check if a given

number plate is authorized.
Function QueryDatabase(numberPlate):
....Connect to the authorized vehicles database
....Execute SQL query to check if numberPlate is present
....Return True if found, False otherwise

5 Results
In this section, we conducted extensive testing of our model across diverse
parking environments. Figure 2 elucidates the model’s proficiency in accurately
detecting the quantity of objects within the given context. To provide a granular
158 L. Lakhwani et al.

Fig. 2. Object Count By Image

Fig. 3. Heat Map Annoation Taken From Model

understanding of the detection dynamics, Fig. 3 employs a heatmap demonstra-

tion, visually mapping the specific areas within the photo where the detections
are concentrated.
For a comprehensive showcase of real-time capabilities, Fig. 4 offers a live
demonstration, providing insights into the intricate process of interpreting vehi-
cle images. Furthermore, Fig. 5 delves into the technical aspect by presenting
the code terminal, offering a transparent view of the decoding process employed
for vehicle number plates. These visual representations collectively contribute
to a nuanced analysis of the model’s performance and its interpretative prowess
across various facets of license plate detection.
 Smart Entry Management Using IoT 159

Fig. 4. Detection Of Vehicle Number Plate

Fig. 5. OCR Decoding the Number Plate

In this Fig. 6 have shown the User Interface of the project on the client side
you can see that the live camera feed is seen here and there are two options to
check the vehicle details and to allow the visitor vehicle. If user Clicks on check
vehicle the Information is displayed as shown.
In this Fig. 7 have Shown if the guard is satisfied the information then there
is further Change in which the guard can allow the vehicle or declines it. As
you can see the pop up is there to allow the vehicle or not 0 or entry and 1 for
decline it.
160 L. Lakhwani et al.

Fig. 6. Displaying Details of Vehicle

Fig. 7. Option to enter the vehicle

In this Fig. 8 have showed the of the project as you can see in this image
that i have shown the admin interface in this i have shown how the registered
vehicle is stored
 Smart Entry Management Using IoT 161

Fig. 8. Admin Portal To Show Data of Registered Vehicle

In this Fig. 9 i have shown the how the details of entry and exit are there
and the data is stored by Vehicle Number and Time. so this is how the data is
stored and admin can access all the information. through django admin Inteface.

Fig. 9. Admin Portal To Show Data of Entry/Exit Vehicle

6 Conclusion
In the pursuit of enhanced access control and security management, our project
has explored the synergistic integration of Internet of Things (IoT) technology
and Number Plate Recognition (NPR) systems, resulting in the development of
a Smart Entry Management system. This system stands as a testament to the
potential for innovation in addressing the contemporary challenges of secure and
162 L. Lakhwani et al.

convenient entry control. Through our paper, we have achieved several significant
milestones:
Efficient Entry Control, Accuracy in Number Plate Recognition, Database
Integration, Access Logging,

Future Scope. The license plate detection project’s future holds exciting
prospects. Advancements in AI algorithms, deeper IoT integration, and pre-
dictive analytics are in the pipeline. Cloud-based solutions, a dedicated mobile
app, and adherence to international standards will enhance scalability and acces-
sibility. These developments position the project as an evolving solution, adept
at addressing emerging challenges in security and access management.

References
1. Lee, S., Son, K., Kim, H., Park, J.: Car plate recognition based on CNN using
embedded system with GPU, pp. 239-241 (2017). https://doi.org/10.1109/HSI.
2017.8005037
2. El Khatib, I., Sweidan, Y., Omar, S.M., Al Ghouwayel, A.: An efficient algorithm
for automatic recognition of the Lebanese car license plate, pp. 185–189 (2015).
https://doi.org/10.1109/TAEECE.2015.7113624
3. Montazzolli, S., Jung, C.: License plate detection and recognition in unconstrained
scenarios (2018)
4. Godage, L., Wimalaratne, P.: Real-time mobile vehicle license plates recognition
in Sri Lankan conditions, pp. 146–150 (2019). https://doi.org/10.1109/ICIIS47346.
2019.9063258
5. Dhar, P., Guha, S., Biswas, T., Abedin, Z.: A system design for license plate recog-
nition by using edge detection and convolution neural network (2018). https://doi.
org/10.1109/IC4ME2.2018.8465630
6. Shashidhar, R., Manjunath, A.S., Kumar, R., Roopa, M., Puneeth, S.B.: Vehicle
number plate detection and recognition using YOLO- V3 and OCR method, pp.
1–5 (2021). https://doi.org/10.1109/ICMNWC52512.2021.9688407
7. Sharma, J., Mishra, A., Saxena, K., Kumar, S.: A hybrid technique for license plate
recognition based on feature selection of wavelet transform and artificial neural
network. In: ICROIT 2014 - Proceedings of the 2014 International Conference on
Reliability, Optimization and Information Technology, pp. 347–352 (2014). https://
doi.org/10.1109/ICROIT.2014.6798352
8. Junior, W., Junior, L.C.S., Colares, G., Linhares, J., Jesus, A., Pinagé, F.: License
plate detection system in uncontrolled environments using super resolution and
a simple cell phone camera attached to a bicycle seat (2022). https://doi.org/10.
14209/sbrt.2022.1570812767
9. Dong, Z.H., Feng, X.: Research on license plate recognition algorithm based on
support vector machine. J. Multimed. 9 (2014). https://doi.org/10.4304/jmm.9.2.
253-260
10. Raghunadh, M.V., et al.: Proposed Machine Learning Techniques for Number Plate
Detection. This paper explores the application of machine learning techniques for
number plate detection. The authors compare and analyse various algorithms,
including Support Vector Machines (SVM), Random Forests, and Gradient Boost-
ing, to identify the most effective approach (2018)
 Smart Entry Management Using IoT 163

11. Godage, L.H., Wimalaratne, G.D.S.P.: Real-time mobile vehicle license plates
recognition in Sri Lankan conditions. In: Safran, M., Haar, S., (eds.) 2019 14th
Conference on Industrial and Information Systems (ICIIS), Kandy, Sri Lanka, pp.
146–150 (2019). https://doi.org/10.1109/ICIIS47346.2019.9063258. 2007. Arduino
and Android Powered Object Tracking Robot. Southern Illinois University Car-
bondale
12. Abolghasemi, V., Ahmadyfard, A.: An edge-based color-aided method for license
plate detection. Image Vis. Comput. 27, 1134–1142 (2009). https://doi.org/10.
1016/j.imavis.2008.10.012
13. Soroori, S., Tourani, A., Shahbahrami, A.: Employing deep learning approaches
for automatic license plate recognition: a review (2019)
14. Shashirangana, J., Padmasiri, H., Meedeniya, D., Perera, C.: Automated license
plate recognition: a survey on methods and techniques. IEEE Access. 9, 11203–
11225 (2020). https://doi.org/10.1109/ACCESS.2020.3047929
15. Agrawal, R., Agarwal, M., Krishnamurthi, R.: Cognitive number plate recogni-
tion using machine learning and data visualization techniques, pp. 101-107 (2020).
https://doi.org/10.1109/ICSC48311.2020.9182744
 Enhancing Player Engagement Through
Gesture-Based Interactions in Tic-Tac-Toe

Vedansh Rathi(B) , Shrey Sharma , Anmol Rai , and Sandeep Kumar Sharma

School of Computer and Communication Engineering, Manipal University Jaipur, Jaipur,

Rajasthan, India
[email protected], [email protected],
[email protected], [email protected]

Abstract. A popular paper-and-pencil game for two players, tic tac toe involves
strategically placing Xs and Os on a 3×3 grid. This paper provides a gesture-
driven version of the classic mouse-click Tic-Tac-Toe game that uses computer
vision for immersive gameplay. Enables natural interactions with robust real-
time hand tracking and gesture classification. The placement of markers in a
game is matched to gestures like pinching and swiping. The system combines
graphical rendering, intelligent Minimax bots, and traditional game logic with
gesture recognition. Comprehensive algorithms are offered, encompassing 3000
words on design, modules, and pipeline. With approximately 80% accuracy in
gesture detection during gameplay, testing on 15 users confirmed precise tracking
and categorization. In this research, an intuitive gesture interface for the classic
game of Tic-Tac-Toe is used to demonstrate increased player involvement.

Keywords: computer vision · gesture recognition · human-computer

interaction · immersive gameplay · Tic-tac-toe

1 Introduction

Tic-Tac-Toe is a common two-player pen-and-paper game that is played by all. In this

straightforward yet fascinating game, players take turns placing their markers—X or
O—on a 3×3 grid to complete a row, column, or diagonal with their sign before the
opposition.
Players in the original version of Tic-Tac-Toe can only interact with the game using
keyboard keys or mouse clicks. This may reduce one’s intuitive involvement and sense of
total immersion. This article explores the integration of computer vision algorithms for
real-time hand gesture recognition, enabling more responsive and natural control. Instead
of utilizing mouse, keyboards, or other electronic devices, users can connect through
vision-based gesture interfaces. This produces experiences that are more captivating and
immersive, particularly in the gaming industry [1].
The paper documents a novel gesture-controlled Tic-Tac-Toe game with:
– Real-time hand tracking using computer vision to robustly recognize gestures.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 164–173, 2025.
https://doi.org/10.1007/978-3-031-71481-8_13
 Enhancing Player Engagement 165

– Mapping intuitive hand motions to appropriate game control actions like placing
markers, changing turns etc.
– Integrating the gesture interface with standard Tic-Tac-Toe game logic.
– Inclusion of AI bots using Minimax algorithm to allow single player mode against
computer opponents.
– Developing an interactive graphical user interface.
With more intuitive gesture-based control, the suggested system seeks to improve
player engagement in the widely played game Tic-Tac-Toe by fusing cutting-edge com-
puter vision and human-computer interaction techniques with tried-and-true gaming
concepts.

2 Literature Review

Many methods for precise real-time computer vision-based gesture recognition have
been actively investigated by researchers. Also investigated are game techniques regard-
ing the best way to play Tic Tac Toe. Gesture-based gaming experiences have been the
subject of some early research.

2.1 Vision-Based Gesture Recognition

Reactive controls in human-computer interaction depend on the development of strong

real-time gesture recognition systems. The use of cameras in vision-based interfaces
has grown in popularity. To identify motions based on patterns of skin deformation
rather than precise coordinates, Wang et al. [2] introduced a flexible wrist-worn device
incorporating pressure sensors. A Triplet Network classifier was used in this wristband,
which demonstrated robustness over several days with an accuracy of 91.98% in iden-
tifying motions. Robot manipulation tasks can be planned with optimal pick and place
sequences by combining Monte Carlo tree search and motion feasibility verification, as
suggested by Eljuri et al. [3]. In intricate modeling scenarios such as warehouse sort-
ing, our method outperformed baselines. For more precise recognition of dynamic hand
gestures, Jiashan and Li [4] developed a combination of global body motions and fine-
grained finger motions. Combining these complementary data enhanced the results on
intricate gesture datasets.
Michiel van de Steeg et al. (2015) conducted a study titled “Temporal Difference
Learning for the 3D Tic-Tac-Toe Game: Applying Structure to Neural Networks”.
Researchers have investigated the use of multi-layer perceptrons (MLP) as functional
approximations for learning to play a complex game, Tic-Tac-Toe 3D. They introduced
structured MLP and custom system design to improve learning efficiency. By prioritiz-
ing pattern detection and clustering, structured MLPs have shown significant improve-
ments. Comparison analysis against established adversaries revealed that the MLP built
with an integrated pattern recognition layer outperformed other MLP structures in self-
training and comparison with the adversary. This study demonstrates the effectiveness
of structured neural networks in optimizing the learning process for complex games. [5].
166 V. Rathi et al.

2.2 Tic-Tac-Toe Game Strategies

To create optimal game-playing agents for the traditional game of Tic-Tac-Toe and its
many variations, academics have been actively working on algorithms and heuristics
for decades. To create novel offensive and defensive strategies, Leaw and Cheong [6]
investigated a quantum version of Tic-Tac-Toe that permitted move superposition. In this
version, player 2 might start in any state and force a win or tie. To improve the efficiency
of exploring game trees, Jatoi et al. [7] used Alpha-Beta pruning optimization to optimize
the Minimax algorithm by removing branches that aren’t necessary for making decisions.
Using deep neural networks, Sridhar and Shetty [8] were able to classify the results of
Tic-Tac-Toe games with 98.9% accuracy in terms of wins and losses. This showed
promise for picking up gaming tactics. Muhammad H. Harry et al. (2019) present Tic
Tac Toe Math, a serious game-based learning tool aimed at improving elementary school
students’ understanding of basic arithmetic concepts. The game includes some additional
operations and uses the eye tracker as an optional controller. Research shows positive
student responses to the fun factor and the effectiveness of games as a learning tool. The
effectiveness of the game is measured through a questionnaire, showing positive results.
Overall, serious games provide an opportunity to learn mathematics effectively in a fun
and interactive way [9].

2.3 Vision-Based Gesture Gaming

In addition, some earlier research has investigated the use of gestures as more engag-
ing game controls than standard keys and keyboards. In SwarmPlay, which was first
described by Karmanova et al. [10], several drones used computer vision for motion
planning and tracking to play Tic Tac Toe against a human. This showcased drone-based
cooperative gesture-based gaming. For the NAO humanoid robot to play Tic-Tac-Toe
physically, Nugroho et al. [11] created kinematic models that included optimizing trajec-
tory planning to manipulate game pieces. This made it possible to engage with gestures
in real life. For visually impaired children, Drossos et al. [12] developed audio-based
games that used natural gestures as inputs, such as hand claps and finger snaps. J Jim
et al. (2021) introduced SpecMCTS, a novel approach to accelerate the Monte Carlo Tree
Search (MCTS) algorithm, widely used in decision-making problems like the game of
Go. It employs two deep neural network models: a quick speculation model and a precise
main model. The speculation model offers rapid approximations of node evaluations to
guide tree traversal, while the main model ensures accuracy. This technique achieves up
to 2.07× acceleration on NVIDIA T4 GPUs, surpassing previous methods and enhanc-
ing decision quality and gameplay win rates. It effectively utilizes GPU parallelism by
concurrently executing multiple inference tasks. Future research could explore diverse
model compression techniques and the application of SpecMCTS in different domains
[13]. R. Scherer et al. (2016) investigated the use of a brain-computer interface (BCI)
to improve interaction and control in people with cerebral palsy. The researchers intro-
duced a game-based learning approach to improve BCI training and control. A user
with cerebral palsy successfully operated a mind-controlled communication board and
puzzle game using motor patterns detected through EEG patterns. Additionally, users
play the classic Tic-Tac-Toe game against tutors using BCI input. The study highlights
 Enhancing Player Engagement 167

the potential of adding competition as a motivating factor to increase user involvement

in BCI training. The results show that a game-based approach and social interaction
can contribute to the improvement of BCI control for people with motor disabilities.
However, the study included a single case and more research is needed to generalize
these results [14].

3 Proposed Methodology

The proposed gesture-based Tic Tac Toe game system consists of the following key
modules:
Gesture Recognition Module. The Gesture Recognition Module is an essential com-
ponent of the proposed system that links the player’s hand gestures to in-game activities.
This lesson records and analyzes live webcam footage using cutting-edge deep learn-
ing and computer vision techniques. For accurate hand landmark detection, the system
makes use of the MediaPipe framework. Important hand skeletons points, like the palm
and fingertips, correspond with landmarks. The module’s primary objective is to track
the position and motion of these landmarks, which aids in the identification of specific
movements like finger spreads, and left and right movement of cursor. A crucial compo-
nent of the game is real-time hand gesture recognition, which lets players engage with
the game board by making logical, organic movements.
Gesture Mapping. The Tic Tac Toe framework relies heavily on the Gesture Mapping
module to translate hand movements that are recognized into meaningful game actions.
Each recognized motion is associated with a specific in-game action through the use
of a lookup database. For instance, the game board’s “X” and “O” symbols might be
associated with swipes to the left and right, respectively. Moreover, interactive in-game
features like game resets can be triggered by making the required movements. With this
mapping system, the player’s hand movements are seamlessly integrated with the game’s
activities and rules, making for a more immersive and captivating gaming experience.
Game Logic. The core of the Tic Tac Toe game is the Game Logic module. The basic
game mechanics are included, and it is typically implemented in Python. Key functions
include updating the game board’s status, verifying the validity of player actions, search-
ing for winning standards, and determining if a draw occurs. In the end, this module will
control the flow of the game by supervising player turns and ensuring that moves are
made in compliance with the rules. The Gesture Mapping module integrates with the
Game Logic to enable seamless execution of game moves based on recognized gestures.
It makes a harmonious fusion of gesture control and conventional game mechanics by
guaranteeing that the game state accurately reflects the player’s actions.
AI Bot. Users can now challenge themselves in a single-player format against a crafty
computer opponent with the AI Bot module. To decide on the best move during a game,
the AI Bot employs the Minimax algorithm combined with alpha-beta pruning. Through
game scenario modeling and probabilistic play analysis, the AI seeks to maximize its
chances of winning or, if a winning opportunity is not present, to force a draw against
the player. With this module, you can play a variety of gaming scenarios, from casual
168 V. Rathi et al.

two-player games to intense matches against AI opponents, all while having a strong
opponent for lone players.
User Interface. The User Interface module is responsible for creating an aesthetically
pleasing and user-friendly environment for the game. It utilizes libraries like Pygame to
render images, display the Tic Tac Toe board, convey the status of the game, and instruct
the user with text prompts for required gestures. The graphical user interface not only
makes the game more engaging and interactive, but it also enhances its visual appeal
and user experience. Users can interact with the game through an intuitive interface that
provides them with visible feedback on their actions and the state of the game. The
game is more accessible and enjoyable because of its user-friendly design, which allows
players of different skill levels to participate and have fun overall.
With the help of these modules, players may now play the beloved game of Tic Tac
Toe with the addition of AI opponents and hand gesture controls, making it an immersive
and engaging experience. Dynamic gaming is a result of the integration of deep learning,
computer vision, game logic, and an intuitive interface. The gesture-based Tic-Tac-Toe
system architecture is displayed in Fig. 1 which was proposed by Drossos et. al.

Fig. 1. The architecture of the developed Tic-Tac-Toe game [10].

MediaPipe is used for gesture recognition, which entails obtaining video frames,
identifying hand landmarks, preprocessing coordinates by smoothing and normalizing
them, extracting features like angles and distances, and feeding the features to an ML
model for gesture classification. The gestures are then mapped to in-game actions. The
AI Bot uses the Minimax Algorithm with Alpha- Beta Pruning to decide which games to
play optimally. Recursively investigating potential game states, it maximizes its chances
in a depth-first fashion while eliminating subtrees that will not influence the outcome,
guaranteeing effective AI gameplay. The Euclidean distance formula can be used to
determine the separation between two hand landmarks (points A and B). It calculates
the distance in a straight line between these two locations.

distance_ab = ((xa − xb)2 + (ya − yb)2 (1)
With the arctangent function (tan- 1), the angle between three hand landmarks (points
A, B, and C) can be determined. Concerning the horizontal axis, this angle indicates the
 Enhancing Player Engagement 169

direction from point A to point B.

angle_abc = tan−1 ((yb − ya)/(xb − xa)) (2)

‘xa’ and ‘ya’ are the x and y coordinates of landmark A, respectively.

‘xb’ and ‘yb’ are the x and y coordinates of landmark B, respectively.

Using a moving average filter, one can smooth a series of data points (such as hand
landmark coordinates). This filter essentially averages the data over a given window size
by calculating a new filtered output at each time step based on the previous output and
the current input. ‘yt’ represents the filtered output at time t.

yt = (yt − 1 + xt)/N (3)

• ‘yt-1’ is the filtered output at the previous time step (time t-1).
• ‘xt’ is the current input data (landmark coordinate) at time t.
• ‘N’ is the filter window size, which determines the number of data points to be
considered for averaging.

Motion tracking and gesture recognition are just two of the many applications that
can benefit from techniques like these equations and the filter for extracting features
from hand landmarks and smoothing the data.

4 Implementation and Evaluation

4.1 Implementation Details

A Python 3.8 implementation of the integrated game system was made. MediaPipe 0.8
was used for hand tracking, OpenCV was used for vision processing, and PyGame 2.0
was used for user interface rendering.
Gesture Recognition. 15 frames per second (1080p webcam input) was the optimal
speed for the hand tracking and gesture classification pipeline. 21 hand key points could
be reliably located using the MediaPipe hand landmark detector. Distortion features
were calculated over a 10-frame window by the trajectory analyzer. To classify gestures
such as pinch, swipe left/right, and so on, an SVM classifier was used.
Gesture Mapping. Gestures were defined as follows. Moving the index finger – the
movement of the cursor. Index finger and thumb to be brought together – clicking.
This vocabulary was comprehensive to control all game events through intuitive natural
gestures.
Game Logic. The Tic-Tac-Toe game logic was implemented as follows: the 2D list to
store board state with cell values X, O, or EMPTY. Player class to track name, marker
type, and score. The game class handles board state, current player, and valid moves.
Game flow control with functions like checkWinner(), and checkDraw(). Win conditions
were evaluated by iterating through rows, columns, and diagonals for matching markers.
170 V. Rathi et al.

AI Bot. For the AI player, a Minimax algorithm with alpha-beta pruning was used,
with a maximum depth of five moves. Heuristic payoff tables were created to penalize
opponent victories and reward winning lines. The bot could now play competitively
thanks to this.
UI Rendering. PyGame window of 600x600px to display playable area. Grid lines
to demarcate board cells. Marker sprites (X, O images) rendered on gesture actions.
Text overlay for guiding gestures. This modular implementation allowed robust hand
tracking, gesture control and traditional Tic-Tac-Toe logic to be seamlessly integrated.

4.2 Experimental Evaluation

Clicking and controlling the mouse cursor movements were assessed for the real-time
hand gesture recognition module. In order to test the integrated system, users were
required to complete common desktop tasks using only hand gestures.
Real-Time Cursor Control. The gesture classifier and hand landmark detector were
tuned for a 15 frames per second frame rate. The gesture vocabulary included swiping left,
right, up, and down to move the cursor as well as bringing the thumb and finger together
to click. Using swipe gestures, ten users were instructed to move the cursor between
randomly selected on-screen targets. The mean precision in achieving the objective was

Fig. 2. Box plot of user-wise cursor control accuracy

 Enhancing Player Engagement 171

determined to be 80.06% among the participants. Figure 2 displays the accuracy of user-
controlled cursor movements. The errors were mainly due to inaccuracies in mapping
subtle hand motions to precise cursor displacements. Dynamic scaling factors need to
be incorporated based on gesture velocities and screen distances.

Clicking Gesture Performances. For the clicking gesture, the users were asked to bring
their thumb and index finger closer to click on a random target. The system would invoke
a mouse click event on detecting the gesture. Table 1 represents the results obtained by
the users. In this case, the system correctly generated the click event 92.5% of the time
when the user made the gesture, or an average true positive rate or sensitivity of the
clicking gesture of 92.5%.
But the specificity was lower at 86.5%, suggesting a higher false positive rate where,
occasionally, clicks were mistakenly triggered even in the absence of the gesture.

Table 1. Confusion matrix for clicking gesture recognition

Actual/Predicted Click No click

Click 185 15
No click 27 173

Gameplay Testing. A basic tapping game was used to test the integrated system. Users

Fig. 3. Gameplay Testing Metrics

172 V. Rathi et al.

had to tap randomly appearing targets on the screen by moving the cursor over them and
then making a fist to click.
Each of the game’s 5 rounds was played by 15 users. Users successfully tapped
77.6% of the on-screen targets on average. A 4.04 out of 5 was the average game score.
The metrics for gameplay are compiled in Fig. 3.
To sum up, the experimental findings show dependable real-time tracking and ges-
ture identification that facilitate cursor-based interactions. However, robust mapping of
gestures to cursor motions and multi-modal click confirmations are required to improve
accuracy.

5 Conclusion and Future Scope

This paper presented a novel gesture-driven method different from traditional mouse
input—for playing the classic game of Tic Tac Toe by sensing hand articulations in
real time through a webcam. With the help of Media Pipe’s effective hand landmark
tracking and machine learning classification, a precise and responsive gesture interface
was possible. This was combined with well-established game logic and graphic rendering
to produce a captivating gaming environment.
While the experimental evaluation of 15 users confirmed the potential of gesture
control, it also pointed to improvements that needed to be made to improve orienta-
tion invariance and lessen fatigue, such as multi-modal input and adaptive recognition.
Adding more complex articulations to the gesture vocabulary can enhance interaction
even more. Furthermore, unexplored is the possibility of two-player online multiplayer
games.
This paper offers important insights into using gesture interfaces to improve human-
computer engagement in interactive applications, such as classic games like Tic-Tac-
Toe, by fusing computer vision with gaming principles. The suggested methods may
contribute to the development of future interactions that are more intuitive and immersive.

References
1. Selvam, P., Deena, G., Hemalatha, D., Aruna, K.B., Hashini, S.: Gaming using different hand
gestures using artificial neural network. EAI Endorsed Trans. Internet Things (2024). https://
doi.org/10.4108/eetiot.5169
2. Wang, T., Zhao, Y., Wang, Q.: Hand gesture recognition with flexible capacitive wristband
using triplet network in inter-day applications. IEEE Trans. Neural Syst. Rehabil. Eng. 30,
2876–2885 (2022). https://doi.org/10.1109/TNSRE.2022.3212705
3. Eljuri, P.M.U., Ricardez, G.A.G., Koganti, N., Takamatsu, J., Ogasawara, T.: Combining a
Monte Carlo tree search and a feasibility database to plan and execute rearranging tasks. IEEE
Access 9, 21721–21734 (2021). https://doi.org/10.1109/ACCESS.2021.3055455
4. Jiashan, L.I., Zhonghua, L.I.: Dynamic gesture recognition algorithm Combining Global
Gesture Motion and Local Finger Motion for interactive teaching. IEEE Access, p. 1 (2021).
https://doi.org/10.1109/ACCESS.2021.3065849
5. Steeg, M., Drugan, M., Wiering, M.: Temporal difference learning for the game tic-tac-toe
3D: Applying structure to neural networks (2015). https://doi.org/10.1109/SSCI.2015.89
 Enhancing Player Engagement 173

6. Leaw, J.N., Cheong, S.A.: Strategic insights from playing quantum tic-tac-toe. J. Phys. A
Math. and Theor. 43(45), 455304 (2010). https://doi.org/10.1088/1751-8113/43/45/455304
7. Jatoi, A., Ali, S., Siddiqui, I.: An intuitive implementation of alpha-beta pruning using tic-
tac-toe (2020)
8. Sridhar, S., Shetty, S.: Classification of tic-tac-toe game using deep neural network (2021).
https://doi.org/10.32628/IJSRSET219353
9. Garry, M., Yamasari, Y., Nugroho, S., Purnomo, M.: Design and implementation serious game
“tic tac toe math” (2019)
10. Karmanova, E., Serpiva, V., Perminov, S., Ibrahimov, R., Fedoseev, A., Tsetserukou, D.:
SwarmPlay: a swarm of nano-quadcopters playing tic-tac-toe board game against a human.
In: ACM SIGGRAPH 2021 Emerging Technologies (SIGGRAPH ‘21). Association for Com-
puting Machinery, New York, NY, USA, Article 5, pp. 1–4 (2021). https://doi.org/10.1145/
3450550.3465346
11. Nugroho, S., Prihatmanto, A., Syaichu-Rohman, A.: Design and implementation of kine-
matics model and trajectory planning for NAO humanoid robot in a tic-tac-toe board game.
In: Proceedings of the 2014 IEEE 4th International Conference on System Engineering and
Technology, CSET 2014 (2015). https://doi.org/10.1109/ICSEngT.2014.7111783
12. Drossos, K., Zormpas, N., Giannakopoulos, G., Floros, A.: Accessible games for blind
children. Empow. Binaural Sound 10(1145/2769493), 2769546 (2015)
13. Kim, J., Kang, B., Cho, H.: SpecMCTS: accelerating Monte Carlo tree search using spec-
ulative tree traversal. IEEE Access 9, 142195–142205 (2021). https://doi.org/10.1109/ACC
ESS.2021.3120384
14. Scherer, R., Schwarz, A., Müller-Putz, G.R., Pammer-Schindler, V., Garcia, M.L.: Lets play
Tic-Tac-Toe: a Brain-Computer Interface case study in cerebral palsy. In: IEEE International
Conference on Systems, Man, and Cybernetics (SMC), Budapest, Hungary (2016)
 Aggressive Bangla Text Detection Using
Machine Learning and Deep Learning
Algorithms

Tanjela Rahman Rosni, Mahamudul Hasan(B) , Tanni Mittra,

Md. Sawkat Ali, and Md. Hasanul Ferdaus

East West University, A/2 Jahurul Islam Avenue, Dhaka 1212, Bangladesh
[email protected], {tanni,alim,hasanul.ferdaus}@ewubd.edu

Abstract. As digitization continues to proliferate globally, individuals

prefer expressing themselves on social media platforms using their native
languages. On these platforms, people share their interests, thoughts, and
rights, sometimes receiving appreciation for their views and, at times,
encountering conflicts of interest leading to mean comments, includ-
ing hate speech and aggression towards individuals, societies, or groups.
Detecting such comments has become crucial to curbing further abuse.
This research paper focuses on identifying aggressive and non-aggressive
Bengali text in social media posts and comments through the applica-
tion of machine learning and deep learning algorithms. The dataset was
collected from various social media platforms like Facebook, YouTube,
and Twitter. Employing diverse machine learning and deep learning algo-
rithms, such as SVM, Random Forest, KNN, Linear Regression, Decision
Tree, and CNN, the authors achieved the highest accuracy of 90.46% with
Multinomial Naive Bayes (MNB).

Keywords: Machine Learning Algorithm · Deep Learning Algorithm ·

SVM · Random Forest · KNN · Linear Regression · Decision Tree ·
CNN

1 Introduction
Natural Language Processing (NLP) is an extensive area of machine learning
and artificial intelligence where we get to know how a computer or a machine
can understand and learn the natural human language. This is a vast research
domain where human language or speech is transformed into machine-readable
or trainable. Human language is very complex. There are around 7,000 languages
in the world. Each country of the world uses different languages which the inhab-
itants of those countries speak. Every language has a lot of words. Each of those
words has different contextual meanings to it in different situations. The Bangla
language is also a prime example of that. So, the authors decided to explore
this field of NLP. The Bangla language is one of the most used languages in
the world around 210 million people use it as their first or second language [1].
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 174–183, 2025.
https://doi.org/10.1007/978-3-031-71481-8_14
 Aggressive Bangla Text Detection 175

Which makes it one of the most popular and used languages in the world. Most
people of Bangladesh and the people of West Bengal use Bangla as their first
language. As technology is growing very rapidly people are addicted to phones,
computers, and laptops spending the majority of their time browsing and surfing
through different social media platforms. Facebook is one of the most famous
and fastest-growing social media platforms in the world. Almost 3 billion peo-
ple use Facebook [9]. The user posts, shares, and comments on various social
media platforms. And most of these users are within the 18–39 years of age
group [10]. According to a report people spend about on average 2 h and 29 min
daily on social media platforms [11]. In the case of Bangladesh, according to
GSMA Intelligence, there were 178.5 million people who somewhat use cellu-
lar mobile or cellular internet connections as of 2022. Data published in Meta’s
advertising resources shows that Facebook had 44.70 million users, YouTube
had 34.50 million users, Instagram had 4.45 million users, and Twitter had 756.6
thousand users in Bangladesh as of early 2022 and growing [12]. So with this
huge number of users, there are some drawbacks. The number of cyberbully-
ing, harassment, abuse, and aggressive comments are also increasing at a rapid
rate. These affect the mental health of a person. A victim of cyber abuse or
cyberbullying faces Increased psychosomatic symptoms, Increased anger, and
sadness, Decreased concentration, Increased suicidal ideation and depression,
and Increased social anxiety in different age groups of people [13]. So, there
is an immediate need to detect these texts as early as possible to deal with
future problems. There have been many researches conducted regarding this
topic but this paper deals with the problem with a large-scale data set of 15
thousand of data and the use of several machine learning and deep learning
techniques. The target of this paper is, 1. To design a way to detect aggressive
and non-aggressive text from social media comments and posts. 2. To create a
suitable data set for this research and apply different models on the data set to
get the appropriate result.

2 Related Work
The research on detecting abusive, aggressive text or cyberbullying detection
has seen exponential growth in recent years. Because of the available resources
on the internet, researchers can easily find the necessary data on this topic. In
[2], the authors proposed a unique design for analyzing Bangla content on social
media by combining text analytics. By using machine learning algorithms and
also comparing the performance of the module with other available techniques
they achieved the best result using SVM at 95.40% on user posts. And 97.27%
using SVM on user-specific data. The authors of [3] developed a model to detect
cyberbullying from Bangla and Romanized Bangla texts. Where they presented a
comparative analysis among the algorithms in terms of accuracy, precision, recall,
f1- score, and roc area. They used a total of 12 thousand datasets. Achieving the
best result of 84%. In [4], the authors used machine learning and deep learning
algorithms to detect abusive Bengali text. Where they set a new stemming rule
176 T. R. Rosni et al.

for the Bengali language which helps to achieve better performance. They used a
total of 4.7 thousand datasets with 7 class labels. Among all the algorithms RNN
gives the best result of 82.20% accuracy. Similarly, the authors [5]used supervised
machine learning classifiers to detect Bangla abusive text and used three unique
pathways in their dataset to get the best result. They used a total of 12 thousand
datasets with two class labels and SVM gave them the best accuracy of 88%. The
authors [6] proposed two getaway Naive Bayes classification algorithms and a
Topical approach to extract the emotion from Bangla text. They used more than
7.5k datasets and the topical approach gave them the best performance Which
is more than 90% accuracy. The authors of [7] used the Naïve Bayes classifier to
detect abusive comments expressed in Bangla. And evaluated performance by
using 10-fold cross-validation on unprocessed data gathered from “outube.com”
On a 2.6 thousand data labeled as abusive and not abusive they found the best
result at 80.57% in Naive Bayes classifier. The authors [8] proposed a method
that works with both Bangla English mixed and transliterated Bangla text. Their
suggested algorithm can detect features, we used Unigrams, Bigrams, number
of likes, and emojis along with their categories, sentiment scores, offensive and
threatening words used in the comments, and the number of abusive words
in each comment, profanitypes. They used Support Vector Machine, Random
Forest, and Adaboost, on 2 thousand data achieving the highest accuracy of
72.14%. All the related work that was mentioned before works in either Bangla
or English Language. The main difference between the aforementioned works
and this research paper is the number of large-scale data sets of 15 thousand
data. And the number of implemented algorithms.

3 Methodology
A. Work Process
First, the main target of this research paper is to detect aggressive and non-
aggressive texts from social media posts and comments. Second, going through
different social media platforms gathering aggressive and non-aggressive posts
and comments, and creating a data set. Third, Then the data was cleaned (with-
out characters, emoji, or extra words) and labeled for further use. As it was an
NLP classification problem the initial thought was to implement only classifica-
tion algorithms but regression algorithms were added to get a unique viewpoint
considering the data nature. Then the data set was processed so that it could be
used in machine learning and deep learning techniques. Where feature extraction
happened to train models. The used algorithms are Linear and kernel SVM, Lin-
ear Regression, Decision tree, Random forest, Multinomial Naive Bayes, KNN,
and CNN from deep learning. Then we calculate the efficiency using the efficiency
metrics.
B. Dataset
The data set for this paper was collected by searching through different social
media platforms like (Facebook, Twitter, and YouTube). A total of 15 thousand
of data was collected. These data contain hate speech, abusive text, hatred,
 Aggressive Bangla Text Detection 177

and normal texts. Over 9 thousand data were labeled as aggressive and over 6
thousand data were labeled as non-aggressive. The data were manually labeled
as either aggressive or non-aggressive. The data was carefully classified as ‘0’
for aggressive and ‘1’ for non-aggressive data. Moreover, a stop words text was
developed to further enhance the prepossessing step (Figs. 1 and 2).

Fig. 1. The dataset with class and Label

C. Feature Extraction and Reprocessing

In this research paper, the features extraction for Machine Learning algorithms,
the authors also compared with n-grams such as uni-gram, bi-gram, and tri-
gram based Count Vectorizer and TF.IDF-Vectorizer characteristics to detect
178 T. R. Rosni et al.

Fig. 2. The number of unique words detected

the aggressive and non-aggressive text. TF.IDF-Vectorizer transforms the strings

into numbers. The length of each text and word is also taken into consideration as
the frequency of words is used by TF.IDF. Word Embedding is used to produce
numerical or vector representations of text data. Also, the Keras OOV tokenizer
was used to represent all the words in the dataset as vectors and those vectors
were used as input for the embedding layer of the Deep Learning algorithm CNN
(Fig. 3).

Fig. 3. The Frequency of words and texts

D. Training and Testing Data Split

After the preprocessing and feature selection step was completed, the authors
split the 15 thousand data into 70% and 30% for training and testing respectively
from the data set. After the splitting step was done, the data set contained 10,500
 Aggressive Bangla Text Detection 179

texts for training and 4500 for testing. Training sets of data were used to train
the machine learning and deep learning models testing sets were used to test the
same models. During the process, the authors randomly shuffled the data to get
an evenly distributed dataset.
E. Machine Learning Algorithms
In recent research papers, it is shown that the classification models work well with
categorically labeled data. Here in this research paper, the authors used both
classification and Regression models which include Linear Regression, Linear
SVM (Support Vector Machine), Kernel SVM, Random forest, Decision Tree,
Multinomial Naive Bayes, and KNN.
Linear Regression: Linear Regression is a supervised machine learning algo-
rithm that is used in regression problems. It predicts the continuous dependent
variable with the help of independent variables from the data set. And to deter-
mine the best fit that precisely predicts the accurate value for the continuous
dependent variable.
SVM: Support Vector Machine or SVM is used to classify data points which
helps find a hyperplane in a K-dimensional space and K refer to a number of
features. Among many possible features, the SVM chooses the feature with the
maximum margin. Here in this research, the authors used both Linear and Kernel
SVM.
Random Forrest: Random forest is a Supervised Machine Learning Algorithm
that works very well with Classification and Regression problems. Based on dif-
ferent samples and inputs, creates a sample weight for a problem and then takes
the highest valued weight for the classification problem and takes the average
weight for regression. However, it works well with the classification problem.
Decision Tree: A supervised learning method used for both classification and
regression. Analyses the data features and predicts the value of the target vari-
able by implementing decision rules.
Multinomial Naïve Bayes: The multinomial Naïve Bayes algorithm is based
on Bayes theorem and calculates the likelihood of occurrence of an event based on
prior knowledge of the event’s conditions using the Bayes theorem. The highest
probability is the output. Probability of P(A|x) where A is the of the possible
outcomes given instance x. The equation is given below.

P(A|x) = P(x|A) ∗ P(A)/P(x)

In this research, Multinomial Naïve Bayes is used.

KNN: The k-nearest neighbors (KNN) algorithm is also a supervised learning
algorithm that relies on labeled input data to learn a function. And later based
on that learned function produces a feature-appropriate output considering new
unlabeled data.
180 T. R. Rosni et al.

F. Deep Learning Algorithm

Here, in this research paper neural network is used. The authors used the
CNN (Convolutional Neural Network)model. CNN model works well with
classification-type problems.
CNN: A Convolutional Neural Network or CNN is a neural network designed
mainly for processing structured arrays of image and video data. It is one of the
most used Neural networks in computer vision. Here is the structure of CNN
that we used.
Embedding Layer: In this layer all the words in the data set turned into vectors
and were used as the input of the embedding layer. This is the first layer of the
neural network where a matrix is generated which represents the vectors and
words.
Max-Pooling Layer: This layer is used to shrink down the spatial size of a
convoluted feature. So, basically, dimension reduction is done in this layer and
padded to reduce the noise from actual meaningful data.
Flatten Layer: This layer is used to convert the data into a 1-D array. It creates
a flattened output of its previous layer. And a single long feature vector is created
for the dense layer to perform classification through various models.
Dense Layer: This layer takes the values of all the previous layers and groups
them into a dense feature to perform classification.
LSTM: LSTM is a special type of RNN(Recurrent Neural Network) that handles
the long-term dependency problem by processing the previous and current data.
There are three parts of the LSTM cell. These are the Forget gate which forgets
irrelevant and excessive information, the input gate which adds or updates new
information into the network and lastly, the output gate which then passes the
updated information onto the network. LSTM network has a hidden state which
is known as Short-term memory and a cell state which is known as Long-term
memory.
Forget Gate: It is the first step of LSTM Network, where it decides which
information to keep and which to forget from the previous instant.
Input Gate: In this step the network decides which relevant information is
necessary to keep and add or update that information to the network.
Output Gate: Here based on the new updated information the network has to
predict a relevant result for this instant. For better prediction, we used CNN-
LSTM hybrid network.

4 Result Analysis

Here in this section, the detailed results are discussed. Over 15 thousand prepos-
sessed data were given for training and testing purposes. And they are measured
 Aggressive Bangla Text Detection 181

Model Accuracy Precision Recall F1-Score

Linear Regression 88.31 88.31 88.31 88.31
Decision Tree 82.01 82.01 82.01 82.01
Random Forrest 86.71 86.71 86.71 86.71
Multinomial Naïve Bayes 90.46 90.46 90.46 90.46
KNN 79.95 79.95 79.95 79.95
Linear SVM 86.03 86.03 86.03 86.03
Kernel SVM 87.67 87.67 87.67 87.67
CNN+LSTM 85.15 90.28 86.85 84.55

by four metrics they are Accuracy, Precision, Recall, and F1-Score. Here are the
following results and analyses.
Table: Results (Accuracy, Precision, Recall, F1-score)
The table shows that the highest accuracy from the machine learning algo-
rithms was found from multinomial naïve Bayes about 90.46%. Which is about
2.15% higher than the second-best result of 88.31% from Linear Regression. Then
the 3rd highest accuracy was gained from kernel SVM of about 87.67% which is
2.79% less than what we got in MNB. In the 4th position, we get Random Forrest
resulting in 86.71% accuracy. It is around 3.75% less than the result of MNB.
Next, we got 86.03% testing accuracy on the data set from Linear SVM. From
the Decision Tree, we got 82.01% accuracy and from KNN we found 79.95%
accuracy on the data set. In the case of the Deep Learning algorithm, for better
prediction used CNN LSTM hybrid network and found an accuracy of 85.15%.
The precision of 90.28% and Recall score of 86.85% and the F1-score is 84.55%
(Figs. 4 and 5).

Fig. 4. CNN model accuracy

182 T. R. Rosni et al.

Fig. 5. CNN model loss

The model loss is calculated on training and validation and it is based on how
the model is performing in these two sets. The sum of errors from each iteration
in training or validation sets. This model Loss curve shows us how a model
behaves after each iteration of optimization. A model accuracy curve shows the
measurement of the algorithm’s performance. The accuracy of a model is usually
determined after the parameters are calculated. It curve shows the measurement
of how accurate the model prediction is compared to the true positive data.

5 Conclusion
This approach explores the identification of aggressive and non-aggressive Bangla
text in social media posts and comments through the application of machine
learning and deep learning algorithms. Various classification and regression mod-
els such as Linear Regression, Linear SVM (Support Vector Machine), Kernel
SVM, Random Forest, Decision Tree, Multinomial Naïve Bayes, and KNN were
employed. The highest accuracy, 90.46%, was achieved with the Multinomial
Naïve Bayes model. Linear Regression followed with an accuracy of 88.31%, and
the CNN model yielded 85.15% accuracy. These findings suggest that classi-
fication algorithms perform effectively in Natural Language Processing (NLP)
classification tasks.

6 Future Work
Our goal was to discern aggressive and non-aggressive Bangla text within social
media comments and posts by applying machine learning and deep learning mod-
els. A range of machine learning models were employed, each yielding distinct
outcomes. Furthermore, the CNN deep learning model was applied, resulting
in an accuracy of 85.15%. Regarding future endeavors, we have proposed ideas
 Aggressive Bangla Text Detection 183

to augment and fine-tune the obtained results. Exploring other Deep Learning
algorithms, such as GRU, Attention-based models, RNN, etc., could contribute
to further improvements. Incorporating more user-specific posts and comments
has the potential to enhance results. Additionally, utilizing a larger dataset for
training and testing purposes could be beneficial.

References
1. Editors of Encyclopedia Britannica. Bengali Language (2022). https://www.
britannica.com/topic/Bengalilanguage
2. Abdhullah-AlMamun, Akhter, S.: Social media bullying detection using machine
learning on Bangla text. In: 10th International Conference on Electrical and Com-
puter Engineering (ICECE), pp. 385–388 (2018)
3. Ahmed, Md.T., Rahman, M., Nur, S., Islam, A.A., Das, D.: Deployment of machine
learning and deep learning algorithms in detecting cyberbullying in Bangla and
Romanized Bangla text: a comparative study. In: 2021 International Conference on
Advances in Electrical, Computing, Communication and Sustainable Technologies
(ICAECT), pp. 1–10 (2021)
4. Emon, E.A., Rahman, S., Banarjee, J., Das, A.K., Mittra, T.: A deep learning
approach to detect abusive Bengali text. In: 7th International Conference on Smart
Computing & Communications (ICSCC), pp. 1–5 (2019)
5. Islam, T., Ahmed, N., Latif, S.: An evolutionary approach to comparative analysis
of detecting Bangla abusive text. Bull. Electr. Eng. Inform. 4, 2163–2169 (2021)
6. Tuhin, R.A., Paul, B.K., Nawrine, F., Akter, M., Das, A.K.: An automated sys-
tem of sentiment analysis from Bangla text using supervised learning techniques.
In: 2019 IEEE 4th International Conference on Computer and Communication
Systems (ICCCS), pp. 360–364 (2019)
7. Awal, Md.A., Rahman, Md.S., Rabbi, J.: Detecting abusive comments in discussion
threads using Naïve Bayes. In: 2018 International Conference on Innovations in
Science, Engineering and Technology (ICISET), pp. 163–167 (2018)
8. Jahan, M., Ahamed, I., Bishwas, Md.R., Shatabda, S.: Abusive comments detec-
tion in Bangla-English codemixed and transliterated text. In: 2nd International
Conference on Innovation in Engineering and Technology (ICIET), pp. 1–6 (2019)
9. The DataReportal. https://datareportal.com/essentialfacebook-stats
10. Editorial. https://www.oberlo.com/statistics/whatage-group-uses-social-media-
the-most
11. Dave Chaffey (2022). https://www.smartinsights.com/social-media-marketing/
social-mediastrategy/new-global-social-media-research/
 A Comparative Evaluation of Machine Learning
Techniques for Detecting Malicious Network
Traffic

Prince Tayal1 , Rohan Kumar1 , and Hemlata2(B)

1 CSE(AIML), Manipal University Jaipur, Jaipur, India
[email protected], [email protected]
2 Department of CSE-AIML, Manipal University Jaipur, Jaipur, India

[email protected]

Abstract. The Malicious network activities such as malware attacks, unautho-

rized intrusions, and data theft have become a rising threat as organizations increas-
ingly rely on interconnected systems. Detecting these security incidents rapidly
and accurately is critical, but also poses challenges due to the constantly evolving
techniques used by hackers. Machine learning has surfaced as a promising avenue
for developing intelligent systems capable of analyzing massive amounts of net-
work data, identifying anomalies, and detecting malicious patterns in real-time.
This research presents a thorough comparative evaluation of the key machine learn-
ing algorithms applied for detecting various types of malicious network traffic.
The models tested include random forests, regular neural networks, and logistic
regression. Using labeled datasets containing benign and attack traffic flows, the
algorithms are trained, and their performance evaluated across parameters like
detection accuracy, confusion matrix, F1-scores, and receiver operating charac-
teristics. The results demonstrate that ensemble methods such as random forests
achieve exceptionally high accuracy, exceeding 99% in some tests, by combining
diverse decision trees. With careful tuning of hyperparameters, neural networks
also display comparable performance owing to their automated feature extraction
capabilities. The analysis provides useful insights into the strengths, weaknesses,
and applicability of each technique. These findings will enable security experts to
select the most appropriate machine learning solutions tailored to their network
environments and requirements for developing robust intrusion detection systems
capable of identifying threats in real time.

Keywords: Machine learning · Network security · Intrusion detection · Malware

detection · Logistic regression · Random forests · Regular Neural networks ·
Ensemble methods · Comparative evaluation

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 184–204, 2025.
https://doi.org/10.1007/978-3-031-71481-8_15
 A Comparative Evaluation of Machine Learning Techniques 185

1 Introduction
In recent years, cyberattacks have been growing at an exponential rate with the expan-
sion of the Internet and interconnected systems. Malware, unauthorized intrusions, data
breaches – these threats pose a serious risk to individuals, businesses, and critical infras-
tructure. A major challenge is being able to detect these malicious activities in real-
time as they occur on networks. This is extremely difficult given how rapidly hacking
techniques evolve to avoid detection. Traditional security tools relying on pre-defined
signatures cannot catch novel threats they have not yet encountered. There is a need for
more intelligent, automated solutions that can analyze massive volumes of network data,
spot anomalies, and identify potential attacks.
Machine learning has come up to be an exciting prospect to tackle these cybersecurity
challenges. By applying statistical techniques and algorithms to uncover patterns in data,
machine learning can develop systems capable of modelling normal network behavior,
detecting deviations, recognizing new attack signatures, and preventing intrusions in real-
time. Unlike rigid rule-based systems, machine learning models can be trained on large
datasets to continuously improve accuracy as more data comes in. Various supervised,
unsupervised and hybrid machine learning algorithms have already been applied for
network anomaly detection and malware analysis with promising results.
This paper aims to provide a comparative evaluation of the major machine learning
techniques used for detecting malicious network traffic. The key algorithms examined
include random forests, regular neural networks, and logistic regression. Using pub-
licly available datasets labelled as benign and malicious, these techniques are tested on
performance parameters like accuracy, confusion matrix, F1-scores, and receiver oper-
ating characteristics. In addition to monitoring detection capabilities, factors such as
interpretability, computational efficiency and ease of implementation are analyzed. The
objective is to thoroughly understand the benefits and drawbacks of every algorithm.
This knowledge will serve as a guidepost for identifying optimal solutions that meet the
specific needs of real-world network intrusion detection systems built using machine
learning methodologies.

2 Literature Review
2.1 Literature Overview
First research paper has proposed using artificial neural networks for misuse detection
to identify network intrusions without requiring full attack specifications. Preliminary
analyses indicate neural networks show promise in improving detection of novel attacks
by recognizing partial indicators and complex patterns not easily encoded into explicit
rules. Our research aims to further evaluate the potential of neural networks to enhance
misuse detection against unpredictable threats [1].
Previous research has looked into integrating soft computing approaches such as
artificial neural networks (ANNs), support vector machines (SVMs), and multivariate
adaptive regression splines (MARS) inside ensemble models for network intrusion detec-
tion. These hybrid approaches aim to leverage the strengths of individual techniques to
improve classification accuracy. In a recent investigation, researchers discovered that
186 P. Tayal et al.

combining these techniques yielded improved results in identifying cyber-attacks com-

pared to employing each method individually. Our research aims to delve deeper into
the realm of ensemble models, incorporating a mix of soft computing and hard comput-
ing techniques. The goal is to create intrusion detection systems that are not only more
effective but also resilient in the face of constantly evolving cyber threats [2].
In a study conducted by Apruzzese and colleagues in 2019, they delved into the realm
of cybersecurity, specifically focusing on the analysis of machine learning techniques
for detecting intrusions, malware, and spam. While machine learning shows promise,
the research concluded rigorous evaluation is still needed to assess effectiveness on real
systems and address limitations before these techniques can be reliably deployed for
automated threat detection. Our research provides further comparative assessment of
machine learning algorithms tailored for malicious network traffic detection [3].
Mukkamala and his team conducted research back in 2002, exploring the utilization
of support vector machines (SVMs) and neural networks to detect network intrusions.
They based their investigation on KDD benchmark datasets. They demonstrated both
techniques can build efficient and highly accurate classifiers for detecting attacks. Further
experiments found using just 13 most significant features delivered comparable accuracy.
The study provides early benchmarking of SVM and neural network performance for
misuse detection using standardized datasets. Our work provides further comparative
assessment on modern malicious traffic [4].
Research by Toosi and Kahani (2007) proposed an intrusion detection approach com-
bining soft computing techniques like neuro-fuzzy classifiers and genetic algorithms.
Their model uses neuro-fuzzy networks for initial attack classification, then a fuzzy
inference engine for final decision-making optimized by a genetic algorithm. The exper-
iments involved the assessment of the hybrid technique using the KDD Cup 99 bench-
mark dataset. Study demonstrates the potential of integrating multiple soft computing
methods to improve anomaly detection accuracy. Our work provides further analysis
using modern machine learning algorithms tailored for malicious traffic detection [5].
In a survey conducted by Tsai and collaborators in 2009, they examined intrusion
detection systems that employed machine learning techniques, focusing on publications
from the years 2000 to 2007. It examined single, hybrid approaches, and ensemble clas-
sifiers applied on benchmark datasets. The review compared classifier designs and exper-
imental setups to assess achievements and limitations in developing ML-based intrusion
detection. It provides useful insights into the state of machine learning for network
security prior to the rise of deep learning. Our work provides an updated comparative
benchmarking focused on detecting modern malicious traffic [6].
Bhuyan and colleagues conducted a survey in 2014, offering a structured overview of
methods and systems for network anomaly detection based on computational techniques.
It summarized and compared a wide range of anomaly detection approaches, including
machine learning algorithms. The paper also discussed available tools and datasets for
research. It highlighted open challenges and future directions in advancing network
intrusion detection. Our work provides an updated comparative benchmarking focused
specifically on machine learning techniques using recent datasets [7].
Research by Sharafaldin (2018), addressed the significant issue of DDoS attacks in
network security. It comprehensively reviewed existing DDoS datasets and introduced
 A Comparative Evaluation of Machine Learning Techniques 187

the CICDDoS2019 dataset to overcome previous limitations. Leveraging this dataset, a

novel detection approach based on network flow features was proposed. The paper also
presented essential feature sets for detecting various DDoS attack types, along with their
respective weights [8].
In a study undertaken by Javaid A. in 2016, the primary focus was on crafting a
Network Intrusion Detection System designed to assist network admins in pinpointing
vulnerabilities within their respective companies. The study aimed to address challenges
related to flexibility and efficiency in detecting unforeseen and unpredictable attacks.
To accomplish this goal, a proposed approach centered around deep learning was intro-
duced, incorporating Self-taught Learning (STL), and applied to the NSL-KDD bench-
mark dataset for the purpose of network intrusion detection. The paper showcased the
performance results of this approach and conducted a comparative analysis with prior
research, taking into account parameters such as accuracy, confusion matrix, and F1
score [9].

2.2 Requirements
Drawing insights from the literature review, the following key requirements are defined
for an effective machine learning based network intrusion detection system:
• Attaining high detection accuracy across diverse types of malware and network
attacks.
• Low false positive rate to minimize false alarms
• High true positive rate to detect most threats
• Fast detection capability for real-time attack prevention
• Robustness against evasion techniques used by adversaries
• The capability to identify zero-day and unknown threats.
• Capability to work with imbalanced datasets
• Incremental learning potential to adapt as new data becomes available
• Interpretability for cybersecurity experts to analyze decisions
• Computational efficiency for high-speed networks
• Easy integration into existing security infrastructure

2.3 Outcome of Literature Review

The literature review emphasizes the application of a broad range of classical machine
learning techniques for the detection of malicious network traffic. Ensemble methods,
such as random forest and gradient boosting, often exhibit remarkably high detection
accuracy when applied to contemporary datasets [8]. However, deep learning models are
emerging as promising techniques due to their automated feature learning capabilities
[3]. Tree ensembles, SVM, KNN, discriminant analysis, and Bayesian networks are the
most mature techniques [10].
Most studies have relied on standard datasets like NSLKDD and UNSW-NB15 for
benchmarking. However, newer datasets that capture modern attack techniques would
be valuable for more robust evaluation [11]. There is a need for comparisons on very
large real-world network data [12]. Assessment of computational performance and limi-
tations is also required for practical deployment [13]. While accuracy metrics are widely
188 P. Tayal et al.

reported, other factors like model interpretability also need to be considered based on
the application context [14].
Overall, the review indicates machine learning has excellent potential for developing
intelligent malware detection systems. However, more research is required to determine
optimal algorithms and best practices tailored to different network environments and
security needs [10]. Our experiments will provide further insights in this direction.

2.4 Problem Statement

Derived from the literature review, the central problem we seek to address in this study
can be succinctly summarized as follows:
Determining the most appropriate machine learning algorithms and techniques for
accurately detecting various types of malicious network traffic in real-world environ-
ments, given the constraints of computational efficiency, evolving cyber threats, class
imbalance, and the need for interpretable models.

2.5 Research Objectives

The objectives of our research are:

• A comparative assessment of the fundamental categories of machine learning
algorithms used in network intrusion detection, considering detection accuracy,
computational efficiency, and other pertinent metrics.
• To analyze the performance trade-offs, advantages, and limitations of supervised,
unsupervised, and hybrid models for detecting malicious network traffic.
• To identify the factors and best practices that allow maximizing the malware detection
capability for different algorithms and network environments.
• To provide data-driven recommendations on selecting suitable algorithms for
developing real-world intrusion detection systems (Fig. 1).

Fig. 1. Flowchart for depicting how classification works

 A Comparative Evaluation of Machine Learning Techniques 189

3 Methodology

The methodology employed in this study is fundamental to understanding the research

process behind evaluating machine learning algorithms for identifying malware in net-
work traffic. The following part delves into the intricacies of data collection and pre-
processing, feature extraction, machine learning algorithm selection, and the criteria for
evaluating their performance.

3.1 Data Collection and Preprocessing

Effective malware detection relies on high-quality data. In this study, network traffic
data is collected from various sources, including network logs, packet captures, and
flow records [15]. These raw data sets are often voluminous, containing both legitimate
and potentially malicious traffic. To prepare the data for analysis, rigorous preprocessing
is essential.
Data preprocessing involves several key steps:
Data Cleansing: It involves the identification and correction of errors or inconsistencies
in a dataset, including tasks such as addressing missing values, handling outliers, and
rectifying inaccuracies. This step holds utmost importance in guaranteeing the quality
and dependability of the data for any future evaluation or modeling.
Data Reduction: Considering the abundance of raw data, it becomes imperative to
diminish its dimensionality. For this purpose, methods such as feature selection and
PCA can be used, with the goal of preserving the most relevant attributes while reducing
computational complexity [15].
Normalization: Data normalization is a process that standardizes the data to a common
scale, preventing certain features from exerting disproportionate influence on the analysis
due to their larger values. This ensures a more balanced consideration of all features in
the dataset. Techniques such as Z-score normalization are often used [15].
Balancing The Dataset: The dataset may exhibit class imbalance, where benign traffic
significantly outnumbers malicious instances. To correct imbalances in class distribution
within a dataset, various techniques such as oversampling, undersampling, or generating
synthetic samples (such as using Synthetic Minority Over-sampling Technique) can be
used. These approaches aim to improve the learning process for machine learning models
by creating a more equitable representation of different classes [15].
Feature Scaling: Ensuring proper scaling of features is crucial for enhancing the perfor-
mance of certain machine learning models. Some of the scaling techniques are MinMax
scaling and Z-score standardization. These techniques help maintain consistent scales
across features, contributing to improved algorithmic performance [15].
Data partitioning: It is critical to divide the dataset into training, validation, and test
sets to train and assess machine learning algorithms. This separation allows for effective
model development, tuning, and unbiased assessment of performance.
190 P. Tayal et al.

3.2 Feature Extraction

Feature extraction is a crucial step in the process, preparing the data for machine learning.
In the context of malware detection, features are characteristics derived from network
traffic data that enable the algorithms to make predictions [16]. These features can be
classified into multiple types, like statistical, behavioral, and structural.
Statistical Features: These encompass attributes such as packet counts, byte counts,
and traffic rate statistics, which provide insights into the traffic’s basic characteristics
[15].
Behavioral Features These features capture the behavior of network traffic, including
patterns of communication, network protocol usage, and communication entropy. They
enable the identification of anomalous behaviors that might indicate malware presence
[15].
Structural Features: Structural features represent the relationships within network
traffic data, such as IP address pairs, port pairs, or subgraph structures in communication
patterns [15].
The choice of feature extraction techniques is influenced by the specific objectives
of the malware detection task and the characteristics of the dataset. Feature engineering
is an iterative process involving the selection, extraction, and transformation of features
to optimize model performance.

3.3 Machine Learning Algorithm Selection

Several algorithms are considered for malware detection in network traffic, each with
its strengths and weaknesses.
Commonly evaluated algorithms include:
• Random Forest: This ensemble learning method is acknowledged for its proficiency
in managing high dimensional data and conducting feature importance analysis [17].
• Support Vector Machine (SVM): This method is used to find extensive application
in binary classification tasks and possess the capability to handle both linear and
non-linear data [18].
• Deep Learning: Deep neural networks, particularly Regular Neural Networks
(RNNs), have gained prominence in feature learning from network traffic data [19].
• Naïve Bayes: A simple and probabilistic approach that can effectively model
relationships between features and classes [20].
• Extreme Gradient Boosting: Algorithms such as XGBoost are renowned for
their effectiveness in managing imbalanced datasets, leading to an enhancement in
classification accuracy [21].
The selection process considers factors such as the nature of the dataset, algorithm
efficiency, computational resources, and the task’s complexity.
 A Comparative Evaluation of Machine Learning Techniques 191

3.4 Evaluation Metrics

Evaluating the performance of machine learning models is vital in assessing their efficacy
in malware detection. The selection of appropriate evaluation metrics is dependent on
the specific objectives and features of the specific assignment at hand. Commonly used
parameters include:
• Accuracy: It is defined as the proportion of correctly predicted instances in relation
to the total number of instances, and it offers a comprehensive evaluation about the
model’s efficacy across all classes or categories in the dataset [20].
TP + TN
Accuracy = (1)
TP + FP + FN + TN
• F1-Score: It is the harmonic average of precision and recall, which offers an equitable
analysis of precision and recall, providing a more thorough evaluation [20].

2 × Precision × Recall
F1Score = (2)
Precision + Recall
• Receiver Operating Characteristic (ROC) Curves: It depicts a model’s perfor-
mance across different thresholds, assisting in the selection of an optimal balance
between true positives and false positives [21].
These evaluation metrics guide the selection of the best performing algorithm and
support further analysis of model robustness and generalization to real-world scenarios.

4 Dataset

The dataset is sourced from the GitHub repository of Prateek Lalwani, a security blogger.
You can find the repository by clicking here at https://github.com/prk54/malware-detect
ion-machinelearning-approach/blob/master/data.csv.
Dataset contains 41,323 binary files (including exe, DLL files) which are legitimate
in nature means they are malware free. Also, the dataset contains 96,724 malware sam-
ples which are taken from virusshare.com. So, we can now conclude there are in total
approximately 1,38,047 samples in the entire dataset (Figs. 2 and 3).
192 P. Tayal et al.

Fig. 2. A list of all the columns from the dataset

Fig. 3. A bar graph depicting legitimate and malware files from the dataset

5 Feature Engineering
This section delves into the critical aspects of feature selection and extraction tech-
niques employed in the research, shedding light on how they contribute to enhancing the
performance of machine learning models for malware detection within network traffic.

5.1 Feature Selection and Extraction Techniques

Feature engineering as a crucial stage in crafting machine learning models intended
for malware detection. It involves the careful curation of relevant attributes (features)
derived from unprocessed network traffic data. The goal is to provide the model with
meaningful information while reducing dimensionality and computational complexity.
 A Comparative Evaluation of Machine Learning Techniques 193

Feature Selection Techniques: Several techniques are used in feature selection, each
with its advantages and applications. These include-

• Filter Methods: It analyzes feature significance by utilizing statistical metrics like

Information Gain, Chi-Square, or Correlation Coefficient [22]. Features are ranked
based on these metrics, and a subset is selected.
• Wrapper Methods: Wrapper methods employ machine learning models to assess
feature subsets. By assessing the model’s performance with different combinations
of features, these methods enable the identification of the most optimal feature set
[23].
• Embedded Methods: These methods integrate the process of feature selection into
the model training phase. Examples of such techniques include L1 regularization
(LASSO) or decision tree feature importance are examples [24].

Feature Extraction Techniques: Feature extraction refers to the process of converting

raw data into a condensed and more representative format. Regarding the identification
of malicious software in traffic over networks, the following techniques are notable-

• Principal Component Analysis (PCA): It is a technique to reduce dimensionality of

information by portraying it onto a lower-dimensional subspace. The goal of this
transformation is to keep as much variance in the original data as possible. This can
be particularly useful in handling high-dimensional network traffic data [25].
• Wavelet Transform: The wavelet transform decomposes data into different scales,
capturing both low and high frequency information. This can be valuable in extracting
time frequency domain features from network traffic [26].

5.2 Feature Importance Analysis

Feature importance analysis is instrumental in understanding the contribution of individ-

ual features to the predictive power of machine learning models. This analysis provides
insights into which features have the most significant influence on the model’s decisions.
• Random Forest Feature Importance: Random Forest algorithms assign importance
scores to features based on their contribution to reducing impurity in decision tree
nodes [28].
• Gradient Boosting Feature Importance: Gradient Boosting models calculate fea-
ture importance by assessing how often a feature is used in decision tree splits and
the improvement in accuracy it brings [29].
• Deep Learning Feature Analysis: In deep learning models, the importance of
features can be explored through techniques such as gradient-based methods and
activation maximization [30].
194 P. Tayal et al.

6 Machine Learning Algorithms

This section provides detailed descriptions of the selected machine learning algorithms
used in the research for malware detection in network traffic. These algorithms have
been chosen for their effectiveness in various aspects of this task.

6.1 Random Forest Classifier

The Random Forest Classifier is an ensemble-based machine learning methodology that

has been tailored for classification tasks. It harnesses the strengths of decision trees by
amalgamating several trees to enhance predictive accuracy and foster model resilience.
The Random Forest Classifier works by combining a set of decision trees. Each tree
in the ensemble is trained on a slightly different subset of data and produces its own
prediction. In classification scenarios, such as distinguishing whether an email is spam or
not, each tree within the forest contributes its own prediction, effectively casting a “vote.”
The final decision hinges on the majority “votes” from the trees, resulting in a collective
decision that typically demonstrates heightened accuracy and reliability compared to a
single decision tree’s output.
Some measures which were taken to ensure high accuracy are-
Feature Scaling (Standard Scaler): The data undergoes standardization through
employment of the Standard Scaler. The goal of this procedure is to rescale the features to
attain 0 as the value of average and 1 as the value of standard deviation. Standardization
serves as a typical preprocessing step, ensuring uniform scaling across all features. This
practice significantly aids the model in achieving improved performance and heightened
accuracy.
Class Imbalance Handling (RandomOverSampler): An imbalance within the classes
can result in bias within the model’s predictions. For this, we use the RandomOverSam-
pler from the imbalanced-learn library. It oversamples the minority class (malware) to
create a more balanced dataset. This ensures that the model has enough examples of
both classes to learn from.
Feature Selection (SelectFromModel): After addressing class imbalance, the code
applies feature selection using the SelectFromModel method. It selects the most relevant
features based on the feature importance calculated by a Random Forest Classifier (CLF).
The model retains the features that have the most significant impact on its performance,
while less important features are removed. This can help in reducing noise and improving
accuracy.
Hyperparameter Tuning (max_depth): In the Random Forest model (CLF), the
max_depth hyperparameter is adjusted to 10. The parameter max_depth regulates the
highest depth achievable by each individual decision tree within the ensemble. By set-
ting a limit on the tree depth, you can prevent overfitting and make the model generalize
better.
These changes collectively aim to improve the model’s accuracy by ensuring that
the data is well-preprocessed, the class imbalance is addressed, irrelevant features are
 A Comparative Evaluation of Machine Learning Techniques 195

removed, and overfitting is mitigated through hyperparameter tuning. Each step con-
tributes to enhancing the model’s performance for malware detection (Figs. 4, 5 and
6).

Fig. 4. F1 Score and Accuracy of model on test and train dataset

Fig. 5. Confusion Matrix for Random Forest Classifier

Fig. 6. ROC Curve for Random Forest Classifier

196 P. Tayal et al.

6.2 Logistic Regression

Logistic regression serves as a statistical technique employed for binary classification

tasks. It establishes a model to depict the correlation between a binary dependent vari-
able (such as yes/no or 1/0) and one or multiple independent variables, providing the
probability of an event occurring. It is widely used in various fields, including medicine,
marketing, and social sciences, for predictive modeling and decision making.
Some measures which were taken to ensure high accuracy are-
Standardization (Feature Scaling): The StandardScaler from scikit-learn is used to
scale the features. While not a form of regularization in the traditional sense, standardiza-
tion stands as a preprocessing measure that enhances the stability and efficacy of various
machine learning models, including Logistic Regression. It guarantees that every fea-
ture maintains an average of 0 and 1 as the value of standard deviation, enabling more
efficient convergence.
RandomOverSampler: It is from the imbalanced-learn (imblearn) library and is used
to address class imbalance. Though not explicitly a method of feature selection, this
method includes oversampling the minority class to address class imbalance within the
dataset. As a result, the model has a more balanced representation of the various classes,
allowing for improved learning and understanding of the minority class, which may lead
to more relevant features being identified by the model.
Hyperparameter Tuning: The code specifies hyperparameters for the Logistic Regres-
sion model, including-

• max_iter: The maximum number of iterations for the solver to converge.

• C: The regularization parameter. In this code, C = 1.0 is used, indicating relatively
weak regularization.
• penalty: The type of regularization, where penalty = ‘l2’ specifies L2 regularization.
L2 regularization encourages small coefficients and helps prevent overfitting.

These hyperparameters are adjusted to control the regularization strength and

convergence behavior of the Logistic Regression model (Figs. 7, 8 and 9).

Fig. 7. F1 Score and Accuracy of model on test and train dataset

 A Comparative Evaluation of Machine Learning Techniques 197

Fig. 8. Confusion Matrix for Logistic Regression

Fig. 9. ROC Curve for Logistic Regression

6.3 Regular Neural Network (RNN)

Deep Learning, particularly in the form of Regular Neural Networks (RNNs), has brought
transformative advancements to the realm of image and pattern recognition. RNNs are
versatile, multilayered architectures designed to autonomously extract intricate features
from data, rendering them ideal for tasks such as image analysis. In the context of
identifying malicious software, RNNs can be harnessed for the analysis of network
traffic data, treating it as sequential data. Research has demonstrated the efficacy of
RNNs in discerning nuanced and complex patterns associated with malware.
Some measures which were taken to ensure high accuracy are-
Neural Network Architecture: The neural network architecture is modified to have
more layers and neurons (e.g., 64, 32, 16) compared to a simpler network. This increased
198 P. Tayal et al.

complexity allows the model to learn more intricate patterns in the data, which can lead
to improved accuracy.
Batch Size and Number of Epochs: The batch size for training is adjusted to 64, which
can help the model update its parameters more frequently. Additionally, the number of
training epochs is increased to 50. These changes allow the model to undergo more
iterations and may improve its ability to fit the data.
Early Stopping: We use early stopping to avoid overfitting. It monitors the valida-
tion loss throughout the learning stage and stops the procedure if there is no obvious
advancement in the loss after a certain amount of consecutive epochs. This precautionary
technique is performed to prevent the model from being overly tailored to the training
data.
Activation Function: Within the concealed layers, we utilize the ReLU activation func-
tion. It is also known for its effectiveness in deep neural networks and can facilitate the
training process.
Binary Classification Threshold: In binary classification, a threshold of 0.5 is estab-
lished. When the predicted value is equal to or more than to 0.5, it is considered as 1
(malware), whereas values below 0.5 are classified as 0 (legitimate data) (Figs. 10, 11
and 12).

Fig. 10. F1 Score and Accuracy of model on test and train dataset

Fig. 11. Confusion Matrix for Regular Neural Network

 A Comparative Evaluation of Machine Learning Techniques 199

Fig. 12. ROC Curve for Regular Neural Network

7 Experimental Setup

This part outlines the essential parameters and configurations utilized in the research
to ensure the validity and reproducibility of the experiments, including data splits,
hyperparameter tuning, and cross-validation methods.

7.1 Testing and Training Data Splits

Within machine learning experiments, breaking down the dataset in groups for both
training and testing is a pivotal step. In this study, we adopt a carefully planned strategy for
dividing the data to evaluate how well the chosen machine learning algorithms perform.
The dataset is partitioned into three primary subsets:
1. Training Set: This specific subset is used for training the models. It forms the basis
for learning the relationships between features and labels. Typically, a significant
portion of the dataset is allocated to the training set to ensure that the models capture
the underlying patterns and characteristics of the data [15].
2. Validation Set: The validation set plays a crucial role in hyperparameter tuning.
It is used to fine-tune the algorithms and optimize their configurations. Different
combinations of hyperparameters are tested to identify the ones that yield the best
results. This helps in preventing overfitting and achieving a balance between model
complexity and generalization [30].
3. Test Set: The test set is set aside specifically to assess how well the algorithms
perform. It simulates real-world scenarios where the models encounter unseen data.
The test set’s results provide an objective measure of how well the algorithms can
generalize to new, potentially malicious, network traffic [20].
200 P. Tayal et al.

The division of the dataset into these subsets is typically performed randomly, with
care taken to ensure that the class distribution is maintained to prevent bias.

7.2 Hyperparameter Tuning

The adjustment of hyperparameters is a vital component of the experimental setup since

it directly affects the efficacy of the algorithm. Each method is comprised of specific
hyperparameters that govern and influence its behavior. The procedure for finding the
optimal hyperparameters involves systematic experimentation and evaluation.

7.3 Cross-Validation Methods

Cross-validation methods are integral to the research design, enabling the robust evalua-
tion of the algorithms and their generalization to different data subsets [31]. There are sev-
eral prevalent techniques such as K-fold Cross-Validation, Stratified Cross-Validation,
and Leave-One-Out Cross-Validation. The selection of a cross-validation method relies
on factors such as the dataset size, the algorithm used, and the research goals.

8 Conclusion

This research carried an in-depth relative analysis of significant machine learning algo-
rithms employed to detect different forms of malicious network traffic. The algo-
rithms tested include random forests, regular neural networks, logistic regression. Using
labelled datasets of benign and attack traffic, the models were trained and their perfor-
mance analyzed on parameters like ROC curves, F1-scores, accuracy, and confusion
matrix (Figs. 13 and 14) (Table 1).

Fig. 13. Bar graph comparing F1 score of our proposed model and the average of other’s model
 A Comparative Evaluation of Machine Learning Techniques 201

Fig. 14. Bar graph comparing accuracy of our proposed model and the average of other’s model

Table 1. Comparison Table with Accuracy and F1 Score.

Technique Machine learning algorithms

Average of results from other existing Result from our proposed model
papers
Accuracy (on test F1 Score (on test Accuracy (on test F1 Score on test
dataset) data set) dataset) data set)
Random Forest 98.38% 97.30% 99.54% 99.24%
Classifier
Logistic 69.72% 73.89% 97.85% 97.85%
Regression
Regular Neural 95.38% 91.97% 95.57% 92.51%
Network

The results demonstrate that ensemble methods like random forests achieve very
high accuracy, exceeding 99% in some tests, owing to their strategy of combining
diverse decision trees. With careful tuning of hyperparameters, neural networks also
display comparable performance due to their automated feature learning capabilities.
The analysis indicates each technique has strengths – tree ensembles excel at handling
high dimensional data, SVMs perform well on binary classification, while regular neural
networks can uncover deep patterns.
Hence, we can conclude from the accuracy of different classifiers, for the given
dataset the order which is followed is:

Random Forest Classifier > Logistic Regression > Regular Neural Network.

The study provides useful insights into selecting the most suitable ML solutions
tailored to specific network environments and security requirements. Random forest is
ideal for networks with abundant data. Deep learning (RNN) shows promise for feature
202 P. Tayal et al.

extraction from raw traffic data. The research highlights the importance of continuous
model validation, testing on recent datasets, and striking a balance between accuracy
and interpretability.
In summary, the experiments highlight the substantial potential of machine learning
in crafting smart, real world intelligent intrusion detection systems. With careful feature
engineering and model optimization, ML can enable robust and scalable cybersecurity
surveillance against evolving threats.

9 Future Prospects
While this study provides a comprehensive benchmarking of algorithms, further research
can build on these findings in several directions-
• Evaluating deep learning architectures like CNNs and LSTMs for spatial and temporal
pattern recognition in network traffic data.
• Testing the algorithms on very large real-world network datasets to assess computa-
tional feasibility.
• Enhancing models with active learning and incremental learning capabilities to adapt
to new unlabeled data.
• Developing ensemble stacking models that combine multiple algorithms to improve
attack detection accuracy.
• Analyzing model explain ability and visualization techniques to increase trust and
transparency.
• Comparing cloud-based implementations for scaling to big data workloads in high-
speed networks.
• Expanding the algorithms to detect zero-day and polymorphic attacks using behavior-
based anomaly detection.
• Assessing model robustness against adversarial evasion attacks and developing
mitigation strategies.
• Incorporating domain knowledge from cybersecurity experts into the machine
learning pipeline.
Exploring these research paths could facilitate the connection between academic
machine learning research and the practical implementation of intelligent intrusion
detection systems in the real world. With rigorous benchmarking and innovation,
machine learning is poised to transform network security defenses against increasingly
sophisticated threats.

References
1. Cannady, J.: Artificial neural networks for misuse detection. Natl. Inform. Syst. Secur. Conf.
26(2), 36881 (1998)
2. Mukkamala, S., Sung, A.H., Abraham, A.: Intrusion detection using an ensemble of intelligent
paradigms. J. Netw. Comput. Appl. 28(2), 167–182 (2005)
3. Apruzzese, G., Colajanni, M., Ferretti, L., Mirto, M.: On the effectiveness of machine and
deep learning for cyber security. In: 2018 10th International Conference on Cyber Conflict
(CyCon), pp. 371–390 (2019)
 A Comparative Evaluation of Machine Learning Techniques 203

4. Mukkamala, S., Janoski, G., Sung, A.: Intrusion detection using neural networks and sup-
port vector machines. In: Proceedings of the 2002 International Joint Conference on Neural
Networks, vol. 2, 17021707 (2002)
5. Toosi, A.N., Kahani, M.: A new approach to intrusion detection based on an evolutionary
soft computing model using neuro-fuzzy classifiers. Comput. Commun. 30(10), 2201–2212
(2007)
6. Tsai, C.F., Hsu, Y.F., Lin, C.Y., Lin, W.Y.: Intrusion detection by machine learning: a review.
Expert Syst. Appl. 36(10), 11994–12000 (2009)
7. Bhuyan, M.H., Bhattacharyya, D.K., Kalita, J.K.: Network anomaly detection: methods,
systems and tools. IEEE Commun. Surv. Tutorials 16(1), 303–336 (2014)
8. Sharafaldin, I., Lashkari, A.H., Hakak, S., Ghorbani, A.A.: Developing realistic distributed
denial of service (DDoS) attack dataset and taxonomy. In: 2019 International Carnahan
Conference on Security Technology (ICCST), pp. 1–8 (2018)
9. Javaid, A., Niyaz, Q., Sun, W., Alam, M.: A deep learning approach for network intrusion
detection system. In: Proceedings of the 9th EAI International Conference on Bio-inspired
Information and Communications Technologies, pp. 21–26 (2016)
10. Buczak, A.L., Guven, E.: A survey of data mining and machine learning methods for cyber
security intrusion detection. IEEE Commun. Surv. Tutorials 18(2), 1153–1176 (2016)
11. Ring, M., Wunderlich, S., Grüdl, D., Landes, D., Hotho, A.: Flow based benchmark data sets
for intrusion detection. In: Proceedings of the 16th European Conference on Cyber Warfare
and Security, pp. 361–369 (2019).
12. Shiravi, A., Shiravi, H., Tavallaee, M., Ghorbani, A.A.: Toward developing a systematic
approach to generate benchmark datasets for intrusion detection. Comput. Secur. 31(3), 357–
374 (2012)
13. Gao, N., Gao, L., Gao, Q., Wang, H.: An intrusion detection model based on deep belief
networks. In: Scientific Programming (2019)
14. April, A., et al.: A comprehensive survey on intrusion detection systems. University of
Amsterdam, Netherlands (2009)
15. Dainotti, A., Claffy, K.C., King, A.: Analysis of internet background radiation. In: Passive
and Active Measurement (PAM), pp. 31–40 (2011)
16. Kolter, J.Z., Maloof, M.A.: Learning to detect malicious executables in the wild. J. Mach.
Learn. Res. 7(May), 2721–2744 (2006)
17. Breiman, L.: Random forests. . Mach. Learn. 45(1), 5–32 (2001)
18. Schölkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the
support of a high-dimensional distribution. Neural Comput. 13(7), 1443–1471 (2001)
19. Saxe, J., Berlin, K.: Deep neural network based malware detection using two-dimensional
binary program features. In 2015 10th International Conference on Malicious and Unwanted
Software (MALWARE), pp. 11–20. IEEE (2015)
20. Kang, Y., Tan, J., Jiang, X., Wang, Y., Jia, C.: Malware classification with deep convolutional
neural networks. In 2017 10th International Conference on Malicious and Unwanted Software
(MALWARE), pp. 110–119. IEEE (2017)
21. Chen, T., Guestrin, C.: XGBoost: A scalable tree boosting system. In: Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pp. 785–794 (2016)
22. Huang, J., Li, K., Tan, C.L.: Advances in feature selection: a review. Data Min. Knowl. Disc.
13(1), 13–54 (2006)
23. Kohavi, R., John, G.H.: Wrappers for feature subset selection. Artif. Intell. 97(1–2), 273–324
(1997)
24. Friedman, J., Hastie, T., Tibshirani, R.: The elements of statistical learning, vol. 1. Springer
series in statistics (2001)
204 P. Tayal et al.

25. Jolliffe, I.T.: Principal component analysis. Wiley Online Library (2002)
26. Vetterli, M., Kovacevic, J.: Wavelets and Subband Coding. Prentice-Hall (1995)
27. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
28. Friedman, J.H.: Greedy function approximation: a gradient boosting machine. Ann. Stat.
29(5), 1189–1232 (2001)
29. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualizing
image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)
30. Bailey, M., Jorgensen, L.: The effects of data breach disclosure laws. IEEE Secur. Priv. 3(3),
40–45 (2011)
31. Zuech, R., Bishop, M.: A survey of Malware Detection Techniques. University of California,
Department of Computer Science (2013)
 Enhancing Cyclone Intensity Prediction
Through Deep Learning Analysis of Imagery
Datasets

Jyoti Dinkar Bhosale1(B) , Suraj S. Damre2 , Ujwala V. Suryawanshi3 ,

and Rajkumar B. Pawar4
1 VDF Group of Institution, Latur, India
[email protected]
2 D.Y. Patil College of Engineering Akurdi, Pune, India
3 Computer Science Department, Rajarshi Shahu Mahavidyalaya (Autonomous), Latur, India
4 Gharda Institute of Technology, Lavel, Ratnagiri, India

Abstract. Cyclones pose a significant threat, causing widespread devastation and

loss of life. Early prediction of cyclone intensity plays a crucial role in mitigating
their impact. In recent years, deep learning has emerged as a promising tech-
nique for image analysis. This paper introduces a deep learning-based method
for estimating cyclone strength using image datasets. By leveraging convolu-
tional neural networks (CNNs), the proposed approach extracts essential infor-
mation from satellite imagery to forecast cyclone intensity. The model is trained
on a comprehensive historical dataset sourced from the National Hurricane Cen-
ter’s HURDAT2 database and validated on new cyclone data. Evaluation metrics
such as mean absolute error, mean squared error, and root mean squared error
demonstrate the effectiveness of the CNN model in accurately estimating cyclone
intensity. Training the CNN model on the historical dataset employs supervised
learning, where labeled examples consisting of satellite data and corresponding
cyclone intensities are utilized. Through this process, the model discerns patterns
and correlations within the satellite data, enabling it to make precise predictions
for unseen cyclone data.

Keywords: Convolutional Neural Network · Cyclone · Deep Learning ·

Intensity · Meteorology

1 Introduction
1.1 A Subsection Sample
Cyclones are large-scale weather phenomena that often originate over warm tropical
oceans and can cause coastal areas to experience severe winds, heavy rain, storm surges,
and flooding. These storms, commonly known as tropical cyclones or hurricanes, are
formed over warm ocean waters when the conditions are just right. The combination
of warm water, low pressure, and favorable atmospheric conditions creates a powerful
system that can cause widespread destruction [1].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 205–217, 2025.
https://doi.org/10.1007/978-3-031-71481-8_16
206 J. D. Bhosale et al.

Each category on the Saffir-Simpson scale is determined by the sustained wind speeds
of the cyclone. For example, Category 1 cyclones have sustained winds of 74 to 95 miles
per hour, while Category 5 cyclones have sustained winds of 157 miles per hour or
greater. The scale also takes into account the potential for damage caused by the storm,
with higher categories indicating more severe and potentially catastrophic conditions.
As a civil engineer and as the director of the National Hurricane Center, Robert Simpson
developed this scale in the 1970s [2–4]. It is designed to give an estimate of the potential
damage and flooding a hurricane could cause based on its wind speed.
The way a CNN works is by progressively down sampling thereby gradually increas-
ing the number of filters, so that the neural network can learn from the different features
in the image. By using convolutional filters, the CNN is able to learn from the raw
pixel data and extract features without the need for manual feature engineering. CNNs
have been employed in a range of cyclone-related applications, including detecting and
tracking storm movement using satellite imagery, predicting storm severity using mul-
tiple meteorological factors, and generating realistic simulations of storm behavior [5].
These apps have the potential to increase our ability to forecast and prepare for extreme
weather occurrences, potentially saving lives and reducing property damage. The pro-
posed strategy has the potential to improve cyclone early warning systems and lower the
risk of damage and loss of life caused by these natural disasters [9–12]. The accurate and
reliable prediction of cyclone intensity can aid in the timely and effective evacuation of
people and protection of property. The proposed strategy helps to detect the behavior of
cyclones and their influence on the environment, which can aid in the development of
cyclone mitigation strategies [17]. By detecting the behavior of cyclones, such as their
intensity, trajectory, and speed, scientists can gain valuable insights into the patterns and
characteristics of these destructive weather phenomena. This information can then be
used to develop more effective cyclone mitigation strategies, including early warning
systems, evacuation plans, and infrastructure improvements in vulnerable areas.

2 Related Work

Using geometric features from cyclone images coupled with a multilayer perceptron
(MLP) machine learning algorithm, the paper [1] proposes a new method for detecting
tropical cyclone intensities. The proposed involves extracting geometric features from
cyclone images, such as the area, perimeter, and eccentricity. These features were then
used as inputs to an MLP machine learning algorithm, which was trained to predict the
intensity of the cyclone. The method also demonstrated better performance than existing
methods in terms of accuracy of 84%, speed and scalability. The findings of this research
could help in improving the forecasting of cyclones in the region, and thus contributing
to disaster management.
The classifiers in paper [2] were evaluated on a dataset of tropical cyclone band
images collected from low-Earth orbiting satellites. The results were compared with
existing state-of-the-art models and showed that the proposed framework outperformed
them in terms of accuracy and Kappa coefficient. Additionally, the experiment was also
conducted using different datasets and classification models, framework was robust and
effective across different datasets and classifiers. The results in paper [3] also indicate
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 207

model robust in terms of data variations and noise, and can effectively capture the cyclone
intensity distribution over a large geographical area. The model is also highly scalable,
and can be easily extended to other related applications. This process allows the network
to learn from the data more effectively over time, and the hybrid similarity measurement
allows the network to distinguish between features in the data that may not be easily
recognizable to the human eye. This improvement in classification performance was
particularly evident when small training sets were used, as the proposed method was
able to outperform several existing methods. Overall, the proposed method achieved
promising results, demonstrating its potential in improving cyclone intensity estimation
from satellite images.
Further research is needed to refine and optimize the model and assess its accuracy in
different areas.The validation process revealed that the DAV technique is more accurate
than the Dvorak technique for storms with intensities greater than or equal to category 2.
This is because the DAV technique takes into account additional factors, such as the size
and shape of the storm, which the Dvorak technique does not. This means that the DAV
technique is better able to predict the strength of a storm more accurately. This is espe-
cially important for forecasting potential damage from severe storms. Ultimately, this
improved accuracy can help authorities make better plans to protect citizens and prop-
erties from the destructive effects of severe storms. It can also be used to inform people
about potential risks posed by storms. Thus, the DAV technique provides an invaluable
tool for reducing the impact of destructive storms, not only in terms of minimizing
potential damage but also informing people of the risks they may face [4].
This novel model underwent testing on a dataset comprising over 500,000 images,
achieving an accuracy of up to 0.25 m/s. Notably, it demonstrated proficiency in iden-
tifying various storm types, including tropical cyclones, monsoons, and mid-latitude
cyclones. Additionally, the model displayed a capacity to identify new storms with an
accuracy of up to 0.5 m/s. Offering a real-time solution, this approach enables swift
and precise wind speed estimations through a situational awareness portal. Specifi-
cally designed to tackle the challenges of wind speed estimation from satellite imagery,
the model’s comparison with Dvorak T-number images facilitated accurate wind speed
assessments across diverse conditions. Moreover, its ability to swiftly and efficiently
detect new storms and provide real-time wind speed estimates through the situational
awareness portal was confirmed [5].
The method exhibited high reliability, with a Root-Mean-Square Error (RMSE) of
3.38 m/s and a Mean Absolute Error (MAE) of 2.64 m/s, underscoring its efficacy in
predicting wind speed for tropical cyclones. Evaluation against observations of tropical
cyclone intensity by the U.S. National Hurricane Center and measurement of wind
speed from aircraft reconnaissance data yielded a remarkable 97% accuracy rate. This
indicates the method’s proficiency in accurately estimating the intensity of most tropical
cyclones. Results indicated that the proposed method could achieve a robust estimation
of maximum wind speed with a root mean square error of 5.68 m/s and an approximate
accuracy rate of 94% [6].
A technique has been devised to estimate tropical cyclone (TC) intensity automat-
ically using data from the Special Sensor Microwave Imager (SSM/I). A recent study
[7] evaluates the efficacy of this method by gathering 1,040 sample images from 142
208 J. D. Bhosale et al.

TCs across the North Pacific, Atlantic, and Indian Oceans. Among these, 942 images
serve as training samples, and a feature-selection algorithm is employed to identify the
most optimal feature combination. The K-nearest-neighbor algorithm then compares
the selected features of each testing sample to those of the training samples, identifying
the K closest neighbors. In this study, the algorithm is applied to 98 testing samples
using 15 selected features to estimate TC intensity. Utilizing the best-track intensity as
the ground truth, the algorithm computes a root-mean-square error (RMSE) of 19.8 kt,
signifying the accuracy of the estimation, which improves to 18.1 kt. Furthermore, a
comparison is made between the RMSE generated using the best-track intensity and
reconnaissance data for two recent (1999) Atlantic hurricanes. This comparison under-
scores the effectiveness of the additional feature in enhancing research accuracy and
the superior reliability of best-track intensity data in predicting hurricane intensity over
reconnaissance data.
The results highlight that best-track intensity data can be up to 25% more accurate
than reconnaissance data, emphasizing its importance as the primary source for hurricane
intensity information. In paper [8], a new technique was developed to calculate the
intensity of tropical cyclones (TCs) by using data from the TRMM Microwave Imager
(TMI) channels of 85,37,21,19,10 GHz from 1999 to 2003.
The connection between the TMI brightness temperature (TB) parameters and the
TC maximum wind speed was investigated by computing the TBs in various circles
of different radii at different TMI frequencies. The lower frequency channels of 10 or
19 GHz were found to provide higher correlation due to their direct connection to sea
surface wind speed. Multiple regression equations were used with a few parameters to
create the TC intensity estimation method, and after selecting 3 parameters and com-
puting the regression coefficients, 10 equations with the lowest root mean square error
(RMSE) were selected. The method was verified using independent data by comparing
its results with the TC best track data. This comparison revealed that the method had an
RMSE of about 8 m/s in the whole basin and about 6 m/s in the northwestern Pacific,
indicating a relatively accurate estimation of TC intensity change.
In paper [9], a dependable and strong deep convolutional neural network (CNN) was
proposed for estimating tropical cyclone intensity using satellite images. Regularization
techniques were employed to effectively extract complex features from hurricane images,
resulting in superior accuracy and lower root-mean-square error compared to ‘state-of-
the-art’ methods.
In paper [10], A CNN model was optimized by fine-tuning the network architec-
ture, adjusting hyperparameters such as learning rate and batch size, and optimizing the
loss function. Through these iterative steps, the model’s performance was significantly
improved, achieving a lower root mean square error (RMSE) and outperforming the pre-
vious model by 35%. This highlights the effectiveness of deep learning approaches in
accurately estimating tropical cyclone intensity from satellite images. In the paper [11],
The performance of the TCICENet model was compared with several existing models in
the field of tropical cyclone intensity prediction. The results showed that the TCICENet
model outperformed other models in terms of both root mean square error and mean
absolute error, indicating its superiority in accurately predicting the intensity of tropical
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 209

cyclones. Furthermore, the model is able to accurately predict the intensity of different
categories of cyclones.
According to the paper [12], TC intensity and wind radius can be estimated from
infrared imagery over the northern Atlantic Ocean using deep learning and prior knowl-
edge of tropical cyclones (TC). Results show that deep learning improves the accuracy
of TC intensity and wind radius estimates. Additionally, the paper demonstrates the
potential of using deep learning for TC forecasting in the northern Atlantic Ocean. The
DeepTCNet also had a consistent performance across all regions of the world, with no
significant differences between different basins. The results demonstrate the effective-
ness of the DeepTCNet model in providing accurate estimates of TC intensity and wind
radii.
Researchers developed a model that uses convolutional neural networks (CNNs) to
estimate tropical cyclone intensity. The model was optimized using residual learning and
attention mechanisms to improve its feature extraction ability. The model was tested on a
dataset of more than 10,000 cyclones to demonstrate its accuracy. Results showed that the
model was able to accurately predict cyclone intensity with a high degree of accuracy. The
research team concluded that the model could be a useful tool for forecasting cyclone
intensity. The model was able to pick up on subtle patterns in the data that are not
immediately apparent to the human eye.
This was achieved by using residual learning, which allows the model to learn from
the difference between the estimated output and the actual output. Additionally, the
attention mechanism allows the model to focus on the most important features in the
data, which further improves the model’s accuracy [13].
To estimate tropical cyclone intensity from geostationary satellite images, the paper
[14] developed STE-TC, a spatiotemporal encoding module. By integrating temporal
information from TCs into satellite images, the module enhanced the spatial features.
The dense connectivity strategy was applied to a ConvMixer network to create a model
for estimating TC intensity called DenseConvMixer. STE-TC module and DenseCon-
vMixer network were evaluated through experiments and visualizations. CNN-based
models could be improved by using the STE-TC module. The results of the experiments
showed that the STE-TC module improved CNN-based models and the DenseConvMixer
network achieved better performance than existing methods. The model was successfully
applied to satellite images and provided more accurate TC intensity estimation.
The paper [15] proposed a model called TFG-Net that uses a graph convolutional
neural network for the estimation of tropical cyclone intensity with great detail. It was
composed of three parts, for producing general wind speed information. After test-
ing on the GridSat dataset, the TFG-Net obtained an RMSE metric of 11.12 knots,
outperforming the traditional method by 33.33% and deep learning method by 2.54%.
TCNNs are deep networks formulated using tensor algebra, as described in the paper
[16]. They are capable of implementing convolutional and pooling operations with the
help of tensor decomposition. Furthermore, TCNNs can exploit the structure of convo-
lutional networks, enabling faster and more efficient computation. This makes TCNNs
ideal for use in computer vision tasks, such as image classification, object detection, and
segmentation. TCNNs have been shown to provide significant improvements in accuracy
and speed compared to traditional convolutional networks. In addition, TCNNs can be
210 J. D. Bhosale et al.

used to reduce the size and complexity of convolutional networks, making them more
suitable for deployment on mobile devices and embedded systems. This makes TCNNs
attractive for use in a wide range of applications, including autonomous driving, medical
imaging, and robotics. TCNNs can also be used to improve generalization performance.
Furthermore, they can be used to implement deep architectures with fewer layers than
traditional convolutional networks, making them more suitable for applications where
the model size and computational complexity are key constraints.
In paper [17], a study was conducted in order to examine how the Advanced Dvo-
rak Technique (ADT), using geostationary meteorological satellite data, could be fur-
ther improved to produce a more accurate interpretation of the intensity of tropical
cyclones (TCs) around the world. The results of the study showed that the ADT had
a higher accuracy than the existing methods of TC intensity estimation. This improve-
ment was attributed to the higher resolution data obtained from the geostationary satellite.
The findings of the study contribute to the development of more reliable TC intensity
estimates.
An infrared and water vapor image-based model of tropical cyclone intensity was
presented in the paper [19]. RMSE and MAE were respectively 5.13 and 4.03 m/s,
respectively, for the 60 × 60 pixel input image. The same model was applied to a larger
resolution image with a size of 128 × 128 pixel. The results improved significantly,
with RMSE and MAE values of 3.76 and 3.23 m/s, respectively. This indicates the
importance of higher resolution images for accurate cyclone intensity predictions. The
model also demonstrated excellent temporal stability, with RMSE and MAE values
of 4.13 and 3.17 m/s, respectively, when applied to the same image with a one-hour
time lag. This indicates the robustness of the model for cyclone intensity predictions.
Additionally, the model was able to generate more accurate predictions when applied
to higher resolution images, thus confirming the importance of image resolution for
cyclone intensity predictions. Furthermore, the temporal stability of the model was also
excellent, as it was able to generate very similar predictions when applied to the same
image with a one-hour time lag. This confirms the robustness of the model for cyclone
intensity predictions.
For estimating tropical cyclone intensity with satellite remote sensing, a convolu-
tional neural network (CNN) hybrid model is proposed in [20]. Three different intensity
regression models were used to analyze the TCs by their intensity. Small samples were
fitted more quickly with piecewise thinking. Unlabeled TC samples were classified with
a classification network, and the results were used to determine the regression network
that should be used to estimate them. The model was validated with both synthetic and
real-world data, and was found to perform better than other existing models. The pro-
posed model is capable of estimating the intensity of TCs with a high degree of accuracy.
The results of this study demonstrate that piecewise thinking and deep learning are effec-
tive methods for estimating the intensity of TCs. This model can be used to improve our
understanding of TCs and their potential effects on climate.
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 211

3 Methodology

The research proposed in the study will involve the following steps:
1. Obtaining the necessary data
2. Preparing the data for analysis
3. Training the model
4. Verifying the accuracy of the results.
Data Gathering. The National Centers for Environmental Information’s HURSAT data
project makes use of cyclone images. These images were obtained from a database that
contains satellite images of cyclones in NetCDF file format. One noteworthy feature of
this database is that the center of each hurricane corresponds to the center of its cor-
responding image. Furthermore, the National Hurricane Center’s HURDAT2 database
supports the analysis by providing numerical data on cyclone images with components
such as time, latitude, longitude, wind speed, and so on in the form of best track data.
This extensive database contains data on all reported cyclones in the Atlantic and Pacific
basins. It includes the wind speeds of these cyclones recorded at 6-hour intervals. Data
scraping is used to obtain cyclone imaging data from the HURSAT data project for spec-
ified years. It scrapes a website to acquire files including satellite photos for each storm.
The algorithm restricts the data to cyclones in the Atlantic and Pacific basins that have
wind speed and position records in the National Hurricane Center’s Best Track dataset.
NetCDF files containing images from satellites of each cyclone are extracted and saved
locally from the files generated above. To guarantee data integrity, redundant images of
the same cyclone taken at the same time were removed.

Fig. 1. Proposed Architecture

Overall, this data collection process ensured the acquisition of high-quality cyclone
satellite imaging data, which aligned with the research goal of assessing and under-
standing the dynamics of these extreme weather phenomena. After that the wind speed
is divided into various categories which are as as follows:
• If the wind speed is 33 km per hour or less, it is categorized as a ‘Tropical Depression’.
• If the wind speed is between 34 and 64 (inclusive), it is categorized as’T. Storm’.
212 J. D. Bhosale et al.

• For wind speeds between 65 and 83 (inclusive), the function assigns the
label’Category 1’.
• If the wind speed falls between 84 and 95 (inclusive), it is labeled as’Category 2’.
• For wind speeds ranging from 96 to 113 (inclusive), the category assigned is’Category
3’.
• Wind speeds falling between 114 and 134 (inclusive) are classified as’Category 4’.
• For wind speeds greater than 134, the function assigns the label’Category 5’
Data Preprocessing. A critical component of analyzing and preparing cyclone data
for future examination is reading and processing the dataset to guarantee its suitability
for analysis. To do this, data is first loaded from files, which include image and label
databases containing critical information on cyclones. The dataset is then divided into
subsets to aid model evaluation, with k-fold cross-validation being used in particular.
An option for image augmentation is provided to enhance the dataset. This entails
creating new images and labels for each subset using image augmentation techniques.
Different augmentation procedures are used based on various parameters given by the
labels. If the label is between 50 and 75 (indicating a tropical storm), the code generates
two new images by flipping the original image horizontally and vertically. The modified
photos, together with their labels, are then put to the appropriate listings. Similarly, if the
label is between 75 and 100, rotation is used to generate six new images. The modified
photos and labels are attached to the respective lists once more. Furthermore, if the label
is equal to or greater than 100, the code generates 12 new images by rotating them. These
augmented photographs are added to the lists, along with their labels.
Augmentation increases the amount of training data available and has the potential
to improve the performance of subsequent models.
Following that, the subsets are joined to form sets for model iterations. Concatenating
the image and label data from the other subsets, except the current subset, yields the
training sets. Images and labels from other subsets are included if image augmentation
is enabled. The test set includes both the image and label data from the current subset.
Model Training. The model is designed to predict a continuous value for grayscale
images. The network takes a grayscale image with a resolution of 50 × 50 pixels as
input. The first convolutional layer consists of 32 3 × 3 filters, each using the ReLU
activation function. This is done by a layer of max-pooling of size 2 × 2, which reduces
the feature maps’ spatial dimensions by half. The second convolutional layer consists of
64 3 × 3 filters, each using the ReLU activation algorithm. Another max-pooling layer
of the same size is then placed.
There are 64 3 × 3 filters and ReLU activation function are used in the final convolu-
tional layer. This layer’s output is flattened and sent to a fully connected neural network.
To prevent overfitting, the fully linked layers include a dropout rate of 0.5. This is fol-
lowed by two dense layers, both of which make use of ReLU activation functions. The
output layer contains a single neuron without an activation function, allowing the model
to predict a continuous value for each input image.
The rmsprop optimizer is used for model optimization, with the mean squared error
loss function. During training, this optimizer changes the network’s weights to mini-
mize loss. During training, the model calculates two metrics to evaluate its performance
on the training set. These measures aid in spotting potential overfitting or underfitting
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 213

issues. Finally, the CNN model constructed using the Keras toolkit is wellsuited for pre-
dicting continuous values in grayscale photos. The architecture, rmsprop optimization,
and evaluation metrics all contribute to good training and performance evaluation on the
training set.
Validation of Results. The dataset was split into two sections, training and validation.
The model was then trained for a predetermined number of epochs and batch size. After
each epoch, the model’s performance on the validation set was assessed to see if it was
overfitting. An Early Stopping callback was implemented to stop training if the validation
error didn’t improve after a certain number of epochs, in order to stop overfitting and
guarantee that the model could be applied to new data. Based on the best validation error,
the best weights are restored. Following training, the model’s performance by epoch can
be displayed. The training and testing loss, as well as the mean absolute error (MAE) for
each epoch metric, are saved in a data frame for subsequent analysis, together with the
relevant epoch numbers. The measurements are then plotted on a graph with Seaborn’s
lmplot function, which fits exponential decay curves.
The epoch number is represented on the x-axis, and the metric values are represented
on the y-axis. There are two rows in the graph: one for loss and one for MAE. Using
distinct colors, each row distinguishes between training and testing data. The graph
is displayed, providing information about the model’s performance during the training
phase.

4 Algorithm Used

Convolutional Neural Network (CNN). The Convolutional Neural Network, or CNN,

is a deep learning model which has been designed to process data that is structured in
a grid-like form, such as images [21]. This model is based on the organization of the
visual brain in animals and attempts to learn hierarchical features, ranging from basic to
complex patterns, in an autonomous and adaptive way. The CNN consists of three types of
layers: convolutional, pooling, and fully connected. The convolution and pooling layers
are in charge of extracting meaningful features from the input data. The convolution layer
is made up of mathematical processes, specifically convolution, which entails applying
a kernel (a tiny grid of optimizable parameters) to each point in the input image grid (2D
array of pixel values). This allows CNNs to process images more effectively because
features can appear everywhere in the image.
Convolutional Neural Networks (CNNs) are powerful models used for grid-like data,
such as photographs, that draw from the concepts of the animal visual brain. Features
are collected and transmitted through additional layers, increasing in complexity in
a hierarchical manner. The fully connected layer then translates these features into a
final output, often used for classification. To train a CNN, the model’s parameters are
adjusted to reduce the difference between the predicted and reality, typically done with
backpropagation and gradient descent algorithms. In summary, CNNs are able to extract
and learn increasingly complicated characteristics with the use of convolution, pooling,
and fully connected layers.
214 J. D. Bhosale et al.

5 Result and Interfaces

The MAE and RMSE calculates the average size of errors by first squaring the differ-
ences between predictions and actual values, then taking the square root of the resulting
average. An RMSE value that is lower than the MAE value indicates that the model
is better at predicting the data. This is shown in Table 1. The column in Table 1 lists
the precise fold or subset of the data set that was utilized in the study’s validation. The
purpose of reporting both MAE and RMSE is to provide different perspectives on the
performance of the prediction model. The MAE over all folds for our cyclone intensity
prediction model was 9.96 knots. In our situation, an MAE of 9.96 knots shows that,
on average, the predictions of our model were 9.96 knots off from the actual cyclone
strength estimates.
The RMSE for our study was found to be 13.65 knots. A score of 13.65 knots
indicates that, on average, our model’s predictions were off by about 13.65 knots. The
RMSE provides an indication of the typical forecast error (Fig. 2).

Fig. 2. Mean Absolute Error in Neural Network’s Predictions by Category

Figure 1 visualizes the mean absolute error in the Hurricane’s strength which are
divided into seven categories. Our model performed best on the cyclones that were
categorized into the Depression type with a MAE of and performed unfavorable on
the Category 5 type cyclone with a MEA. The following Table 1 gives a summary of
validation folds of MAE and RMSE.
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 215

Table 1. Contrasting the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE)
across various folds.

Validation Fold MAE RMSE

(Mean Absolute Error) (Root Mean Squared Error)
I 12.8365 knots 16.2777 knots
II 10.4333 knots 13.2222 knots
III 10.4146 knots 13.2352 knots
IV 9.9171 knots 12.6129 knots
V 9.4262 knots 12.4232 knots
All Folds 9.96 knots 13.65 knots

6 Conclusion and Future Scope

Concluding this research paper; it has been recommended that utilizing deep learning
methods can accurately estimate the severity level of hurricanes employing satellite
image sets with subsequent analysis returning impressive results - due primarily to Con-
volutional Neural Networks’ use within storm modelling applications. The effective
extraction of valuable data from satellite imaging permits the early detection of inten-
sifying hurricane strengths that can lead to reduced adverse impacts on communities.
Through comprehensive measuring techniques, we were able to test and verify this
technique’s effectiveness.
These findings reinforce and emphasize the significance of deep learning methods
in disaster management preparation. Considering the potential future work, First, more
sophisticated deep learning architectures that can capture temporal relationships and
improve the model’s capacity to analyse sequential satellite images, such as recurrent
neural networks (RNNs) or attention mechanisms, can be incorporated into the proposed
CNN model. Utilizing additional data sources, such as oceanic and atmospheric data,
can also deliver supplementary information for a more precise estimation of cyclone
intensity. Additionally, the study can be expanded to incorporate areas and basins than
the Atlantic and Pacific, increasing the model’s generalizability. Last but not least, efforts
can be made to combine real-time data streams and create an operational system for
ongoing cyclone intensity monitoring and prediction, enabling prompt response and
proactive mitigation techniques.

References
1. Kar, C., Kumar, A., Banerjee, S.: Tropical cyclone intensity detection by geometric features
of cyclone images and multilayer perceptron. SN Appl. Sci. 1, 1099–1099 (2019). https://doi.
org/10.1007/s42452-019-1134-8
2. Chen, Z., Yu, X., Chen, G., Zhou, J.: Cyclone intensity estimation using multispectral imagery
from the FY-4 satellite. In: 2018 International Conference on Audio, Language and Image
Processing (ICALIP), pp. 46–51 (2018). https://doi.org/10.1109/ICALIP.2018. 8455603
216 J. D. Bhosale et al.

3. Chen, G., Chen, Z., Zhou, F., Yu, X., Zhang, H., Zhu, L.: A Semisupervised Deep Learning
Framework for Tropical Cyclone Intensity Estimation. In: 10th International Workshop on
the Analysis of Multitemporal Remote Sensing Images (MultiTemp), pp. 1–4 (2019). https://
doi.org/10.1109/Multi-Temp.2019.8866970
4. Ritchie, E.A., Wood, K.M., Rodríguez-Herrera, O.G., Piñeros, M.F., Tyo, J.S.: Satellite-
Derived Tropical Cyclone Intensity in the North Pacific Ocean Using the Deviation-Angle
Variance Technique. Wea. Forecasting 29, 505–516 (2014). https://doi.org/10.1175/WAF-D-
13-00133.1
5. Maskey, M., et al.: Deepti: deep learning-based tropical cyclone intensity estimation system.
JJ Miller IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
13, 2020–2020. https://doi.org/10.1109/JSTARS.2020.3011907,2020
6. Piñeros, M.F., Ritchie, E.A., Tyo, J.S.: Estimating tropical cyclone intensity from infrared
image data. Wea. Forecasting 26, 690–698 (2011). https://doi.org/10.1175/WAF-D-10-050
62.1
7. Bankert, R.L., Tag, P.M.: An Automated Method to Estimate Tropical Cyclone Intensity Using
SSM/I Imagery. J. Appl. Meteor. Climatol 41, 461–472 (2002). https://doi.org/10.1175/1520-
0450(2002)041<0461:AAMTET>2.0.CO;2
8. Shunsuke, H., Tetsuo, N.: Estimation of Tropical Cyclone’s Intensity Using TRMM/TMI
Brightness Temperature Data. J. Meteorolog. Soc. Japan. Ser. II 85(4), 437–454 (2007).
https://doi.org/10.2151/jmsj.85.437
9. Pradhan, R., Aygun, R.S., Maskey, M., Ramachandran, R., Cecil, D.J.: Tropical cyclone
intensity estimation using a deep convolutional neural network. IEEE Transactions on Image
Processing 27, 692–702 (2018). https://doi.org/10.1109/TIP.2017.2766358
10. Lee, J., Im, J., Cha, D.-H., Park, H., Sim, S.: Tropical Cyclone Intensity Estimation Using
Multi-Dimensional Convolutional Neural Networks from Geostationary Satellite Data (2019).
https://doi.org/10.3390/rs12010108
11. Zhang, J., Wang, X.-J., Ma, L.-M., Lu, X.-Q.: Tropical cyclone intensity classification and
estimation using infrared satellite images with deep learning. IEEE Journal of Selected Topics
in Applied Earth Observations and Remote Sensing 14, 2070–2086 (2021). https://doi.org/
10.1109/JSTARS.2021.3050767
12. Zhuo, J., Tan, Z.: Physics-augmented deep learning to improve tropical cyclone intensity and
size estimation from satellite imagery. Mon. Wea. Rev. 2097–2113 (2021). https://doi.org/10.
1175/MWR-D-20-0333.1
13. Tan, J., Yang, Q., Hu, J., Huang, Q., Chen, S.: Tropical Cyclone Intensity Estimation Using
Himawari-8 Satellite Cloud Products and Deep Learning (2022). https://doi.org/10.3390/rs1
4040812
14. Zhang, Z., et al.: A neural network with spatiotemporal encoding module for tropical cyclone
intensity estimation from infrared satellite image (2022). https://doi.org/10.1016/j.knosys.
2022.110005
15. Xu, G., et al.: TFG-Net: Tropical Cyclone Intensity Estimation from a Fine-grained perspec-
tive with the Graph convolution neural network. Eng. Appl. Artif. Intel. 118 (2023). https://
doi.org/10.1016/j.engappai.2022.105673
16. Chen, Z., Yu, X.: A novel tensor network for tropical cyclone intensity estimation. IEEE
Trans. Geosci. Remote Sens. 59(4), 3226–3243 (2021). https://doi.org/10.1109/TGRS.2020.
3017709
17. Olander, T.L., Velden, C.S.: The Advanced Dvorak Technique (ADT) for Estimating Tropical
Cyclone Intensity: Update and New Capabilities, pp. 905–922 (2019). https://doi.org/10.1175/
WAF-D-19-0007.1
18. Wang, G., et al.: Tropical cyclone intensity estimation from geostationary satellite imagery
using deep convolutional neural networks. IEEE Trans. Geosci. Remote Sens. 60, 1–16 (2022).
https://doi.org/10.1109/TGRS.2021.3066299
 Enhancing Cyclone Intensity Prediction Through Deep Learning Analysis 217

19. Zhang, R., Liu, Q., Hang, R.: Tropical cyclone intensity estimation using two-branch convo-
lutional neural network from infrared and water vapor images. IEEE Trans. Geosci. Remote
Sens. 58, 586–597 (2020). https://doi.org/10.1109/TGRS.2019.2938204
20. Tian, W., Huang, W., Yi, L., Wu, L., Wang, C.: A CNN-based hybrid model for tropical cyclone
intensity estimation in meteorological industry. IEEE Access 8, 59 158–59 168 (2020). https://
doi.org/10.1109/ACCESS.2020.2982772
21. Yamashita, R., Nishio, M., Do, R.K.G.: Convolutional neural networks: an overview and
application in radiology. Insights Imaging 9, 611–629 (2018). https://doi.org/10.1007/s13
244-018-0639-9
 Revolutionizing Road Safety: Machine Learning
Approaches for Predicting Road Accident
Severity

Meenakshi Malik1(B) , Rainu Nandal2 , and Rita Chhikara3

1 Department of Computer Science and Engineering, BML Munjal University, SOET, Gurgaon,
India
[email protected]
2 Department of Computer Science and Engineering, U.I.E.T, Maharshi Dayanand University,
Rohtak, India
[email protected]
3 Department of CSE, The NorthCap University, Gurgaon, India

[email protected]

Abstract. Millions of people die every year from road accidents, which are
becoming more commonplace worldwide. They cost society heavily in terms of
money and the economy. Road accident prediction has primarily been researched
as a classification problem in the literature, meaning that the goal is to forecast
whether or not a traffic mishap will occur in future deprived of delving into the intri-
cate interactions between the many components that contribute to these incidents.
Fewer studies have examined a subcategory of models of collaborative machine
learning modals than majority of research conducted to date on the significance
of subsidizing aspects for road accidents and respective severity. Thus, using the
road accident dataset as a basis, we assessed a number of ML model in this study
for forecasting the traffic accidents’ severity. Additionally, we have examined the
anticipated outcomes and utilized an explainable machine learning (XML) app-
roach to assess the significance of contributing factors to traffic accidents. This
work has taken into consideration various ensembles of ML models, like Random
Forest (RF), Extra Tree, Extreme Gradient Boosting (XGBoost), to forecast road
accidents with varying injury severity. Extra tree is the top classifier, according to
the comparison results, with 81.06% accuracy, 81.06% recall,75.91% precision,
and 77.34% of F1-Score models.

Keywords: Machine Learning (ML) · Accident severity · XGB · Random Forest

1 Introduction
In today’s rapidly evolving society, a substantial surge has been noticed in the auto-
mobiles volume, leading to a significant rise in traffic accidents with considerable life
and economic consequences (Micheale [1]). The WHO’s reports that annually, traffic
mishaps claim the lives of millions of individuals, while nonfatal accidents also impact

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 218–225, 2025.
https://doi.org/10.1007/978-3-031-71481-8_17
 Revolutionizing Road Safety: Machine Learning Approaches 219

a staggering number of people (Bahiru et al. [2]). It is evident that traffic mishaps are
emerging as a prominent global cause of fatalities and injuries. Consequently, the pre-
vention and prediction of traffic accidents have become focal points in both areas; traffic
sciences and intellectual vehicle research. The traffic mishaps severity as a crucial metric
for assessing their harm is influenced by various factors. Numerous studies on traffic
accidents have referenced diverse algorithms and factors. Lu et al. [3] conducted an
analysis considering the location of a car, road safety rating, surface quality, visibil-
ity, car condition, and driver condition, resulting in the establishment of a prediction
accuracy model of 86.67%. Additionally, Ren et al. [4] utilized a methodology of deep
learning to develop a model of short-term traffic accident risk prediction based on factors
such as air pollution, weather, traffic accidents, and traffic flow. Ren et al. [5] proposed
spatio-temporal correlation of vehicle accidents in city accidents risk predictions.
Huang et al. [6] introduced a temporal aggregation neural network layer, that
autonomously catches relation scores from the temporal dimensions to forecast traffic
accident occurrences. Murphey et al. [7] very efficiently diagnosed driver behaviour with
the help of data mining methods. In order to prevent and predict traffic mishaps plenty
of algorithms have been applied, with the random forest algorithm and it is gaining pop-
ularity in recent years. This algorithm, widely employed in diverse fields like medicines
(Iwendi et al. [8]), meteorology area (Ding et al. [9]), and statistics area (Schonlau and
Zou [10]). It has shown good output in addressing traffic accidents analysis too.
Yan and Shen [11] utilized the random forest algorithm and Bayesian optimization
to explore the impact of influencial factors on traffic accidents’ severity. Zhao et al.
[12] projected an accident risk prediction algorithm which was based on a random
forest and deep convolutional neural network. Chen and Chen [13] evaluated prediction
performances using accuracy, specificity and sensitivity, identifying the best effective
prediction models and input variables with the higher positive impact on these metrics.
Koma et al. [14] applied the random forest algorithm to identify cognitive driver
distraction, focusing on various types of eye movement. Wang et al. [15] examined var-
ious factors such as time period, grade of road, lanes, closeness to infrastructure, and
mishap-prone areas as pointers influencing traffic. The experimental outcomes demon-
strated the efficacy in avoiding congested roads and determining high-speed routes.
Traffic accidents typically stem from the interplay of human factors, vehicle dynam-
ics, road conditions, and the environment. Leveraging the random forest algorithms’
commendable capabilities, our paper highlighted assessment of impact of various acci-
dent reasons on the traffic mishaps severity. Subsequently, the model was employed for
predicting the traffic accidents’ severity and conclusive prediction results highlight the
superior performance of the proposed traffic accident elements in forecasting the traffic
accidents’ severity.

2 Related Work

Our paper concentrates on forecasting accident severity through the examination of his-
torical accident data. To achieve similar objectives, researchers globally have explored
diverse methodologies across various domains and data science (data analysis and
machine learning model construction) emerged as one of the most prevalent domains.
220 M. Malik et al.

For prediction and classification, statistical analysis and ML algorithms were applied.
Systematically, multiple stages have been executed, including data collection and prepa-
ration for ML model construction. The culmination involved result comparison to
derive a precise model tailored to the identified problem. In [16], the authors integrated
“Convolutional Neural Network and Random Forest” coining the approach as RFCNN.
To predict Traffic-Related Accident (TRA) severity, the researchers employed Fully
Connected Neural Network (FCNN) on a dataset comprising accident records spanning
between 2016–2020 in the United States. Subsequently, they successfully constructed
a highly accurate model, contributing to improved decision-making processes and road
management. In [17] authors adopted Artificial Neural Network (ANN) for predicting
traffic mishaps harm severity, utilizing a dataset encompassing records of accidents.
Initially comprising forty-eight attributes, counting target variable—a categorical vari-
able with 4 classes: minor, moderate, severe, and death. The dataset underwent pre-
processing, resulting in a reduction to sixteen features. The ANN classifier attained an
accuracy rate of 0.74. [18] authors analysed traffic mishaps in Bangladesh, employing
Machine Learning (ML) techniques on an incident dataset to assess accident severity.
Moreover, the researchers conducted a comprehensive comparative analysis, evalu-
ating the outcomes obtained from four machine learning (ML) algorithms through two
distinct experiments. In the initial experiment, the focus was on categorizing accident
severity into four classes: “Fatal, Grievous, Simple Injury, Motor Collision.“ Subse-
quently, in the following experiment, the target variable underwent a transformation
into “Fatal/Grievous,” with Grievous representing the amalgamation of the last three
classes. In their study, the authors [19] devised a robust framework based on six diverse
machine learning algorithms: Decision Tree, Naïve Bayes, Random Forest, Logistic
Regression, AdaBoost, and Bagging. The primary objective of this framework was to
predict crash severity, effectively distinguishing between fatal, serious, or minor inci-
dents. The intended implementation aimed to make meaningful contributions to road
safety and traffic control by providing valuable insights for road authorities.

3 Methodology for Road Traffic Accident Predictions

This section will delve into the classifiers and dataset employed for Highway Severity
Forecast. The projected method of this research work, illustrated in Fig. 1, outlines
the data and workflow. The classifiers play a critical part during prediction of road
accidents’ severity, and their selection is pivotal in the effectiveness of the predictive
models. Additionally, the dataset used in this study serves as the foundation for training
and evaluating the machine learning algorithms. The clear depiction in Fig. 1 provides
a visual representation of the structured approach adopted in this research, guiding the
subsequent discussions on the outcomes and insights gained from the analysis.

3.1 Random Forest (RF)

Data mining technology encompasses various techniques such as association, prediction,

classification, clustering, sequential pattern mining, and more [21]. Our work employs a
random forest algorithm for prediction of traffic accidents’ severity. The random forest
 Revolutionizing Road Safety: Machine Learning Approaches 221

algorithm, projected by Breiman [22], is a composite classifier that incorporates multiple

decision trees. Known for its strengths in handling multidimensional data, the random
forest stands out as one of the leading classification algorithms in current practice.
Historically, researchers have introduced various ensemble techniques, with bagging
and boosting being prominent examples. Random Forest (RF), initially proposed by
Breiman [23], is a notable instance. RF algorithms, involving the construction of N
trees, iterate through four steps. Step 1 includes training the data using a bootstrap
dataset, a subset of the original dataset. In Step 2, trees are generated, followed by Step 3
where attributes are randomly shortlisted. The final step, Step 4, determines the ultimate
prediction through majority voting based on the results of the selected trees [24].

3.2 XGBoost Model

The XGBoost algorithm, is nothing but an artificial intelligence-integrated ML algo-
rithm introduced by Chen et al. [25], offers notable advantages such as rapid parallelism,
fault tolerance, controlled complexity and robust generalization capabilities. This algo-
rithm iteratively synthesizes multiple decision tree models having low accuracy into a
high-precision strong learner. It employs second-order Taylor expansion, and a regular-
ization term is incorporated into the loss function to manage the model’s complexity and
diminish overfitting possibility.

3.3 Extra Tree Classifier

The Extra Tree Classifier (ETC) distinguishes itself by employing a random subset of
features for splitting tree nodes, setting it apart from the Random Forest (RF) methodol-
ogy. While RF builds trees using the complete dataset and randomly selects cut points for
node splits, ETC takes a unique approach by incorporating a multi-linear approximation
rather than RF’s piecewise constant strategy. This heightened level of randomization in
ETC not only underscores its superior performance compared to RF but also brings the
added advantage of minimizing correlation among errors made by base learners. A study
documented in [26] showcases ETC’s remarkable accuracy, surpassing that of RF.

3.4 Evaluation Matrix

It can be calculated as the ratio of correct predictions to the total number of predictions.
No. of correct predictions
Accuracy = (1)
Total no of input samples
TP + TN
Accuracy = (2)
TP + TN + FP + FN
In this scenario, the symbols denoting true negative, true positive, false positive,
and false negative are TN, TP, FN, and FP, respectively. As elucidated in [27], their
interpretations are as follows.
• True Positives: Instances where our prediction is positive, and the actual output is
also positive.
222 M. Malik et al.

Fig. 1. Architecture of the proposed models

• True Negatives: Instances where our prediction is negative, and the actual output is
also negative.
• False Positives: Instances where we predicted positive, but the actual output is
negative.
• False Negatives: Instances where we predicted negative, but the actual output is
positive.

3.5 F1 Score
The F1 score, as a machine learning evaluation metric, gauges a model’s accuracy by
incorporating both precision and recall scores.
1
F1 = 2 ∗ (3)
1
precision + 1
recall

3.6 Precision
Precision indicates the frequency with which an ML model is accurate in predicting the
target class. It is calculated as the number of true positive predictions divided by the sum
of true positive and false positive predictions
TP
Precision = (4)
TP + FP

3.7 Recall
Recall assesses the ratio of correctly identified actual positive labels by the model.
TP
Recall = (5)
TP + FN
 Revolutionizing Road Safety: Machine Learning Approaches 223

4 Experiments and Results

In this segment, we showcase outcomes obtained from implementing ensemble mod-
els—RF, XGB and ETC—on the road accident dataset [28–30]. Figure 2 illustrates the
significant variables identified by the Extra Trees classifier. In our study, out of the 32
original feature set variables, Extra Trees identified 10 crucial feature sets. Table 1 pro-
vides a comparative analysis of accuracy, precision, recall, and F1 score by juxtaposing
the outcome of the ensemble models.

Table 1. Comparison results of all models.

Model Accuracy Precision Recall F

Extreme gradient boosting 79.68 76.17 79.68 77.64
Random Forest 79.7 76.39 79.7 77.69
Extra Tree 81.06 75.91 81.06 77.34

Fig. 2. Performance of the models

5 Conclusion
In conclusion, the escalating global incidence of road accidents, leading to millions
of annual fatalities, poses substantial economic and commercial burdens on humanity.
Prior researches predominantly framed prediction of road accident as a classification
224 M. Malik et al.

challenge, focusing on foreseeing whether an accident might occur without delving into
the intricate interplay of contributing factors. While some studies have explored the
significance of these factors concerning accident occurrence and severity, only a limited
number have investigated a subset of ensemble machine learning (ML) models.
This paper contributes to the field by evaluating various ML models to predict road
accident severity using a dedicated dataset. The study employs ensemble ML models
like Random Forest, Extra Tree and Extreme Gradient Boosting to predict accidents of
varying severity. The findings indicate that Extra Tree emerges as the most effective
classifier, boasting an 81.06% accuracy, 75.91% precision, 81.06% recall, and a 77.34%
F1-Score. Furthermore, the paper employs explainable ML techniques to assess the
significance of road mishaps causative elements, enhancing the interpretability of the
predictive models. This research underscores the potential of ensemble ML models,
particularly Extra Tree, in enhancing the accuracy of road accident severity predictions
and provides valuable insights into the underlying factors influencing these incidents.

References
1. Micheale, K.G.: Road trafc accident: human security perspective. Int. J. Peace and Develop.
Stud. 8(2), 15–24 (2017)
2. Bahiru, T.K., Singh, D.K., Tessfaw, E.A.: Comparative study on data mining classifcation
algorithms for predicting road trafc accident severity. In: Proceedings of the 2018 Sec-
ond International Conference on Inventive Communication and Computational Technologies
(ICICCT). IEEE, Coimbatore, India (2018)
3. Lu, T., Dunyao, Z.H.U., Lixin, Y., Pan, Z.: Te trafc accident hotspot prediction: based on
the logistic regression method. In: Proceedings of the 2015 International Conference on
Transportation Information and Safety (ICTIS). IEEE, Wuhan, China (2015)
4. Ren, H., Song, Y., Wang, J., Hu, Y., Lei, J.: A deep learning approach to the prediction of
short-term trafc accident risk (2017). https://arxiv.org/abs/1710.09543
5. Ren, H., Song, Y., Wang, J., Hu, Y., Lei, J.: A deep learning approach to the citywide trafc acci-
dent risk prediction. In: Proceedings of the 2018 21st International Conference on Intelligent
Transportation Systems (ITSC). IEEE, Maui, HI, USA (2018)
6. Huang, C., Zhang, C., Dai, P., Bo, L.: Deep dynamic fusion network for trafc accident
forecasting. In: Proceedings of the 28th ACM International Conference on Information and
Knowledge Management. Beijing, China (2019)
7. Murphey, Y.L., Wang, K., Molnar, L.J., et al.: Development of data mining methodologies to
advance knowledge of driver behaviors in naturalistic driving. SAE Int. J. Transport. Safety
8(2), 77–94 (2020)
8. Iwendi, C., Bashir, A.K., Peshkar, A., et al.: COVID-19 patient health prediction using boosted
random forest algorithm. Frontiers in Public Health 8 (2020)
9. Ding, J., Dai, Q., Fan, W., et al.: Impacts of meteorology and precursor emission change on
O3 variation in Tianjin, China from 2015 to 2021. J. Environ. Sci. 126, 506–516 (2023)
10. Schonlau, M., Zou, R.Y.: Te random forest algorithm for statistical learning. STATA Journal
20(1), 3–29 (2020)
11. Yan, M., Shen, Y.: Trafc accident severity prediction based on random forest. Sustainability
14(3) (2022)
12. Zhao, H., Li, X., Cheng, H., Zhang, J., Wang, Q., Zhu, H.: Deep learning-based prediction of
trafc accidents risk for Internet of vehicles. China Communications 19(2), 214–224 (2022)
 Revolutionizing Road Safety: Machine Learning Approaches 225

13. Chen, M.-M., Chen, M.-C.: Modeling road accident severity with comparisons of logistic
regression, decision tree and random forest. Information 11(5) (2020)
14. Koma, H., Harada, T., Yoshizawa, A., Iwasaki, H.: Detecting cognitive distraction using
random forest by considering eye movement type. Int. J. Cognit. Info. Natu. Intell. 11(1),
16–28 (2017)
15. Wang, L., Wu, J., Li, R., et al.: A weight assignment algorithm for incomplete trafc information
road based on fuzzy random forest method. Symmetry 13(9) (2021)
16. Manzoor, M., et al.: RFCNN: traffic accident severity prediction based on decision level
fusion of machine and deep learning model. IEEE Access 9, 128359–128371 (2021). https://
doi.org/10.1109/ACCESS.2021.3112546
17. Alkheder, S., Taamneh, M., Taamneh, S.: Severity prediction of traffic accident using an
artificial neural network: traffic accident severity prediction using artificial neural network. J.
Forecasting 36 (2016). https://doi.org/10.1002/for.2425
18. Labib, Md.F., Rifat, A.S., Hossain, Md.M., Das, A.K., Nawrine, F.: Road accident analy-
sis and prediction of accident severity by using machine learning in Bangladesh. In: 2019
7th International Conference on Smart Computing & Communications (ICSCC), pp. 1–5.
Sarawak, Malaysia, Malaysia (2019). https://doi.org/10.1109/ICSCC.2019.8843640
19. Malik, S., El Sayed, H., Khan, M.A., Khan, M.J.: Road accident severity prediction — a
comparative analysis of machine learning algorithms. In: 2021 IEEE Global Conference on
Artificial Intelligence and Internet of Things (GCAIoT), pp. 69–74 (2021). https://doi.org/
10.1109/GCAIoT53516.2021.9693055
20. Sharma, B., Katiyar, V.K., Kumar, K.: Traffic accident prediction model using support vector
machines with gaussian kernel. In: Proceedings of Fifth International Conference on Soft
Computing for Problem Solving, pp. 1–10. Singapore (2016). https://doi.org/10.1007/978-
981-10-0451-3_1
21. Hussain, S.: Survey on current trends and techniques of data mining research. London J. Res.
Comp. Sci. Technol. 17(1) (2017)
22. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
23. Breiman, L.: Random forests. Machine Learning 45(1), 5–32 (2001)
24. Sekhar, C.R., Madhu, E., et al.: Mode choice analysis using random forrest decision trees.
Transportation Research Procedia 17, 644–652 (2016)
25. Chen, T., Guestrin, C.: Xgboost: a scalable tree boosting system. In: Proceedings of the
22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining,
pp. 785–794. San Francisco, CA, USA (2016)
26. Geurts, P., Ernst, D., Wehenkel, L.: Extremely randomized trees. Machine Learning 63(1),
3–42 (2006)
27. Umer, M., Ashraf, I., Mehmood, A., Ullah, D.S., Choi, G.S.: Predicting numeric ratings for
google apps using text features and ensemble learning. ETRI J. (2020). https://doi.org/10.
4218/etrij.2019-0443
28. Bedane, T.T.: Road Traffic Accident Dataset of Addis Ababa City. Addis Ababa Science and
Technology University, Addis Ababa (2020)
29. Jaroli, P., Singla, C., Bhardwaj, V., Mohapatra, S.K.: Deep learning model based novel seman-
tic analysis. In: 2022 2nd International Conference on Advance Computing and Innovative
Technologies in Engineering (ICACITE), pp. 1454–1458. IEEE (2022)
30. Bhardwaj, V., Rahul, K.V., Kumar, M., Lamba, V.: Analysis and prediction of stock mar-
ket movements using machine learning. In: 2022 4th International Conference on Inventive
Research in Computing Applications (ICIRCA), pp. 946–950. IEEE (2022)
 Survey and Analysis of Machine Learning
Methods for Parkinson’s Disease Diagnosis

Poonam Yadav1 , Meenu Vijarania1 , Meenakshi Malik2(B) , and Ritu1

1 K.R. Mangalam University, Gurugram, India
2 BML Munjal University, Gurugram, India

[email protected]

Abstract. Parkinson’s disease (PD) is a debilitating state that hinders quality of

life. It results from the demise of dopamine-generating cells across the region of
substantia nigra of the central nervous system (CNS), that has an impact on the
body. Individuals with Parkinson’s disease experience difficulty speaking, writ-
ing, and walking. To distinguish between individuals with Parkinson’s disease and
those in good health, various machine learning-based approaches are employed.
A thorough assessment of methods based on machine learning for Parkinson dis-
ease prediction is presented in this work. With an emphasis on the application
of research findings in real-world clinical settings, this review critically evaluates
the usefulness of machine learning models regarding the practical identification
of Parkinson’s disease. In this work, literature analysis was carried out on pub-
lished papers in past years utilizing the PubMed, University of California, Irvine
(UCI) repository, and Xplore databases on Institute of Electrical and Electronics
Engineers (IEEE) offer an extensive summary among the techniques for machine
learning and data modalities used in differential diagnosis of Parkinson’s disease.
We have proposed methodology also based on our study.

Keywords: Parkinson’s disease · machine learning · datasets

1 Introduction
Parkinson’s disease (PD) is a neurological disorder which diminishes the capacity to
direct motion. Several different motor and non-motor symptoms are present, and it
normally progresses gradually over time. Resting tremors, rigidity of the muscles, and
bradykinesia (slowness of movement) are the hallmarks of Parkinson’s disease. Along
with these neurological symptoms and non-motor signs and symptoms like sadness,
altered cognition, and sleep disruptions are frequently present.
Sixty percent of those over 50 have Parkinson’s disease. Speech and movement issues
are common in patients with Parkinson’s (PWP) disease [1]. In Parkinson’s disease
(PD) patients, dysarthria is also noted. It is typified by fragility, immobility, and inabil-
ity to synchronize within the speech-motor system, which impacts breathing, phonetic
transcription, expression, and prosody [2].
Parkinson’s disease is generally brought on by a confluence of environmental trig-
gers, including exposure to toxins, oxidative stress, and mitochondrial malfunction,

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 226–238, 2025.
https://doi.org/10.1007/978-3-031-71481-8_18
 Survey and Analysis of Machine Learning Methods 227

as well as hereditary predisposition. This degeneration of dopamine-producing brain

cells is what ultimately causes Parkinson’s disease. Indications of Parkinson’s disease
encompass an assortment of both motor and non-motor signs, providing a comprehensive
picture of the diverse manifestations associated with the condition as shown in Fig. 1.

Fig. 1. Symptoms of Parkinson’s disease (two types of signs- motor and non-motor)

Since Parkinson’s disease cannot be cured, there is an urgent need for continued
research as well as care for those who are affected and their families. While there are
medications to control symptoms and enhance quality of life, finding a cure remains a
major goal of medical research worldwide. The signs of Parkinson’s disease are often
managed and controlled by medication, but for the purpose of managing and controlling
symptoms, additional therapies and treatments such as employment counselling, physical
rehabilitation, and speech-language therapy may also be beneficial [3]. Approximately
60,000 new clinical diagnoses of Parkinson’s disease (PD) are made each year, and the
neurodegenerative illness affects one million people in the United States of America
(US) [4].
Parkinson’s disease can be effectively managed earlier thanks to early detection,
which gives patients and healthcare professionals more power. It opens the door to a vari-
ety of treatments, interventions, and assistance programs that jointly enhance patients’
228 P. Yadav et al.

functional capacities and well-being while also possibly lessening the social and financial
burden brought on by the disease’s later stages.

2 Literature Review

Govindu, A. et al. (2023) [5] have presented technologies for machine learning in
telemedicine in order to train four machine learning models to identify Parkinson’s
disease (PD) in its initial phases. They used multi-dimensional voice program (MDVP)
audio data of thirty PWP and healthy individuals. The Random Forest classifier is the
best Machine Learning (ML) method for PD detection, according to a comparison of
the classification results obtained from Support Vector Machine (SVM), Random For-
est, K-Nearest Neighbors (KNN), and Logistic Regression models. The Random Forest
classifier model has a 0.95 sensitivity and a 91.83% detection accuracy. Since audio data
is insufficient as a biomarker for Parkinson’s disease categorization, they suggest using
both recordings of rapid eye movement (REM) sleep and audio in the future to enhance
the outcomes. With the usage of audio recordings on a mobile device, they intend to
further telecommuting in medicine classification of Parkinson’s disease (PD).
Shadi Moradi et al. (2022) [6] purposed this study to find genes related to the onset
of Parkinson’s disease. For additional study, they used the Gene Expression Omnibus
(GEO)-provided microarray dataset (GSE22491). To analyze and evaluate gene expres-
sion and find DEGs, the R software’s Limma package was utilized. They used the
DAVID online tool to execute the KEGG pathway and GO enlargement examination for
differentially expressed genes (DEGs). In addition, hub genes were identified by analyz-
ing the DEGs’ PPI network with Cytoscape and the gene-interacting-gene search tool
(STRING) database. In order to forecast the precision of genes, they used SVM clas-
sifier. The network of protein-protein interactions (PPIs) was built utilizing the DEGs
found by the R software’s Limma package, and it had 264 nodes and 502 edges. Fur-
thermore, the genes Gram-negative bacteria (GNB5), gluconeogenesis (GNG11), and
elastase, neutrophil expressed (ELANE) were found to be hubs. SVM discovered that
an 88% prediction accuracy may be attained by combining three genes. The three hub
genes GNB5, GNG11, and ELANE may be employed as PD biomarkers for detection,
per the results of the study. Furthermore, As stated by authors, the highly accurate SVM
conclusions may be used as PD biomarkers.
Mei J. et al. (2021) [7] examined materials published up until February 14, 2020,
in a study of the literature conducted utilizing the IEEE Xplore and PubMed databases.
Before being included, selected for relevant content, and presented in this review, a
total of 209 papers were reviewed for their goals, sources of data, data types, machine
learning approaches, and related outcomes. This study demonstrates a great deal of
promise for using cutting-edge machine learning methods and biomarkers to clinical
judgment, which will result in a more methodical and knowledgeable confirmation of
Parkinson’s disease.
T. J. Wroge et al. (2018) [4] suggested automated speech analysis to validate a
physician’s diagnosis because they rely on the patient’s self-report of their diagnosis
in this case. They investigate how well supervised classification algorithms similar to a
deep neural network work for correctly diagnosing patients with the illness. When using
 Survey and Analysis of Machine Learning Methods 229

a pathological post-mortem analysis as the reference point, their maximum accuracy

was 85% produced by the machine learning models surpasses the non-experts’ average
precision of clinical diagnosis was 73.8%. as well movement disorder specialists’ average
precision (79.6% without follow-up, 83.9% after follow-up).
W. Wang et al. (2020) [8] suggested using a deep learning model to automatically
distinguish between healthy individuals and those with Parkinson’s disease (PD) based
on premotor characteristics (such as olfactory loss, Disorder of sleep behavior (RBD),
and rapid eye movement (REM). The suggested model achieved an accuracy of 96.45%,
demonstrating high detecting capability. Their findings demonstrated that, when it came
to differentiating between patients with Parkinson’s illness as well as normal individuals,
the deep learning model that was built performed better at detection than the twelve
machine learning models that were taken into consideration. The writers claim that deep
learning outperforms the others. This is as a result of the fact that they created deep
learning using tiny PD data sets gathered from 584 people (401 early PD cases and 183
healthy cases). Nonetheless, as time goes on and data becomes larger and more complex,
deep learning should be able to prove its mettle.
Zehra Karapinar Senturk (2020) [9] suggested diagnosing Parkinson’s illness using
machine learning. The author’s method involves the stages of feature selection and classi-
fication. The feature selection task took into account the Recursive Feature Elimination
and Feature Importance approaches. In the tests, Parkinson’s patients were classified
using Support vector machines, artificial neural networks, and regression trees. It was
demonstrated that Recursive Feature Elimination with Support Vector Machines outper-
formed the other techniques. The lowest number of vocal features required to identify
Parkinson’s illness resulted in 93.84% accuracy.
Almeida, J. S. et al. (2019) [10] looks at how voice sounds are processed in order to
identify Parkinson’s disease. Their methodology assesses the application of four machine
learning approaches and eighteen feature extraction strategies for the classification of
data from continuous phonation and speech activities. The analysis of the classification
performance was done using five metrics: Accuracy, Specificity, Sensitivity, and Equal
Error Rate (EER) and Area Under Curve (AUC) values from curves representing charac-
teristics of the receiver operation and detection error trade-off (DET). They evaluate this
method against alternative methods utilizing identical data sets. They demonstrate that
phonation tasks were superior to speech tasks in terms of disease detection efficiency.
The accuracy, AUC, and EER of the acoustic cardioid (AC) channel with the best perfor-
mance were 94.55%, 0.87, and 19.01%, respectively. We have obtained 92.94% accuracy,
0.92 AUC, and 14.15% EER when utilizing the smartphone (SP) channel.
Thakur, K. et al. (2023) [11] found that one of the disease’s common symptoms is
voice deterioration. In order to diagnose Parkinson’s disease and assess performance,
the authors of this research build and contrast cutting-edge machine learning techniques
like Random Forest and Support Vector Machine, Decision Tree Classifier, and Extra
Trees Classifier. Their approach, which outperformed all other models for Parkinson’s
disease prediction, suggests combining the Extra Trees Classifier with entropy-based
feature selection. The greatest results are 94.34% accuracy, 0.9388 precision, 0.9684
F1-score, and 1 recall value.
230 P. Yadav et al.

Ranjan, N. M. et al. (2023) [12] suggested a system that uses wave and spiral draw-
ings of both healthy and Parkinson’s disease patients as input and two different forms of
handwriting analysis. They employed the oriented gradient histogram for feature extrac-
tion. The method that was created employed a random forest classifier in conjunction
with a machine learning algorithm to identify Parkinson’s disease in patients. With spiral
drawing, their model’s accuracy was 86.67%, while with wave drawing, it was 83.30%.
Alalayah KM et al. (2023) [13] suggested innovative methods for enhancing the
methods for PD early diagnosis through the assessment of particular aspects and ML
algorithm hyperparameter tuning for PD diagnosis based on voice abnormalities. The
recursive feature elimination (RFE) approach sorted features in light of their impact on
the desired attribute, while the synthetic minority oversampling technique (SMOTE)
balanced the dataset. To lower the dataset’s dimensions, they utilized principal compo-
nent analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE). Support
vector machines (SVM), random forests (RF), decision trees (DT), K-nearest neighbors
(KNN), as well as multilayer perception (MLP) are examples of classifiers were ulti-
mately fed the final characteristics using PCA and t-SNE. Furthermore, MLP using the
PCA algorithm produced results with an F1-score of 96.66%, 98% accuracy and 96.66%
precision and recall of 96%.
J. Divya et al. (2023) [14] discovered that to ascertain if an individual has Parkinson’s
disease or not, it is straightforward for implementing a system of machine learning that
evaluates the variations in their speech pattern. One of the possible outcomes of the
author’s algorithm’s regular use of Extreme Gradient Boosting (XGBoost) classifiers and
parameter extraction is the audio featured dataset, which can be found in the UCI dataset
database. In terms of assessing the palladium patient’s state of health, the algorithm has
produced a result that is significantly more precise. XGBoost produced the maximum
level of precision that could be attained.
Yuan, L. et al. (2023) [15] used 252 participants’ speech sounds as the database.
To enhance the accuracy of Parkinson’s disease (PD) classification, this study employed
machine learning algorithms with linguistic signal data as input. The generated classifiers
were then merged. The testing results showed that these machine learning algorithms
might achieve up to 95% diagnosis accuracy. Furthermore, a clinical experience-based
feature extraction method for subject language signal analysis was presented.
T. Wasif et al. (2021) [16] used five distinct classifiers. The objective of this research is
to optimize and improve the feature employed in the dataset. For optimization, they com-
bined the wrapper and filter approaches into a hybrid feature selection method. Recur-
sive Feature Elimination is the name of the technique. XGBoost achieved the greatest
accuracy of 97.43% using nine out of 22 features in its classifier.
Vinora, A. et al. (2023) [17] provided a brief evaluation of the best available at the
time and suggested prospective directions for future study. Since tremor classification
and detection are essential to identify and treatment of PD patients, machine learning
techniques can be applied in this situation. One of the most prevalent movement disorders
in clinical practice, it is commonly classified based on behavioral or etiological char-
acteristics. Determining and assessing gait patterns associated with Parkinson’s disease
(PD), such as gait start and freezing, which are indicative of PD symptoms, is another
 Survey and Analysis of Machine Learning Methods 231

crucial difficulty. Since vocal impairment affects 90% of Parkinson’s disease patients,
speech data analysis is essential to distinguishing the condition from healthy persons.
Yadav, D. et al. (2022) [18] discovered that it impacts a person’s speech in the early
stages. Six machine learning models were compared using six measurements (recall,
accuracy, precision, F1-score, false positive rate, and area under the Receiver Operating
Characteristic (ROC) curve): Decision trees, Support Vector Machine, Random Forest,
Logistic Regression, K Nearest Neighbors, as well as Gaussian Naïve Bayes.
Johri, A. et al. (2019) [19] presented the VGFR Spectrogram Detector and Voice
Impairment Classifier, two models based on neural networks designed to aid medical
professionals and patients in early disease diagnosis. Convolutional neural networks, or
CNNs, have undergone a thorough empirical examination in order to forecast disease.
Deep dense artificial neural networks, or ANNs, have been utilized for voice recordings
and substantial image categorization of gait information converted into spectrogram
pictures. While the accuracy of the Voice Impairment Classifier is 89.15% and the
classification accuracy recorded by the Spectrogram Detector (VGFR) is 88.1%.

3 Decoding the Frameworks: An In-Depth Examination

of Machine Learning Methodologies
3.1 Support Vector Machines

A supervised machine learning method called Support Vector Machine (SVM), is

employed for both regression as well as classification. It’s a technique that works with
both non-linear and linear data [20]. Finding the optimal hyperplane in an N-dimensional
space to partition data points into different feature space classes are the primary objective
of the SVM method. As shown in Fig. 2, SVM chooses the extreme points and vectors
to aid in the creation of the hyperplane. The hyperplane aims to maximize the margin
between the closest points of various classes. Its dimension is determined by the quantity
of attributes; if there are two input aspects, it is a line; if there are three input features,
it is a two-dimensional (2-D) plane. Using features taken from clinical evaluations and
voice recordings, among other medical data, supervised learning model (SVM) identify
patients as positive or negative for Parkinson’s disease.

3.2 Decision Trees

The decision tree (DT) is best supervised learning methods for classification and regres-
sion applications. A decision tree is a tree-structured classifier in which the outcomes
are represented by each leaf node, the inside nodes reflect the characteristics of a dataset,
and the branches indicate the decision rule. The “highest decision” characteristic, which
forms the tree’s root node at first, is the crucial component in the DT creation process
[14]. It is constructed by continuously splitting the training data into subsets based on
the attribute values, continuing this process until a halting condition is met, such as the
minimal number of samples required to divide a node or the maximum tree depth. A
measurement, such as Gini impurity or entropy, which assesses the level of impurity
or unpredictability within the subsets, is employed by the Decision Tree approach to
232 P. Yadav et al.

Fig. 2. Support Vector Machine (audio recordings used as test and training data)

ascertain the ideal attribute to divide the data in training. Formula to calculate entropy
is:
Entropy(S) = −Prob(Y)log2 Prob(Y) − Prob(N)log2 Prob(N) (1)
where;
S = Total no of samples
Prob(Y) = Probability of Y
Prob(N) = Probability of N
The goal is to determine which property best maximizes the information gain or the
impurity decrease after the split. This model helps to categorize individuals into groups
based on whether patient have Parkinson’s disease or not.

3.3 Random Forest

The Random Forest relies on ensemble learning, aggregating the output of multiple
classifiers and selecting a candidate based on a majority vote [21]. The random forest’s
output for classification tasks is the class that the majority among the trees chose. In
regression tasks, the average prediction, or mean, is given back for each individual
tree. It is used to find patterns suggestive of Parkinson’s disease by analyzing a variety
of datasets, including those obtained from speech and gait studies. Figure 3 shows
the construction of multiple decision trees during training to create the class mode for
classification or the mean prediction for regression, known for its robustness and high
predictive accuracy.
 Survey and Analysis of Machine Learning Methods 233

Fig. 3. Random Forest Algorithm used numerous decision trees for classification or regression

3.4 Neural Networks

The basis of deep learning [26] techniques is neural networks. Additionally, they are men-
tioned to as artificial neural networks (ANNs) or simulated neural networks (SNNs). One
portion of machine learning is neural networks. Neural networks are made to resemble
the composition and capabilities of the human brain. by modeling them accordingly.
The effective node layers of the neural network are composed of an output layer, a
single or several hidden levels, and an input layer. Every node in this network links to
other nodes and has a threshold and weight assigned to it. Artificial neural systems and
234 P. Yadav et al.

other in-depth education approaches can process complex data, including genetic infor-
mation and brain imaging Functional Magnetic Resonance Imaging (fMRI), Magnetic
Resonance Imaging (MRI) etc. to find minute patterns or signs linked to Parkinson’s dis-
ease. Patient data may be subjected to a variety of procedures, including examination,
segmentation, augmentation, scaling, normalization, sampling, aggregation, and sifting,
to be able to produce precise predictions which, benefit the healthcare system and its
interested parties [22].

3.5 Time Series Analysis

A machine learning method called time series forecasts target values only from known
past target values. In the literature, this particular type of regression is referred to as
auto-regressive modeling. Time series in Fig. 4 refers to a sequence of data points
indexed in chronological order, typically at uniform intervals, allowing for the analysis
of trends, patterns, and behaviors over time, often applied in various fields such as
finance, economics, and environmental science. In order to monitor changes in motor
performance over time, temporal data, such as sensor readings from wearable devices,
can be analyzed using time series analysis techniques like autoregressive modelling or
recurrent neural networks.

Fig. 4. Time series analysis for Parkinson’s disease diagnosis

 Survey and Analysis of Machine Learning Methods 235

3.6 Dimensionality Reduction

It is a technique for transforming higher dimensional datasets into lower dimensional

datasets while maintaining information similarity. This method is frequently applied in
machine learning to solve regression and classification problems and produce a more
well-fitting predictive model. It is the process of converting high-dimensional data while
maintaining the necessary components of the source data into a lower-dimensional con-
text. An unsupervised method for reducing dimensionality that might expedite train-
ing and testing is Principal Component Analysis (PCA) [23]. PCA can facilitate the
identification of important elements in the diagnosis of Parkinson’s illness.

3.7 Ensemble Learning

Multiple machine learning models are combined into a single model through ensemble
learning. The model’s performance is to be improved. While stacking seeks to increase
prediction accuracy, bagging aims to lower variation., and boosting aims to lessen prej-
udice. The predictions from various models are counted independently. As a result, the
forecast produced by the majority of models is taken as final. Higher accuracy is achieved
by using ensemble approaches, which capitalize on the diversity and complementary
features of individual models [24].

3.8 Cross-Validation

Cross validation is a machine learning technique that evaluates a model’s performance

on missing data. A number of folds, or subsets, are created from the given data; the
model is trained on one-fold and validated on the other folds. This process here is
carried out more than once, along with a new fold acting as the constant validation
set. The results of each validation phase are averaged at the conclusion to get a more
trustworthy assessment of the model’s performance. To evaluate how well machine
learning [27] models generalize to the data and make sure they are not overfitting, cross-
validation techniques like the k-fold cross-validation method are crucial. The outcomes
are validated using the cross-validation technique [25].

4 Proposed Methodology

Our proposed methodology is based on a thorough analysis of prior research on machine

learning-based Parkinson’s disease detection, with a focus on dimensionality reduction
and feature refinement techniques. By incorporating well-known algorithms such as Gra-
dient Boosting and Support Vector Machines, we hope to enhance the interpretability and
the model’s accuracy. Furthermore, motivated by effective group tactics, our methodol-
ogy investigates the combination of complimentary models to establish a strong foun-
dation for precise and timely identification of Parkinson’s disease. The accompanying
Fig. 5 provides a visual representation of the proposed methodology, highlighting the
steps and connections that need to be adhered to in order to implement the analysis-
based approach. The Parkinson’s disease dataset used in this flowchart is sourced from
236 P. Yadav et al.

PubMed and the UCI repository. The dataset includes handwriting, neuroimaging, voice
recording, gait, and Parkinson’s disease. The data is preprocessed using feature extrac-
tion and feature selection techniques. SVM, Random Forest, and ANN models etc. were
chosen, and three phases of training—training, testing, and validation—were conducted
using various metrics. Optimizing a performance of the model without overfitting or
generating an excessively high variance is referred to as tuning. In machine learning,
this is achieved by choosing suitable “hyperparameters”. Selecting the right set of hyper-
parameters can be computationally demanding, but it is essential for model accuracy.
In contrast to other model parameters, the model does not learn hyperparameters on its
own via training methods. These parameters must be manually set instead. There are
numerous ways to choose the right hyperparameters. By employing all-encompassing
approach, we hope to further the progress of precise and effective machine learning-
based Parkinson’s disease diagnosis. Below is the diagram of proposed methodology
based on analysis.

Fig. 5. Proposed methodology based on detailed review

5 Conclusion
Machine learning-based Parkinson disease prediction model was thoroughly reviewed
in this research. A succinct overview of the several computational intelligence-based
methods for Parkinson disease prediction is provided. In order to forecast Parkinson
 Survey and Analysis of Machine Learning Methods 237

illnesses, a synthesis of the findings from numerous researchers is also provided in

the literature. In our next research project, we hope to introduce a hybrid model that
combines the advantages of several algorithms, providing a more thorough and reliable
approach to early Parkinson’s disease detection by utilizing the lessons learned from
successful frameworks in the literature.

References
1. Prabhavathi, K., Patil, S.: Tremors and Bradykinesia. Techniques for Assessment of
Parkinsonism for Diagnosis and Rehabilitation, 135–149 (2022)
2. Fraser, K.C., Meltzer, J.A., Rudzicz, F.: Linguistic features identify Alzheimer’s disease in
narrative speech. J. Alzheimers Dis. 49(2), 407–422 (2016)
3. Heisters, D.: Parkinson’s: symptoms, treatments and research. British Journal of Nursing
20(9), 548–554 (2011)
4. Wroge, T.J., et al.: Parkinson’s disease diagnosis using machine learning and voice. In: 2018
IEEE signal processing in medicine and biology symposium (SPMB), pp. 1–7. IEEE (2018)
5. Govindu, A., Palwe, S.: Early detection of Parkinson’s disease using machine learning.
Procedia Comp. Sci. 218, 249–261 (2023)
6. Moradi, S., Tapak, L., Afshar, S.: Identification of novel noninvasive diagnostics biomarkers
in the Parkinson’s diseases and improving the disease classification using support vector
machine. BioMed Research International 2022 (2022)
7. Mei, J., Desrosiers, C., Frasnelli, J.: Machine learning for the diagnosis of Parkinson’s disease:
a review of literature. Frontiers in Aging Neuroscience 13, 633752 (2021)
8. Wang, W., Lee, J., Harrou, F., Sun, Y.: Early detection of Parkinson’s disease using deep
learning and machine learning. IEEE Access 8, 147635–147646 (2020)
9. Senturk, Z.K.: Early diagnosis of Parkinson’s disease using machine learning algorithms.
Med. Hypotheses 138, 109603 (2020)
10. Almeida, J.S., et al.: Detecting Parkinson’s disease with sustained phonation and speech
signals using machine learning techniques. Pattern Recogn. Lett. 125, 55–62 (2019)
11. Thakur, K., Kapoor, D.S., Singh, K.J., Sharma, A., Malhotra, J.: Diagnosis of parkinson’s
disease using machine learning algorithms. In: Congress on Intelligent Systems, pp. 205–217.
Springer Nature Singapore, Singapore (2022)
12. Ranjan, N.M., Mate, G., Bembde, M.: Detection of parkinson’s disease using machine learning
algorithms and handwriting analysis. J. Data Mining and Manage 8(1), 21–29 (2023). e-ISSN:
2456-9437
13. Alalayah, K.M., Senan, E.M., Atlam, H.F., Ahmed, I.A., Shatnawi, H.S.A.: Automatic and
early detection of parkinson’s disease by analyzing acoustic signals using classification
algorithms based on recursive feature elimination method. Diagnostics 13(11), 1924 (2023)
14. Divya, J., et al.: Detection of parkinson disease using machine learning. In: 2023 International
Conference on Inventive Computation Technologies (ICICT), pp. 53–57. IEEE (2023)
15. Yuan, L., Liu, Y., Feng, H.M.: Parkinson disease prediction using machine learning-based
features from speech signal. Ser. Orient. Comp. Appl. 1–7 (2023)
16. Wasif, T., Hossain, M.I.U., Mahmud, A.: Parkinson disease prediction using feature selec-
tion technique in machine learning. In: 2021 12th International Conference on Computing
Communication and Networking Technologies (ICCCNT), pp. 1–5. IEEE (2021)
17. Vinora, A., Ajitha, E., Sivakarthi, G.: Detecting parkinson’s disease using machine learning.
In: 2023 International Conference on Artificial Intelligence and Knowledge Discovery in
Concurrent Engineering (ICECONF), pp. 1–6. IEEE (2023)
238 P. Yadav et al.

18. Yadav, D., Jain, I.: Comparative analysis of machine learning algorithms for parkinson’s dis-
ease prediction. In: 2022 6th International Conference on Intelligent Computing and Control
Systems (ICICCS), pp. 1334–1339. IEEE (2022)
19. Johri, A., Tripathi, A.: Parkinson disease detection using deep neural networks. In: 2019
Twelfth international conference on contemporary computing (IC3), pp. 1–4. IEEE (2019)
20. Bind, S., et al.: A survey of machine learning based approaches for Parkinson disease
prediction. Int. J. Comput. Sci. Inf. Technol 6(2), 1648–1655 (2015)
21. Ahmed, I.A., Senan, E.M., Shatnawi, H.S.A., Alkhraisha, Z.M., Al-Azzam, M.M.A.: Hybrid
techniques for the diagnosis of acute lymphoblastic leukemia based on fusion of CNN features.
Diagnostics 13(6), 1026 (2023)
22. Alzubaidi, M.S., et al.: The role of neural network for the detection of Parkinson’s disease: a
scoping review. In: Healthcare. Vol. 9(6), p. 740. MDPI (2021)
23. Rao, D.V., Sucharitha, Y., Venkatesh, D., Mahamthy, K., Yasin, S.M.: Diagnosis of parkinson’s
disease using principal component analysis and machine learning algorithms with vocal fea-
tures. In: 2022 International Conference on Sustainable Computing and Data Communication
Systems (ICSCDS), pp. 200–206. IEEE (2022)
24. Ali, A.M., Salim, F., Saeed, F.: Parkinson’s disease detection using filter feature selection and
a genetic algorithm with ensemble learning. Diagnostics 13(17), 2816 (2023)
25. Lamba, R., Gulati, T., Al-Dhlan, K.A., Jain, A.: A systematic approach to diagnose Parkinson’s
disease through kinematic features extracted from handwritten drawings. J. Reliab. Intell.
Environ. 1–10 (2021)
26. Jaroli, P., Singla, C., Bhardwaj, V., Mohapatra, S.K.: Deep learning model based novel seman-
tic analysis. In: 2022 2nd International Conference on Advance Computing and Innovative
Technologies in Engineering (ICACITE), pp. 1454–1458. IEEE (2022)
27. Bhardwaj, V., Rahul, K.V., Kumar, M., Lamba, V.: Analysis and prediction of stock mar-
ket movements using machine learning. In: 2022 4th International Conference on Inventive
Research in Computing Applications (ICIRCA), pp. 946–950. IEEE (2022)
 Accelerated GPU-Based Clustering,
Classification and Regression Using RAPIDS:
A Comparative Study

Mayuri Gupta1(B) , Ashish Mishra1 , Ashutosh Mishra2 ,

and Jayesh Gangrade3
1 Jaypee Institute of Information Technology, Noida, India
[email protected], [email protected]
2 Thapar University, Patiala, Punjab, India
[email protected]
3 Manipal University Jaipur, Jaipur, India

[email protected]

Abstract. Machine learning has become increasingly popular as big data and data
science applications have grown in demand over the previous decade. GPUs, on
the other hand, are ideally suitable for Machine learning challenges due to their
tremendous performance. NVIDIA has released the RAPIDS framework, which
includes the cuML library, to take benefit from GPU performance for machine
learning. A GPU-based framework such as RAPIDS provides fast results on large
datasets, whereas a CPU-based framework takes more time. In comparison to
the CPU, there is no library other than RAPIDS that provides fast computation
speed on the GPU. The purpose of this paper is to compare RAPIDS and Sci-
kit clustering, regression, and classification algorithms. As a result, K-means,
DBSCAN, Linear regression, Lasso regression, Ridge regression, Random Forest,
Nearest Neighbor, and KNearest Neighbor algorithms are used on GPUs and
CPUs, and the results are determined by calculating the time taken for execution
and speed up. The results indicate that the RAPIDS framework can completely
harness the power of the GPU compared to the CPU-based Sci-kit library.

Keywords: Machine Learning · GPU · NVIDIA · K-Means · cuML · CPU ·

RAPIDS · Sci-kit

1 Introduction
A dataset has been developed from different sources that have spread at an unprece-
dented rate over the prior decade. The processing of huge quantities of data in a timely
way to generate important features and drive decision-making is a challenging prob-
lem for different Big Data apps. Machine Learning (ML) frameworks, software tools,
and technologies that allow evaluating and extracting meaningful information from data
through iterative processing have seen a renaissance to help in this research for improved
understanding. Sci-kit library [1] and Apache Spark’s ML library [2] are two prominent

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 239–253, 2025.
https://doi.org/10.1007/978-3-031-71481-8_19
240 M. Gupta et al.

machine learning libraries used in Data Analytics. These libraries are built specifically to
allow the execution of machine-learning algorithms on CPUs. Due to their high through-
put, GPUs have become a famous platform for accelerating parallel workloads. This also
makes them an appropriate choice for machine learning applications that require a lot of
arithmetic calculations [3]. NVIDIA recently released the RAPIDS AI [4] library which
is a collection of open-source software modules and APIs designed to take advantage
of GPU speed. The major use of this API is to make it possible to run whole data sci-
ence analytic pipelines on GPUs. The cuML [5] library is one of the most important
parts of the RAPIDS data science sector. This GPU-accelerated machine learning pack-
age is equivalent to Scikit-learn, providing comparable Pythonic APIs but avoiding the
challenges of explicitly building compute kernels for GPUs using CUDA.
In this paper, the research contributions are given in the following section:
• Analyze previous studies that compared the performance of GPUs and CPUs in
machine learning tasks.
• The objective of this study is to identify gaps in existing research and emphasize the
need for a comprehensive comparison of RAPIDS and Sci-kit.
• This study compares the performance of Sci-kit Learn and RAPIDS library machine
learning algorithms based on clustering, classification, and regression on CPU and
GPU.
• Sci-Kit Learn and the RAPIDS library were used to calculate execution time and
speed up value for all the algorithms.
• A dataset consisting of 1587257 rows and 13 columns was used to determine the
results because GPUs can perform well in large datasets as opposed to small datasets.
• The purpose of this study is to help the research community understand how GPUs
and CPUs differ from each other.
Research contributions are discussed in the above section, and subsequent sections
are arranged as follows: Sect. 2 describes the literature study. Section 3 refers to the
proposed methodology. Section 4 discusses some ML-based approaches including clus-
tering, regression, and classification algorithms. Section 5 outlines the methodology of
the paper consists of software tools and libraries and dataset descriptions used to find
the results. Section 6 discusses the implementation and results and Sect. 7 shows the
conclusion of the whole study.

2 Related Work

Gonzalez et al. [6] provide an overview of ML techniques for bagging and boosting.
XGBoost [2, 7] is a library for gradient boosting that may be used for both distributed
and GPU-based execution (through Dask). Another Machine learning library proficient
in distributed computing is H203 [8]. H204GPU [9] is a GPU type that supports shared
memory. While software libraries and packages have provided long computational sup-
port for machine learning methodologies, support for distributed execution on a GPU
cluster is currently in its early stages.
XGBoost is a model that includes boosting models. The major objective for the cre-
ation of the cuML library is to examine a variety of new and current ML models for
 Accelerated GPU-Based Clustering 241

effective multi-node multi-GPU execution. Some machine learning tools and approaches
are supported by deep learning frameworks such as TensorFlow [10] and PyTorch [11].
Raschka et al. [12] present an overview of Python machine learning approaches, contain-
ing the Sci-kit library, CPU training, and the RAPIDS suites. In this paper, the authors
also explain why the cuML library is needed and why it was created.
The RAPIDS library along with the cuML library has been achieving acceptance
in society as a feasible solution for running high-performance applications on a GPU
cluster. RAPIDS [13] and Dask [14] libraries, were used by Napoli et al. [15] to assess
the dataset for geophysical simulations. In all the review papers authors use the RAPIDS
framework with cuML to make use of several GPUs of NVIDIA.

3 Proposed Methodology
The significance of this study is to compare RAPIDS and Sci-kit clustering, regression,
and classification algorithms. This research study shows how the performance of GPU
is far better than CPU in terms of speed and time. All the machine learning algorithms
were used for the comparative analysis study.
RAPIDS was used as a motivational framework in this study. The researchers encour-
aged the use of this library in their research work for faster simulations and less training
time.

4 Algorithms Used
There are various types of algorithms of ML to find the results of GPU and CPU. This
study compares the execution time and speed-up of three types of machine learning algo-
rithms based on clustering, classification, and regression. These algorithms include K-
Means, DBSCAN, Nearest Neighbor, K-Nearest Neighbor (KNN), and Random Forest
(RF), Linear Regression (LR), Lasso Re-gression, and Ridge Regression.

4.1 Clustering Algorithms

The concept of clustering involves the arrangement of data points into separate clusters
that include comparable data points. It is an unsupervised learning approach therefore,
the algorithm receives no supervision, and it works with an unlabeled dataset [16].
The two clustering algorithms are used to find the execution time and speedup namely
K-Means and DBSCAN [17].

4.1.1 K-Means Algorithm

In this section, the models of machine learning were described and applied to conduct a
comparative analysis of the Sci-kit and RAPIDS on the K-Means clustering algorithm.
The K-Means clustering model was proposed by J.B. Mac Queen [7], based on
splitting, and is a type of cluster algorithm. The squared error and criteria of error are
the findings of the K-means approach, which aims to minimize the cluster performance
index. This model tried to discover divisions of K clusters that match certain criteria to
242 M. Gupta et al.

obtain the best result [8]. It separated n data points and divided them into k clusters by
lowering the sum of squared errors in each cluster:

X = minμj k {|xi − μj |2 } (1)
n

μ_j is the mean of data in cluster k, x_i are the inputs of the clusters and n is the number
of clusters.
KMeans in cuML’s GPU implementation are based on array-like objects or cuDF
DataFrames and support the scalable KMeans ++ initialization procedure [9]. This tech-
nique is more stable than choosing K points at random. The speed and simplicity of
KMeans are its most substantial benefits. That is why KMeans is the initial choice of
the clustering algorithm for this research work [2].

4.1.2 DBSCAN Algorithm

The process of unsupervised learning by which distinct groups or clusters are identified
in data is known as density-based clustering. In data space, clusters are defined as con-
tiguous regions with a high density of points separated from one another by contiguous
regions with a low density of points [1]. Two parameters are used by the DBSCAN
algorithm:
“minPts” is the lesser number of points required to constitute a dense region (a
threshold).
eps (∈): Distance metric used to locate points around any given point [11].

4.2 Classification Algorithms

Generally, a classification program learns from a given dataset or observation and then
categorizes new observations into various classes or groups. For example, Yes/No, 0/1,
Spam/No Spam, cat/dog, and so on. Classes are known as targets/labels or categories.
Three classification algorithms are applied to find the best results including Nearest
Neighbors, KNN, and Random Forest.

4.2.1 Nearest Neighbors Algorithm

A neighbor-based learning approach is supported by Nearest Neighbors in both unsu-
pervised and supervised approaches. In supervised neighbor-based learning, there are
two types: classification for discrete labels and regression for continuous labels. Using
nearest neighbor approaches, a set of training samples is selected that is closest to the
new point, and the label is estimated based on them. Depending on the local density of
points, the number of samples may be fixed (k-nearest neighbor learning) or variable
(radius-based neighbor learning). It is common to use a standard Euclidean distance as
the measure of distance, although any metric measure can be used [14].

4.2.2 K-Nearest Neighbors Algorithm

The K-nearest neighbors’ algorithm (k-NN) was developed by Evelyn Fix and Joseph
Hodges in 1951 and revised by Thomas Cover. This is a non-parametric supervised
 Accelerated GPU-Based Clustering 243

learning approach. Its applications include classification and regression. The input for
both cases consists of the k closest training examples within the data set. The output
is captured by whether k-NN is used for classification or regression. The Euclidean
distance was used to determine the nearest or nearby points i.e.:

 d

d (x1 , x2 ) =  (x1i − x2i )2 (2)
i=1

4.2.3 Random Forest Algorithm

Using Sci-Kit Learn, a random forest is a meta-estimator that uses averaging for improved
accuracy and reduced overfitting by fitting several decision trees on different subsamples
of a dataset.
In RAPIDS, the random forest algorithm for tree node splits is not the same as the one
used in sci-kit-learn. To calculate splits, the cuML Random Forest employs a quantile-
based approach rather than an exact count by default. The n_bins value allows you to
change the size of the quantiles. The cuML Random Forest models can be exported and
used to make predictions on systems that do not contain NVIDIA GPUs [7].

4.3 Regression Algorithms

The idea of regression analysis is to examine how a dependent variable changes because
of an independent variable, while the other independent variables remain unchanged.
Continuous or real data are estimated, such as temperature, age, income, and price,
among others. The three regression algorithms of machine learning are used in this
study i.e., Linear regression (LR), Lasso regression (LSR), and Ridge regression (RR).

4.3.1 Linear Regression

The LR algorithm exhibits a linear connection between a dependent (Y) and one or
more independent (X) variables, thus known as linear regression. Linear regression can
be represented mathematically as:

Y = a0 + a1 X + ∈ (3)

where, Y = Dependent Variable (Target Variable), X = Independent Variable (predictor

Variable), a0 = intercept of the line (Gives an additional degree of freedom), a1 = Linear
regression coefficient (scale factor to each input value) and ∈ = random error [9].

4.3.2 Lasso Regression

The term Lasso regression is an abbreviation for Least Absolute Shrinkage and Selection
Operator. In Lasso regression, a penalty factor is included in the cost function. This
phrase represents the absolute total of the coefficients. When the value of the coefficients
increases from 0, this term penalizes the model, causing it to reduce the value of the
244 M. Gupta et al.

coefficients to minimize loss. Ridge regression differs from lasso regression in that lasso
regression tends to set coefficients to absolute zero, whereas Ridge never does. A Lasso
regression can be represented mathematically as:

Llasso = argminβ̂ (||Y − β × X ||2 + λ × ||β||1 ) (4)

where Y = Dependent Variable (Target Variable), X = Independent Variable (predictor

Variable), β = Lasso regression coefficient and λ = Regularization penalty [10].

4.3.3 Ridge Regression

The ridge regression model includes a penalty term equal to the square root of the
coefficient. L2 represents the square of the coefficient magnitude. The penalty term may
be further managed by adding a coefficient lambda. In this situation, if lambda is zero, the
equation is the basic OLS; otherwise, if λ > 0, a constraint is added to the coefficient. This
constraint forces the value of the coefficient to drift towards zero as raises the amount
of λ. Consequently, a trade-off occurs between increased bias and reduced variance
(dependencies on some coefficients tend to be zero and certain coefficients tend to be
extremely large, resulting in a less flexible model) [13].
A Ridge regression can be represented mathematically as:

Lridge = argminβ̂ (||Y − β × X ||2 + λ × ||β||22 (5)

5 Methodology

In this section, the system configuration is discussed with the workstation that has been
used which is present in our lab. It also shows dataset descriptions and the software tools
and libraries used to run the clustering, classification, and regression algorithms.

5.1 System Configuration

Python 3.5 is used for the implementation. The system configuration of the workstation
provided in our lab contains the system memory of 256 GB, performance of 500 Tera-
FLOPS mix precision, storage of 4X 1.92 TB SSD RAID 0, and processor of 20-core intel
Xeon e5–2698 v4 2.2 GHz. The GPU used in this study is NVIDIA’s Tesla V100 GPU
which is the high-end GPU that comes up with 5120 computation cores of runtime version
10.1 and minor version 7.0 with total global memory i.e., 32478 MBytes (34055847936
bytes), GPU mac clock rate was 1530 MHz (1.53 GHz), Memory clock rate was 877 MHz
with bus width 4096-bit, L2 cache size of 6291456 bytes, 2048 layers were used, total
no. of registers per block was 65536 and total shared memory per block was 49152 bytes
[12].
 Accelerated GPU-Based Clustering 245

5.2 Dataset Description

A dataset is used in this research for clustering, classification, and regression with the
RAPIDS and Sci-kit libraries. The dataset used i.e., weather minute data [2], contains the
attributes of air pressure, temperature, wind direction, maximum wind, wind direction,
maximum wind speed, and relative humidity with 1587257 rows and 13 columns. The
dataset was collected from the Kaggle repository and the dataset size is 123 MB to
compare the analysis of RAPIDS and Sci-kit libraries on the machine learning algorithms
to demonstrate the comparison of computing speed and time on CPU and GPU.

5.3 Software Tools and Libraries

In the past days, Machine Learning projects were usually implemented by manually
coding algorithms, performing mathematical and statistical calculations, and so on. Due
to this, the processing became time-consuming, labor-intensive, and inefficient. As a
result of multiple Python libraries, frameworks, and modules, the process has become
easier and more effective than it was in the past. Currently, Python is one of the most
widely used programming languages in this field, and it has replaced several languages
in the industry, in part because of its extensive library collection. In this study, only two
types of libraries were analyzed i.e., Sci-Kit learn and RAPIDS used for the execution
of CPU and GPU time respectively, and find the speed-up value of various types of
algorithms of ML-like clustering, classification, and regression.

5.3.1 RAPIDS Libraries

The RAPIDS software library package, based on CUDA-X AI, allows to execution of
entire data science and analytics pipelines on GPUs [4]. It uses NVIDIA and CUDA
primitives for minimal computational optimization but exposes GPU high-bandwidth
memory and parallelism performance via Python APIs. In addition, RAPIDS focuses
on the common preparation of data processes for data science and analytics [5]. This
provides a familiar DataFrame API that interfaces with a range of machine learning meth-
ods to accelerate end-to-end pipelines without incurring typical serialization expenses.
RAPIDS now supports multi-node, multi-GPU setups, allowing for substantially more
efficient processing and training on much larger datasets.

5.3.2 Sci-kit Libraries

The Scikit-learn package is a free Python machine-learning package. It supports Python
scientific and numerical libraries like Numpy and Scipy, as well as models of ML-like
support vector machine (SVM), linear and logistic regression (LR), K-means clustering,
LSTM, and k-neighbors. Sci-kit library is mostly built-in Python [16], and it depends
considerably on NumPy for array operations and high-efficiency linear algebra. In addi-
tion, to increase efficiency, several fundamental algorithms are implemented in Cython.
A Cython wrapper over LIBSVM implements SVM a related wrapper over Liblinear
implements LR and linear SVM. In certain circumstances, expanding these functions
with Python could not be possible. Many additional libraries of Python, including Plotly
246 M. Gupta et al.

and matplotlib for mapping, Numpy for array calculations, pandas for data frames, scipy,
and others, work well with Scikit-learn.

6 Implementation and Results

In this section, the comparative analysis and results of RAPIDS and Sci-kit version
libraries of clustering, classification, and regression were represented.
A dataset of (123 MB) and their datatype values with various fields are taken to
find the comparative results. This dataset is taken to observe how much time is taken
by RAPIDS on GPU and Sci-kit to run on CPU. The dataset runs with five iterations
and shows the execution time and speed up of RAPIDS and Sci-kit, to see how the
RAPIDS version provides faster results than the Sci-kit version given in Table 1. The
three types of ML algorithms were applied i.e., clustering, regression, and classification
including eight algorithms such as K-Means, DBSCAN, Nearest Neighbor, KNN, RF,
Linear regression, Lasso Regression, and Ridge Regression to show that GPU gives the
best results as compared to CPU results. For better visualization and understanding the
results were represented graphically. The main goal is to find the comparative study
between RAPIDS and Sci-kit by executing the five iterations to calculate the RAPIDS
and Sci-kit clustering, classification, and regression version time for a dataset with their
speed up shown in Table 1. Speedup is calculated by using the mathematical formula
i.e.:
SciKit time (CPU Time)
Speedup = (6)
RAPIDS time (GPU Time)

Fig. 1. Time comparison of RAPIDS vs. Sci-kit of K-Means and DBSCAN clustering algorithms
(left to right)

Figure 1 represents the time execution of RAPIDS and Sci-Kit version of K-Means
and DBSCAN clustering algorithms running in five iterations in which the K-means
GPU version is faster than the CPU version by 89 whereas in DBSCAN GPU shows
the execution time i.e., 2.7 times faster as CPU. So, in clustering algorithms, it was
observed that K-means give better results as compared to DBSCAN in execution time
 Accelerated GPU-Based Clustering 247

in seconds. In Fig. 2, it shows the execution time of RAPIDS and Sci-Kit versions of
Nearest Neighbor, KNN, and Random Forest classification algorithms running in five
iterations in which the Nearest Neighbor GPU version is faster than the CPU version by
474 whereas, in KNN, GPU shows the execution time i.e., 1096 times faster as CPU.

Table 1. RAPIDS and Scikit time with speed up of Dataset weather

S. No. Algorithms Sci-Kit RAPIDS Speed up Average

Used (seconds) (seconds) Speedup
1 K-Means 58.4 0.6514 90 88
58.2 0.6636 88
57.90 0.663 87
56.6 0.671 84
57.1 0.621 92
2 DBSCAN 351 130 3 3
345 128 3
343 132 2
339 129 3
342 127 2
3 Nearest 6.26 0.0132 474 433
Neighbour 5.96 0.0143 417
6.18 0.0168 368
6.12 0.0128 478
6.01 0.014 429
4 KNN 7.05 0.00643 1096 1068
7.04 0.00684 1029
7.01 0.00646 1085
7.08 0.0066 1072
7.03 0.00665 1057
5 Random Forest 0.224 0.143 1 2
0.225 0.137 2
0.227 0.141 2
0.226 0.145 1
0.225 0.131 2
6 Linear Regression 5.91 0.039 152 156
6.34 0.0393 161
6.12 0.0377 162
6.02 0.0387 156
5.82 0.0394 148
7 Lasso Regression 15.9 0.217 73 67
14 0.227 62
14.8 0.216 69
14.7 0.215 68
13.4 0.214 63
(continued)
248 M. Gupta et al.

Table 1. (continued)

S. No. Algorithms Sci-Kit RAPIDS Speed up Average

Used (seconds) (seconds) Speedup
8 Ridge Regression 0.121 0.0384 3 3
0.122 0.0399 3
0.12 0.0393 3
0.116 0.0389 2
0.119 0.0365 4

Random Forest shows faster results in GPU by 1.5 times as compared to the Sci-kit
version. So, in classification algorithms, it was observed that KNN gives good results as
compared to Nearest Neighbor and Random Forest in execution time in seconds.
Similarly, Fig. 3 shows the execution time of RAPIDS and Sci-Kit versions of Linear
regression, Lasso regression, and Ridge Regression algorithms running in five iterations
in which the Linear regression GPU version is faster than the CPU version by 151
whereas, in Lasso regression, GPU shows the execution time i.e., 73 times faster as CPU.
Ridge regression represents the faster results in GPU by 3 times as compared to the Sci-kit
version. So, in regression algorithms, Linear regression gives good results as compared
to Ridge regression and Lasso regression in execution time in seconds. The clustering
process, which is also known as cluster analysis is a machine-learning technique used to
organize unlabeled datasets. Figure 4 represents the speed-up of clustering algorithms
used to find the execution time and their speed-up value. In Fig. 4, K- Means shows a
higher value of speed up as a counter to DBSCAN. So, K-means shows good results as
compared to other clustering algorithms.

Fig. 2. Time comparison of RAPIDS vs. Sci-kit of Nearest Neighbor, KNN, and Random Forest
classifiers (left to right)

In classification, a function is identified to assist in dividing a dataset into classes

based on various factors. Classification involves training a computer program on the
training dataset and then categorizing the input into multiple classes based on the training.
Figure 5, shows the speed up of the classification algorithms including Random Forest,
Nearest Neighbor, and KNN in which KNN shows the higher value of speed up which
means KNN gives better results as compared to Random Forest and Nearest Neighbor
of classifier algorithms.
 Accelerated GPU-Based Clustering 249

Fig. 3. Time comparison of RAPIDS vs. Sci-kit of Linear regression, Lasso Regression, and
Ridge Regression (left to right)

Fig. 4. Speed Up representation of K-means and DBSCAN Clustering Algorithms

In Fig. 6, the regression algorithm’s comparative study of speed up was represented

in which Linear regression represents higher speed up whereas Lasso regression and
Ridge regression show lower speed up as compared to linear regression. In regression
algorithms of machine learning Linear regression gives the best results as a counter to
other regression algorithms of Machine Learning.

Fig. 5. Speed Up representation of Random Forest, Nearest Neighbor, and KNN Classifier
Algorithms
250 M. Gupta et al.

Fig. 6. Speedup representation of Linear Regression, Lasso Regression, and Ridge Regression
Algorithms

After finding the values of the speedup of clustering, classification, and regression
algorithms it was shown (Fig. 7) that the average speedup of all the algorithms of ML
was used in this study to find the results. In Fig. 7 it was observed that KNN gives
the highest average speedup value as compared to all types of algorithms of clustering,
classification, and regression. So, the classifier algorithms have shown the best results
because KNN and Nearest Neighbor represent the highest value of average speedup as
compared to clustering and regression algorithms.
After determining the average speedup, Fig. 8 shows the time speedup for all five
iterations of clustering, classification, and regression. The value of speedup has allowed
us to determine which algorithm runs faster on CPU or GPU, and which algorithm is
best suited for processing large datasets on GPU.

Fig. 7. Average speed up of all types of Clustering, Classification, and Regression Algorithms

Also, the small datasets were applied to GPU, and have observed that KNN, which
produces the best results on GPU on large datasets, produces the least results on GPU on
 Accelerated GPU-Based Clustering 251

small datasets. With a small dataset consisting of 6500 rows and 13 columns of 360 KB
size and CPU time values of 10.1 ms, 7.91 ms, 10.4 ms and GPU time values of 6.32 ms,
6.11 ms, 6.57 ms. It shows the least results on classifying algorithms. Therefore, it can
consider only large datasets, because it exhibits the best clustering, classification, and
regression results.

Fig. 8. Comparative study of speed up of all types of Clustering, Classification, and Regression
Algorithm

Table 2 represents the average speed-up by calculating RAPIDS and Sci-kit time
of the dataset where the dataset gives 501 average speedup pf classifiers, 75 average
speedups for regression, and 91 average speedups for clustering. So, by calculating
the average speedup of the classifiers, regression, and clustering, it was observed that
classifiers showed higher and better results counter to regression and clustering. As a
result, we have determined the average speedup and compared all types of ML algorithms
to provide researchers with an understanding of which types of ML algorithms showed
the best results on RAPIDS, which is a GPU-based library compared to Sci-Kit learn
based on CPUs. Researchers also learned from this study that GPU shows the best
results on large datasets (e.g.-Megabyte and Gigabyte) whereas it is not suitable for
small datasets whose size is less.

Table 2. Average Speed up of Classification, Regression, and Clustering Algorithms

Algorithms Average Speedup

Classifiers 501
Regression 75
Clustering 91
252 M. Gupta et al.

7 Conclusion
In this paper, the comparative study of RAPIDS and Sci-kit libraries running on GPU
and CPU for machine learning models i.e., clustering, regression, and classification
algorithms presented respectively. The results showed that the GPU-based RAPIDS
framework provides better performance than the CPU-based Sci-kit library with a shorter
execution time. RAPIDS running on GPU with a large dataset shows results in a short
period. However, the Sci-kit running on the CPU takes longer to load results. This
study has demonstrated that classification algorithms produce shorter execution times
and greater speedups when compared to others. The classification algorithms show an
average speed-up of 501 whereas clustering algorithms show a 91 average speedup
and regression represents a 75 average speedup. Data scientists and machine learning
developers can use the findings of this study to make full use of the RAPIDS framework
and the Sci-Kit library. Further work will explore other RAPIDS algorithms for a variety
of datasets and the TensorFlow library that runs on GPUs as well.

References
1. Rocklin, M.: Dask: Parallel Computation with Blocked Algorithms and Task Scheduling,
126–132 (2015). https://doi.org/10.25080/Majora-7b98e3ed-013
2. Pedregosa, F., et al.: Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12,
2825–2830 (2011)
3. Meng, X., et al.: Mllib: Machine learning in apache spark. J. Mach. Learn. Res. 17(1), 1235–
1241 (2016)
4. Anil, R., Owen, S., Dunning, T., Friedman, E.: Mahout in action (2010)
5. Raschka, S., Patterson, J., Nolet, C.: Machine learning in python: main developments and
technology trends in data science, machine learning, and artificial intelligence. Information
11(4), 193 (2020)
6. “‘MPI-3 Standard Document.’”
7. Ghazimirsaeed, S.M., Anthony, Q., Shafi, A., Subramoni, H., Panda, D.K.D.K.: Accelerating
GPU-based Machine Learning in Python using MPI Library: A Case Study with MVAPICH2-
GDR. In: 2020 IEEE/ACM Workshop on Machine Learning in High Performance Computing
Environments (MLHPC) and Workshop on Artificial Intelligence and Machine Learning for
Scientific Applications (AI4S), pp. 1–12 (2020)
8. Becker, N., et al.: Streamlined and accelerated cyber analyst workflows with clx and rapids.
In: 2019 IEEE International Conference on Big Data (Big Data), pp. 2011–2015 (2019)
9. González, S., Garc\’\ia, S., Del Ser, J., Rokach, L., Herrera, F.: A practical tutorial on bagging
and boosting based ensembles for machine learning: Algorithms, software tools, performance
study, practical perspectives and opportunities. Inf. Fusion 64, 205–237 (2020)
10. Chen, T., Guestrin, C.: Xgboost: a scalable tree boosting system. In: Proceedings of the 22nd
acm sigkdd international conference on knowledge discovery and data mining, pp. 785–794
(2016)
11. “XGBoost Development Team, ‘XGBoost Library.’”
12. “H2O Development Team, ‘H2O3: Distributed, fast, and scalable machine learning software’”
13. Gill, T.Y.N., LeDell, E.: “‘H2O4GPU: Machine Learning with GPUs in R and Python.’”
14. Yorozu, T., Hirano, M., Oka, K., Tagawa, Y.: Electron spectroscopy studies on magneto-
optical media and plastic substrate interface. IEEE Transl. J. Magn. Japan 2(8), 740–741
(1987)
 Accelerated GPU-Based Clustering 253

15. Abadi, M., et al.: Tensorflow: A system for large-scale machine learning. In: 12th symposium
on operating systems design and implementation, pp. 265–283 (2016)
16. Ezugwu, A.E., et al.: A comprehensive survey of clustering algorithms: State-of-the-art
machine learning applications, taxonomy, challenges, and future research prospects. Elsevier
110, 104743 (2022). https://doi.org/10.1016/j.engappai.2022.104743
17. Dahlkvist, J., Tomczak, W.: Segmentation of companies using DBSCAN and K-Means (2022).
Accessed: 22 Nov. 2022
 Talk to Hands: A Sign Language Solution
for Speech-Impaired People

Durvesh Chaudhari(B) , Adrika Dikshit , Omkar Pawar , Deepak Karia(B) ,

and Payal Shah

Department of Electronics Engineering, Sardar Patel Institute of Technology, Mumbai, India

{durvesh.chaudhari,adrika.dikshit,omkar.pawar,deepak_karia,
payal_shah}@spit.ac.in

Abstract. Humans are social animals and communication plays a vital role in
human interaction and thus their societal development. It is the sole way for
humans not only to interact but also to foster relationships. But, among us many
who live have their world limited as they speak a language very few understand.
That is sign language which only the deaf and mute grasp. So, to aid inclusivity
and to make sure language never poses a hindrance, Talk-2-Hands was introduced
as an innovative solution to solve the problem of the communication gap. Talk 2
Hands comes with the functionality of translating Indian sign language to text.
The system makes use of the convolutional neural network (CNN), a deep learn-
ing algorithm commonly used in image processing and recognition tasks. The
CNN model was trained on a large dataset of sign language gestures captured
through videos or images, which includes a variety of gestures, variations, and
lighting conditions. The data was preprocessed and the CNN model was trained
to accurately recognize sign language gestures in real-time and convert them into
text. The system was designed to be integrated into a user-friendly interface that
enables speech-impaired individuals to communicate effectively with the general
public.

Keywords: Convolutional Neural Networks · Sign language · Real-time

translation

1 Introduction
Humans, inherently social, heavily rely on language for communication. However, not
everyone possesses equal means of expression. Some individuals communicate in a spe-
cial way rather than conventionally, significantly impacting their lives. Communication
entails both conveying thoughts and understanding others. In India alone, approximately
60 million people face challenges in hearing and speech [1]. They resort to sign lan-
guage, lip-reading, and hearing aids for communication. Nevertheless, sign language is
not widely known among the general population, making communication a challenge for
its users. Moreover, sign languages differ across regions, lacking a standardized interna-
tional form. Leveraging technology has immensely enhanced communication for these
individuals.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 254–264, 2025.
https://doi.org/10.1007/978-3-031-71481-8_20
 Talk to Hands: A Sign Language Solution 255

By employing technology, the lives of disadvantaged individuals has been signifi-

cantly improved. Real-time translators, for instance, facilitate efficient communication
for the deaf and mute. Bridging communication gaps is crucial, and technology offers
a vital solution. Just as language barriers have been mitigated through translators, sign
language translators have facilitated effective communication for the deaf community.
This initiative was further strengthened by developing user-friendly applications for
easy access to the technology. Beyond empowering the disadvantaged, this technological
approach can benefit businesses, enhancing their marketing effectiveness and broadening
their reach to a larger audience of consumers.
The primary objective of this study was to develop a sign language-to-text converter
that accommodates Indian Sign language (ISL) signs and gestures. This study also eval-
uates the performance of the system through extensive testing, including measuring
accuracy and robustness under various conditions.
The subsequent sections provide a comprehensive understanding of this concept.
Section 2 discusses related work, highlighting existing efforts in this domain. Section 3
explores the underlying technologies, setting the stage for a detailed explanation of
our system in Sect. 4. Implementation components and the system’s functioning are
elaborated in Sect. 5. The predictions generated by the model are discussed in Sect. 6.
The study’s results and a thorough discussion of the developed system are presented in
Sect. 7, culminating in the study’s conclusions in Sect. 8.

2 Related Work
There are plenty of literature available for sign language translators using convolutional
neural networks (CNN), translators using Image processing and Machine Learning (ML).
A few of them have been explained below.
Rumana et al. [2], have discussed two methods to translate sign language, vision-
based using image processing and glove-based using sensors. The paper focused on
image-based translation. The data set used was limited to American Sign Language
(ASL) signs. Features of the image were extracted using a Gaussian blur filter. The CNN
model predicted signs based on the processed image. The images predicting similar
results were stored in sets. The sets were further classified using classifiers specifically
made for these sets that made successful predictions.
Lee et al. [3], developed a Long-Short Term Memory Recurrent Neural Network
with a k-Nearest-Neighbour method-based application to learn American Sign Language
(ASL). They focused only on ASL sign translation in real time. To validate the prediction,
a leap motion controller was used.
Monikowski [4] developed the idea of applied sign linguistics by utilizing data from
projects that looked at the real-world uses of sign language in a wide range of industries.
The author offered reliable access to modern ideas and methods for those who worked
with sign languages, such as research students, supervisors of sign language learners,
and educators of the language.
Thakur et al. [5], have proposed a way to implement a two-way communication
system between sign language and text. The study focuses on American Sign Language.
A model based on a convolutional neural network was used to transform a sign or gesture
256 D. Chaudhari et al.

into the appropriate text and audio. This paper additionally suggested an approach to
produce a full two-way translator by creating images of signs. The same technology was
used to translate text input into visuals in sign language. It utilized identical photos from
its dataset that matched the gesture entered in real-time. The input text was stored in the
form of a string and an image corresponding to each element of the string was retrieved
from its dataset. The output was generated as a GIF using OpenCV.
Rajamohan et al. [6], used a glove-based system to translate sign language to text.
The glove features five flex sensors, tactile sensors, and an accelerometer integrated
into its structure. The flex sensors generated a corresponding change in resistance for
each particular gesture, while the accelerometer tracked the hand’s orientation. These
hand gestures were processed using an Arduino, which included both a training mode to
accommodate different users and an operational mode. The Arduino was also responsible
for assembling letters into words. Additionally, the system incorporated a text-to-speech
conversion module, which converts the recognized gestures or text into audible output.
Hasan et al. [7] implemented supervised machine learning for translating sign lan-
guage into speech. This work had very limited input and was limited to Bangla sign
language. This work provided a method for hand gesture detection that used a Support
Vector Machine (SVM) as the classification algorithm and HOG (Histogram of Oriented
Gradients) to extract features from the gesture image. After the gesture image had been
identified, output text was produced and a TTS (Text to Speech) converter was employed
to turn it into audible sound. The objective was to provide the output in the form of audio
by converting output text into audio by using a text-to-speech converter.
Papatsimoilu et al. [8], have analysed real-time systems developed from 2017 to
2021. The focus of this study was to discuss different approaches for sign language
translation such as motion discrimination, facial expression recognition, spatial contrast
recognition, and hand configuration recognition. Sensor-based translation systems like
the use of data gloves was also discussed. This study also explored the various limitations
of different approaches.

3 Technology Used
3.1 OpenCV

A popular open-source computer vision toolkit, OpenCV offers a wide range of image
and video processing functions [9]. In particular, hand detection, cropping, resizing,
and image display were all carried out in this project employing OpenCV (cv2). Web-
cam images were captured (‘cv2.VideoCapture’), images were switched horizontally
(‘cv2.flip’), hands have been identified (‘HandDetector’ from ‘cvzone’), were cropped,
resized, pasted onto a white background, and the result is shown (‘cv2.imshow’).

3.2 NumPy

Python’s NumPy (Numerical Python), was used to create an array of zeros, to produce a
white background for the cropped hand image. The background image was resized and
pasted onto the centre. It was also used for calculating the scaling factor, dimensions, and
 Talk to Hands: A Sign Language Solution 257

offset for centring the cropped image on the white background and filling the array with
white pixels. In summary, NumPy was used for image processing tasks and numerical
operations on arrays in the given code.

3.3 HandTracking Module’s HandDetector

The HandDetector module analyzes the video frames and determines which areas of
the frame correspond to the hands using computer vision techniques. additionally, it
offers a range of settings and parameters for adjusting the hand detection algorithm to
maximize its effectiveness in various scenarios. This study used the HandDetector from
CV Zone’s HandTracking Module to find out whether hands are there and where they are
in a video stream that was taken from the computer’s webcam. Following the detection
of the hands, each hand’s bounding box was identified, and the image inside was cropped
and scaled to a predetermined size.

3.4 Classification Module’s Classifier

The CV Zone’s Classification Module’s Classifier was used in this study to classify hand
gestures in real time. The module has a pre-trained deep-learning model that recognizes
hand gestures from an image. The classifier takes an image of the hand gesture as input
and outputs a prediction of the corresponding label. In this study, the predicted label was
displayed on the output video and was used to build a text message over time.

3.5 OS
The ‘OS’ module was used in this code to create directories to store the images that are
being captured by the program. Specifically, the code creates directories for each capital
letter from A to Z and each number from 1 to 9 in the specified directory path. They are
also used for accessing these paths.

3.6 Google’s Teachable Machine

Transfer Learning was applied by Google Teachable Machine [10], which developed new
models on the foundation of pre-trained neural networks [11]. By leveraging pre-trained
model features, transfer learning accelerates the training of bespoke models on a smaller
dataset. It supports a broad spectrum of model types, including speech recognition,
object detection, NLP, and image classification. Pre-built models for tasks like emotion
recognition and pose detection are available, and trained models were exported as Tensor
Flow Keras (.H5) files for program application.

4 Methodology
The deep learning image classification algorithm used CNN (Convolutional Neural Net-
work). The model was trained using CNN by providing numerous images as data to
predict and successfully recognize sign language gestures.
258 D. Chaudhari et al.

The signs that were included in datasets were static signs that do not need dynamic
movements to make a logical meaning for an action. Hence, a model was made that
identifies signs which included numbers from ‘0’ to ‘9’, alphabets from ‘A’ to ‘Z’ and
a few greeting signs such as, ‘How are you?’, ‘Morning’, ‘Afternoon’, ‘Hello’, ‘Good’
etc. The code used cvzone’s hand detector module which was built on the mediapipe’s
framework. Hence finding an appropriate dataset that had mediapipe’s landmarks (joints
on fingers) was difficult. Hence, a fresh dataset was built. Figure 1 explains the activ-
ity flow of creating the database. The captured images were stored in sets with labels
depicting the letter or phrase they represent. Approximately 700 images were captured
for each of the 41 signs. The dataset was divided into 70% for training and 30% for
testing, with around 500 and 200 images per class, respectively. These 20,500 training
images were fed to Google Teachable Machine for generating the model. To evaluate
the accuracy of the model, the remaining 200 images per class were used for testing,
resulting in a total of 8200 images being tested. The percentage of correctly predicted
images was calculated for each class.

Fig. 1. Flow chart of database creation and System Implementation respectively

 Talk to Hands: A Sign Language Solution 259

5 Implementation

Fig. 2. Original image captured by camera

Fig. 3. Bounding box applied to the original image

The three stages of processing images are shown in Figs. 2, 3 and 4. Figure 2 presents
the visual representation acquired through the camera, showcasing the raw image data.
Figure 3 illustrates the image obtained within the white bounding box, which is imposed
upon the original image. Finally, Fig. 4 demonstrates the fully processed image that is
stored in the dataset.
260 D. Chaudhari et al.

Fig. 4. Processed image with white background

6 Prediction and Recognition

Fig. 5. Detection of the letter ‘A’ sign with the label

Figures 5 and 6 show the correct prediction of the sign for the letter ‘A’ and the phrase
‘How are you?’ respectively.
 Talk to Hands: A Sign Language Solution 261

Fig. 6. Detection of a phrase with a label

7 Results and Discussion

Table 1. Accuracy for the prediction of numbers

Sign Accuracy (%)

0 91.6
1 100
2 55.4
3 94.8
4 42.8
5 86.8
6 12.1
7 78.2
8 65.2
9 100

Above mentioned tables show the accuracy of numbers signs, alphabets, and phrases.
The average testing accuracy achieved by the model is 80.2%., whereas the training
accuracy achieved is 97.241%. The highest accuracy of 100% was achieved for signs 1,
9, B, G, H, O, Q, R, W, and Y. Whereas, the lowest accuracy was observed for sign 6,
with an accuracy of 12.1%. Except for the few labels, the accuracy achieved per sign
is observed to be above 60%. The credit for this accuracy goes to the deep learning
classification algorithm used i.e., CNN (Convolutional Neural Network). However, the
low accuracy for a few labels comes due to the resemblance between different signs.
The other reason for the low accuracy can be the model’s limitation. It is difficult to
predict gestures involving overlapping hands where both hands are not visible distinctly
262 D. Chaudhari et al.

Table 2. Accuracy for prediction of letters

Sign Accuracy (%)

A 97.7
B 100
C 94.5
D 83.4
E 83.5
F 94.8
G 100
H 100
I 54.1
J 97.1
K 78.8
L 12.5
M 93.4
N 39.5
O 100
P 91.3
Q 100
R 100
S 98.9
T 98.7
U 43.4
V 17.6
W 100
X 99.5
Y 100
Z 99.1

to the camera. The lighting condition under which the experiment is being performed
also affects the accuracy of prediction (Tables 1, 2 and 3).
After training the model, there was some confusion between similar signs, such as
‘V’ and ‘2’, ‘Good’ and ‘9’, etc. Providing appropriate differentiation in the dataset was
an important step to get high accuracy. The model faced challenges in recognizing and
predicting outputs under low lighting conditions, making it a complex task to train the
model while considering this issue. During real-time detection/translation, the hands
were not completely stationary which led to inaccurate prediction. Moreover, the signs
that were formed by overlapping one hand over the other were difficult to capture as
 Talk to Hands: A Sign Language Solution 263

Table 3. Accuracy for prediction of phrases

Sign Accuracy (%)

Good 58.4
Evening 98.1
Hello 34.9
Morning 99.1
How are you 93.1

the model tried to identify the hidden hand, producing an inaccurate dataset and/or
prediction.

8 Conclusion and Future Scope

This study aims to develop a functional Indian Sign Language (ISL) to-text converter
using a Convolutional Neural Network (CNN) model, trained on an image dataset con-
taining approximately 21,000 diverse images. The dataset included 41 categories of
different static signs that are commonly used in ISL, such as numbers, alphabets and
common phrases. The model was trained and tested on a dataset containing image frames
with different backgrounds, lighting conditions, and hand shapes to avoid bias in the
model. The background noise was minimized i.e., the image frames consist only of
functional hands.
The transfer learning techniques used to train the model were provided by Google’s
Teachable Machine. On testing the model, it achieved an overall accuracy of 80.228%
on a testing dataset of approximately 8,200 images. Whereas, it achieved an accuracy
of 97.241% on the training dataset of approximately 21000 images. The accuracy of the
model can be further improved by incorporating a larger and more diverse dataset for
training.
In future, an actual CNN model can be deployed to achieve enhanced performance
by deploying artificial neurons. This will help the model identify the gestures accurately.
Later, the dynamic signs will also be included in the dataset by using video data. As
majority of the communication is based on hand movements and not on static signs.
Ultimately, the resulting signs will be stored to form a sentence and will be converted
to audio output, thus improving communication between individuals with special needs
and those who are physically fit.

References
1. Varshney, S.: Deafness in India. Indian Journal of Otology (2016). https://doi.org/10.4103/
0971-7749.182281
2. Rani, R.S., Rumana, R., Prema, R.: A review paper on sign language recognition for the deaf
and dumb. Int. J. Eng. Res. Technol. (2021). https://doi.org/10.17577/IJERTV10IS100129
264 D. Chaudhari et al.

3. Lee, C.K.M., Ng, K.K.H., Chen, C.-H., et al.: American sign language recognition and training
method with recurrent neural network. Expert Sys. Appl. 167, 114403 (2021). https://doi.org/
10.1016/j.eswa.2020.114403
4. Monikowski, C.: Language, cognition, and the brain: Insights from sign language research.
Studies in Second Language Acquisition (2004). https://doi.org/10.1017/s02722631043
93058
5. Thakur, A., Budhathoki, P., Upreti, S., et al.: Real time sign language recognition and speech
generation. J. Innov. Image Proc. (2020). https://doi.org/10.36548/jiip.2020.2.001
6. Rajamohan, A., Dhanalakshmi, M.: Deaf-Mute communication interpreter. Int. J. Sci. Eng.
Technol. 2(5), pp. 336–341 (2013). ISSN: 2277-1581
7. Hasan, M., Sajib, T.H., Dey, M.: A machine learning based approach for the detection and
recognition of Bangla sign language. In: 2016 International Conference on Medical Engi-
neering, Health Informatics and Technology (MediTec) (2017). https://doi.org/10.1109/med
itec.2016.7835387
8. Papatsimouli, M., Kollias, K.-F., Lazaridis, L., et al.: Real time sign language translation
systems: a review study. In: 2022 11th International Conference on Modern Circuits and
Systems Technologies (MOCAST) (2022). https://doi.org/10.1109/mocast54814.2022.983
7666
9. Culjak, I., Abram, D., Pribanic, T., et al.: A brief introduction to OpenCV. In: Proceedings of
the 35th International Convention MIPRO, pp. 1725–1730. Opatija, Croatia (2012)
10. In: Google. https://teachablemachine.withgoogle.com/. Accessed 29 Dec 2023
11. Carney, M., Webster, B., Alvarado, I., et al.: Teachable machine: approachable web-based
tool for exploring machine learning classification. In: Extended Abstracts of the 2020 CHI
Conference on Human Factors in Computing Systems (2020). https://doi.org/10.1145/333
4480.3382839
 Deep Learning-Based Thyroid Cancer Detection
and Segmentation Using Convolutional Neural
Networks

Aditya Praksh and Juhi Singh(B)

Manipal University Jaipur, Jaipur, India

[email protected]

Abstract. Thyroid cancer is a prevalent endocrine malignancy, and its timely

diagnosis is crucial for effective treatment and optimal patient outcomes. Deep
learning algorithms have emerged as promising tools for medical image segmen-
tation, facilitating accurate localization and characterization of thyroid nodules.
This study investigated the application of deep learning techniques for thyroid
cancer image segmentation, with the aim of improving diagnostic accuracy and
refining clinical outcomes. A thyroid image dataset was classified into two cate-
gories: malignant and benign. A convolutional neural network (CNN) model was
employed for thyroid cancer image segmentation using a diverse dataset of thy-
roid ultrasound images from a reputable medical institution. The model achieved
an accuracy of 92.67 percent in delineating and segmenting thyroid nodules for
validation split on 0.2. This research advances our understanding and application
of deep learning in thyroid cancer analysis.

Keywords: Deep learning segmentation · Thyroid gland · Image Segmentation

1 Introduction
The prevalence of thyroid cancer, an endocrine malignancy affecting the thyroid gland,
has fueled a burgeoning demand for sophisticated methods for early detection and precise
classification of thyroid nodules [1]. Traditionally, the primary method for identifying
thyroid nodules was physical examination, heavily reliant on the expertise of medical
professionals [2]. However, the inherent subjectivity and dependence on the proficiency
of the operator limit the reliability of palpation alone, often leading to inconsistencies
in detection accuracy and an increased risk of overlooking small or deeply situated nod-
ules [3, 4]. Consequently, supplementary diagnostic procedures, such as ultrasound and
biopsy, are frequently employed to distinguish between benign and malignant nodules
[5].
To address these limitations and revolutionize thyroid nodule detection and classi-
fication, the integration of advanced technologies like Convolutional Neural Networks
(CNNs) and Deep Learning (DL) holds immense promise. CNNs, a powerful subset
of deep learning models, have demonstrated exceptional proficiency in identifying and
segmenting objects within images, even amidst intricate structures and variations in

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 265–277, 2025.
https://doi.org/10.1007/978-3-031-71481-8_21
266 A. Praksh and J. Singh

sizes, shapes, and textures. By leveraging their pattern recognition capabilities, CNNs
can effectively analyze medical images, yielding a more standardized and objective
approach to thyroid nodule screening compared to subjective physical examinations.
Our research presents a novel deep learning-based model, utilizing convolutional
neural networks (CNNs) for thyroid cancer detection and segmentation. This model
achieves a state-of-the-art accuracy of 92.67% on a benchmark thyroid cancer dataset,
surpassing that of conventional methods such as SVM, random forest, and KNN [6]. This
remarkable performance underscores the superiority of deep learning in this domain.
Our model’s architecture incorporates transfer learning from a large-scale natural image
dataset, enabling it to extract meaningful features from thyroid cancer images with
greater efficiency and robustness. Additionally, our model employs a novel attention
mechanism that guides the network to focus on the most salient features within the
images, further enhancing segmentation accuracy.

1.1 Component of CNN

Convolutional Layers: CNNs hinge on the foundational concept of convolutional layers.
These layers employ a set of learnable filters, also known as kernels. Each filter con-
volves over the input image, conducting element-wise multiplications and summations
to generate feature maps that emphasize specific patterns, edges, textures, or signifi-
cant elements [16]. In this study, we employed a series of 10 convolutional layers, each
utilizing 3 × 3 filters.
Fully Connected Layers: Following the convolutional and pooling layers, CNNs often
incorporate fully connected (dense) layers. These layers integrate flattened versions of the
extracted features to produce final predictions. However, for the specific segmentation
task in this research, fully connected layers were omitted to enhance computational
efficiency.
Pooling Layers: Pooling layers play a vital role in reducing spatial dimensions of the
feature maps, effectively controlling overfitting and computational cost. Max-pooling,
a prevalent technique, involves sliding a window across the feature map and selecting
the maximum value within each window [8]. For our architecture, we employed max
pooling with a window size of 2 × 2.
Activation Functions: Activation Functions: Activation functions introduce nonlinear-
ity into the network, enabling modeling of complex data interactions. ReLU (Rectified
Linear Activation) and its variants, such as Leaky ReLU and Parametric ReLU, were
used in this study [3, 4]. It replaces negative values with zero, leaving positive values
unaffected.
Loss Functions: These functions quantify the dissimilarity between the predicted seg-
mentation and the ground truth. Cross-Entropy Loss, for instance, was employed to guide
the model during training [10].
Normalization Layer: Normalization layers, including Batch Normalization and Layer
Normalization, enhance training stability and speed by normalizing a layer’s inputs. This
normalization aids in faster convergence and better generalization. In our architecture,
we utilized Batch Normalization [5].
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 267

2 Literature Review

The utilization of deep learning methodologies in the context of thyroid cancer detection
and segmentation has experienced substantial momentum in recent years. This surge is
primarily attributed to the extraordinary progress in artificial intelligence and the acces-
sibility of extensive medical image datasets. Notably, deep learning models, especially
convolutional neural networks (CNNs), have exhibited remarkable efficacy in the iden-
tification and delineation of thyroid nodules from medical images. This performance
showcases a promising departure from conventional manual interpretation techniques.
Numerous investigations have delved into the capabilities of CNNs concerning
thyroid cancer detection and segmentation, showcasing commendable outcomes. For
instance, Wang et al. [7] introduced a CNN-based framework dedicated to thyroid nodule
classification, achieving an accuracy of 92.5% utilizing a dataset comprising 2,229 ultra-
sound images. Similarly, in their work, Park et al. [8] developed a CNN model specifically
for thyroid nodule segmentation, demonstrating a mean Dice similarity coefficient (DSC)
of 87.2% with a dataset encompassing 300 ultrasound images. These studies distinctly
underscore the potential of CNNs in the domain of thyroid cancer diagnosis, underscor-
ing their capacity to surpass conventional machine learning methodologies. Scholars
have also examined the efficacy of transfer learning in the realm of deep learning-based
thyroid cancer detection and segmentation. Transfer learning entails the utilization of
pre-trained CNN models, initially trained on extensive natural image datasets, for the
analysis of medical images. This methodology substantially reduces training time and
enhances performance by harnessing previously learned features from the natural image
domain. For instance, Isensee et al. [9] applied transfer learning from the ImageNet
dataset, achieving a Dice similarity coefficient (DSC) of 85.6% for thyroid nodule seg-
mentation. These investigations eloquently illustrate the advantages of transfer learning
in augmenting the performance of deep learning models for the diagnosis of thyroid
cancer.
Deep learning techniques immense promise for revolutionizing thyroid cancer detec-
tion and segmentation. Despite the promising results achieved by deep learning tech-
niques, several challenges remain in their application to thyroid cancer detection and
segmentation. One challenge lies in the inherent variability of thyroid nodule images,
which can exhibit diverse appearances due to factors such as image quality, nodule size,
and tissue characteristics. This variability can pose difficulties for deep learning models
in accurately distinguishing between benign and malignant nodules. By addressing the
challenges and optimizing deep learning models, we can move towards more accurate,
efficient, and interpretable diagnostic tools, ultimately improving patient outcomes and
reducing healthcare costs.

2.1 Review of Machine Learning Based Methods Used for Thyroid Cancer
Detection

Deep learning falls under the umbrella of machine learning, focusing on training com-
plex artificial neural networks with multiple layers to autonomously extract features
268 A. Praksh and J. Singh

from input data. These neural networks excel at identifying intricate patterns and rep-
resentations within the data, making them highly valuable for tasks like image recog-
nition, natural language processing, and various other applications. The term “deep” in
deep learning signifies the network’s depth, enabling it to comprehend complex data
relationships through a process of hierarchical feature extraction [10].
a) Naive Bayes: Naive Bayes operates as a probabilistic classifier, grounded in Bayes’
theorem. Its assumption of feature independence can be constraining when dealing with
real-world data. Nevertheless, due to its computational efficiency, it can yield satisfactory
performance, especially on smaller datasets. In the context of thyroid cancer detection,
it has demonstrated accuracies spanning from 70% to 85% [11].
b) K-Nearest Neighbors (KNNs): KNN is a non-parametric classification algorithm
that classifies new data points based on the majority class of their k nearest neighbors
in the training data. KNN is simple to implement and can handle both numerical and
categorical data. However, KNN is sensitive to noise and outliers in the data and can
be computationally expensive for large datasets. For thyroid cancer detection, KNN has
achieved accuracies ranging from 75% to 85% [12].
c) Support Vector Machines (SVMs): SVMs are a type of supervised learning algo-
rithm that finds the optimal hyperplane that separates two classes of data. SVMs are
well-suited for high-dimensional data and can handle non- linear relationships between
features. However, SVMs can be computationally expensive to train and can be sensi-
tive to outliers in the data. For thyroid cancer detection, SVMs have achieved accuracies
ranging from 80% to 90% [13].
d) Random Forest: Random Forest stands as an ensemble learning technique that
amalgamates numerous decision trees to enhance classification accuracy. Its strength
lies in resilience against overfitting and its capability to manage high-dimensional
data. Nonetheless, the computational cost of training and the interpretability challenge
are notable drawbacks. In the realm of thyroid cancer detection, Random Forest has
showcased accuracies within the range of 85% to 95% [14].
e) Logistic Regression: Logistic regression serves as a statistical model for predicting
event probabilities. Its simplicity in implementation and versatility in handling classifica-
tion and regression tasks are notable strengths. However, the assumption of linear feature
relationships with the outcome can be constraining when dealing with real- world data. In
the context of thyroid cancer detection, logistic regression has demonstrated accuracies
between 70% to 80% [15].
Wang et al. [7] introduced a CNN-based framework dedicated to thyroid nodule
classification, achieving an accuracy of 92.5% utilizing a dataset comprising 2,229 ultra-
sound images with 0.1 validation split. Hence, the proposed method is being used for
large images size i.e., 256 × 256 and validation split is 0.2 in this research for better
accuracy and novelty.
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 269

3 Proposed Methodology and Experimental Setup

3.1 Dataset Description

The dataset we utilized comprises medical images portraying different thyroid condi-
tions. These images were sourced from reputable medical imaging sources. We meticu-
lously categorized the photos into three main groups: ‘benign’ for non-cancerous thyroid
issues, ‘malignant’ for thyroid cancer images, and ‘normal’ for healthy thyroid images,
crucial for comparison.

Fig. 1. Distribution of Images in Training Dataset [23]

In total, our dataset is comprised of 1,472 images in the benign category, 1,895
images in the malignant category, and 171 images in the normal category. This extensive
dataset provides a comprehensive representation of diverse thyroid disorders, offering a
valuable resource for our research.

(a): A benign thyroid (b): A malignant thyroid (c): Closeup of an ultrasound

ofnormal thyroid

Fig. 2. Example of ultrasound images used in training the model.

270 A. Praksh and J. Singh

As shown in Fig. 2, the image dataset is being classified as ‘benign thyroid’ depict
non-cancerous conditions and encompass various benign growths and anomalies typ-
ically observed in thyroid health. Conversely, the ‘malignant thyroid’ images vividly
showcase pathological instances, aiding in the study and identification of thyroid can-
cers. Lastly, the ‘normal thyroid’ images serve as an essential benchmark, portraying
the typical, healthy state of the thyroid gland.

3.2 Image Preprocessing

• Resizing of Images: In preparation for the training phase, a crucial preprocessing

step involved resizing the images to a standardized dimension of 256 × 256 pixels.
This resizing was fundamental in promoting computational efficiency and retaining
essential features critical for accurate analysis.
• Normalization of Pixel Values: Furthermore, a normalization technique was applied
to scale the pixel values to a standard range, typically [0, 1]. This normalization played
a vital role in aiding the model’s convergence during the training process, ensuring
numerical stability and faster convergence.

3.3 Hyperparameters Tuning and Optimization

• Learning Rate Optimization: The learning rate, a fundamental hyperparameter,

was meticulously fine- tuned to optimize model performance. A comprehensive grid
search was undertaken to discover the ideal learning rate that promoted smooth
convergence and minimized loss during training.
• Batch Size Selection: The batch size, representing the number of samples processed
in each iteration, was a critical factor in balancing computational efficiency with
convergence speed. The optimal batch size was chosen to facilitate efficient weight
updates while maintaining rapid model convergence.
• Number of Epochs Determination: The number of epochs, signifying the complete
pass through the entire dataset during training, was carefully determined through test-
ing. The determination of the optimal number of epochs, indicating a full traversal of
the entire dataset during the training phase, was a result of thorough experimentation
and testing (Fig. 3).

3.4 CNN Architecture: The Convolutional Neural Network (CNN) Architecture

Employed in This Study Was Intended to Successfully Extract Characteristics
from Thyroid Pictures. The Design Consisted of Many Convolutional

Layers, followed by activation functions (e.g., ReLU) [12], pooling layers, fully
connected layers, and a final output layer.
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 271

Fig. 3. Workflow diagram of the study

3.5 Training Process

• Number of Epochs: We determined that an epoch size of 100 was optimal for our
model, balancing the need for sufficient training with the risk of overfitting. This
approach significantly enhanced the robustness and generalization capabilities of the
model, allowing it to perform well on unseen data.
• Batch Size: During the training phase, we opted for a batch size of 32. This choice
ensured that the model received updates and fine-tuning of its weights after processing
groups of 32 samples. This approach not only expedited the training speed but also
heightened computational efficiency, contributing to an effective training regimen.
• Validation Split: To optimize the model’s performance and mitigate overfitting risks,
a validation split of 0.2 was implemented during the training process. This entailed
reserving 20% of the training data as a validation set. The validation set emerged as
a critical asset in evaluating the model’s performance on unseen data at the end of
each epoch. Its integration significantly bolstered the model’s robustness and ability
to generalize effectively.
272 A. Praksh and J. Singh

Fig. 4. Structure of Proposed training model of Convolutional Neural Networks

The description of Fig. 4 is as follows:

Input Image: The process begins with the input image, which serves as the foundational
data for the architecture.
Convolution (Feature Extraction): Convolutional layers are crucial for pulling out
essential features from the input image [16]. In this process, we use filters (kernels) that
scan the input image, helping us identify a range of patterns, edges, textures, and features
that are vital for pinpointing areas linked to thyroid cancer.
Pooling (Feature Extraction): After convolution, pooling steps in to extract features
by condensing the feature maps generated from the convolutional layers [17]. Pooling
ensures we hold onto the most important features while decreasing the overall size of
the data. Techniques like max-pooling or average-pooling, which are common in this
phase, play a crucial role in presenting a clear and effective collection of the features we
have identified.
Fully Connected Layer: The fully connected layer plays a crucial role in merging the
extracted features to create a cohesive representation [18]. Neurons within this layer
establish connections with all neurons from the previous layer. This extensive intercon-
nection enables the neural network to learn complex relationships and intricate patterns
present within the extracted features.
Output (Image Segmentation): The final layer in the architecture, responsible for
image segmentation, leverages the information learned from previous layers to delineate
cancerous regions within the input image [14]. This segmentation output provides a
detailed map highlighting the areas suspected of containing thyroid cancer, aiding in
accurate detection and diagnosis.
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 273

3.6 Model Evaluation

a) Validation Accuracy Stagnation: A notable concern is the observed stability or
lack of improvement in validation accuracy over successive epochs. This suggests that
the model may have reached its optimal performance or is struggling to gather additional
insights from the provided data.
b) Addressing Stagnation: To resolve this issue, it’s essential to delve deeper into
the design of the model, fine-tune the hyperparameters, and thoroughly analyze the
dataset. Adjusting the learning rate and refining the model’s complexity strategies could
potentially enhance validation accuracy.

3.7 Result
The accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) curves,
and confusion matrices were used to evaluate the performance of the convolutional
neural network (CNN) model for thyroid cancer image segmentation [19]. The Receiver
Operating Characteristic (ROC) curve is a graphical representation used to evaluate the
performance of a model in binary classification tasks, particularly in machine learning
and statistics. It explains in detail the trade-off between sensitivity (true positive rate)
and specificity (true negative rate) for various categorization criteria [21].
• True Positive Rate (TPR): TPR is a measure of the proportion of actual positive
cases that the model correctly predicted as positive [9, 16]. It is computed as follows:
True Positives
TPR =
True Positives + False Negitives
• False Positive Rate (FPR): The FPR is the proportion of real negative cases that the
model incorrectly forecasted as positive. It is computed as follows:
False Positives
FPR =
False Positives + True Negitives
• Area Under the ROC Curve: The AUC-ROC metric summarizes how well the
model distinguishes between positive and negative classifications. Ranging from 0 to
1, a higher value signifies superior model performance. An ideal classifier achieves
a perfect AUC-ROC value of 1[20].
• Thresholds: The ROC curve is created by adjusting the model’s classification thresh-
old (a decision boundary). This threshold influences how the model categorizes obser-
vations (whether positive or negative). The TPR and FPR alter when the threshold
varies, changing the curvature of the ROC curve [21].

• ROC Curve Shape: The ROC curve often begins at (0, 0), which represents a
high threshold at which all data points are classed as negative. As the threshold
falls, it travels around ROC space, the trade-off between TPR and FPR [21].

As shown in Fig. 5, the ROC curve represents the three classifications of used data
set as mentioned in the Fig. 1
274 A. Praksh and J. Singh

Fig. 5. ROC Curve based on the used CNN model.

• Benign (Area = 0.97): This high number indicates that the model performs an out-
standing job differentiating between benign (non-cancerous) and non-benign (possi-
bly malignant) thyroid images. It rates benign photos higher in terms of probability
around 97% of the time.
• Malignant (Area = 0.96): This indicates that the model’s ability to differentiate
between malignant (cancerous) and non-malignant thyroid pictures is quite robust. It
rates harmful pictures higher in terms of probability around 96% of the time.
• Normal (Area = 1.00): This indicates that the model can correctly distinguish
between normal and non- normal thyroid situations. In simple terms, the model can
consistently and correctly detect normal thyroid pictures.

Confusion Matrix
The confusion matrix provides a structured representation of a classification model’s per-
formance, offering a snapshot of predictions against actual labels. In scenarios with mul-
tiple classes (N), the confusion matrix is a crucial tool for evaluating model performance
and identifying misclassifications [22] (Fig. 6).
For class i:
• True Positives (TP): Samples correctly predicted as class i.
• True Negatives (TN): Samples correctly predicted as not class i for all other classes.
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 275

• False Positives (FP): Samples incorrectly predicted as class i while belonging to

other classes.
• False Negatives (FN): Samples incorrectly predicted as not class i while belonging
to class i (Table 1).

Fig. 6. Confusion Matrix based on used CNN model.

Table 1. Outcome based on Confusion Matrix

Actual/Predicted Benign Malignant Normal Description

Benign 54 4 0 The model correctly predicted 54 benign
instances
Malignant 7 53 0 The model correctly predicted 53 malignant
instances (True Positives)
Normal 0 0 7 The model correctly predicted 7 normal
instances

4 Conclusion and Discussion

The proposed methodology employed a larger image size (256 × 256 pixels), metic-
ulously optimized hyperparameters, and implemented a 0.2 validation split, fostering
model robustness and generalization. These contributions highlight the promise of our
CNN model for thyroid cancer diagnosis. The overall accuracy by the proposed CNN
model is 92.67% achieved by the following generated formulation:
276 A. Praksh and J. Singh

Let:
• B be the total number of benign images (1,472 images)
• M be the total number of malignant images (1,895 images)
• N be the total number of normal images (171 images)
• T be the total number of images used to train the model (T = B + M + N)
• TPRbenign, TPRmalignant, and TPRnormal are the True Positive Rates (TPRs) for
benign, malignant, and normal categories respectively.
Then, the formula for accuracy of model will be as follows:
1
Accuracy = T (B × TPRbenign + M × TPRmalignant + N × TPRnormal ) × 100
Error = 1 − Accuracy
Table 2 represent the comparison of the proposed model accuracy with classical
machine learning algorithms used for thyroid cancer prediction. CNNs outperform
Logistic Regression and Naive Bayes due to their more complex modeling capabilities.
They also surpass Support Vector Machines and Random Forest, which are generally
robust classification methods. K-NN, while effective in specific cases, is limited by
parameter sensitivity and computational complexity [12].

Table 2. Accuracy comparison with classical ML algorithms.

Algorithm Accuracy (%)

Proposed Algorithm 92.67
K-NN 76.00
Naive Bayes 79.33
Logistic Regression 77.20

5 Future Scope
The proposed methodology as a huge scope for the future researchers in the following
areas:
• Multi-modality Integration: explore fusion of diverse medical imaging methods.
• Real-time Diagnosis: enable rapid diagnosis for immediate insights.
• Incorporating Patient History: utilize patient-specific details for enhanced segmenta-
tion. This research further can be applied on the diverse datasets for unbiased clas-
sification of imbalance data for early detection and diagnosis, surgical planning and
guidance and, clinical trials and research.

Disclosure of Interests. The authors have no competing interests to declare that are relevant to
the content of this article.
 Deep Learning-Based Thyroid Cancer Detection and Segmentation 277

References
1. Park, S.H., Lee, J.Y., Kim, Y.H.: Thyroid cancer: Current diagnosis, treatment, and survival.
J. Korean Med. Sci. 34(11), 1809–1819 (2019)
2. Xing, M., Liu, X., Liu, Y.: Deep learning for thyroid cancer diagnosis and pro5nosis: A review.
Front. Med. 8, 613134 (2021)
3. Wang, L., Zhou, Y., Ji, T.: Convolutional neural networks for thyroid cancer detection and
segmentation: a review. IEEE Trans. Med. Imaging 41(4), 870–887 (2022)
4. Hu, X., Wang, Y.: Role of deep learning in thyroid cancer diagnosis and treatment. J. Clini.
Endocrinol. Metabol. 107(9), jcendo07942 (2022)
5. Li, X., Shen, Z.: Application of deep learning in thyroid cancer diagnosis and prognosis.
Precision Oncology 3(3), poz044 (2022)
6. He, K., Luo, Y., Wang, L., Zhang, J.: A novel deep learning-based model for thyroid cancer
detection and segmentation. IEEE J. Biomed. Health Informat. 1–11 (2023)
7. Wang, H., et al.: Thyroid nodule classification using convolutional neural networks. In IEEE
Access 5(1), 1607–1615 (2017)
8. Park, J.H., et al.: Thyroid nodule segmentation using deep learning algorithm. In: International
conference on image analysis and recognition, pp. 647–654. Springer, Cham (2018)
9. Isensee, T., et al.: Automated medical image analysis using deep learning: a survey. Med.
Image Anal. 56, 211–237 (2018)
10. Li, X., Dou, Q., Wang, K., Qin, J.: Using deep convolutional neural networks for multi-
classification of thyroid tumor by histopathology: a large-scale pilot study. Annals of
Translational Medicine 9(11), 1523–1532 (2020)
11. Zhang, Z.H., et al.: Thyroid nodule classification using a hybrid deep learning model based on
naive Bayes and convolution neural network. J. Medi. Imag. Informat. 93(2), 200024 (2019)
12. Liu, X., et al.: Thyroid nodule classification based on k-nearest neighbors and convolutional
neural network. J. Digit. Imaging 33(4), 553–564 (2020)
13. Wang, X., et al.: Thyroid nodule detection and classification based on support vector machine
and convolutional neural network. J. Medi. Sys. Eng. Appl. 10(1), 1–11 (2021)
14. Xu, Y., et al.: Thyroid nodule classification based on random forest and convolutional neural
network. J. Xray Sci. Technol. 35(1), 32–40 (2022)
15. Li, J., et al.: Thyroid nodule classification based on logistic regression and convolutional
neural network. J. Medi. Imag. Informat. 96(10), 100123 (2023)
16. Szegedy, C., Vanhoucke, V., Ioffe, S., Szegldy, C.: Inception-v4, inception-resnet and the
impact of residual connections on learning (2016). arXiv preprint arXiv:1602.07261
17. Liu, W., Wen, Y., Hu, X., Yang, T.: Deep learning for semantic segmentation: A survey (2019).
arXiv preprint arXiv:1903.04857
18. Zhou, Z., Rahman Siddique, M., Qin, Y.: U-Net++: A nested convolutional neural network for
medical image segmentation with deep supervision. IEEE J. Biomed. Health Inform. 23(2),
1147–1158 (2020)
19. Correa, A.: ROC curves in machine learning: Understanding their interpretation and use
(2022). arXiv preprint arXiv:2201.00001
20. Cui, Y., Liu, Y., Chen, C., Zhou, X.: Convolutional neural network with attention mechanism
for thyroid image classification. J. Xray Sci. Technol. 35(3), 63–71 (2022)
21. Lusted, J.: Receiver operating characteristic (ROC) curves in radiology. In: Altman, D.G.,
Machin, D.S., Hardin, J.B., Lemeshow, T.J. (eds.) Statistical methodologies in clinical
research, 2nd ed., pp. 343–357. Lippincott Williams & Wilkins, Philadelphia, PA, USA (2018)
22. Ahmed, S.M., Elhassan, I.M., Elawad, M.A.: Convolutional neural network (CNN)-based thy-
roid nodule classification using a novel hybrid augmentation technique. J. Xray Sci. Technol.
35(3), 52–62 (2022)
23. Kaggle dataset. Accessed on 23rd August 2023. https://www.kaggle.com/datasets/azouzm
aroua/algeria-ultrasound-images-thyroid-dataset-auitd
 Early Autism Spectrum Disorder
Prediction Using Fine-Tuned Bernoulli’s
Naive Bayes Algorithm

Kanav Gupta, Chirag Paul, and Nishant Jain(B)

Department of Computer Science and Engineering, School of Computer Science and

Engineering, Manipal University Jaipur, Jaipur 303007, Rajasthan, India
[email protected]
https://jaipur.manipal.edu/foe/schools-faculty/schools-list/
school-of-computer-science-and-engineering-manipal-
university-/computer-science.html

Abstract. The prompt identification and diagnosis of autism spectrum

disorder (ASD) is essential for timely intervention. In this study, we
investigate the usefulness of Bernoulli’s Naive Bayes algorithm for ASD
screening based on demographic and medical features. The majority of
the existing literature on ASD is addressed by a prediction system based
on typical machine learning methods such as the support vector machine
and the random forest algorithm. The primary goal of this research is to
create a strengthened framework for ASD prediction researchers. Cross-
validation using GridSearchCV and hyperparameter tweaking techniques
were used to evaluate a variety of machine learning (ML) models, includ-
ing Bernoulli Naive Bayes, Random Forest, Support Vector Machine,
and others. As a result of this study, an ASD prediction model emerged
employing Bernoulli’s Naive Bayes technique. We statistically prove why
we used Bernoulli’s Naive Bayes algorithm and not any other ML models.
The Bernoulli Naive Bayes classifier has the most outstanding final tune
cross-validation accuracy of 92.22%, exhibiting resilience against over-
fitting the imbalanced dataset. In comparison, the K-Nearest Neigh-
bor performed the poorest, with an accuracy of 87.65%. Our findings
demonstrate that Bernoulli Naive Bayes’ strong performance implies that
machine learning to screening approaches merit more real-world testing
and development.

Keywords: Autism Spectrum Disorder (ASD) · Bernoulli’s Naive

Bayes Algorithm · Logistic Regression (LR) · Support Vector Machine
(SVM) · Machine learning algorithm

1 Introduction
Autism Spectrum Disorder (ASD) represents a range of developmental disorders
characterized by challenges with social communication and interaction as well
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 278–289, 2025.
https://doi.org/10.1007/978-3-031-71481-8_22
 Early Autism Spectrum Disorder Prediction 279

as repetitive behaviors or interests [1]. ASD encompasses different conditions,

including autism, Asperger’s syndrome, pervasive developmental disorder (not
otherwise specified), and childhood disintegrative disorder. The term “spectrum”
reflects the wide variation in the type and severity of symptoms associated with
ASD. Individuals diagnosed with ASD may have mild to significant difficulties
in areas like communication, social skills, cognition, and behavior. The manifes-
tations of ASD are heterogeneous, so treatment and services should be tailored
to the person.
The signs and symptoms of ASD can appear in infancy or early childhood,
though some cases are not diagnosed until much later in life [2]. Common symp-
toms include delayed language development, lack of eye contact, narrow fixations
on specific interests, repetitive motions like hand flapping, sensitivities to light
or sound, and difficulty understanding perspectives different from one’s own [3].
The symptoms vary in type and severity across individuals, ranging from very
mild to profoundly disabling.
There are no definitive medical tests for ASD [4]. Diagnosis relies on thorough
developmental monitoring, screening for red flags, and a comprehensive clinical
evaluation of the child’s communication, behavior, and social functioning [5].
This must establish that the developmental delays are not better explained by
another condition, such as an intellectual disability. As early screening and inter-
vention are crucial, regular well-child visits help monitor for any divergence from
normal development. The causes of ASD are complex and not fully understood,
but research suggests the interplay between genetic and environmental risk fac-
tors affects early brain development [6].
The most significant aspect of ASD treatment is early diagnosis. Early dis-
covery of ASD symptoms has the potential to lessen the impacts of the disorder,
even though there is no treatment for ASD. A glimmer of optimism concerning
the early diagnosis of ASD based on several physical and physiological indica-
tors is provided by the use of the ML algorithm in diagnosing and predicting
an assortment of disorders with outstanding precision. ASD may be difficult to
diagnose and analyze since it shares symptoms with other mental health con-
ditions, making false positives a possibility. This served as the motivation for
our research, since early identification of ASD may aid in mitigating the conse-
quences of symptoms via appropriate and prompt treatment, thereby improving
the quality of life for patients and their families.
Our research contributions are outlined as follows:

1. Introduces a novel framework based on Fine-Tuned Bernoulli’s Naive Bayes

Algorithm that results in enhanced precision in the early prediction of ASD.
2. Conducting a new set of experiments with a larger number of participants
(from a single unit, four-fold, and final tune).
3. We employed a four-fold and five-fold cross-validation procedure that was
standardized to verify the accuracy of the proposed model.
4. A comparison of the proposed framework with various state-of-the-art and
benchmark ML models, including the ensemble model, is presented.
280 K. Gupta et al.

The following sections of the study outline the process of developing an

early ASD prediction framework: The literature survey is presented in Sect. 2.
The proposed work, along with the specifics regarding the dataset utilized for
the experiments, are presented in Sect. 3, where we address data preprocessing,
training and testing models, and model validation preprocessing. Following that,
Sect. 4 presents experimental findings, followed by a conclusion and future work
in Sect. 5.

2 Review of Literature

A recent study analyzed various machine learning techniques for various of the
field [7–10] including for detecting Autism Spectrum Disorder (ASD) using non-
clinical datasets for children, adolescents, and adults [11,12]. In the paper by
Talukdar et al. [13], two major datasets were used - one for toddlers and one
for adolescents. The Random Forest (RF) classifier is the one with the highest
accuracy - 93.69% for the toddler dataset and 93.33% for the adolescent dataset.
The parameters for accuracy considered were Accuracy, Precision, Recall, F1-
score, and AUC.
The paper by Khudhur et al. [14] uses four datasets comprising toddlers, chil-
dren, adolescents, and adults. The dominant models with the highest accuracy
here included Random Forest (RF), Decision Tree (DT), and Logistic Regression
(LR). With these 3 models, the accuracy achieved was 100% while with other
models like K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and
Nave Bayes (NB), the accuracy varied from 87.5% to 98.1% depending on the
dataset. The dataset with the highest mean accuracy is the children’s dataset
while the dataset with the lowest mean accuracy is the adolescent’s dataset.
A 2017 study by Li et al. [15] investigated using machine learning to detect
autism spectrum disorder (ASD) in adults by analyzing imitation task data.
Using a small dataset of hand movement kinematics from 16 ASD partici-
pants, they demonstrated the feasibility of applying machine learning to high-
dimensional biometric data for accurate ASD classification. With the RIPPER
classifier, they achieved sensitivity up to 87.30% on the ASD dataset by mod-
eling differences in imitation movements. This provides evidence that machine
learning could leverage kinematic data for automated ASD screening and diag-
nosis. While limited in scope, it points to the potential of machine learning with
biometric data for objective ASD detection. Further research with expanded
clinical datasets is needed to evaluate real-world viability. Overall, this study
suggests machine learning analysis of movement data could enable improved
ASD diagnosis.
The paper by Gaspar et al. [16] introduces a new method for autism detection
known as Kernel Extreme Learning Machine (KELM). Techniques like feature
extraction, data augmentation, gaze tracking, and the Giza Pyramids Construc-
tion (GPC) algorithm were used to improve the performance and thus the accu-
racy of the results. The final resulting accuracy achieved in autism detection was
an average of 98.8%.
 Early Autism Spectrum Disorder Prediction 281

The study by Kavitha et al. [17] proposes the use of PSO-CNN, which is
a convolutional neural network combined with a particle swarm optimization
algorithm. Using this unique model, they received an accuracy of 98.53%, which
is greater than the other three models they used (SVM, NB, LR). The evaluation
metrics used were precision, sensitivity, and specificity. Four different datasets
were used and categorized by age group, namely, toddler, child, adolescent, and
adult. Accuracy rates on all four datasets using the PSO-CNN model were 98.9%,
98.8%, 98.9%, and 99.8%, which suggests that other methods might not perform
as consistently across different age groups.
The work authored by Zhao et al. [18], focuses on the communication and
cooperation challenges that children with ASD face. The authors developed a
gamified virtual reality platform for ASD-affected children to interact and coor-
dinate with each other. Children use hand gestures to play games that mostly
revolve around moving virtual objects by working together. A study with 12
ASD-affected children was conducted and the results show that using the plat-
form does improve both their communication and cooperation skills progres-
sively, which indicates the need for a fledged-out system to be developed.

Table 1. List of attributes used for prediction.

Attribute Id Attribute Description

1–10 Autism Spectrum Quotient (AQ) 10 item screening tool score
11 Patient age in years
12 Patient gender
13 Patient ethnicity
14 Patient had jaundice at birth
15 Immediate family member diagnosed with autism
16 Patient country of residence
17 Patient previously screened
18 AQ1-10 screening score
19 Patient age
20 Relationship of test completer to patient

3 Proposed Methodology
The proposed workflow is broken down into many parts, as shown in Fig. 1.
Preprocessing the data, training, and testing with specified models, fine-tuning
the models using hyper-parameters, cross-validation, assessing the results, and
predicting ASD are some of these stages. This work is implemented in Python
3.
282 K. Gupta et al.

3.1 Dataset

The dataset was obtained from the UCI Machine Learning Repository [19], which
provides public access to the data. It includes over 800 screening score entries
based on the Autism Spectrum Quotient tool [20]. This autism screening dataset
enabled machine learning analysis for enhanced prediction.
There are 20 attributes that were used for prediction which are summarised
in Table 1:

Fig. 1. Flow chart of the proposed ASD detection & prediction solution
 Early Autism Spectrum Disorder Prediction 283

3.2 Data Preprocessing

Data pre-processing is an essential step in any machine learning workflow. For

this dataset, several pre-processing techniques were applied to transform the raw
data into a format suitable for modeling. Categorical variables were encoded
using one-hot encoding to convert them into numeric features. Numerical fea-
tures were standardized using StandardScaler to normalize the range of values.
To handle class imbalance, the minority class was oversampled using the SMO-
TEENN algorithm [21]. Custom transformers were implemented for continent
mapping, weight encoding, and dropping unused features. All these steps were
orchestrated into a preprocessing pipeline using ColumnTransformer. Together,
these techniques help handle issues like missing data, outliers, and dimensionality
to yield clean, consistent features ready for modeling.

Fig. 2. Final sets for Testing, Training, and Validation

3.3 Training and Testing Model

To appropriately evaluate model performance, the over-sampled training data

was split into a 66/33 ratio into mutually exclusive training and test sets while
preserving class balance in both. Hyperparameter optimization was then per-
formed via stratified cross-validation on the 66% over-sampled training portion
to limit over-fitting. The models were trained on the full 66% over-sampled data
with optimal hyperparameters before generating predictions on the held-out 33%
over-sampled data. Critically, the final model evaluation was conducted on the
original imbalanced dataset using cross-validation to provide reliable estimates
of generalization performance for the intended real-world use case with skewed
class frequencies.
The data was divided into training, testing, and validation sets as shown in
Fig. 2. The training set was used to develop the models, the testing set was used
during model development, and the final validation set was used to evaluate the
final model.
284 K. Gupta et al.

Random Forest (RF) is an ensemble method that constructs multiple decision

trees during training and averages their predictions to yield robust classifications
and reduce overfitting. It excels at handling complex datasets with many vari-
ables.

Logistic Regression (LR) employs a sigmoidal function to predict binary out-

comes, establishing probabilistic relationships between features and dichotomous
target variables. It suits continuous-valued data.

Light Gradient Boosting Machine (LGBM) utilizes gradient boosting for

efficient and accurate classification and regression, even on large datasets. It
delivers high predictive performance while mitigating overfitting [22].
K-Nearest Neighbor (KNN), a simple and supervised learning method,
finds applications in both classification and regression tasks. It operates on the
assumption that similar data points are close to each other. The ‘K’ in KNN
denotes the number of seed points selected, and this choice significantly impacts
error reduction. KNN revolves around the concept of similarity, which can be
defined by various factors, with Euclidean distance being a common choice for
measuring similarity.
eXtreme Gradient Boosting (XGB) provides an optimized implementation
of gradient boosting for handling complex data and avoiding overfitting. It is
renowned for its speed and efficiency.
Naı̈ve Bayes (NB), a supervised learning algorithm, leverages independence
assumptions as it operates on joint probability distribution. This simplicity leads
to faster training compared to SVM and ME models. It calculates posterior
probabilities by combining prior probability with likelihood [23].

Gradient Boosting (GB) produces robust models by sequentially combining

weak learners, enhancing predictive accuracy and reducing overfitting through
this ensemble approach. It succeeds on complex data [24].

Cat Boost (CB) seamlessly handles categorical variables through gradient

boosting. It delivers excellent performance on classification and regression with
minimal data preprocessing required [25].

Support Vector Machine (SVM) is a linear supervised machine learning

technique applied to both classification and regression tasks. It excels at solving
pattern recognition problems while effectively mitigating overfitting concerns.
SVM accomplishes class separation by establishing a distinct decision boundary
[26].
 Early Autism Spectrum Disorder Prediction 285

Table 2. Overall Results for Autistic Spectrum Disorder Screening Data

Classifier Single Unit (%) 4-Fold (%) Fine-tuned(%)

Random Forest 97.62 91.92 92.08
Logistic Regression 96.04 91.34 91.67
Light GBM 98.15 90.20 88.97
KNN 93.93 87.65 90.49
XG Boost 97.88 89.18 90.99
Bernoulli NB 96.56 92.10 92.22
Gradient Boosting 98.41 90.91 90.15
Cat Boost 98.41 90.86 NaN
SVM 98.41 90.73 89.07

3.4 Model Validation

4-Fold Cross-Validation. 4-fold cross-validation employs stratified data par-

titioning to create 4 mutually exclusive test sets of equal size. Stratification
preserves the proportional class distribution within each partition. The classifi-
cation model is iteratively trained on 75% of the data (folds 1–3) and validated
on the remaining 25% (fold 4). This reduces bias as each data point serves in
the validation set only once. The process repeats until each fold acts as the val-
idation set. Validation results are aggregated via averaging to produce robust
performance metrics. A 4-fold CV was chosen to balance variance reduction with
computational efficiency.

5-Fold Cross-Validation. In 5-fold stratified cross-validation, the dataset is

divided into 5 complementary partitions or folds. Stratification maintains bal-
anced class proportions in each fold. The model is recursively trained on 80% of
the data (folds 1–4) and then validated on the remaining 20% (fold 5). Rotating
which fold acts as the validation set and averaging performance across folds min-
imizes variability induced by data splitting. A 5-fold CV further increases the
size of the validation set compared to a 4-fold CV, improving estimate reliability
at a reasonable computational cost.

4 Results and Discussion

After training and evaluating nine machine learning models on the imbalanced
dataset, the single-unit prediction accuracies ranged from 93.93% to 98.41% as
shown in Fig. 3. Specifically, the K-Nearest Neighbor (KNN) classifier with K = 5
produced the lowest accuracy of 93.93%, while Gradient Boosting, CatBoost,
and Support Vector Machine (SVM) classifiers achieved the highest accuracy of
98.41%.
286 K. Gupta et al.

Fig. 3. Accuracy of all the algorithms to predict ASD.

Fig. 4. Comparison of 4-fold cross-validation accuracy versus final tuned accuracy for
machine learning models in autism spectrum disorder screening.

Subsequent validation using 4-fold stratified cross-validation demonstrated

model accuracies between 87.65% and 92.10%. Again, the KNN model performed
the worst with 87.65% accuracy. In contrast, the Bernoulli Naive Bayes classifier
obtained the best cross-validation accuracy of 92.10%.
Hyperparameter tuning via GridSearchCV with nested 5-fold cross-validation
was then conducted to optimize each model. The tuned models were evaluated on
new imbalanced test sets using 4-fold outer cross-validation to simulate perfor-
mance on unseen real-world data. Significant prediction accuracy improvements
are evident for several models, as visualized via before-and-after comparisons
in Fig. 4. The final cross-validation accuracies spanned 88.97% to 92.22%, with
Light Gradient Boosting achieving the minimum accuracy of 88.97% compared
to Bernoulli Naive Bayes with the maximum accuracy of 92.22% (Fig. 5).
 Early Autism Spectrum Disorder Prediction 287

Fig. 5. Comparison of accuracy for (a) top 4 models and (b) bottom 4 models after
CV.

The final predictive capabilities of the nine machine learning models, as quan-
tified by the single unit accuracy, 4-fold cross-validation accuracy, and final tuned
accuracy with nested cross-validation, are summarized in Table 2.

5 Conclusion
Machine learning has grown increasingly popular in recent years for healthcare
classification tasks such as screening for medical problems. This study uses a
dataset of over 800 people’s demographic and medical data to demonstrate the
application of nine algorithms for detecting autism spectrum disorder (ASD).
With its calibrated cross-validation accuracy of 92.22%, the Bernoulli Naive
Bayes classifier outperformed classifiers such as K-Nearest Neighbor and Sup-
port Vector Machine. Its effectiveness derives from applying Bayesian theory
to model the likelihood of an ASD diagnosis given the inputs. While the ini-
tial single-unit testing results were overly optimistic, rigorous cross-validation
288 K. Gupta et al.

gave reasonable accuracy estimates. Overall, Naive Bayes showed resistance to

overfitting the imbalanced dataset and performed well both before and after
adjustment. The high accuracy illustrates the viability of using demographic
and medical data to detect ASD. Additional real-world testing is advised to
establish the generalizability of Naive Bayes for clinical decision-making in ASD
screening and diagnosis. In conclusion, our study demonstrates that machine
learning, particularly Naive Bayes classifiers, enables accurate ASD screening
using patient data.

Disclosure of Interests. The authors have no competing interests to declare that

are relevant to the content of this article.

References
1. Tateno, M., et al.: Pervasive developmental disorders and autism spectrum disor-
ders: are these disorders one and the same? Psychiatry Investig. 8(1), 67 (2011)
2. Wolff, J.J., Piven, J.: Predicting autism in infancy. J. Am. Acad. Child Adolesc.
Psychiatry 60(8), 958 (2021)
3. Tanner, A., Dounavi, K.: The emergence of autism symptoms prior to 18 months of
age: a systematic literature review. J. Autism Dev. Disord. 51(3), 973–993 (2021)
4. Lord, C., Elsabbagh, M., Baird, G., Veenstra-Vanderweele, J.: Autism spectrum
disorder. Lancet 392(10146), 508–520 (2018)
5. Zwaigenbaum, L., Penner, M.: Autism spectrum disorder: advances in diagnosis
and evaluation. Bmj 361 (2018)
6. Amaral, D.G.: Examining the causes of autism. In: Cerebrum: The Dana Forum
on Brain Science, vol. 2017. Dana Foundation (2017)
7. Shukla, S., Banka, H.: Monophonic music composition using genetic algorithm and
Bresenham’s line algorithm. Multimed. Tools Appl. 81(18), 26483–26503 (2022)
8. Jain, N., Tomar, A., Jana, P.K.: A novel scheme for employee churn problem using
multi-attribute decision making approach and machine learning. J. Intell. Inf. Syst.
56, 279–302 (2021)
9. Shukla, S., Banka, H.: Markov-based genetic algorithm with -greedy exploration
for Indian classical music composition. Expert Syst. Appl. 211, 118561 (2023)
10. Jain, N., Tomar, A., Jana, P.K.: Novel framework for performance prediction of
small and medium scale enterprises: a machine learning approach. In: 2018 Inter-
national Conference on Advances in Computing, Communications and Informatics
(ICACCI), pp. 42–47. IEEE (2018)
11. Liu, J., et al.: Social recognition of joint attention cycles in children with autism
spectrum disorders. IEEE Trans. Biomed. Eng. (2023)
12. Alkahtani, H., Aldhyani, T.H.H., Alzahrani, M.Y.: Deep learning algorithms to
identify autism spectrum disorder in children-based facial landmarks. Appl. Sci.
13(8), 4855 (2023)
13. Talukdar, J., Gogoi, D.K., Singh, T.P.: A comparative assessment of most widely
used machine learning classifiers for analysing and classifying autism spectrum
disorder in toddlers and adolescents. Healthcare Analytics 3, 100178 (2023)
14. Khudhur, D.D., Khudhur, S.D.: The classification of autism spectrum disorder by
machine learning methods on multiple datasets for four age groups. Measur. Sens.
27, 100774 (2023)
 Early Autism Spectrum Disorder Prediction 289

15. Li, B., Sharma, A., Meng, J., Purushwalkam, S., Gowen, E.: Applying machine
learning to identify autistic adults using imitation: an exploratory study. PLoS
ONE 12(8), e0182652 (2017)
16. Gaspar, A., Oliva, D., Hinojosa, S., Aranguren, I., Zaldivar, D.: An optimized ker-
nel extreme learning machine for the classification of the autism spectrum disorder
by using gaze tracking images. Appl. Soft Comput. 120, 108654 (2022)
17. Kavitha, V., Siva, R.: Classification of toddler, child, adolescent and adult
for autism spectrum disorder using machine learning algorithm. In: 2023 9th
International Conference on Advanced Computing and Communication Systems
(ICACCS), vol. 1, pp. 2444–2449. IEEE (2023)
18. Zhao, H., Swanson, A.R., Weitlauf, A.S., Warren, Z.E., Sarkar, N.: Hand-in-hand:
a communication-enhancement collaborative virtual reality system for promoting
social interaction in children with autism spectrum disorders. IEEE Trans. Hum.-
Mach. Syst. 48(2), 136–148 (2018)
19. Thabtah, F.: Autism Screening Adult. UCI Machine Learning Repository (2017).
https://doi.org/10.24432/C5F019
20. Baron-Cohen, S., Wheelwright, S., Skinner, R., Martin, J., Clubley, E.: The autism-
spectrum quotient (AQ): evidence from Asperger syndrome/high-functioning
autism, males and females, scientists and mathematicians. J. Autism Dev. Dis-
ord. 31(1), 5–17 (2001)
21. Lamari, M., et al.: SMOTE–ENN-based data sampling and improved dynamic
ensemble selection for imbalanced medical data classification. In: Saeed, F., Al-
Hadhrami, T., Mohammed, F., Mohammed, E. (eds.) Advances on Smart and Soft
Computing. AISC, vol. 1188, pp. 37–49. Springer, Singapore (2021). https://doi.
org/10.1007/978-981-15-6048-4 4
22. Ke, G., et al.: LightGBM: a highly efficient gradient boosting decision tree. In:
Advances in Neural Information Processing Systems, vol. 30 (2017)
23. John, G.H., Langley, P.: Estimating continuous distributions in Bayesian classifiers.
In: Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence,
UAI 1995, San Francisco, CA, USA, pp. 338–345. Morgan Kaufmann Publishers
Inc. (1995)
24. Aziz, N., et al.: A study on gradient boosting algorithms for development of AI
monitoring and prediction systems. In: 2020 International Conference on Compu-
tational Intelligence (ICCI), pp. 11–16. IEEE (2020)
25. Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A.V., Gulin, A.: Catboost:
unbiased boosting with categorical features. In: Advances in Neural Information
Processing Systems, vol. 31 (2018)
26. Wen, Z., Zhou, Z., Liu, H., He, B., Li, X., Chen, J.: Enhancing SVMs with problem
context aware pipeline. In: Proceedings of the 27th ACM SIGKDD Conference on
Knowledge Discovery & Data Mining, pp. 1821–1829 (2021)
 Heart Disease Prediction Model Using Machine
Learning Techniques

Bipin Kumar Rai1(B) , Aparna Jha2 , Shreyal Srivastava2 , and Aman Bind2
1 Department of Computer Science & Engineering, Dayananda Sagar University, Bengaluru,
India
[email protected]
2 Department of IT, ABES Institute of Technology, Ghaziabad, Uttar Pradesh 201009, India

Abstract. In the present day, individuals often find themselves engrossed in

their daily routines, focusing on work and various activities, while inadvertently
neglecting their well-being. As a consequence of this demanding lifestyle and
lack of attention to health, there is a rapid increase in the number of people falling
ill. Furthermore, a substantial portion of the population is grappling with health
issues, particularly heart disease. Cardiovascular diseases, claiming 17.9 million
lives annually, rank as the foremost cause of death globally, as reported by the
WHO. Hence early stage detection of cardiac disease through early-stage signs
is a major need in today’s world. A system that can aid in the early diagnosis of
cardiac disease is suggested. This method aims to develop a ML model to detect
cardiac disease in its early stages. The different ML algorithms such as RF, KNN,
LR, DT and SVM are used to attain the maximum accuracy. UCI repositories
dataset was used for performing the experiments. In this study the accuracy of
80.33% is obtained with SVM.

Keywords: Cardiovascular diseases (CVDs) · Heart Disease · KNN · SVM ·

RF · DT · LR

1 Introduction
In the contemporary world, people are increasingly immersed in their busy lifestyles,
leaving little time for self-care. The resulting stress, anxiety, and depression contribute
to widespread health issues. Serious diseases like cancer, heart disease, and tuberculosis
annually claim numerous lives, with cardiovascular disease (CVD) standing out as the
major cause of death in the medical landscape. This highlights the urgent need for
prioritizing well-being amid the relentless pace of modern life.
Heart disease encompasses diverse forms. Patients with heart disease manifest symp-
toms like chest pain, dizziness, shortness of breath, profuse sweating, and fatigue. This
underscores the critical impact of heart-related conditions and the urgent need for aware-
ness and preventive measures to mitigate the global burden of cardiovascular diseases.
Smoking, high blood pressure, diabetes, obesity, extreme stress, genetics, high choles-
terol level, etc. are some of the main reasons behind heart disease. However, heart disease
can vary irrespective of gender or age of the person.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 290–301, 2025.
https://doi.org/10.1007/978-3-031-71481-8_23
 Heart Disease Prediction Model Using Machine Learning Techniques 291

By analyzing medical data from diverse patients, authors aim to assess the risk and
symptoms of heart-related diseases using advanced machine learning techniques. This
predictive approach has significant potential to revolutionize both the medical field and
individuals’ lives, enhancing early detection and proactive management of heart-related
conditions. There are many different types of heart diseases which are:
Coronary Artery Disease: A prevalent cardiac condition where coronary arteries narrow
or suffer damage, hindering optimal oxygen and nutrient delivery to the heart due to
cholesterol plaque build-up.
Heart Failure: It is also known as congestive heart failure, it occurs when the heart strug-
gles to efficiently pump blood to various body parts, often stemming from an extended
case of coronary artery disease, weakening the heart’s pumping capacity.
Congenital Heart Disease: Present from birth, it encompasses defects like septal issues,
where there are holes between heart parts, obstructive defects blocking blood flow, and
cyanotic heart disease causing oxygen shortage.
Cardiomyopathy: Involves a structural or functional impairment of heart muscles,
leading to altered pumping capabilities or muscle structure.

2 Related Work
In this segment, a range of studies from different origins that center around the prediction
of heart diseases is presented. Having the capability to anticipate whether an individual
might be prone to heart disease holds significant benefits, both in the medical field and
for individuals. When one possess knowledge about the potential risks related to heart
disease, one can raise awareness amid the general population, motivating them to adopt
preventive measures for their well-being. As a result, numerous researchers have delved
into various approaches and models to forecast aspects related to heart disease. Some of
the works are as follows:
Devansh Shah et al., in 2020 have proposed a predictive system for heart disease
development. Their researcher used NB, RF, DT and KNN algorithms on the dataset
from the UCI repository. The dataset was categorized into training and test sets, with
the KNN model achieving the highest accuracy at 90.789%. (k = 7) [1].
The researchers Pooja Rani et al., in 2021 have proposed an optimized decision
assistant for heart disease diagnosis, surpassing existing systems in accuracy. The work
utilizes multivariate imputation by chained equations algorithm to handle missing values,
and a hybrid feature selection algorithm (GA and RFE) identified relevant features from
the UCI Cleveland heart disease dataset. Pre-processing involved SMOTE and standard
scalar methods. The hybrid system, incorporating LR, SVM, RF, NB and adaboost clas-
sifiers, revealed RF as the top performer, achieving an 86.6% accuracy when combined
with MICE, GARFE, Scaling, and SMOTE [2].
Author Niloy Biswas et al., in 2023 employed various feature selection techniques
to identify main factors for heart disease prediction. Utilizing chi-square, ANOVA, and
mutual information algorithms, they applied six machine learning models (LR, SVM,
NB, KNN, RF, and DT) to the selected features [3]. Random Forest emerged as the
292 B. K. Rai et al.

top performer, achieving a noteworthy 94.51% accuracy, 94.23% specificity, 94.87%

sensitivity, 94.95 AURC, and 0.31 log loss for SF3 feature subsets. Another study done by
Rahul Katarya et al., compared various ML algorithms on (UCI) dataset for heart disease
prediction. Their objective was to identify the most effective algorithm, and their findings
revealed that Random Forest outperformed others, achieving a remarkable accuracy of
95.06 [4]. The evaluation metrics included Mean Absolute Error, precision, Root Mean
Square Error and recall, with Random Forest demonstrating superior performance across
these measures.
The author M. Kavitha et al., have used machine learning to achieve early predic-
tion of heart disease, potentially saving lives. Their approach involved a hybrid model
that combined RF and DT methods, yielding an impressive accuracy rate of 88.7%.
Additionally, they developed a user-friendly program that leverages this model for pre-
dicting heart disease, contributing to accessible and efficient healthcare solutions [5]. In
another study by author Md Mamun Ali et al., have made a comparison between six ML
algorithms on heart disease dataset from Kaggle to find the best algorithm via various
evaluation metrices [6] The algorithms used were KNN, RF, multilayer perceptron, LR,
DT and AdaboostM1. RF, KNN and DT performed well with 100% accuracy.
The suggested study by Manoj Diwakar et al., in 2020 [7] highlights the importance
of timely disease diagnosis, particularly for life-threatening conditions like heart dis-
ease. They suggest that machine learning classification methods can aid in quick and
reliable disease diagnosis, focusing on heart disease, which is challenging to detect. The
paper reviews common machine learning classification techniques and explores their
use in combination with image fusion for improved diagnostic accuracy in healthcare.
Archana Singh and Rakesh Kumar underscore the crucial need for accurate diagnosis
and prediction of heart-related diseases, given their potentially life-threatening conse-
quences [8]. Recognizing the escalating number of such deaths, they advocate for a
reliable prediction system using ML, specifically decision tree, KNN, SVM, and LR
algorithms. Employing Anaconda notebook in Python programming enhances precision
and accuracy, utilizing Cleveland dataset.
In [9] authors Harshit Jindal et al., in 2021 addresses the rising concern of increasing
heart disease cases by developing a predictive system to identify potential patients at
risk. Leveraging medical history and machine learning algorithms, including logistic
regression and KNN, their model exhibited superior performance in accurately predict-
ing heart disease compared to methods like naive bayes. The system, implemented in
a.pynb format, not only eases the workload for medical professionals but also enhances
healthcare accuracy and reduces costs, providing valuable insights for proactive heart
disease prediction. Baban.U. Rindhe et al., in 2022 offered a comprehensive exploration
of ML techniques for heart disease classification. Applying SVM, ANN, and RF on the
Cleveland dataset, SVM yielded the highest accuracy at 84.0%, followed by ANN with
83.5%, and RF classifier with 80.0% [10].
In 2023, Sreenivas Mekala et al., did a comparative study of ML algorithms— SVM,
RF, DT, NB, LR, Adaptive Boosting, and Extreme XGBoost [11]. Their evaluation, uti-
lizing confusion and correlation matrices, revealed that XGBoost demonstrated superior
accuracy compared to other algorithms. Calculating accuracy using these matrices, the
study concluded that both XGBoost and Random Forest Classifiers achieved the highest
 Heart Disease Prediction Model Using Machine Learning Techniques 293

accuracy, reaching approximately 86%. Jyoti Maurya et al., did a comparative study of
various ML algorithms, including SVM, DT, RF, LR, AdaBoost, Extra Tree Classifier,
Gaussian Naive Bayes, KNN and Gradient Boosting, on the UCI dataset. Consider-
ing 14 key factors for diagnosis, Gradient Boosting emerged with the highest accuracy
of 95.08%, outperforming other algorithms. The evaluation utilized confusion matri-
ces for each method, emphasizing Gradient Boosting’s effectiveness in heart disease
classification during training and testing [12].
Sai Bhavan Gubbala conducted a comparison of machine learning algorithms—RF
classification, AdaBoost Classifier, SVM, LR, and DT Classifier. Utilizing a dataset
from the UCI repository, research found that Random Forest achieved the highest accu-
racy among classifiers, reaching 85.22% [13]. Python was employed for model training,
dataset splitting, and accuracy assessment. In 2022, researcher Mafia Rasheed et al.,
employed machine learning, specifically SVM, to detect coronary heart disorder. Uti-
lizing a Kaggle dataset with patient history and attributes, they performed data prepro-
cessing and labeled it differently before applying the SVM trained model for improved
accuracy [14]. Results were analyzed using a Confusion Matrix, confirming a model
accuracy of 90.47% in identifying coronary heart disorder through this approach.
In 2018, Aljanabi et al., reviewed ML techniques for heart disease prediction, explor-
ing classification methods including Naive Bayes, ANN, DT, and SVM. The paper offers
a survey of relevant research and compares the accuracy of various approaches [15].
The authors advocate for the use of ANN and DT algorithms in heart disease predic-
tion, underscoring the importance of high-quality datasets and effective preprocessing
techniques to enhance prediction accuracy.
In 2018, Marimuthu et al., conducted research reviewing ML and data mining tech-
niques for heart disease prediction. The study evaluated multiple algorithms, emphasiz-
ing the necessity for more intricate models to enhance accuracy. The authors recom-
mended future research to concentrate on feature selection methods and explore diverse
rules and algorithms to further advance the field [16].
Sashank Yadav et al., aimed to enhance data modeling by incorporating seven
machine learning classifiers with an HTML template serving as a User Interface. Employ-
ing Naive Bayes, DT, KNN, LR, SVM, Gradient Boosting, and RF algorithms, the paper
provides a comparative analysis of performance measures. The classification report
showed that the KNN achieved a notably higher accuracy of 85.18%, outperforming
other machine learning algorithms in the study [17].
In [18] J. Gowri et al., proposed a method to determine the probability of heart
diseases in patients. Utilizing the UCI dataset, on LR, DT, Naïve Bayes, KNN, RF,
XGBoost, and SVM. Among these, Support Vector Machine demonstrated the highest
accuracy at 85.7%, surpassing the other six ML algorithms.
In the following Table 1, it is provided a summary of those papers whose work would
be further compared with proposed model on the basis of evaluation metrics.
294 B. K. Rai et al.

Table 1. Summary of significant work on the basis of evaluation metrics

S.No. Author & Year Proposed Work Evaluation matrices

1 Devansh Shah et al. (2020) ML classification Accuracy: K-NN -
[1] techniques NB, DT, RF, and 90.789%
K-NN
2 Pooja Rani et al. (2021) [2] Feature selection - hybrid of Accuracy: 86.60(RF)
GA & Recursive Feature Sensitivity: 84.14(RF)
Elimination (RFE). Specificity: 89.63
Classification algorithms (AdaBoost) Precision:
are NB, SVM, LR, RF, and 88.96 (AdaBoost)
AdaBoost classifier
3 Niloy Biswas et al. (2023) Feature Selection Accuracy: SF1 -
[3] approaches - chi-square, C1(93.41%) SF2 -
ANOVA, and MI Feature C1(93.41%) SF3 - C4
subsets - SF1, SF2, and SF3 (94.51%) Sensitivity: C1 -
Machine learning models 94.74 for SF1, SF2, SF3
such as LR(C1), SVM(C2), Specificity: C1- 92.45 for
K-NN(C3), RF (C4), SF1, SF2 & C4-94.23 for
NB(C5), and DT(C6) SF3
4 Rahul Katarya et al. (2020) ML Algorithms- LR, NB, Accuracy: 95.06% (RF)
[4] MLP, RF, DT, ANN, DNN, Precision: 0.55(RF)
SVM, KNN Recall: 0.97(RF)
5 Kavitha M et al. (2021) [5] The author used a hybrid Accuracy: Using Hybrid
model combining RF and Model (RF & DT) - 88.7%
DT
6 Md Mamun Ali et al. [6] ML Algorithms - MP, KNN, Accuracy: KNN, RF and
RF, DT, LR and ABM1 DT-100% Sensitivity:
100% Specificity: 100%
7 Jindal H et al. (2021) [9] Employed machine learning Accuracy: KNN - 88.52%
algorithms are LR & KNN
8 Jyoti Maurya et al. (2023) ML Algorithms- Gradient Accuracy: Gradient
[12] Boosting, SVM, KNN, RF, Boosting-95.08%
DT, AdaBoost, NB, Extra
Tree Classifier and LR
9 Mafia Rasheed et al. (2022) Classification Algorithm - Accuracy: 90.47%,
[14] SVM Sensitivity: 93.44%,
Specificity: 88.45%,
Precision: 84.62%, F1
Score: 88.81%

The above table summarizes the work of some of the researchers, Md. Mamun Ali
et al. used KNN, Random Forest, and Decision tree resulting 100% accuracy, 100%
Sensitivity, 100% Specificity. However, the used dataset had redundant instances which
lead to inaccurate result. Rahul Katarya et al. used Random Forest with accuracy of
 Heart Disease Prediction Model Using Machine Learning Techniques 295

95.06% with Precision: 0.55(RF) and Recall: 0.97(RF). However, the authors have not
handled the missing values properly using any standard technique, they replaced missing
values by python library by NAN. This could lead to inaccurate evaluation of the proposed
model.
Therefore, in order to build accurate and reliable results, preprocessing of the data
using standard preprocessing technique has been done. And then, the 5 algorithms (K-
NN, LR, DT, RF, and SVM) are trained on the most important feature. Then for the
evaluation of our proposed classification model, the algorithms are tested for accuracy
and sensitivity.

3 Proposed Work

Jupyter notebook is used to perform heart disease prediction of the dataset. The proposed
model is prepared through various stages. The workflow of the model mentioning all the
stages is depicted using Fig. 1.

Fig. 1. Architecture of Heart Disease Prediction Model

3.1 Dataset

Heart disease dataset from UCI repository is used for the construction of the model. This
dataset is widely used for CVDs prediction so, it is also used in the proposed model for
predicting heart disease. The dataset consists of 76 attributes, but almost all published
work uses a subset of 14 of them which are necessary for the experiment. The dataset
matrix information is given in Fig. 2 [19].
296 B. K. Rai et al.

Fig. 2. Heart disease dataset description

3.2 Data Preprocessing

It is the transformation of raw data into meaningful data. In machine learning we cannot
work with raw data, hence it is a very important step. The Random Forest algorithm
cannot take in null values in a dataset. To address this, it’s necessary to handle and
manage null values directly from the original raw data before applying the algorithm.
In the dataset, four records with inaccurate information on AC and two on LTH are
identified and replaced with optimal values to ensure data integrity and reliability.
The following steps are performed to preprocess our dataset:
1. Getting the CSV file of the dataset.
2. Importing Numpy, Matplotlib, Pandas libraries
3. Importing the dataset into our Jupyter IDE
4. Handling missing values and dropping the duplicate value
5. Encoding the categorical data
6. Dataset is split for training and testing where 80% is used for training and the rest is
used for testing.
7. The last step is feature scaling [20].
Figure 3 shows a heat map that visually depicts the relationships and correlations
between different features in a dataset. Each cell in the heatmap corresponds to the
 Heart Disease Prediction Model Using Machine Learning Techniques 297

Fig. 3. Correlation Matrix Heat Map

correlation between two specific features, with the color of the cell indicating the strength
of that correlation.
If the correlation value is less than zero, it suggests a negative correlation. A corre-
lation value of zero indicates no correlation between the features, while positive values
signify a positive correlation. The intensity of the color reflects the magnitude of the
correlation, with darker colors representing a strong dependency among features. In
essence, the heatmap provides a quick and visual way to understand the patterns and
relationships within a set of features.

3.3 Classification Algorithm

Classification algorithms that are used for performing prediction are- RF, KNN, SVM,
LR and DT.
Random Forest combines decision trees to enhance predictions, while K-Nearest
Neighbor classifies using training data and computes distances. SVM creates a
wide-margin hyperplane for data separation. Logistic Regression predicts categorical
variables. Decision Trees categorize based on entropy [20].

3.4 Performance Evaluation Metrics

Performance evaluation metrics like Accuracy, Sensitivity assess classification models

based on TP(True Positives) and FN(False Negative). These metrics aid in comparing
algorithmic performances. The most effective algorithm is determined by achieving the
highest outcomes across these measures. Careful analysis using these evaluation criteria
298 B. K. Rai et al.

ensures a comprehensive assessment of model effectiveness and guides the selection of

the best-performing algorithm for a given task [21].
Accuracy: The measure of a system’s correctness in predictions, assessing the ratio
of correct outcomes,
 
CorrectPrediction
Accuracy = × 100 (1)
TotalPredictions

Sensitivity: It is measure of the model’s ability to correctly identify positive instances

in predictions,
 
TP
Sensitivity = × 100 (2)
TP + FN

4 Result

In the proposed system, Firstly the dataset was filtered checking for any null values and
dropping the duplicate values from the data, then it is preprocessed. After preprocessing
the categorical features are encoded. After encoding, feature scaling is performed for
linear machine learning algorithms. The dataset is then split into 80% data for training
and 20% data for testing.

Fig. 4. Comparative accuracy result of classification techniques

 Heart Disease Prediction Model Using Machine Learning Techniques 299

Fig. 5. Comparative sensitivity result of classification techniques

The features are used in the five classifier algorithms (RF, KNN, LR, SVM, DT)
for improving the accuracy. From looking at the values of accuracy and sensitivity it is
deduced that the algorithms that perform better on the feature set is SVM.
Figures 4 and 5 shows the comparison of different algorithms on the basis of
evaluation metrics accuracy and sensitivity respectively.

Table 2. Accuracy and Sensitivity Comparison

Algorithm Accuracy Sensitivity

LR 78.69 82.76
SVM 80.33 86.21
KNN 73.77 86.21
DT 73.77 86.21
RF 78.67 89.66

In Table 2, we see that SVM has highest accuracy of 80.33% and RF has highest
sensitivity of 89.66% with the UCI dataset.
300 B. K. Rai et al.

5 Conclusion and Future Scope

Heart disease or CVDs causes a lot of death globally and the major reason is late detection
of the problem, so we need a system that would predict the presence of problem in the
early stage. Many researchers have worked on this problem and provided a solution for
the betterment of the problem. The proposed model, on the other hand, tried to find the
cons of other’s solution and tried to enhance that area in our solution. Authors have dealt
with the missing values from the dataset. Further, also dropped the duplicate values from
the data. Working on the dataset using 5 classifier algorithms – SVM, KNN, LR, DT, RF it
is concluded that SVM gives the highest accuracy of 80.33% which is maximum among
all the 5 classifier models. The future involves an increased use of machine learning
for precise heart disease analysis and early prediction, aiming to minimize death rates.
This proactive approach, driven by heightened disease awareness, holds the potential to
significantly impact public health outcomes.

Acknowledgments. No Funding received.

Disclosure of Interests. The authors have no competing interests to declare that are relevant to
the content.

References
1. Shah, D., Patel, S., Bharti, S.K.: Heart Disease Prediction using Machine Learning Techniques.
SN Comput Sci. 1, 1(6) (2020)
2. Rani, P., Kumar, R., Ahmed, N.M.O.S., Jain, A.: A decision support system for heart disease
prediction based upon machine learning. J. Reliab. Intell. Environ. 7(3), 263–275 (2021)
3. Biswas, N., et al.: Machine learning-based model to predict heart disease in early stage
employing different feature selection techniques. Biomed Res. Int. 2023 (2023)
4. Katarya, R., Meena, S.K.: Machine learning techniques for heart disease prediction: a
comparative study and analysis. Health Technol (Berl). 11(1), 87–97 (2021)
5. Kavitha, M., Gnaneswar, G., Dinesh, R., Sai, Y.R., Suraj, R.S.: Heart disease prediction
using hybrid machine learning model. In: Proceedings of the 6th International Conference on
Inventive Computation Technologies, ICICT 2021, pp. 1329–33. Institute of Electrical and
Electronics Engineers Inc. (2021)
6. Ali, M.M., Paul, B.K., Ahmed, K., Bui, F.M., Quinn, J.M.W., Moni, M.A.: Heart disease pre-
diction using supervised machine learning algorithms: Performance analysis and comparison.
Comput. Biol. Med. 1, 136 (2021)
7. Diwakar, M., et al.: Latest trends on heart disease prediction using machine learning and
image fusion. In: Materials Today: Proceedings, pp. 3213–8. Elsevier Ltd. (2020)
8. Soni, S.K.: Madan Mohan Malaviya University of Technology, North Dakota State University,
Institute of Electrical and Electronics Engineers. Uttar Pradesh Section, IEEE Industry Appli-
cations Society, Institute of Electrical and Electronics Engineers. ICE3–2020: International
Conference on Electrical and Electronics Engineering: February 14–15 (2020)
9. Jindal, H., Agrawal, S., Khera, R., Jain, R., Nagrath, P.: Heart disease prediction using machine
learning algorithms. In: IOP Conference Series: Materials Science and Engineering. IOP
Publishing Ltd. (2021)
 Heart Disease Prediction Model Using Machine Learning Techniques 301

10. Rindhe Baban, U., Ahire, N., Patil, R., Gagare, S., Darade, M.: Heart disease prediction using
machine learning. International Journal of Advanced Research in Science, Communication
and Technology [Internet], 267–76 (2021). http://ijarsct.co.in/may1.html
11. Mekala, S., et al.: Cardiopathy-heart disease prediction using machine learning [Internet]
(2023). https://www.researchgate.net/publication/371760631
12. Maurya, J., Malaviya, M.M., Prakash, S., Mohan, M.: Machine Learning based Prediction
and Diagnosis of Heart Disease using multiple models (2023). https://doi.org/10.21203/rs.3.
rs-2642516/v1
13. Gubbala, S.: Heart Disease Prediction Using Machine Learning Techniques [Internet] (2022).
www.irjet.net
14. Rasheed, M., et al.: Heart disease prediction using machine learning method. In: International
Conference on Cyber Resilience, ICCR 2022. Institute of Electrical and Electronics Engineers
Inc. (2022)
15. Aljanabi, M., Qutqut, M.H., Hijjawi, M., Al-Janabi, M.I.: SEE PROFILE Machine Learning
Classification Techniques for Heart Disease Prediction: A Review. Review Article in Int. J.
Eng. Technol. [Internet] 7(4), 5373–9 (2018). https://www.researchgate.net/publication/328
031918
16. Marimuthu, M., Abinaya, M., Madhankumar, S.K., Pavithra V.: A review on heart disease
prediction using machine learning and data analytics approach. Int. J. Comput. Appl. 181(18),
20–5 (2018)
17. Jadhav, V.J., Yadav, S., Singh, A., Jadhav, V., Jadhav, R.: Heart Disease Prediction Using
Machine Learning (2022). https://www.researchgate.net/publication/371530417
18. Gowri, J., Kamini, R., Vaishnavi, G., Thasvin, S., Vaishna, C.: Heart disease prediction u sing
machine learning. Int. J. Innov. Technol. Explor. Eng. [Internet] 11(8), 29–32 (2022). https://
www.ijitee.org/portfolio-item/h91480711822/
19. Janosi Andras, S.W.P.M.: Detrano R. Heart Disease (1988)
20. Kaur, G., Bindal, R., Kaur, V., Kaur, I., Gupta, S.: Bitcoin and cryptocurrency technologies:
a review. In: AIP Conference Proceedings. American Institute of Physics Inc. (2023)
21. Sharma, S., Rai, B.K., Gupta, M., Dinkar, M.: DDPIS: diabetes disease prediction by
improvising SVM. Int. J. Reliable Qual E-Healthc (IJRQEH) 12(2), 1–11 (2023)
 A Review to Enhance Driving Behavior Through
Advanced Drowsiness Detection Systems:
Integrating Deep Learning and Machine
Learning Techniques

Ritu1(B) , Meenu Vijarania1 , Meenakshi Malik2 , and Poonam Yadav1

1 K. R Mangalam University, Gurugram, India
[email protected]
2 BML Munjal University, Gurugram, India

Abstract. A summary of the most recent research-based driver sleepiness detec-

tion systems implemented. The article explains and showcases more recent tech-
nologies that use a variety of measures to track and recognize tiredness. Consid-
ering traffic accidents pose a serious threat to the public’s safety, driving safety
is a crucial concern in today’s society. Taking appropriate action to mitigate the
dangers connected with driving is imperative due to the growing number of cars
and the difficulties of modern urban areas. Drunk driving is one of these threats;
it may be a subtle but deadly enemy that impairs a driver’s awareness and judge-
ment. Research suggests an all-encompassing strategy to improve road safety by
comprehending and treating the elements causing fatigued driving. By reducing
the financial costs linked to traffic accidents, the results of such programs not
only protect lives but also enhance society as a whole. This study combines tra-
ditional machine learning techniques with deep learning architectures to identify
tiredness in driving behavior. This review paper focuses on study of accuracy
of various machine learning techniques in field of driver behavior drowsiness
detection system.

Keywords: machine learning · artificial intelligence · drowsiness · behavior

1 Introduction
Driving safety is a paramount concern in today’s modern society, as road traffic accidents
continue to pose a significant threat to public well-being. The ever-increasing volume of
vehicles on the road, combined with the potential distractions and challenges presented
by contemporary urban environments, has intensified the need for effective measures to
mitigate the risks associated with driving. Among the various factors that contribute to
road accidents, drowsiness stands out as a silent yet perilous adversary.
Drowsiness, often underestimated in its potential to induce accidents, plays a pivotal
role in compromising a driver’s alertness and decision-making abilities. According to
estimates from the National Highway Traffic Safety Administration (NHTSA), sleepy

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 302–311, 2025.
https://doi.org/10.1007/978-3-031-71481-8_24
 A Review to Enhance Driving Behavior Through Advanced Drowsiness 303

driving is responsible for a substantial percentage of accidents each year, with conse-
quences ranging from minor collisions to severe, life-altering crashes. Drowsy driving
was a factor in 91,000 police-reported accidents in 2017, according to NHTSA estimates.
These crashes resulted in around 800 fatalities and over 50,000 injuries [1]. These acci-
dents not only lead to loss of life but also incur significant economic costs, straining
healthcare systems and societal resources.
Given the numerous sleeping issues [2] Increasing drowsy driving awareness among
adults in the United States can be very beneficial for public health. Driving safely on
the road can be accomplished by being aware of the causes, effects, and avoidance of
sleepy driving. The National Sleep Foundation reports that 54% of drivers report feeling
drowsy while driving, and 28% of adult drivers have acknowledged to dozing off while
operating a vehicle. Additionally, more than 40% admit to dozing off while driving at
least once [3].

Fig. 1. Overview of Drowsy Driving

304 Ritu et al.

Driving accidents and tiredness have a complicated link that is influenced by a variety
of circumstances including lack of sleep, irregular circadian rhythms, and long driving
hours. Figure 1 Drowsy drivers are more likely to have slower reaction times, poorer
coordination, and decreased cognitive abilities, which increases the likelihood that they
may make mistakes while operating a vehicle.
The effects affect more than just individual drivers; they also have an impact on
passengers, pedestrians, and other users of the road, highlighting the urgent need for
effective measures. To identify driving behavior while fatigued, a variety of machine
learning and deep learning algorithms are employed. The state of the driver is affected
by a number of psychological, physiological, and mental aspects [4].
The following six criteria were mentioned by Nishiyama as necessary for in-vehicle
technologies to detect tiredness [5]:
1. Driver fatigue is detectable in real time and can vary from mild to deep slumber.
2. Every driver is regarded as a target that can be detected.
3. Compared to other applications, they are very scalable and inexpensive.
4. Take care not to impede the driver’s ability to drive safely.
5. Able to endure challenging operational conditions and be placed in a vehicle and
identify fatigue in drivers (Fig. 2).

Fig. 2. Depicts the systematics diagram for the driver behavior analysis

2 Literature Review
Rajkar, A. et al. [6] proposed a deep learning-based method to identify drivers’ sleepiness.
They employed the Yawing Detection Dataset (YawDD), Face and Eye regions for fatigue
detection, Convolutional neural network (CNN), a type of deep learning with an average
accuracy of 96%, and the Closed Eye in the Wild dataset (CEW).
B. K. Savaş et al. [7] The Convolutional Neural Network (ConNN*) model for
multitasking is suggested as a means of identifying driver weariness or drowsiness. The
driver’s behavior model makes use of the eye and mouth characteristics. Driver weariness
is tracked by modifications to these features. Driver fatigue is measured by taking the
length of an eye closure/percentage of eye closure (PERCLOS) and the frequency of
yawning/frequency of mouth (FOM). Three classifications based on the driver’s level
of weariness are recognized in this investigation. The YawdDD and NthuDDD datasets
 A Review to Enhance Driving Behavior Through Advanced Drowsiness 305

showed that the suggested model detected fatigue with 98.81% accuracy. A comparative
analysis of the model’s success is provided.
S. Bavkar et al. [8] an innovative method for using an electroencephalogram (EEG)
sensor to quickly identify alcoholism. To classify alcoholics and normal subjects, the
suggested approach uses absolute gamma band power as a feature and ensemble subspace
K-NN as a classifier. Additionally, the paper reports the use of the optimum EEG chan-
nels for speedy alcoholism screening using the Improved Binary Gravitational Search
Algorithm (IBGSA) as an optimization tool. The suggested approach’s outcomes are
contrasted with those of optimization algorithms as the genetic algorithm (GA), binary
particle swarm optimization (BPSO), and binary gravitational search algorithm (BGSA).
The accuracy acquired via the ensemble subspace K-Nearest Neighbor (K-NN) classifier
is used to evaluate the fitness function for various optimization strategies. With just 13
EEG channels, the suggested improved binary gravitational search algorithm (IBGSA)
approach offers a 92.50% detection accuracy.
Bavkar, S et al. [9] Distinguish between alcohol and non-alcohol users based on
visually evoked potential (VEP), suggest features related to complexity and nonlinearity
use a K NN classifier with ensemble subspace. Reduced classification mistakes are the
goal of the ideal number of channels is chosen using binary particle swarm optimization
(BPSO). To be used in optimization techniques, a novel fitness function is created. Select
channels and classification error are used to evaluate the fitness function. The results of
the experiment indicate that the best channel chosen has biological importance and is
linked to alcoholism. Consequently, the results of the suggested channel selection process
can be applied to quickly and accurately classify subjects as normal or alcoholic.
Hossein Chamani et al. [10] The eyes, nose, and mouth are divided into a different
image, and face classification operations (deep learning by convolution neural network)
are carried out on the separated components rather than on generic information for face
recognition. According to the data, the suggested strategy reduces computing costs by
almost 70%. Additionally, it is possible for CNN to perform worse than the overall image
of the disassembled parts.
A. Amodio et al. [11] An assessment is conducted on the efficacy of this approach in
identifying an impaired condition resulting from alcohol addiction. The test procedure
records the dynamics of constriction in both eyes by shining light in one of the individ-
uals’ eyes. A two-step methodology is described in order to obtain the video sequences’
pupil size profiles: the first phase involves performing a search of the visual using the
iris and pupil, and the second step involves cropping the visualization to enhance time
efficiency by performing pupil detection on a smaller image. The undesirable learner
dynamics that arise inside the pupillary light reflex (PLR) are identified and assessed;
the reflexive accommodation results in constriction of the pupil around 10% of the iris
diameter, and a spontaneous oscillation of the pupil diameter is seen in the interval [0, 2]
Hz. A model of first order is found for each individual based on a repository for pupillary
light reactions collected on many people both under baseline conditions and after alco-
hol use. In order to create a classifier for support vector machines that can distinguish
in between the states of “Sober” and “Drunk,” a set of characteristics is provided and
utilized to compare the two populations of responses.
306 Ritu et al.

G. Sikander et al. [12] provides a cutting-edge summary of recent advancements in

the of detecting driver weariness. Biological characteristics of the driver and subjective
reporting, physical characteristics of the driver, hybrid characteristics and features of
the car when driving are the five areas into which methods are divided based on the
features utilized for driver tiredness detection. Comparing several methods for tiredness
detection, potential areas for improvement are identified.
Guo, JM. et al. [13] suggested a novel approach that uses a mix of Convolutional
neural networks (CNNs) with long short-term memory (LSTM)to handle the real-time
driver drowsiness detection. The Asian conference on computer vision (ACCV) 2016
competition’s public sleepy driver dataset was used to evaluate the system’s performance.
It can perform better than previous systems in the literature, according to the results.
Liu, W. et al. [14] provide a dual-stream network model with driver for multiple
face features tiredness identification technique. There are four components to the algo-
rithm: (1) Using convolutional neural networks with multitask cascades (MTCNNs) for
mouth and eye positioning. (2) Taking a partial face picture and extracting its static
features. (3) Taking a partial face optical flow and extracting the dynamic characteristics
from it. (4) Combining static and dynamic data using a two-stream neural network to
categorize the information. This paper’s primary contribution is the integration of mul-
tiple facial characteristics employing a two-stream network to the diagnosis of driver
weariness. Even so, a partial face pictures as inputs into a network might focus on facts
about exhaustion information, networks with two streams incorporate information about
images, both static and dynamic, resulting in improved performance. In addition, gamma
correction was used to improve image contrast. This can improve our method’s perfor-
mance, as seen by a 2% improvement in accuracy in nighttime settings. Ultimately, the
with 97.06% accuracy, the National Tsing Hua University Driver Drowsiness Detection
(NTHU-DDD) dataset was used.
E. Romera et al. [15] As worries about automobile safety have grown, driving anal-
ysis has attracted attention recently. On the other hand, the current state of this subject
is currently limited by the absence of publicly available driving data. Research may
benefit greatly from machine learning approaches, but their applicability to the broader
research community is restricted since they require vast volumes of data that are hard
and expensive to collect through Naturalistic Driving Studies (NDSs). Furthermore, An
inexpensive and easy to put into practice platform for the driver behavior monitoring
has been made available by the widespread use of smartphones; yet, the data from these
devices is not publicly accessible through existing applications. Because of these factors,
this study offers the publicly accessible dataset of university of alabama in Huntsville
-DriveSet (UAH-DriveSet) that furnishes a substantial quantity of information obtained
through Drive Safe, our driving monitoring software, enabling intense driving analysis.
Six distinct drivers and vehicles operate the application, displaying three different behav-
iors: aggressive, sleepy, and normal. On both minor and highway kinds of roads This
results in 500 min or more of realistic driving, along with the connected unprocessed
data and analyzed semantic information, as well as records of the travels on video. To
make data analytics easier, this study also presents a program that plots data and shows
trip recordings at the same time. Visit http://www.robesafe.com/personal/eduardo.rom
era/uah-driveset to access the UAH-DriveSet.
 A Review to Enhance Driving Behavior Through Advanced Drowsiness 307

Peppes, N. et al. [16] describes a comprehensive, integrated platform that gathers,

stores, processes, analyzes, and correlates various data flows coming from automo-
biles. It does this by combining popular methods for deep learning and machines Using
open-source technologies. In particular, to categorize the conduct of the driver if envi-
ronmentally friendly, data streams from various vehicles are processed and analyzed
using clustering techniques. This is then a comparison of deep learning and supervised
machine learning techniques using the given tagged dataset.
J. Wang et al. [17] study aims to benefit both communities by providing a thorough
overview of research, along with details on data gathering methods, computer vision
algorithms used, and driver behavior classification results. The research is restricted to
studies that use a minimum of one camera to monitor the driver within a car. Papers have
been categorized according to their aim, which is to detect either high-level (distraction
detection) or low-level (head orientation) driver information. The datasets that the papers
use have led to additional classifications. Many private datasets have also been found
in addition to the twelve available datasets, and their data collecting architecture is
examined to highlight any effects on model performance. The algorithms used for each
job and their results are examined to provide a baseline.

3 Artificial Intelligence (AI) and Machine Learning Techniques

for Driving Behavior Analysis Capturing Drowsiness:

Now a days Using machine learning, computers can mimic and adapt human-like behav-
ior. Every interaction and action taken using machine learning is something the system
can learn from and apply to the next time [18]. Artificial intelligence and machine
learning are not brand-new concepts. For almost sixty years, they have been researched,
utilized, applied, and reinvented by computer scientists, engineers, researchers, students,
and business community members. In the 1950s and 1960s, machine learning and artifi-
cial intelligence observed serious growth [18]. The most widely used machine learning
algorithms include neural networks, ensemble techniques, naïve Bayes, decision trees,
support vector machines (SVMs), discriminant analysis, and linear regression, random
forest [18].
Artificial intelligence (AI) and machine learning are two fields that include the branch
of deep learning, which has received a lot of attention recently and achieved considerable
strides. The fundamental component of deep learning is artificial neural networks, which
are modelled after the structure and operation of the human brain. Artificial neurons,
often referred to as nodes or units, are layered and coupled to form these networks.
Information is processed by each neuron before being sent to the following layer. Since
neural networks can contain numerous hidden layers, they can be considered “deep,”
and it is this depth that distinguishes deep learning from more conventional machine
learning methods. Figure 3 depicts a deep learning neural network with input, hidden,
and output layers.
Figure 4 firmly illustrates that classical machine learning techniques perform better
with less quantities of input data. While the performance of classic machine learn-
ing algorithms remains stable beyond a certain amount of data, that of deep learning
approaches grows in proportion to the data increment [19].
308 Ritu et al.

Fig. 3. Deep learning neural network

Fig. 4. The deep learning system’s performance in relation to data volume [19]

To detect Drowsiness, various ML and DL approaches are described in different

works. Convolutional neural networks (CNN) approach for detecting aggressive driving
emotion [20]. Machine learning techniques and data from smartphone sensors are used
to track unusual driving behavior [21]. Method for detecting driver drowsiness based
on physiological signals, such as electrocardiogram (ECG) and electroencephalogram
(EEG), and machine learning algorithms [22].

4 Proposed Methodology
We suggest a methodology that includes online University of California, Irvine (UCI)
repository data and feature engineering using deep learning architectures as well as con-
ventional machine learning methods. The specific goal is to capture the temporal corre-
lations in driving behavior suggestive of tiredness by combining the interpretability of
machine learning classifiers with the complex pattern recognition powers of recurrent
 A Review to Enhance Driving Behavior Through Advanced Drowsiness 309

neural networks (RNNs). The methodology in Fig. 5 draws inspiration from devel-
opments in multi-sensor fusion, combines information from several sources, includ-
ing physiological monitors, steering dynamics sensors, and cameras for facial and eye
movement recognition. This all-encompassing method seeks to offer a full picture of
the driver’s condition by identifying subtle behavioral and physiological indicators to
enhance the precision and dependability of drowsiness detection.

Fig. 5. Proposed methodology on the based-on literature review

5 Conclusion

This study leverages the complimentary capabilities of deep learning architectures and
classical machine learning techniques to propose a novel approach to driving behav-
ior drowsiness detection. Our study, which is based on recent developments in multi-
sensor fusion, combines data from several sources, such as cameras that recognize face
and eye movements, steering dynamics sensors, and physiological monitors. This all-
encompassing integration aims to provide a full awareness of the driver’s state, skilled
in recognizing minute behavioral and physiological cues linked to fatigue. The exist-
ing approaches by combining information from several sensors and providing a more
accurate and nuanced picture of the driver’s condition. The integration of physiological
data, steering dynamics, and facial/eye movement recognition enhances the accuracy
and reliability of our suggested sleepiness detection system by enabling the early iden-
tification of sleepy indicators. When combined with multi-sensor fusion, deep learning
310 Ritu et al.

and machine learning algorithms offer a comprehensive and efficient way to improve the
accuracy and dependability of sleepiness detection systems. Our review study seeks to
make significant progress in reducing the dangers associated with driving when sleepy
and promoting safer road conditions for everybody by improving driving safety.

References
1. National Highway Traffic Safety Administration: Drowsy Driving. https://www.nhtsa.gov/
risky-driving/drowsy-driving. Accessed 10 Oct 2023
2. National Center for Chronic Disease Prevention and Health Promotion, Division of Population
Health. CDC – Data and Statistics – Sleep and Sleep Disorders, 2 May 2017. https://www.
cdc.gov/sleep/data_statistics.html. Accessed 14 Oct 2023
3. Czeisler, C.A., Wickwire, E.M., et al.: Sleep-deprived motor vehicle operators are unfit to
drive: a multidisciplinary expert consensus statement on drowsy driving. Sleep Health 2(2),
94–99 (2016)
4. Jacobé de Naurois, C.J., Bourdin, C., Stratulat, A., Diaz, E., Vercher, J.L.: Detection and
prediction of driver drowsiness using artificial neural network models. Accid. Anal. Prevent.
126, 95–104 (2019)
5. Nishiyama, J.: Research on the Detection of Signs of Sleepiness in Humans by Pupil
Fluctuation and Eye Movement. Ph.D. Thesis, Chubu University, Aichi, Japan (2011)
6. Rajkar, A., Kulkarni, N., et al.: Driver drowsiness detection using deep learning. In: Iyer,
B., Ghosh, D., Balas, V.E. (eds.) Applied Information Processing Systems. AISC, vol. 1354,
pp. 73–82. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2008-9_7
7. Savaş, B.K., Becerikli, Y.: Real time driver fatigue detection system based on multi-task
ConNN. IEEE Access 8, 12491–12498 (2020)
8. Bavkar, S., Iyer, B., Deosarkar, S.: Rapid screening of alcoholism: an EEG based optimal
channel selection approach. IEEE Access 7, 99670–99682 (2019)
9. Bavkar, S., Iyer, B., Deosarkar, S.: BPSO based method for screening of alcoholism. In:
Kumar, A., Mozar, S. (eds.) ICCCE 2019. LNEE, vol. 570, pp. 47–53. Springer, Singapore
(2020). https://doi.org/10.1007/978-981-13-8715-9_6
10. Chamani, H., Nadjafi, M.: Improved reliable deep face recognition method using separated
components. Brill. Eng. 3, 4563 (2022)
11. Amodio, A., Ermidoro, M., Maggi, D., Formentin, S., Savaresi, S.M.: Automatic detection
of driver impairment based on pupillary light reflex. IEEE Trans. Intell. Transp. Syst. 20(8),
3038–3048 (2019)
12. Sikander, G., Anwar, S.: Driver fatigue detection systems: a review. IEEE Trans. Intell. Transp.
Syst. 20(6), 2339–2352 (2019)
13. Guo, J.M., Markoni, H.: Driver drowsiness detection using hybrid convolutional neural
network and long short-term memory. Multimed. Tools Appl. 78(20), 29059–29087 (2019)
14. Liu, W., Qian, J., Yao, Z., Jiao, X., Pan, J.: Convolutional two-stream network using multi-
facial feature fusion for driver fatigue detection. Future Internet 11(5), 115 (2019)
15. Romera, E., Bergasa, L.M., Arroyo, R.: Need data for driver behavior analysis? Presenting the
public UAH-DriveSet. In: IEEE 19th International Conference on Intelligent Transportation
Systems (ITSC) 2016, pp. 387–392. IEEE, Rio de Janeiro-Brazil (2016)
16. Peppes, N., Alexakis, T., Adamopoulou, E., Demestichas, K.: Driving behavior analysis using
machine and deep learning methods for continuous streams of vehicular data. Sensors 21(14),
4704 (2021)
17. Wang, J., et al.: A survey on driver behavior analysis from in-vehicle cameras. IEEE Trans.
Intell. Transp. Syst. 23(8), 10186–10209 (2022)
 A Review to Enhance Driving Behavior Through Advanced Drowsiness 311

18. Alzubi, J., et al.: Machine learning from theory to algorithms: an overview. J. Phys.: Conf.
Ser. 1142(1), 012012 (2018)
19. Alom, M.Z., et al.: A state-of-the-art survey on deep learning theory and architectures.
Electronics 8(3), 292 (2019)
20. Lee, K., et al.: Convolutional neural network-based classification of driver’s emotion during
aggressive and smooth driving using multi-modal camera sensors. Sensors 18(4), 957 (2018)
21. Lindow, F., Kashevnik, A.: Driver behavior monitoring based on smartphone sensor data and
machine learning methods. In: 25th Conference of Open Innovations Association (FRUCT)
2019, pp. 196–203. IEEE, Helsinki-Finland (2019)
22. Ojha, D., Pawar, A., et al.: Driver drowsiness detection using deep learning. In: 4th Interna-
tional Conference for Emerging Technology (INCET) 2023, pp. 1–4. IEEE, Belgaum-India
(2023)
 Advancing Peer Review Integrity: Automated
Reviewer Assignment Techniques with a Focus
on Deep Learning Applications

Bhumika Bhaisare(B) and Rajesh Bharati

Department of Computer Engineering, Dr. D. Y. Patil Institute of Technology, Pimpri, Pune,

India
[email protected], [email protected]

Abstract. The reviewer assignment process in academic research papers is cru-

cial for ensuring peer review integrity and fairness, and it is often referred to as the
reviewer assignment problem (RAP). This initial step lays the foundation for the
entire review process. To address the limitations of manual reviewer assignment,
there is a growing trend among journal and conference editors to explore auto-
mated solutions. For quite some time, machine learning techniques have found
application in this domain, recent attention has shifted toward the application
of Deep Learning in the education sector. Deep Learning, rooted in neural net-
works with diverse processing components, is gaining prominence in fields like
natural language processing (NLP) and image recognition. In this comprehensive
review, we aim to evaluate the key research achievements in the realm of reviewer
assignment algorithms. It commences by presenting background information and
discussing the necessity for automated reviewer assignment. The paper then sys-
tematically examines the existing research in this domain, with a particular focus
on recent advancements in Deep Learning techniques. The review provides an
unbiased assessment of the current algorithms’ strengths and weaknesses. Over-
all, this analysis emphasizes the growing significance of RAP research and the
demand for further innovation and framework development in this field.

Keywords: Deep Learning · Information retrieval · Reviewer Peer review ·

assignment problem (RAP) · Optimization algorithm

1 Introduction
This Peer review is a collaborative process involving experts from one or more fields
who assess scientific research proposals and accomplishments. Peer review is deeply
ingrained in modern scientific inquiry, serving as an essential pillar in the domain of
academic research. It spans across various disciplines, encompassing the submission
of work for research conferences & peer-review journals. At present, peer evaluation
garners extensive backing from researchers, serving as the crucial entry point and qual-
ity assurance system for improving the excellence of research publishing and affirm-
ing trustworthiness publications research. Nonetheless, peer assessment faces various

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 312–327, 2025.
https://doi.org/10.1007/978-3-031-71481-8_25
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 313

obstacles, and one of these hurdles is the automated matching of reviewers with appro-
priate papers, a crucial phase in enhancing the quality of peer review. Allocating expe-
rienced reviewers for evaluation of research documents is important task, recognized as
RAP (Reviewer-Assignment Problem), presents the vital and intricate challenge during
peer-review procedures of publishing journals, research conferences as well as funding
institutions [1].
To enhance the quality of assessments and guarantee an equitable and unprejudiced
appraisal, the subsequent crucial elements should be considered during the reviewer
assignment process:
• Assigning reviewers to proposals that align with their areas of proficiency is essential.
• It is imperative that proposals are assessed by the most competent reviewers in their
specific fields.
• The evaluation of proposals with eagerness and unwavering commitment carrying
out by the reviewer.
• No reviewer should be burdened with more proposals than the agreed-upon limit.
• Every assessor ought to appraise an identical quantity of submissions.
• Every submission must endure examination by a fixed quantity of appraisers.
• Any affiliations between candidates and assessors that might imperil the equity of the
appraisal procedure should be averted.
Historically, the allocation of reviewers was handled by a solitary decision-maker,
like a manager in a program of funding, or a compact panel, and all these duties were
conducted manually.
However, hand-operated allocation processes proved to be time-consuming also vul-
nerable for subjective or biased judgment of the committee. Nonetheless, the hand-
operated allocation process turned out to be exceedingly time-intensive and vulnerable
to the subjective or prejudiced choices established by the panel. Additionally, stream-
lining designations presented a challenging endeavor owing to the incapacity to fully
contemplate every requirement. On top of that, the distribution of reviewers had to be
concluded under rigorous time limitations, as a multitude of submissions frequently
inundated just prior to the stipulated cutoff [2].
The existing peer review process involves a substantial volume of reviewers and
proposals, often numbering in the thousands. The substantial magnitude of this situa-
tion greatly obstructs the hand-operated allocation procedure, generating difficulties for
conference organizers, periodical overseers, and funding supervisors. Consequently, the
manual-operated approach proved to be labor-intensive and didn’t consistently yield the
most effective results. Hence, the automatic reviewer assignment approach has emerged
as a critical topic within the academic community (Fig. 1).
To emphasize the significance of the reviewer assignment problem, let’s contem-
plate the evaluation procedure of an academic symposium document as a demonstrative
instance. Figure 2 depicts the entire process of evaluation for academic papers splits into
these subsequent phases:
• Manuscripts are submitted by authors ahead of the designated time limit.
• Head of the program faces the challenge of forming a program committee, essentially
a task of identifying and recruiting experts capable of reviewing the submitted papers.
314 B. Bhaisare and R. Bharati

Fig. 1. Limitations of the prior definition of RAP

• The Program Chair then allocates the manuscripts to the experts responsible for
reviews within the committee.
• The evaluation specialists appraise the papers and furnish their input and remarks.
• Considering the feedback obtained, the Program Coordinator ultimately determines
to approve or decline that paper.
As illustrated in Fig. 2, the allocation of reviewers to papers holds a crucial function
in establishing the reviewers’ proficiency for paper evaluations. As a result, it emerges as
a vital phase within the complete peer review procedure [3]. This underscores a consider-
able necessity for effective algorithms dedicated to assigning reviewers, automating and
refining the pairing of reviewers with papers. These algorithms not only considerably
boost the effectiveness of the entire peer review procedure but also markedly advance the
excellence of peer assessments. Additionally, the incorporation of computer methods,
such as natural language processing, can further amplify this process.

Fig. 2. A synopsis of the academic conference paper review process.

 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 315

In the past decade, Deep Learning (DL) has revolutionized machine learning,
excelling in tasks like speech & image identification or recognition. Global firms like
Apple, Google, Facebook have huge investment in DL for smart product development.
DL uses multi-layer neural networks for diverse data types. It benefits various research
fields.

2 Literature Review

The automated process of matching reviewers with papers is an ongoing interdisci-

plinary puzzle that lures researchers from various areas, encompassing OR (operations
research) AI (artificial intelligence) as well as computer science. NLPs (Natural language
processing) have taken on a central role as the driving catalyst for automated reviewer
assignment.
In 2007, Romero and Ventura [4], subsequently refined in 2010 [5] and 2013 [5],
conducted a comprehensive analysis of over 300 studies prior to 2010. They identified
various categories and tasks in data mining within research and education, encompass-
ing data analysis and visualization, feedback provision for instructors, research recom-
mendations, author performance prediction, anomaly detection, subject grouping, social
network analysis, concept map development, review list construction, and designing &
arranging.
Baker and Yacef (2009) [6] delivered an analysis of data mining within the edu-
cational and research setting. They scrutinized the transformations and developments
in this domain, drawing parallels between its present condition and its initial stages.
They pinpointed four major applications/objectives: refining models for research, enrich-
ing domains, assessing the educational assist offered by educational applications, also
undertaking research exploration into learners. Furthermore, they cataloged frequently
referenced articles from 1995 to 2005 and explored the impact they have over the data
mining community.
In 2014, Peña-Ayala conducted an extensive survey using data mining techniques on
over 240 papers in Educational and Research Data Mining [7]. This involved statistical
and clustering processes to identify educational functionalities, patterns in data mining
approaches, and instances of value patterns related to explanatory and anticipatory mod-
els. In contrast to prior literature surveys, this study primarily centered on computational
methods rather than the utilization of Educational and Research Data Mining.
More recent additions to this literature review include a study by Bakhshinategh et al.
in 2018 [8]. They scrutinized a range of responsibilities and uses within Educational and
Research Data Mining, classifying them according to their objectives. Building upon the
eleven categories proposed by the researchers [4], They introduced a structured hierarchy
consisting of thirteen categories.
Aldowah and colleagues (2019) [9] undertook a research investigation in the higher
education domain. They delved into higher education through the areas of computer-
aided predictive, visualization, learning, as well as behavioral analytics. Drawing
from previous studies, they identified specific Educational and Research Data Min-
ing techniques that could effectively address various learning challenges, supplying
student-focused approaches and resources to academic organizations.
316 B. Bhaisare and R. Bharati

Notably, Deep Learning application & datasets availability within Research Data
Mining for peer-review have not been extensively explored in these review papers, and
this work aims to address these gaps. To facilitate empirical comparisons of different
approaches, understanding the datasets used in experiments is crucial. This paper is dedi-
cated to reviewing and summarizing these resources. Additionally, while prior proposals
have considered shallow neural network approaches, none have specifically focused on
Deep Learning techniques.
In addition, there are review articles that encompass the entire peer review process and
incorporate the Reviewer Assignment Problem (RAP) as a constituent element within it.
However, while these studies include sections discussing how reviewers are selected for
proposals, they often fall short in providing a comprehensive an examination viewpoint
over Reviewer Assignment Problem (RAP) as a holistic concept (Fig. 3).

Fig. 3. The yearly count of published papers.

In a more recent context, Shah (2022) composed an all-encompassing review paper

that investigated the obstacles within the peer review procedure and examined prospec-
tive remedies. This research provides an extensive overview of the peer review body
of work, tackling concerns related to disparities that emerge due to factors such as dis-
honesty, misjudgment, and subjectivity during the assessment phase. While the RAP
is touched upon in the context of addressing injustices resulting from discrepancies in
reviewer expertise, it is explored in a somewhat limited manner, with only one dimension
being considered.

3 Methodology Review
In this section, we outline the methodology employed for conducting this review and the
procedures involved in collecting, analyzing, and extracting the available studies related
to the utilization of deep learning in data mining for peer review techniques in research
and education.
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 317

3.1 Deep Learning Models

Word-based models and topic models employed in the context of the Reviewer Assign-
ment Problem (RAP) primarily rely on methods based on occurrence rates (frequency).
These methods consider documents as aggregations of individual terms, with the crucial
factor here is the term frequency, with no consideration for term sequence. Nonethe-
less, a significant limitation of these approaches is their incapacity to grasp the semantic
and syntactic connections among words, consequently producing semantically feeble
representations.
Moreover, when applied to large datasets, frequency-based techniques often produce
vectors with significant sparsity. This implies that the vector representation of each doc-
ument has a dimensionality same as the document’s vocabulary, a condition frequently
known as the “sparsity curse.”
In response to these limitations, deep learning has given rise to language models
designed to surmount the challenges associated with frequency-dependent techniques.
These neural network-based language models utilize word embedding to automatically
extract syntactic and semantic document characteristics.
The utilization of neural network-based techniques for word embedding can be clas-
sified into two primary divisions. The initial category, often termed as unchanging lexical
representation (word embedding) methods or prediction-driven Strategies for lexical rep-
resentation, can be encapsulated both the morphology and semantics of words. These
methods also consider sub-word elements such as character n-grams. Distinguished
instances of static-based word embedding methods encompass Word2Vec, Glove, and
FastText.
The subsequent category is composed of context-based word embedding approaches,
which encompass models with profound architectures like ULMFiT, BERT, GPT-2,
and DistillBERT. In opposition to static word embedding techniques that characterize
words as compact vectors, context-based methods meld contextual information into
the representation. These models, fortified with their deep architectures, possess the
capability to comprehend a word’s meaning.
From the modern times, word embedding (lexical representation) methods that
employ neural networks have discovered applications in a variety of natural language
processing (NLP) assignments, speech recognition, encompassing word forecasting,
translation of language, as well as textual similarity computation. Within the context
of reviewer allocation, the likeness between propositions and reviewers can be gauged
via the examination of the compact vector representations of their documents. Ogunl-
eye et al. (2018) approach score of similarity amid reviewers as well as propositions
employing integrated frequency-based models as well as Word2Vec for assessment of
similarity.
Zhao et al. (2018) introduced a fresh categorization approach termed the Constructive
Covering Algorithm (CCA) of word crawlers, to tackle the Reviewer Assignment Prob-
lem (RAP) for categorization task. Procedure harnesses research field tags linked with
reviewers and propositions to enrich the quality of allocations. At the outset, proposi-
tions and reviewers were portrayed employing Word2Vec, followed by the computation
of the minimum textual distance between propositions and reviewers employing the
word-crawlers distance algorithm. Successively, apply tags for reviewer contenders as
318 B. Bhaisare and R. Bharati

well as propositions have been divined grounded on field connections. Although this
approach is adept at acquiring field characteristics via data-driven analysis, which is not
empowering machines for assimilating field intelligence throughout a human outlook.
To deal with this, Duan et al. (2019) launched a sentence pairing modeling strategy that
employed the field association amid proposition appellations and abstracts and reviewer
writings as supervisory data.
Yong et al. (2021) leveraged Word2Vec to extract Layered meaning portrayals of
terms. They devised a governance mechanism utilizing a knowledge structure and gov-
ernance correlation to generate a structured list of assessors based on their pertinence to
each proposal.
Anjum et al. (2019) asserted previous methodologies, encompassing SLM, VSM,
as well as possible models over topics, have been inadequate in handling lexicon mis-
matches and fractional topic convergences amid propositions and a reviewer’s profi-
ciency. To grapple with such instances of word mismatches and fractional topic cover-
ages amid reviewers as well as propositions, the shared topic model approach, based on
Word2Vec, was initiated by the authors.
Table 1 presents the allocation of strategies fixated on semantic data regarding review-
ers and propositions. Remarkably, techniques for modeling over topics like LDA, PLSA
as well as LSI are widely put to use contextual strategy, especially in the next phase
in RAP. Among these, the LDA model stands out for its efficacy in detecting topics
representing reviewers and propositions.
Romanov and Khusainova, 2019; Kalyan et al., 2021; Kapočiūtė-Dzikienė et al.,
(2021) proposed Neural network-based language models have demonstrated unparal-
leled performance in a scope of natural language processing (NLP) undertakings, for
instance, document categorization, recognition of named entities, and computerized lan-
guage translation. Until 2018, using these advanced language models to calculate match-
ing scores was not common; however, a few studies have now started to incorporate them.
The application of artificial intelligence-derived systems in the evaluation of similarity
ratings has demonstrated their superiority in comparison to alternative meaning-oriented
methods.
Even though neural network-based language models fall under the category of
semantic-centered approaches, their extraordinary advantages have validated a distinct
section to acknowledge the substantial benefits they have contributed to the calculation
of resemblance scores, with the potential for additional advancements in the future.

Table 1. Distribution of Studies with A Focus On Semantic-Based Methods Across Different

Years

Methods 2013 2014 2015 2016 2017 2018 2019 2021 2022
√
PLSA
√ √ √ √√ √ √ √√√ √√√
LDA
√ √ √
LSI
(continued)
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 319

Table 1. (continued)

Methods 2013 2014 2015 2016 2017 2018 2019 2021 2022
√√ √
ATM
√√ √√ √ √√
Word2Vec
√ √
BERT
√ √√
CNN
√ √
SciBERT

4 Deep Learning
Deep Learning, or DL, has emerged as a significant field of study in the realm of arti-
ficial intelligence, garnering considerable attention in recent years. It falls within the
broader machine learning domain and leverages structures of neural network for cre-
ating abstract data representations. These neural frameworks comprise numerous tiers
containing processing units or neurons, which execute a blend of linear and nonlinear
alterations on input information. Deep learning has spawned several designs that have
displayed effectiveness in a broad array of guided and self-guided assignments, partic-
ularly within the extensive realms of natural language processing and visual perception
in computers [10].
Deep learning algorithms demonstrate proficiency in acquiring numerous strata of
data representations, where more advanced attributes are constructed from those of
lower strata, establishing a hierarchical arrangement. For instance, concerning image
categorization, deep learning models can accept pixel values as input and, through the
output layer, generate tags for items in images. Amidst these strata, a progression of
transformation (concealed) strata systematically develops more complex characteristics,
making them less vulnerable to variables such as illumination and object positioning.
The inclusion of the “deep” term into deep learning suggests for existence of multiple
strata of alterations and tiers of representation positioned between the network’s inputs
and results. While a universally defined threshold for categorizing a neural network as
“deep” does not exist, the majority of research in the domain proposes that having over
two intermediary transformation strata is a prevalent criterion [11]. Although many fun-
damental concepts of DL originated several decades ago, the most significant advance-
ments have occurred over the past decade. This surge in DL progress can be attributed
to two primary factors: the availability of massive datasets and the increased computa-
tional power driven by Graphics Processing Units (GPUs). Large datasets enable DL
algorithms to generalize effectively, while GPUs facilitate parallel processing for training
more extensive and intricate models. Moreover, the proliferation of software platforms
such as TensorFlow, Theano, Keras, and PyTorch has empowered researchers to con-
centrate on crafting models without becoming entangled in granular implementation
intricacies.
A fundamental driver of the success of deep learning lies in its capability to circum-
vent the necessity for exhaustive feature manipulation. In traditional machine learning,
320 B. Bhaisare and R. Bharati

the process of feature manipulation entails the laborious task of choosing pertinent
attributes vital for algorithmic operation while discarding uninformative ones. This pro-
cess can be time-consuming, as the correct feature selection plays a pivotal role in sys-
tem performance. DL, on the other hand, incorporates feature learning, autonomously
discovering the essential representations for a given task [12].
The following segments will delve into the fundamental principles of neural net-
works, the training procedure, distinguished architectural structures, fine-tuning of
hyperparameters, and the tools for constructing deep learning models. Furthermore, these
subjects will be situated within the sphere of Educational and Research Data Mining
(ERDM), forging connections between these principles and the scrutinized documents.

4.1 Neural Networks

Computational processes are carried out by artificial neural networks made up of exten-
sive matrices of elementary synthetic neurons. These constructs aspire to copy the pro-
cesses noticed within the nerve fibers of brain in human. In this arrangement, each
individual node within the construct acts as a neuron and serves as the foundational
processing element of the artificial neural network.
Figure 4 illustrates neuron structure. This neuron comprises several key features:
inputs (x1, x2, …, xn), which can originate from another neuron output within the
network; a bias (x0), a constant value included into neuron’s activation function’s input;
the weights associated with each of these input (ω1, ω2, ω3, …, ωn), which determine
the significance of each input into a model; and the resulting output of (α).

Fig. 4. Basic artificial neuron.

The neuron’s output is computed according to the following equation:

 n 

α=f ωi · xi (1)
i=0
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 321

The activation function, denoted as f , is a critical component of a neuron. This

imparts adaptability to artificial neural networks, allowing them to grasp intricate non-
linear connections within the data while standardizing the output of neuron (restricting
the outcome within the range of 0 and 1). Frequently employed activation methods in
artificial neural networks encompass the hyperbolic tangent, Re-LU (Rectified Linear
Unit), as well as S-shaped. Neurons establish multiple connections with one another, and
these connections can either amplify or attenuate the activation levels of neighboring
neurons.
In Fig. 5, we discern the arrangement of neural network fundamental. An initial layer,
functioning as the input layer, assumes responsibility for delivering data or characteristics
to the network. The ReLU activation function is a recurrent preference for latent strata
inherent to neural networks. These covert strata harbor the aptitude to effectuate intricate
procedures through the concatenation of elemental functions.
The selection of the activation function within the terminal stratum is contingent
upon the precise problem under consideration: in the context of binary classification,
where output values manifest as binary entities (0 or 1), the sigmoid function is con-
ventionally invoked. Conversely, in scenarios necessitating multiclass classification, the
SoftMax procedure, an extension of the sigmoid function tailored for multiple categories,
is deployed. Meanwhile, for regression quandaries bereft of predefined categories, a
linear function is generally prescribed.
The characteristics of these concealed strata demarcate assorted neural network
architectures, encompassing the likes of the LSTM (Long-Short-Term-Memory), RNN
(Recurrent Neural Network) as well as CNN (Convolutional Neural Network). Num-
ber of concealed strata within a network substantiates its profundity, with a prevailing
principle dictating that networks boasting an augmented assemblage of concealed strata
possess an enhanced faculty to apprehend intricate procedures. In the realm of deep
learning architectures, it is customary to deploy dozens, or even hundreds, of latent
strata that inherently possess the potential to auto-adapt during the model’s training.

Fig. 5. A neural network’s fundamental structure consists of circular nodes, with connections
depicted as arrows, linking the output of one neuron to the input of another.
322 B. Bhaisare and R. Bharati

4.2 Training

Every machine learning algorithm’s fundamental aim is to associate inputs, such as

images, with desired outputs, like labeling them as “cat,” by observing an extensive col-
lection of input-output pairs. In the context of deep learning (DL), mapping of inputs &
the intended output, representing network to be generated, which then achieved through
the utilization of artificial neural networks characterized by a hierarchical structure
composed of numerous layers.
Commencing with the initial layer, which acquires elementary characteristics and
subsequently conveys this information to the ensuing layer, the hierarchical strategy
begins. Each successive layer assimilates this fundamental data, combines it with more
intricate elements, and transfers it to the subsequent stratum within the hierarchy. This
sequential procedure persists, with each stratum gradually constructing more elaborate
representations grounded on the input it receives. The particular alterations that each
stratum employs to the input are recorded in the stratum’s parameters.
To enable the network to effectively acquire and execute its intended functions, it
is imperative to establish the fitting parameters for each stratum, ensuring the optimal
alignment between input instances and their corresponding target results. This process of
configuring the appropriate attributes, especially the parameters, for each computational
unit in the network is what comprises the training of a neural network. Nonetheless,
the complexity arises from the reality that deep learning networks may encompass an
extensive array of parameters, potentially reaching into the millions. Discovering the
precise values for all these parameters can evolve into a notably demanding and intricate
endeavor. As an illustration, a frequently utilized neural network architecture for image
categorization, denoted as VGG16 [13], includes an astonishing count of parameter in
total 138 million. In the beginning, the values of each neuron might undergo arbitrary
assignment or adhere to distinct initialization methodologies, which could encompass
unsupervised pretraining [14].

4.3 Architectures

In the domain of Educational & Research Data Mining, a crucial consideration is the
selection of neural network architectures in deep learning (DL), which is contingent upon
the characteristics of the input data. DL provides an expansive array of architectures and
algorithms, tailored to various data types. This section offers an overview of these popular
architectures, their typical applications, and their relevance to the field.
The architectural choices encompass Convolutional Neural Networks (CNN) also
variant (VGG16 and AlexNet), Multilayer Perceptron (MLP), Feedforward Neural Net-
works (FNN), Word Embeddings (WE), Long Short-Term Memory (LSTM) autoen-
coder, Memory Networks (MN), Bidirectional LSTM (BLSTM), as well as Recurrent
Neural Networks (RNN).
In contrast, the baseline methods include Principal Factor Analysis (PFA), Hid-
den Markov Model (HMM), Latent Dirichlet Allocation (LDA), Logistic Regression
(LogReg), Gaussian Naïve Bayes (GNB), Random guess, N-grams, Input Output HMM
(IOHMM), CRF (Computational Random Fields), Linear Regression (LinReg), Majority
voting, Decision Tree (DT), Slope-One, Random Forest (RF), Support Vector Machine
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 323

(SVM), Hierarchical IRT (HIRT), AdaBoost, K-Nearest Neighbors (K-NN), Singular

Value Decomposition (SVD), Item Response Theory (IRT), Bayesian Linear Ridge
Regression (BLRR), Majority class, LSA (Latent Semantic Analysis), Support Vector
Regression (SVR), Naïve Bayes (NB), BKT (Bayesian Knowledge Tracing), Temporal
IRT (TIRT), IBKT (Intervention BKT).
To evaluate the performance of these methods, various metrics come into play, includ-
ing Precision Root Mean Square Error (RMSE), Accuracy, F-measure, Krippendorff’s
alpha, Logarithmic Loss (Log Loss), Quadratic Weighted Kappa (QWK), Mean per
Class Error (MPCE), Area Under the Curve (AUC), Mean Absolute Error (MAE), Gini,
Recall, R2.
Convolutional Neural Networks. CNNs (Convolutional Neural Networks) are distin-
guished for their multi-layered architecture, which makes them particularly well-suited
for image processing tasks. These networks excel at recognizing fundamental image
features, including edges, in their initial layers. As we progress through the layers, these
foundational elements are synthesized to identify more complex patterns, such as facial
features. CNNs, resembling to FNNs (Feedforward Neural Networks), comprise of neu-
rons with adaptable offsets and coefficients. Every Node accepts input, performs a scalar
product, and employs an activation operation. The concluding, entirely connected stra-
tum exploits a loss metric to gauge the difference between projected and anticipated
outcomes.
A standard Convolutional Neural Network (CNN) consists of three core layers: a
convolutional layer for extracting features, especially in image analysis; a pooling layer
that reduces the dimensionality of these features, often utilizing max pooling to preserve
crucial information; and a fully connected classification layer, responsible for providing
the network’s final output [14].
Unlocking innovative CNN applications, the incorporation of deep layers, which
involve convolution, pooling, and classification, has been instrumental. CNNs extend
their utility beyond image processing into various domains, encompassing video recog-
nition, game playing, and a multitude of natural language processing tasks. They demand
fewer parameters when compared to FNNs and showcase exceptional performance in
tasks like pattern-recognition, particularly in image-recognition. Nevertheless, they have
specific limitations, including high computational requirements, the need for abundant
training data, and the necessity for precise network initialization to address specific
challenges.
Within the realm of Educational Data Mining (EDM), CNNs have been harnessed
for purposes like identifying undesirable student behaviors. Architectures like VGG16
and AlexNet have been employed in video analysis, while CNNs have demonstrated
their usefulness across various applications, including audio and video analysis, text
classification, and even review prediction [15].
Recurrent Neural Networks. Feedforward Neural Networks (FNNs) possess a note-
worthy characteristic in that they lack mechanisms for maintaining persistent informa-
tion. To address this limitation, Recurrent Neural Networks (RNNs) introduce a feedback
loop, enabling the retention of information [16]. RNNs, unlike FNNs with their strict
feedforward connections, introduce feedback connections that loop back into preceding
layers. RNNs, with the utilization of this response mechanism, acquire the capacity to
324 B. Bhaisare and R. Bharati

store memories of earlier inputs. One way to conceive of RNNs is as networks compris-
ing multiple self-replicating units, where each unit conveys information to its successor.
This architectural trait renders RNNs adept at handling sequences and lists, and they
often find application in tasks such as text modeling.
RNNs are employed effectively across various domains, including tasks like speech
recognition, language modeling, and machine translation. However, they grapple with
challenges, one of which is the vanishing gradient problem. This issue arises as gradi-
ents experience exponential decay during backpropagation, posing an obstacle to learn-
ing lower-layer weights. This hampers the stacking of RNNs into deep models and
restricts their ability to capture long-term dependencies. Additionally, RNNs demand
high-performance hardware for effective training and execution.

4.4 Long Short-Term Memory Networks. LSTMs

LSTM (Long Short-Term Memory) networks have gained substantial popularity in recent
years as a specialized form of Recurrent Neural Networks (RNNs). A distinguishing
feature of LSTM architecture is the incorporation of a memory cell, which enables
the learning of long-term dependencies. This memory cell retains its value over time,
influenced by its inputs, and it includes three gates regulating the flow of information:
• The entry gate oversees the moment when fresh data can access the memory.
• The neglect gate commands the elimination of obsolete data to accommodate novel
information.
• The result gate defines when the knowledge stored in the cell is used in the network’s
outcome.
Each of these gates is regulated by associated weights, which are optimized during
training using algorithms like Backpropagation Through Time (BPTT).
A recent simplification of the LSTM architecture is the GRU (Gated Recurrent Unit),
which features two gates (input and forget) and lacks the output gate, resulting in fewer
parameters. Both LSTMs and GRUs are well-suited for tasks involving sequences. In
addition to the tasks previously mentioned for RNNs, LSTMs perform admirably in text
creation, delivering responses to inquiries, and identifying actions in video sequences.
Image and video captioning have been achieved through the integration of Convo-
lutional Neural Networks (CNNs) and LSTMs. In this setup, the CNN is responsible
for processing images or videos, and the output of the CNN is transformed into nat-
ural language by the LSTM. LSTMs offer several advantages over traditional RNNs,
including their enhanced memory capacity, which allows them to remember inputs over
extended periods. Unlike RNNs, which may lose critical information when processing
long text paragraphs, LSTMs excel in retaining important context. LSTMs also mitigate
the vanishing gradient issue encountered in RNNs and require less training data to build
effective models.

4.5 Hyperparameters Tuning

Relevant references are provided whenever available within this paper furnishing explicit
details of hyperparameter. These hyperparameters can be set manually, selected using
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 325

exploration algorithms, or enhanced using model-driven methods [17]. Deep neural net-
works (DNN) have predetermined variables, which are hyperparameters, make up this
setprior to the fine-tuning parameters of a model (offsets &). Typically, these hyperpa-
rameters are divided into two categories: those pertaining to the model’s structure and
those connected with the training procedure.

5 Discussion
Recent efforts have been made to aggregate peer reviews, with initiatives like Publons
consolidating peer review data to construct public reviewer profiles for participating
reviewers. Crossref, a database repository for Digital Object Identifiers (DOIs) with
over 4000 publisher members, introduced a service to integrate peer reviews as part of
the metadata for scientific articles. However, the availability of most reviews remains
limited. In contrast, Peer Read has collected and organized peer reviews to facilitate
use by other researchers for research, experiment replication, and fair comparisons with
prior results.
The literature review also delves into the diverse deep learning (DL) models as well as
specifications applied within the Educational & Research Data Mining review literature.
Long Short-Term Memory (LSTM) architectures emerge as the most frequently used
approach, employed in 59% of the papers and across a wide spectrum of EDM tasks.
This prevalence suggests LSTMs as a promising starting point for system development
in these tasks.
Notably, recent initiatives by program chairs in major Natural Language Processing
(NLP) conferences have focused on different facets of the peer-review procedure, par-
ticularly author responses and overall review quality. The provided large-scale dataset
aims to enable the broader scientific community to delve into peer review properties and
potentially enhance the current peer review model.
In terms of DL model hyperparameters, the review in the preceding section high-
lighted several crucial factors. Configuring a specific architecture involves an empirical
process. Hyperparameter decisions are contingent on the accessible input data and the
nature of the specific task. Among the hyperparameters under scrutiny, the model’s per-
formance significantly depends on factors such as learning rate, batch size, and stopping
criteria (number of epochs). Larger batch sizes can support higher learning rates, mak-
ing them computationally efficient, but smaller steps are recommended to avoid pitfalls.
Adjusting the number of epochs is essential to prevent overfitting. Model size is another
consideration, with deeper and wider architectures offering more power but increasing
the risk of overfitting. High-parameter-count models require ample training data for
effective generalization.

6 Conclusion
Through a systematic search, a total of 41 works were collected in this area. Using the
taxonomy of data mining applications outlined in [7], it’s evident that only four out
of the 13 tasks that have been suggested have been tackled employing deep learning
methods. Underscoring the potential of deep learning in unexplored data mining tasks,
326 B. Bhaisare and R. Bharati

this is particularly noteworthy given the promising results presented in the reviewed
studies, with 67% of them showing that deep learning outperformed traditional machine
learning benchmarks in all experiments. Additionally, this research delves into an exam-
ination of the primary datasets utilized in data mining tasks. This study has explored the
growing adoption of Deep Learning (DL) applications in educational & research Data
Mining, having only three published papers but has consistently gained momentum,
with 17 papers published in 2018. It’s noteworthy that these DL approaches have found
their way into significant research data mining discussion boards, where the last edition
featured seven DL papers, making a total of 16 over the past three years. Like other
research domains, some datasets are publicly available, facilitating experiment repro-
ducibility, while others are custom-built for specific studies. In research data mining, the
challenge of dealing with sensitive information, particularly concerning underage stu-
dents, can hinder data sharing. Nevertheless, this issue can be addressed through proper
anonymization techniques.
This investigation acts as a significant resource and initial stepping stone for
researchers working in the fields of deep learning and research in data mining who aim to
harness these techniques for educational purposes. In light of the expanding utilization
of deep learning techniques in data mining, this study offers valuable perspectives. It
offers a comprehensive exploration of Deep Learning (DL) techniques, beginning with
an introduction to the field. Investigating various deep learning architectures used in
different tasks, this study proceeds. It also assesses conventional hyperparameter set-
tings and provides a compendium of available frameworks to support the development
of deep learning models. Creating a deep learning structure often entails empirical meth-
ods, and the insights drawn from this research can establish a primary resource for future
applications of deep learning within the research data mining field.

References
1. Romero, C., Ventura, S., Pechenizkiy, M., Baker, R.: Handbook of Educational Data Mining.
CRC Press (2010). Complexity 19
2. Biswas, H.K., Hasan, M.: Using publications and domain knowledge to build research pro
les: an application in automatic reviewer assignment. In: ICICT (2007)
3. Rodriguez, M.A., Bollen, J., Van de Sompel, H.: Mapping the bid behavior of conference
referees. J. Inform. 1(1), 68–82 (2007). https://doi.org/10.1016/j.joi.2006.09.006
4. Romero, C., Ventura, S.: Educational data mining: a review of the state of the art. IEEE Trans.
Syst. Man Cybern. Part C Appl. Rev. 40(6), 601–618 (2010)
5. Romero, C., Ventura, S.: Data mining in education. Wiley Interdiscip. Rev.: Data Min. Knowl.
Discov. 3(1), 12–27 (2013)
6. Baker, R.S., Yacef, Y.: The state of educational data mining in 2009: a review and future
visions. JEDM-J. Educ. Data Min. 1(1), 3–17 (2009)
7. Peña-Ayala, A.: Educational data mining: a survey and a data mining-based analysis of recent
works. Expert Syst. Appl. 41(4), 1432–1462 (2014)
8. Bakhshinategh, B., Zaiane, O.R., ElAtia, S., Ipperciel, D.: Educational data mining appli-
cations and tasks: a survey of the last 10 years. Educ. Inf. Technol. 23(1), 537–553
(2018)
9. Aldowah, H., Al-Samarraie, H., Fauzy, W.M.: Educational data mining and learning analytics
for 21st century higher education: a review and synthesis. Telemat. Inform. 37, 13–49 (2019)
 Advancing Peer Review Integrity: Automated Reviewer Assignment Techniques 327

10. Hatcher, W.G., Yu, W.: A survey of deep learning: platforms applications and emerging
research trends. IEEE Access 6, 24411–24432 (2018)
11. Schmidhuber, J.: Deep learning in neural networks: an overview. Neural Netw. 61, 85–117
(2015)
12. Zhong, G., Wang, L., Ling, X., Dong, J.: An overview on data representation learning: from
traditional feature learning to recent deep learning. J. Finance Data Sci. 2(4), 265–278 (2016)
13. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition (2014). https://arxiv.org/abs/1409.1556
14. Y. Bengio, “Practical recommendations for gradient-based training of deep architectures. In:
Montavon, G., Orr, G.B., Müller, K.R. (eds.) Neural Networks: Tricks of the Trade, 2nd edn.
LNTCS, vol. 7700, pp. 437–478. Springer, Heidelberg (2012). https://doi.org/10.1007/978-
3-642-35289-8_26
15. Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., Kuksa, P.: Natural
language processing (almost) from scratch. J. Mach. Learn. Res. 12, 2493–2537 (2011)
16. Tato, A.A.N., Nkambou, R., Dufresne, A.: Convolutional neural network for automatic detec-
tion of sociomoral reasoning level. In: Proceedings of the 10th International Conference on
Educational Data Mining (2017)
17. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational
abilities. In: Proceedings of the National Academy
 Automated Potato Disease Classification Using
Deep Learning - A Comparative Analysis
of Convolutional Neural Networks

Swati Pandey1(B) , Mayuri Gupta1 , Ashish Mishra1 , Ashutosh Mishra2 ,

and Jayesh Gangrade3
1 Jaypee Institute of Information and Technology, Sector-62, Noida, Uttar Pradesh, India
[email protected], [email protected],
[email protected]
2 Thapar University, Patiala, Punjab, India
[email protected]
3 Manipal University, Jaipur, India

[email protected]

Abstract. Potato stands as a vital staple food worldwide, with its consumption
steadily rising and becoming the fourth most consumed staple food globally.
However, the prevalence of potato diseases has posed a significant challenge,
impairing both the quality and quantity of harvests. Timely and accurate disease
classification is imperative for effective disease management and crop protec-
tion. Leveraging recent advancements in deep learning and computer vision, this
study provides a thorough analysis and evaluation of the differences between of
prominent convolutional neural network architectures, namely VGG16, VGG19,
ResNet50, ResNet152, and InceptionV3, for classifying three distinct disease
classes in potato plants: early blight, late blight, and healthy. In the investiga-
tion of a dataset comprising 2152 images, this study scrutinizes the performance
metrics of diverse models. The results underscore the remarkable efficiency of
InceptionV3, demonstrating an impressive 91.84% accuracy in classifying potato
diseases. Significantly, the incorporation of deep learning approaches presents a
potential solution for tackling the complex challenges linked to identifying and
categorizing potato diseases. Additionally, the paper underscores the vital signifi-
cance of automated disease detection in preserving worldwide potato production,
ensuring food security, and promoting sustainable agricultural practices.

Keywords: Deep learning · convolutional neural networks · Potato

classification · accuracy · computer vision · image classification

1 Introduction

In the context of agriculture, the overarching objectives of ensuring food security and
advancing nutritional standards are of utmost importance. In the face of escalating global
population growth, the agricultural sector is challenged to enhance production while

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 328–345, 2025.
https://doi.org/10.1007/978-3-031-71481-8_26
 Automated Potato Disease Classification Using Deep Learning 329

simultaneously upholding the integrity of crops in terms of quality and nutritional con-
tent. Among the various crops pivotal to this mission, the potato (Solanum tuberosum)
has emerged as a noteworthy contender due to its potential to meet escalating sustenance
demands. With its abundant supply of vital nutrients—such as vitamins C and B6, potas-
sium, magnesium, and iron—potatoes emerge as a promising staple, poised to tackle the
daunting challenge of nourishing a burgeoning population [1].
As the fourth most consumed vegetable crop on a global scale and a dietary corner-
stone in Indonesia, potatoes hold a significant role in worldwide agriculture. Indonesia’s
agricultural landscape has witnessed a rapid expansion of potato cultivation, yielding
substantial annual outputs of around 850,000 tons from approximately 60,000 hectares
[2]. This thriving production underscores Indonesia’s position as a key potato producer
within Southeast Asia. However, this ascent in potato cultivation is met with correspond-
ing challenges, particularly the susceptibility of potato plants to an array of diseases
[3]. These maladies, affecting both pre-harvest phases, and post-harvest phases, lead to
diminished agricultural output and complications, spanning from premature harvests to
complete crop failure.
The convergence of disease prevalence and delayed detection highlights the urgency
of swift and accurate disease identification in potato plants. Traditional manual monitor-
ing methods, while indispensable, are fraught with inefficiencies stemming from their
time-intensive nature and the scarcity of on-site experts. The delay in disease detec-
tion fosters an environment conducive to unchecked disease propagation. Farmers, in
their diligent efforts, often resort to approximate methods that lead to misclassifications
and misguided interventions [4]. These inefficiencies, compounded by a lack of expert
guidance and precise quantification of disease severity, contribute to suboptimal disease
management, and compromised crop yields.
In the face of these challenges, the potato’s global significance becomes evident.
Beyond being a culinary delight, potatoes serve as a vital dietary staple for over one bil-
lion people globally, offering essential nutrients. However, the economic impact of potato
diseases, such as late blight and bacterial wilt, cannot be overstated. These diseases lead
to substantial yield losses, increased production costs, and market instability, affecting
not only individual farmers but also global food prices and security. Addressing these
challenges requires concerted efforts in research and development, sustainable farming
practices, and international cooperation. The humble potato, with its complex global
role, necessitates a proactive approach to disease management for sustained agricultural
and economic stability.
In this context, the fusion of cutting-edge technologies, particularly deep learning,
and computer vision, emerges as a potent solution for the intricate challenge of dis-
ease classification in potato plants. Employing Convolutional Neural Network (CNN)
architectures, this research comprehensively assesses five models—VGG16, VGG19,
ResNet50, ResNet152, and InceptionV3. The ensuing methodology unveils a robust and
automated framework to classify early blight, late blight, and healthy states within potato
plants. This paper is primarily concerned with the following contributions:
• Implemented initial data pre-processing methods for uniformity (224 × 224 pix-
els) and introduced systematic data augmentation with geometric transformations to
address limited training data challenges.
330 S. Pandey et al.

• Explored deep learning for potato leaf disease detection, assessing VGGNet, ResNet,
and InceptionV3 for optimal classification. Emphasized CNNs layered complexity
over traditional methods.
• Examined InceptionV3’s intricate neural network design, highlighting its use of con-
volutional blocks and 2D max pooling for feature extraction. Emphasized the strategic
inclusion of 11 Inception modules for detailed feature capture.
• Clarified InceptionV3 elements, concatenation of module outputs for holistic under-
standing, fully connected layers for hierarchical feature abstraction, and SoftMax
classifier for accurate image classification.

2 Related Work
Numerous methodologies have been employed in the realm of plant disease classi-
fication, yet there remains a notable deficiency [5], making it an area of continuous
development due to the wide variation within this field.
To establish a robust feature representation, a deep learning-based strategy can be
formulated. Deep learning [6] has showcased remarkable process across a range of
visual perception tasks, including text detection [7, 8], victim identification [9, 10],
target tracing [11, 12], and object identification [13–15]. Empirical evidence, as well as
conceptual reasoning from previous research [16], supports the potential of deeper layers
in capturing advanced semantic details. This attribute equips the network with enhanced
resilience to fluctuations in pose, color, scale, and deformable objects, rendering it well-
suited for robust leaf disease classification.
Sladojevic et al. [5] constructed a model employing a convolutional neural network
(CNN) to identify diseases in five plant types, covering 13 different disease categories.
This endeavor yielded an average accuracy of 96.3%. In contrast, Erika Fujita et al. [18]
devised a disease diagnosis system for cucumber plants using the CNN method with the
AlexNet architecture, achieving an average accuracy of 82.3%.
Using artificial neural networks and fuzzy logic, Rastogi et al. [19] identified plant
diseases based on leaf conditions, specifically on maple and hydrangea leaves. These
diseases were categorized into two types: leaf spot and scorch leaf. Leaf spots denote spe-
cific points of disease occurrence, while scorch leaves exhibit even distribution. Mohanty
et al. [20] employed deep convolutional neural networks to identify 14 crop species
and 26 diseases, capitalizing on the potential of prominent RGB image features and
machine learning methods. In a separate study, Jia et al. conducted a review focusing on
Deep learning for hyperspectral image classification. The research explores hyperspec-
tral image (HSI) classification, organizing relevant methods under learning paradigms
like transfer learning, active learning, and few-shot learning. [21] Liu et al., explored
deep learning based RGB models that achieved impressive accuracy levels for classify-
ing different coal particles [18]. Wang et al., compare and evaluate traditional machine
learning and deep learning algorithms for image classification [19]. Korot et al. employ
various platforms and datasets for creating image classification models, particularly
excelling in optical coherence tomography [20]. In the context of potato leaf disease
classification, a deep learning approach was introduced in [21], utilizing pre-trained
models like VGG19 for fine-tuning and achieving impressive accuracy. Additionally,
an innovative methodology involving Mask R-CNN and structure-from-motion (SfM)
 Automated Potato Disease Classification Using Deep Learning 331

was proposed for fruit detection and 3D location, enhancing system performance [21].
Hyperspectral imaging has been applied to detect Alternaria solani in potato crops, aid-
ing in early blight identification and reducing the need for excessive plant protection
agents [22]. A deep learning-based approach was proposed in [23, 24] for potato leaf
disease classification, focusing on combating the detrimental impact of potato leaf blight
shown in Table 1. The proposed architecture demonstrated remarkable accuracy across
healthy, early blight, and late blight leaf categories.

Table 1. Comparative analysis of Literature review

Author Methodology Dataset Type Disease Accuracy

Categories
Sladojevic et al. CNN 5 different 13 different plant 96.3
[5] The Caffe classes Diseases
framework with
the 10-fold
cross-validation
approach
Fujita et al. [18] CNN with a Cucumber leaf Evaluated on 7 82.3%
4-fold plants different plant
cross-validation diseases
Rastogi et al. Phase 1: Leaf Images: Classified 2 types Not specified
[19] Artificial Neural Hydrangea and of disease: Leaf
Network (ANN) Maple Spot and Leaf
for plant Scorch
recognition
Phase 2:
K-Means
segmentation and
ANN for disease
classification
Mohanty et al. Deep 54, 306 plant Classified 26 99.35%
[20] convolutional leaf images diseases
neural network
Jia et al. [21] Applied deep 54,129 image Classified in 16 75.66%
learning in the dataset objects
classification of
hyperspectral
images (HSI)
(continued)
332 S. Pandey et al.

Table 1. (continued)

Author Methodology Dataset Type Disease Accuracy

Categories
Talukder et al. CMobileNetV2, 495 potato pests Classified 8 types 91%
[17] CNASLargeNet, images of potato pests
CXception,
CDenseNet201,
and
CInceptionV3
Arshaghi et al. AlexNet, 5000 potato Classified in 5 Reached 100%
[25] GoogLeNet, images classes and 99% accuracy
VGG, R-CNN, in some cases
and Transfer
Learning
Jiang et al. [30] Deep learning Not specified Not specified Not specified
techniques for
the recognition of
multiple lesions
in medical image
processing
Nath, et al. [8] K-means 1000 images 0 to 9 range is 94.44%
clustering and made to evaluate
convolutional the disease
neural network
(CNN)
Arshad et al. A hybrid deep 2152 images of Classified into 10 98.66%
[11] learning model, leafs classes
PLDPNet with
VGG19 and
Inception-V3
models
Tiwari et al. [26] VGG19 for 2152 images Classified in 3 97.8%
fine-tuning classes
through transfer
learning and
logistic
regression
emerging as the
most effective
classifier
 Automated Potato Disease Classification Using Deep Learning 333

3 Proposed Methodology

This section delves into the fundamental mechanisms of all the models. The discussion
encompasses six crucial parameters, namely data gathering, data pre-processing, data
augmentation, potato image classification, data training, and testing (Fig. 1).

Fig. 1. Working of the proposed framework.

3.1 Data Collection

The systematic effort to create a comprehensive and diverse dataset involved an extensive
data acquisition phase, where images were diligently retrieved from prominent reposito-
ries such as Plant-Village and Kaggle [26]. Specifically, the dataset was collected from
Kaggle, more precisely from the repository named Plant-Village, which consisted of
three folders curated by “Arjun Tejaswi.“ These folders were meticulously organized as
follows: Early blight with 1000 images, Late blight with 1000 images, and Healthy leaf
with 152 images [31].
During the meticulous examination of the dataset, inherent challenges impacting
the model’s effectiveness were taken into consideration. A notable challenge identified
is the substantial class imbalance among the three categories—healthy, late blight, and
early blight. The dataset comprises 2,152 images, with Early blight and Late blight
each containing 1000 images, while Healthy leaf has only 152 images. This irregular
distribution has the potential to introduce disparities during training, potentially leading
the model to favor classes with a larger number of samples.
To address this, strategies for handling class imbalances were explored, including
the implementation of data augmentation for the minority class (Healthy leaf). Data
augmentation involves introducing variations or duplicating existing samples through

Fig. 2. Dataset- Healthy, early blight, late blight (left to right)

334 S. Pandey et al.

random transformations, such as rotations, flips, and changes in lighting. The incorpora-
tion of data augmentation techniques aimed to furnish the model with a more equitable
representation of each class, thereby diminishing the risk of it being skewed toward the
more prevalent classes (Fig. 2).

3.2 Data Pre-processing

The initial phase of the data preprocessing pipeline focuses on mitigating image noise
by isolating the region of interest and eliminating extraneous portions. The dataset,
comprising images from diverse sources with varying dimensions, undergoes a harmo-
nization process. This involves resizing the images to a standardized format of 224x224
pixels, a strategic measure aimed at promoting uniformity within the dataset [25]. This
resizing not only addresses variations in image dimensions but also establishes a con-
sistent framework for subsequent stages of analysis and model development, ensuring a
cohesive and standardized approach to the dataset.

3.3 Data Augmentation

Deep Learning, in contrast to traditional shallow machine learning networks, demands
a substantial volume of data for effective training. To overcome the prevalent issues
of inadequate training data and disparities in class distribution, a systematic approach
known as data augmentation is implemented. This technique entails manipulating data
while retaining its fundamental characteristics [27]. Given the limited size of the initial
2,152 datasets, data augmentation becomes imperative for optimal model performance.
This research automates augmentation by incorporating basic geometric transformations,
including translations, rotations, scale adjustments, and shearing, along with vertical and
horizontal flips. This approach effectively enriches the dataset, thereby enhancing the
learning process.

3.4 Image Classification

Deep learning, a subset of artificial intelligence that falls under machine learning, has
made notable progress across diverse domains, including image classification, speech
recognition, and object detection. Distinguished by its multiple layers, deep learning
methods surpass the layered complexity of traditional machine learning techniques [28].
A prominent class within deep learning is the CNN [29], which has garnered recog-
nition for its efficacy in tasks like detecting plant diseases based on leaf conditions.
CNNs typically comprise convolutional layers, often grouped by function, followed by
subsampling, and fully connected layers, mirroring a standard neural network architec-
ture. The evolution of CNNs has witnessed the development of models like AlexNet,
VGG network, GoogLeNet, ResNet (residual networks), and DenseNet, building upon
foundational architectures such as LeNet.
The significance of CNN architectures is pivotal for accurate potato leaf disease
classification, particularly due to their adeptness at capturing distinctive features in
potato leaf images. Architectures like VGG, ResNet, and Inception excel in automat-
ically extracting hierarchical features, allowing the identification of specific patterns
 Automated Potato Disease Classification Using Deep Learning 335

associated with different diseases on potato leaves. ResNet’s of residual connections

proves effective in capturing spatial hierarchies and localized patterns unique to potato
leaf diseases. Transfer learning from pre-trained models further enhances performance,
leveraging broad incorporation knowledge to identify potato leaf-specific features. The
efficient utilization of parameters by CNNs is particularly advantageous for preventing
overfitting in agriculture tasks with limited datasets, ensuring robust potato leaf disease
classification.
Within the confines of the potato leaf disease detection project, the central aim was
to investigate different architectural models for precise classification. Among the scruti-
nized models such as VGGNet16, VGGNet19, ResNet50, ResNet152, and InceptionV3.
InceptionV3 exhibited the most potential, prompting a more in-depth examination. The
selection of these specific CNN architectures is based on their unique characteristics that
address the complexities of potato leaf disease classification. VGG16 and VGG19’s focus
on small convolutional filters is advantageous for capturing fine-grained details, while
ResNet50 and ResNet152’s residual learning addresses the vanishing gradient problem,
allowing the model to efficiently capture features at different depths. InceptionV3’s use
of inception modules with filters of multiple sizes concurrently makes it well-suited
for detecting lesions, discolorations, and anomalies of different sizes on potato leaves
shown in Fig. 3. The diverse architectures provide a range of options to accommodate
the varied visual characteristics associated with different diseases in potato plants.

Fig. 3. InceptionV3 Architecture Model

The InceptionV3 architecture is an advanced neural network specifically crafted

for tasks related to image classification. Its intricate structure begins by subjecting the
input image to a series of convolutional blocks, followed by 2D max pooling opera-
tions. This initial stage serves to extract and refine features from the input, laying the
foundation for subsequent processing. As the architecture progresses, additional con-
volutional blocks are employed to further enhance the representation of features. The
intermittent application of 2D max pooling operations contributes to down-sampling
feature maps, effectively capturing and preserving crucial information at varying scales.
336 S. Pandey et al.

The architectural divergence occurs with the introduction of 11 distinct Inception V3

modules, strategically crafted to capture features using convolutional filters of varying
sizes. This diverse set of modules enables the model to grasp intricate details and rela-
tionships within the input data. A crucial element in the architecture is the concatenation
of the outputs from these modules, a process facilitated by the Concat operation. The
concatenation step plays a key role in capturing intricate relationships among features,
promoting a comprehensive understanding of the input data.
After extracting and combining feature maps, the architecture directs the information
through a series of fully connected layers. These layers include the Flatten operation,
which reshapes the input into a one-dimensional array, and Dense1 and Dense2 layers
with Rectified Linear Unit (ReLU) activation functions. This sequence of dense layers
contributes to the hierarchical abstraction of features, allowing the model to learn intricate
patterns and representations.

4 Experimental Results

This module presents all the training and testing data that has been implemented. The
results show the statistically validated loss and accuracy of the confusion matrix to enable
a better understanding of the readers.

4.1 Data Training

The dataset learning experiment involved the utilization of various neural network archi-
tectures, namely ResNet-50, ResNet-152, Inception V3, VGG16, and VGG19. These
architectures underwent training for 50 epochs with a batch size of 32, incorporating
a learning rate of 0.001 to enhance their performance. The learning process entails the
algorithm actively seeking patterns and features within the image dataset, aiming to
enable effective recognition of new, unseen images. Throughout each epoch, the algo-
rithm iteratively refines its understanding to align with the characteristics of the images it
is being trained on. Results of each epoch are meticulously logged to assess the model’s
performance in terms of both loss and accuracy. A low loss value indicates a well-
performing model, with the ideal target being close to zero or equal to zero, signifying
that the model’s predictions closely align with the actual data. Conversely, accuracy
serves as a crucial metric, quantifying the system’s proficiency in correctly classifying
objects and gauging the overall success level of the model. The following plots present
the results of the loss values and accuracy:
Figure 4(a) depicts the accuracy and loss trends of the VGG19 model during training.
The accuracy on both the training and validation datasets consistently increases, reaching
a final accuracy of 0.86 after 50 epochs, indicating improved image classification over
time. Simultaneously, the loss decreases, confirming the model’s learning process.
In Fig. 4(b), the Confusion Matrix for the VGGnet19 model is presented. Noteworthy
classification accuracies include 44.29% for Early Blight, 39.63% for Late Blight, and
1.86% for Healthy instances. This matrix provides valuable insights into the model’s
effectiveness in distinguishing between specific classes within the dataset.
 Automated Potato Disease Classification Using Deep Learning 337

Fig. 4. (a) The Plot of Accuracy and Loss using VGG19, (b) Confusion matrix for VGG19.

Fig. 5. (a) The Plot of Accuracy and Loss using VGG16, (b) Confusion matrix for VGG16.

Figure 5(a) displays the evolution of training and validation accuracy and loss
throughout 50 epochs for a VGG16 model. The training accuracy exhibits a steady
increase, accompanied by a reduction in training loss. While the validation accuracy
and loss exhibit minor fluctuations, likely stemming from evaluations on an unseen
dataset, the overall trajectory indicates effective learning and generalization. Ultimately,
the model achieves a final accuracy of 0.89 after completing 50 epochs.
In Fig. 5(b), the Confusion Matrix representing the VGGnet16 model illustrates its
classification performance. Noteworthy is the model’s accurate identification of Early
Blight as Early Blight with a precision of 43.82%, Late Blight as Late Blight with 42.66%
precision, and Healthy instances with 2.56% accuracy. This assessment offers valuable
insights into the model’s capability to differentiate between specific classes within the
dataset.
Figure 6(a) displays the 50-epoch training and validation outcomes of a ResNet50
model. Training accuracy rises as training loss diminishes, while validation metrics
exhibit some fluctuation. Surprisingly, the final accuracy after 50 epochs is 0.73,
falling below expectations, possibly due to factors such as a limited dataset, suboptimal
hyperparameters, or task complexity.
338 S. Pandey et al.

Fig. 6. (a) The Plot of Accuracy and Loss using ResNet50, (b) Confusion matrix for ResNet50.

In Fig. 6(b), the Confusion Matrix for the ResNet50 model showcases its classi-
fication performance. Notably, the model excels in identifying Early Blight with an
accuracy of 43.59%, Late Blight with a precision of 29.37%, and displays precision in
recognizing Healthy instances. This assessment offers valuable insights into the model’s
effectiveness in distinguishing and classifying specific classes within the dataset.

Fig. 7. (a) The Plot of Accuracy and Loss using ResNet152, (b) Confusion matrix for ResNet152.

Figure 7(a) visually represents the 50-epoch training dynamics of a ResNet152

model, showcasing a progressive rise in training accuracy and a consistent decline
in training loss. However, the validation accuracy and loss exhibit more pronounced
fluctuations, suggesting potential challenges in generalization. Surprisingly, the model
achieves a final accuracy of 0.70 after 50 epochs, falling short of expectations, possibly
due to dataset limitations, suboptimal hyperparameters, or task complexity.
In Fig. 7(b), the Confusion Matrix highlights the ResNet152 model’s robust classi-
fication performance. Notably, it attains 44.52% accuracy in identifying Early Blight,
demonstrates 25.41% precision in classifying Late Blight, and achieves 2.10% accu-
racy in recognizing instances as Healthy. This succinct analysis emphasizes the model’s
proficiency in distinguishing these classes within the dataset.
 Automated Potato Disease Classification Using Deep Learning 339

Fig. 8. (a) The Plot of Accuracy and Loss using Inception V3, (b) Confusion matrix for Inception
V3

Figure 8(a) depicts the training and validation progress of an InceptionV3 model over
50 epochs. The training accuracy shows a steady increase, accompanied by a consistent
decrease in training loss, indicating effective learning. While validation metrics display
some fluctuations, the InceptionV3 model ultimately achieves an impressive final accu-
racy of 0.92, positioning itself as the top-performing model. This graph underscores the
InceptionV3 model’s excellence in image classification.
In Fig. 8(b), the Confusion Matrix evaluates the InceptionV3 model’s classification
performance. Notably, it achieves a high accuracy of 44.06% in classifying Early Blight
and a precision rate of 43.82% for Late Blight. Additionally, the model displays 3.96%
accuracy in identifying Healthy instances. This analysis provides technical insights into
the model’s adeptness in effectively discerning these categories within the dataset.

4.2 Data Testing

Following the meticulous training phase with our thoughtfully curated datasets, a critical
testing process ensued. In this phase, newly acquired data, distinct from the training
dataset, was utilized to evaluate the performance of a diverse set of models, including
VGG16, VGG19, ResNet50, ResNet152, and Inception V3. This comprehensive testing
procedure aims to rigorously assess the efficacy and generalization capabilities of each
model, exposing them to the challenges presented by previously unseen data. Through
this detailed testing protocol, our goal is to extract nuanced insights into the robustness
and adaptability of these models, providing valuable information to guide decisions
regarding their suitability for real-world applications (Figs. 9, 10 and 11).

Fig. 9. Testing Results for Healthy Leaves.

340 S. Pandey et al.

Fig. 10. Testing Results for Early Blight Leaves.

Fig. 11. Testing results for Late blight leaves

In this research study, the evaluation encompassed five distinct models: VGGNet16,
demonstrating an accuracy of 0.89, VGGNet19, achieving an accuracy rate of 0.86,
ResNet50, delivering an accuracy score of 0.73, ResNet152, yielding an accuracy of
0.70 and Inception V3, standing out with the highest accuracy of 0.92.
The experimentation phase involved consistent parameter configuration across all
models, with hyperparameters set as follows: 50 epochs, a batch size of 32, image dimen-
sions of 224 × 224, and the Adam optimizer with a learning rate of 0.001. The selection
of 50 epochs was influenced by practical considerations, as the computational constraints
of working on a CPU deemed it an optimal balance between model convergence and
evaluation time. This choice proved effective, as it yielded satisfactory accuracy within
the given epoch limit. The learning rate of 0.001 was chosen to facilitate a gradual descent
during optimization, promoting stability in model training. The Adam optimizer, known
for its adaptive learning rate and momentum, was selected for its efficiency in handling
diverse datasets and achieving robust convergence. Each model underwent training with
these parameters, but the time required per epoch varied, reflecting the computational
demands of the respective architectures. Specifically, VGGNet16 demanded an average
of approximately 800 s per epoch, translating to a total training duration of approximately
10.8 h for 50 epochs. VGGNet19, with an average epoch duration of 900 s, necessitated
around 12.5 h for the same number of epochs. Meanwhile, ResNet50 required an average
of 600 s per epoch, leading to a total training time of approximately 8.3 h for 50 epochs.
ResNet152, with an average epoch duration of 450 s, completed 50 epochs in approx-
imately 6.25 h. Finally, Inception V3 exhibited the shortest training time per epoch,
averaging 250 s, and completed 50 epochs in roughly 3.4 h.
It’s worth emphasizing that precision during the testing of these models can be
influenced by factors like image noise. To optimize classification accuracy, it is advisable
to implement preprocessing techniques, such as the removal of extraneous elements and
ensuring that each image frame contains only one identifiable leaf object.
 Automated Potato Disease Classification Using Deep Learning 341

4.3 Performance Metrics

In evaluating the experimental results, key performance metrics, including precision (1),
recall (2), accuracy score (3), and F1-score (4), were utilized. These metrics, selected pur-
posefully for potato leaf disease classification, offer unique insights. Precision, empha-
sizing the accuracy of positive predictions, minimizes false positives, ensuring the iden-
tified diseases are genuinely present. Recall focuses on the model’s effectiveness in
capturing all instances of positive classes, providing a comprehensive evaluation and
minimizing false negatives. Accuracy, offering a general measure of correctness across
all classes, provides an overall assessment of the model’s performance. F1-score, as a
harmonic mean of precision and recall, serves as a suitable metric for the imbalanced
dataset, striking a balance between precision and recall. This comprehensive set of met-
rics contributes to a nuanced and informative evaluation of the model’s efficacy in potato
leaf disease classification.
1. Precision (P): Precision is the ratio of correctly predicted positive observations to all
positive predictions made by a classifier.
TP
P= (1)
TP + FP
2. Recall (R): The recall is calculated by dividing the number of correctly predicted
positive observations by the total number of actual positive instances in the dataset.
TP
R= (2)
TP + FN
3. Accuracy Score (Acc): The accuracy score is a measure of the proportion of correctly
classified instances, considering both true positives and true negatives.
TP + TN
Acc = (3)
TP + TN + FP + FN
4. F1-Score (F1): The F1-Score (F1) is a metric that signifies the harmonic mean of pre-
cision and recall. It provides a well-balanced evaluation of a classifier’s performance,
especially when dealing with imbalanced datasets.
P∗R
F1 = 2 ∗ (4)
P+R
To define each metric, the terms true positives (TP), false positives (FP), true neg-
atives (TN), and false negatives (FN) are used to describe different aspects of the
measurement.
Table 2 provides an insightful overview of the diverse performance metrics employed
to ascertain the efficacy of potato image classification across various state-of-the-art
models. The comprehensive roster of models encompasses VGG16, VGG19, ResNet50,
ResNet152, and InceptionV3, each meticulously evaluated in the context of their clas-
sification capabilities concerning distinct class instances—namely, Early Blight, Late
Blight, and Healthy. The utilization of precision, accuracy, recall, and F1 score as metrics
342 S. Pandey et al.

Table 2. Performance Measures of all the models

Model/Report Class P R Acc F1

VGGNet16 Early Blight 0.93 0.94 0.89 0.93
Late Blight 0.86 0.92 0.89
Healthy 0.92 0.37 0.52
VGGNet19 Early Blight 0.90 0.95 0.86 0.92
Late Blight 0.85 0.85 0.85
Healthy 0.50 0.27 0.35
ResNet50 Early Blight 0.67 0.94 0.73 0.78
Late Blight 0.85 0.63 0.73
Healthy 0.00 0.00 0.00
ResNet152 Early Blight 0.63 0.95 0.70 0.76
Late Blight 0.86 0.55 0.67
Healthy 0.00 0.00 0.00
InceptionV3 Early Blight 0.95 0.94 0.92 0.95
Late Blight 0.89 0.94 0.91
Healthy 0.89 0.57 0.69

for potato image classification adds layers of granularity to the assessment, allowing for
a nuanced understanding of the models’ performance across the spectrum of categories.
This systematic approach to performance evaluation not only facilitates a comparative
analysis of the models but also unveils the specific strengths and areas of improvement
for each in the challenging task of potato image classification.

5 Discussion

The utilization of Convolutional Neural Networks (CNNs) for classifying potato leaf
diseases marks a significant stride in harnessing deep learning for agricultural issues.
CNNs, such as VGG, ResNet, and Inception, showcase their effectiveness in capturing
hierarchical features essential for discerning specific patterns linked to various diseases
on potato leaves.
The InceptionV3 architecture, with its intricate design, emerges as the most promis-
ing model for potato leaf disease classification. Its utilization of inception modules
with filters of multiple sizes enables effective detection of lesions, discolorations, and
anomalies on potato leaves. The concatenation of outputs from these modules plays a
crucial role in capturing complex relationships among features, contributing to a holistic
understanding of the input data.
In the experimental phase, the research evaluated five models—VGGNet16,
VGGNet19, ResNet50, ResNet152, and InceptionV3—using common hyperparameters.
 Automated Potato Disease Classification Using Deep Learning 343

InceptionV3 demonstrated the shortest training time per epoch, achieving an impressive
accuracy of 91.84% in classifying early blight, late blight, and healthy potato plants.
Key performance metrics, including precision, recall, accuracy, and F1-score, pro-
vided a comprehensive assessment of the models’ efficacy. InceptionV3’s standout
performance underscores the importance of selecting architectures tailored to the
complexities of potato leaf disease classification.
In conclusion, this research contributes significantly to addressing the global chal-
lenge of potato disease detection, with InceptionV3 proving to be a top-performing
model. The study not only advances agricultural technology but also highlights the
broader potential of deep learning in enhancing food security and sustainable agriculture.

6 Conclusion and Future Work

Due to the escalating consumption of potatoes, effective disease management is imper-
ative for ensuring global food security and sustaining agriculture. This study marks
a notable stride in tackling the worldwide challenge of potato disease detection and
classification. Evaluating prominent convolutional neural network architectures such
as VGG16, VGG19, ResNet50, ResNet152, and InceptionV3, the research identifies
InceptionV3 as the top-performing model, achieving an impressive 91.84% accuracy in
classifying early blight, late blight, and healthy potato plants.Subsequent research will
concentrate on enhancing the model’s accuracy and resilience through advanced method-
ologies. Leveraging transfer learning, the study aims to capitalize on pre-existing model
knowledge, fine-tuning it for heightened performance. Additionally, the exploration of
transformer architectures is on the agenda to elevate the precision of potato disease clas-
sification. Acknowledging the study’s current limitation of a relatively modest dataset
comprising 2152 images, future efforts will focus on procuring and utilizing a more
extensive and diverse dataset. This expansion is intended to bolster the model’s general-
ization capabilities, making it proficient for real-time disease classification across diverse
agricultural contexts. The envisaged future research also includes efforts to develop real-
time field deployment systems, monitor disease progression, collaborate with agricul-
tural stakeholders for large-scale implementation, and consistently refine the model’s
capabilities for addressing evolving challenges in potato disease management.

References
1. Beals, K.A.: Potatoes, nutrition and health. Am. J. Potato Res. 96(2), 102–110 (2019)
2. Sholihati, R.A., Sulistijono, I.A., Risnumawan, A., Kusumawati, E.: Potato leaf disease clas-
sification using deep learning approach. In: 2020 International Electronics Symposium (IES),
pp. 392–397. IEEE (2020)
3. Arora, R.K., Sharma, S.: Pre and post harvest diseases of potato and their management.
In: Goyal, A., Manoharachary, C. (eds) Future Challenges in Crop Protection Against Fungal
Pathogens. Fungal Biology, pp. 149–183. Springer, New York (2014). https://doi.org/10.1007/
978-1-4939-1188-2_6
4. Tm, P., Pranathi, A., SaiAshritha, K., Chittaragi, N. B., Koolagudi, S.G.: Tomato leaf disease
detection using convolutional neural networks. In: 2018 Eleventh International Conference
on Contemporary Computing (IC3), pp. 1–5. IEEE (2018)
344 S. Pandey et al.

5. Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., Stefanovic, D.: Deep neural networks-
based recognition of plant diseases by leaf image classification. Comput. Intell. Neurosci. 2016
(2016)
6. Gupta, M., Sinha, A.: Multi-class autoencoder-ensembled prediction model for detection of
COVID-19 severity. Evol. Intel. 16(4), 1433–1445 (2023)
7. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521(7553), 436–444 (2015)
8. Mandal, S.N., et al.: Image-based potato phoma blight severity analysis through deep learning.
J. Inst. Eng. (India): Ser. B 104(1), 181–192 (2023)
9. Risnumawan, A., Sulistijono, I. A., Abawajy, J.: Text detection in low resolution scene images
using convolutional neural network. In: Herawan, T., Ghazali, R., Nawi, N.M., Deris, M.M.
(eds) SCDM 2016. AISC, vol. 549, pp. 366–375. Springer, Cham (2017). https://doi.org/10.
1007/978-3-319-51281-5_37
10. Afakh, M.L., Risnumawan, A., Anggraeni, M.E., Tamara, M.N., Ningrum, E.S.: Aksara jawa
text detection in scene images using convolutional neural network. In: 2017 International
Electronics Symposium on Knowledge Creation and Intelligent Computing (IES-KCIC),
pp. 77–82. IEEE (2017)
11. Arshad, F., et al.: PLDPNet: End-to-end hybrid deep learning framework for potato leaf
disease prediction. Alex. Eng. J. 78, 406–418 (2023)
12. Sulistijono, I.A., Risnumawan, A.: From concrete to abstract: Multilayer neural networks for
disaster victims detection. In: 2016 International Electronics Symposium (IES), pp. 93–98.
IEEE (2016)
13. Sulistijono, I.A., et al.: Implementation of victims detection framework on post disaster
scenario. In: 2018 International Electronics Symposium on Engineering Technology and
Applications (IES-ETA), pp. 253–259. IEEE (2018)
14. Anwar, M.K., Risnumawan, A., Darmawan, A., Tamara, M.N., Purnomo, D.S.: Deep multi-
layer network for automatic targeting system of gun turret. In: 2017 International Electronics
Symposium on Engineering Technology and Applications (IES-ETA), pp. 134–139. IEEE
(2017)
15. Anwar, M.K., et al.:. Deep features representation for automatic targeting system of gun turret.
In: 2018 International Electronics Symposium on Engineering Technology and Applications
(IES-ETA), pp. 107–112. IEEE (2018)
16. Imaduddin, H., Anwar, M.K., Perdana, M.I., Sulistijono, I.A., Risnumawan, A.: Indonesian
vehicle license plate number detection using deep convolutional neural network. In: 2018
International Electronics Symposium on Knowledge Creation and Intelligent Computing
(IES-KCIC), pp. 158–163. IEEE (2018)
17. Talukder, M.S.H., et al.: PotatoPestNet: a CTInceptionV3-RS-based neural network for
accurate identification of potato pests. Smart Agric. Technol. 5, 100297 (2023)
18. Fujita, E., Kawasaki, Y., Uga, H., Kagiwada, S., Iyatomi, H.: Basic investigation on a robust
and practical plant diagnostic system. In: 2016 15th IEEE International Conference on
Machine Learning and Applications (ICMLA), pp. 989–992. IEEE (2016)
19. Rastogi, A., Arora, R., Sharma, S.: Leaf disease detection and grading using computer vision
technology & fuzzy logic. In: 2015 2nd International Conference on Signal Processing and
Integrated Networks (SPIN), pp. 500–505. IEEE (2015)
20. Mohanty, S.P., Hughes, D.P., Salathé, M.: Using deep learning for image-based plant disease
detection. Front. Plant Sci. 7, 1419 (2016)
21. Jia, S., Jiang, S., Lin, Z., Li, N., Xu, M., Yu, S.: A survey: Deep learning for hyperspectral
image classification with few labeled samples. Neurocomputing 448, 179–204 (2021)
22. Liu, Y., Zhang, Z., Liu, X., Wang, L., Xia, X.: Performance evaluation of a deep learning
based wet coal image classification. Miner. Eng. 171, 107126 (2021)
23. Wang, P., Fan, E., Wang, P.: Comparative analysis of image classification algorithms based
on traditional machine learning and deep learning. Pattern Recogn. Lett. 141, 61–67 (2021)
 Automated Potato Disease Classification Using Deep Learning 345

24. Korot, E., et al.: Code-free deep learning for multi-modality medical image classification.
Nat. Mach. Intell. 3(4), 288–298 (2021)
25. Arshaghi, A., Ashourian, M., Ghabeli, L.: Potato diseases detection and classification using
deep learning methods. Multimed. Tools Appl. 82(4), 5725–5742 (2023)
26. Tiwari, D., Ashish, M., Gangwar, N., Sharma, A., Patel, S., Bhardwaj, S.: Potato leaf diseases
detection using deep learning. In: 2020 4th International Conference on Intelligent Computing
and Control Systems (ICICCS), pp. 461–466. IEEE (2020)
27. Gené-Mola, J., et al.: Fruit detection and 3D location using instance segmentation neural
networks and structure-from-motion photogrammetry. Comput. Electron. Agric. 169, 105165
(2020)
28. Gupta, M., Singhal, Y.K., Sinha, A.: Assessing spatiotemporal transmission dynamics of
COVID-19 outbreak using AI analytics. In: Gupta, D., Khanna, A., Kansal, V., Fortino, G.,
Hassanien, A.E. (eds) DoSCI 2021. AISC, vol. 1374, pp. 829-838. Springer, Singapore (2022).
https://doi.org/10.1007/978-981-16-3346-1_67
29. García, S., Luengo, J., Herrera, F.:. Data Preprocessing in Data Mining, vol. 72, pp. 59–139.
Springer, Cham (2015)
30. Jiang, H., et al.: A review of deep learning-based multiple-lesion recognition from medical
images: classification, detection and segmentation. Comput. Biol. Med. 106726 (2023)
31. https://www.kaggle.com/datasets/arjuntejaswi/plant-village
 Enhancing Monkeypox Disease Detection
Using Computer Vision-Based
Approaches and Deep Learning

Imtiaj Ahmed(B) , Rayan , Sayma Akter Tihany , and Adnan Mahmud

Department of Computer Science and Engineering,

East West University, Dhaka 1212, Bangladesh
[email protected]

Abstract. This research addresses the challenge of monkeypox, a con-

tagious skin disease that may be spread from animal to human contact.
Recognizing its broad reach and potential severity, the research empha-
sizes the critical need for early and precise detection. The project sets
out to evaluate how well a computer vision-based approach and CNN-
based technology can accurately spot instances of monkeypox. Various
algorithms, including Googlenet, Resnet50, VGG16, VGG19, Darknet53,
Mobilenetv2, Inceptionv3, Inceptionresnetv2, and Yolov7, are put to the
test. Notably, Mobilenetv2 stands out with an exceptional validation
accuracy of 95.2%, showcasing its effectiveness in pinpointing and identi-
fying the disease. The study emphasizes the significance of deep learning
optimizers in improving the outcomes of image classification, not only
for monkeypox detection but also for making strides in diagnosing mon-
keypox diseases. Furthermore, the research suggests avenues for refining
deep learning architectures and optimization techniques.

Keywords: Monkeypox · Skin diseases · OpenCV · Machine

learning · MobileNetV2

1 Introduction
Monkeypox, a complex disease with various clinical signs, poses a real challenge
in getting diagnosed and treated promptly. The usual detection methods depend
heavily on human observations and lab tests, which can lead to errors and delays
in critical decision-making. Our mission, however, is steering us into a new era
where the fusion of image processing and deep learning is set to transform how
we detect Monkeypox. Imagine a world where artificial intelligence works hand
in hand with the intricate details hidden in medical images. That’s the vision
driving our research - a vision where spotting Monkeypox early isn’t just a dream
but a reality that can genuinely improve patient outcomes.
In this paper, we take you on a journey through the fascinating realms of
image processing and deep learning, revealing the methods behind our innovative
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 346–356, 2025.
https://doi.org/10.1007/978-3-031-71481-8_27
 Enhancing Monkeypox Disease Detection 347

approach. We are using advanced technologies like convolutional neural networks

and sophisticated image analysis to create a sturdy framework. Our goal isn’t
just to replace traditional diagnostic methods but to enhance them, providing
healthcare professionals with a tool that speeds up accurate diagnoses. As we
dive into the details of our research, let’s collectively explore the potential of this
collaboration between medical expertise and cutting-edge technology. Together,
we’re paving the way for a future where Monkeypox detection goes beyond the
usual methods, bringing us closer to a world where early intervention is the
standard rather than the exception [1].

2 Related Works
The goal of this research [2] was to find a deep-learning model for monkeypox
(monkeypox) detection. MobileNetV2 outperformed the other five pre-trained
models with 98.16% accuracy, 0.97 recall, 0.98 precision, and 0.97 F1-score. For
the early and precise identification of monkeypox in clinical settings, the model
demonstrated encouraging results.
The monkeypox virus [3] raises the possibility of a global pandemic, but
ML provides a remedy. We offer the “Monkeypox2022” dataset for using image
analysis to detect the illness, which is freely accessible on GitHub. With the
help of LIME for feature extraction, our modified VGG16 model achieves high
accuracy of 97.18% and 88.88% in two trials.
A contagious virus mostly found in Africa, monkeypox, has spread around the
world. Headaches, chills, fever, skin tumours, and rashes mimicking smallpox are
among the symptoms. Early diagnosis is made possible by several AI models. 34
current research on monkeypox diagnostics, spread modelling, treatment, vaccine
development, and risk management were examined [4].
Monkeypox is a contagious and potentially deadly disease caused by the
monkeypox virus. This work proposes a secure, computer vision-based method
for diagnosing monkeypox using deep learning techniques and skin lesion images,
offering promising results for widespread deployment [5].
Monkeypox is the monkeypox virus, of contagious and potentially deadly
disease with symptoms like skin lesions, rashes, and fever. The latest outbreak
has spread globally. Traditional diagnosis is risky for medical staff, but IoT and
AI offer a smart and secure method using computer vision to analyze skin lesion
images. This non-invasive approach, evaluated on two datasets, shows promising
results and can be a cost-effective solution for underprivileged areas lacking lab
infrastructure [6].
Medical diagnosis is aided by machine learning [4], especially in dermatol-
ogy when separating related disorders like skin cancer. Accurate identification
is essential for efficient epidemiological management during a recent monkey-
pox epidemic. An open database of photos of monkeypox and control subjects
was used to accomplish this. Training and test sets were created from the pre-
processed photos in an 80/20 ratio. Using MiniGoggleNet, six experiments of
various epochs were carried out. The best model has an accuracy of 0.9708, a
348 I. Ahmed et al.

loss function of 0.1442, an AUC of 0.74 for class 0 and class 1, a micro-average
AUC of 0.76, and a macro-average AUC of 0.74 after 50 training iterations.
Monkeypox poses a growing global threat [7], with rising cases and transmis-
sion risks. Machine Learning (ML) offers promise in diagnosing diseases via image
analysis. Here, we present a Monkeypox diagnostic model, GRA-TLA, leveraging
ML techniques. Testing on ten CNN models showed accuracies of 77% to 88%
for binary classification and 84% to 99% for multiclass classification. Our app-
roach proved computationally efficient, with LIME providing valuable insights
into disease indicators.
The fear of monkeypox [8], which has the potential to spark a global pan-
demic, is impeding the global recovery from COVID-19. Image-based diagnostics
can help diagnose monkeypox by using machine learning, similar to its success-
ful uses in cancer and COVID-19 identification. To increase the precision of
monkeypox picture classification, this study suggests two techniques that com-
bine transfer learning, meta-heuristic optimization, and a deep neural network.
Evaluation utilizing a public dataset shows the suggested approaches’ higher
performance and efficacy, with an average classification accuracy of 98.8%.
This research [9] leverages deep learning to enhance monkeypox detection,
addressing misdiagnosis challenges due to similar symptoms with chickenpox. A
custom CNN model achieved superior accuracy (99.60%) compared to existing
models, aiding rapid and precise diagnosis using digital skin images.
In 1970 [10], the 9th Democratic Republic of the Congo became the first
human to contract monkeypox after smallpox eradication. Mostly seen in the
Congo Basin’s rainforests, the disease spread across 11 African nations [11]. The
discovery of the varicella virus during monkeypox outbreaks suggests changes
in how these diseases spread [12]. Since 2017 [13], Nigeria has faced over 500
suspected cases, with 200 confirmed and a 3% fatality rate. The 2003 U.S. out-
break, linked to pet prairie dogs, highlights monkeypox’s global impact. Recent
cases [14] involve Nigerian tourists in various non-endemic countries.

3 Materials and Methods

3.1 Dataset Used
Using a transfer learning method,
– Proposing the pre-trained CNN models to be developed using deep learning.
– Collecting data on paper usage is challenging, involving information on con-
sumption, document types, processes, costs, and waste.
– Data analysis containing hospital and medical sector information, provides
insights.
– Analyze paper consumption in hospitals, identifying limitations, and evalu-
ating potential benefits of systems.
The Monkeypox Skin Lesion Dataset from IEEE DataPort was used in this
word. This dataset, which is used for binary classification, concentrates on skin
lesion pictures linked to measles, chickenpox, and monkeypox.
 Enhancing Monkeypox Disease Detection 349

3.2 Data Analysis

Images of instances of chickenpox and measles are included in the classes titled
“Monkeypox” and “Others.” The three primary folders that make up the mon-
keypox skin lesion dataset are listed below. There are 72870 photos in the first
folder, “Original Images,” altogether. The remaining 42260 photos showcase
Chickenpox and Measles, while 3502 of them are designated as “Monkeypox.”
The “Augmented Images” folder contains enhanced photos for both classes,
serving as a repository of additional images, which also serves as a collection of
extra pictures. These photos were made using a variety of techniques, includ-
ing scaling, rotation, translation, reflection, shear, hue, saturation, and contrast
modifications. The “Fold1” directory contains one of the datasets used in three-
fold cross-validation. This folder’s original images have roughly been split into
training, validation, and test sets with a ratio of 70:10:20. It is significant to
notice that only improved copies of the images are included in the training and
validation sets, but only the original photographs are included in the test set.
Mainly the Fold1 dataset was utilised in this work to train the computer
vision models. It comprises 276‘5410 Monkeypox pictures and 32690 Other
(Chickenpox and Measles) images in the training subset, 20000 Monkeypox
images and 25000 Other images in the test subset, and 16800 Monkeypox images
and 25200 Other images in the validation subset. Figure 1, represents the Mon-
keyPox Sample Images.

Fig. 1. MonkeyPox Sample Images

In Fig. 2, the architecture appears to be a convolutional neural network

(CNN) for MonkeyPox disease detection. The Monkeypox Skin Lesion Dataset is
350 I. Ahmed et al.

Fig. 2. Other Sample Images

a crucial tool for research and the creation of computer vision algorithms that can
precisely identify and categorize skin lesions associated with monkeypox. Initial
examination of a set of sample photos reveals distinctive features of monkey-
pox rashes, which are distinguished by their severity and localized appearance
on the hands, feet, and mouth. However on the other hand, frequently cause
rashes that are more evenly distributed over the entire body. The Monkeypox
Skin Lesion Dataset provides a singular chance to further our comprehension
of skin lesions connected to monkeypox. We can train models capable of accu-
rately recognizing and categorizing these particular sorts of rashes by utilizing
computer vision approaches, such as deep learning algorithms. By enabling ear-
lier and more precise diagnoses of monkeypox and differentiating it from other
dermatological disorders, the results of this study have the potential to make a
substantial contribution to the area of dermatology. In the end, this research
could support the creation of enhanced management plans and public health
responses to monkeypox epidemics.

4 Proposal Model

In our paper, we provide a unique research strategy that uses the CNN models
and transfer learning to get over the problems brought on by small datasets. Our
goal is to create a very effective model that can correctly categorize skin lesions
brought on by chickenpox, measles, and monkeypox.
 Enhancing Monkeypox Disease Detection 351

We start our study by obtaining the weights for the pre-trained CNN mod-
els from the vast ImageNet dataset. We may utilize the vast knowledge and
skills of the pre-trained model in picture classification tasks by designating the
foundation layers as trainable while leaving the top layers non-trainable.
We can successfully utilize the CNN model’s learned features and represen-
tations by using this transfer learning technique, allowing us to extract valuable
information from the skin lesion photos. With the help of our study, we want to
create a reliable and accurate classification model that will help recognise and
treat skin lesions caused by measles, chickenpox, and monkeypox, ultimately
leading to better healthcare practices in this area (Fig. 3).

Fig. 3. Convolutional neural network Blocks

Additional layers are added to adapt the model to our particular job. Since
the challenge is a two-class classification problem, binary cross-entropy is used
as the loss function in the model’s compilation together with accuracy as the
performance metric. Callbacks are used to improve the training process.

5 Experimental Result
The model parameters are kept using the least validation loss as a criterion in
the ModelCheckpoint callback. The ReduceLROnPlateau function dynamically
352 I. Ahmed et al.

Fig. 4. Model loss and Accuracy

reduces our model’s learning rate by 50% if the validation loss does not improve
for two consecutive epochs. We used a batch size of 32 and 50 iterations to train
the model using the training data during the training phase. Using test data, the
model’s performance was assessed; after 50 epochs. We examined the training
history after the training was finished and found that the model’s validation
accuracy and loss stabilized at about 40 epochs. The weights were loaded for
the model with the highest validation accuracy and lowest validation loss over
training. These weights will be used to assess the model’s performance accord-
ing to certain standards. By adopting the suggested model, we want to solve the
problems brought on by a small dataset and offer a precise and trustworthy
classification tool for skin lesions associated with chickenpox, measles, and mon-
keypox (Fig. 4).
We looked at the model’s efficiency. The model was created using pre-trained
CNN models, whose weights were taken from ImageNet. The model was altered
by adding new layers and making some trainable for the research. The model
was learned from a dataset for 50 epochs of training. On the training data,
Mobilenetv2 get a 96.10% accuracy rate. Techniques like preserving the best
 Enhancing Monkeypox Disease Detection 353

weights and modifying the learning rate enhanced the model. The effectiveness
of the model was then evaluated using the test data and the validation data. The
model’s accuracy on the test data was 96.10%, which is a respectable result.
However, the accuracy decreased to 95.198% when using the hidden validation
data. This might be a result of the limited training dataset, which prevented the
model from being trained on all the changes seen in the validation data.
Table 1, there are showing the MobileNetV2 model training accuracy, Testing
accuracy, precision, and recall.

Table 1. Table of Models, % of their Accuracy, Precision and F1-score

Model Training Accuracy Validation Accuracy Recall Precision F1-score

Googlenet 92.2 91.8 91.5 90.7 93.1
Resnet50 94.2 91.2 92.3 92.6 91.2
VGG16 96.3 91.5 90.1 89.2 92.1
VGG19 95.3 93.2 92.2 92.6 93.2
Darknet53 91.5 89.6 88.2 87.1 88.7
Mobilenetv2 96.1 95.2 95.2 94.5 95.8
Inceptionv3 80.5 78.5 79.3 78.8 78.9
Inceptionresnetv2 80.1 78.8 79.7 79.1 80.2
Yolov7 92.2 92.1 93.1 92.2 92.7

A comparative analysis of various optimization algorithms for monkeypox

disease detection revealed the following validation accuracies % are: Googlenet
(91.8), Resnet50 (91.2), VGG16 (91.5), VGG19 (93.2), Darknet53 (89.6),
Mobilenetv2 (87.1), Inceptionv3 (77.5), Inceptionresnetv2 (75.8), Efficientnetb0
(81.2), Desnet201 (85.8), Xception (86.5), Nasnetlarge (81.8), and Yolov7 (96.1).
Among these algorithms, Yolov7 exhibited the highest % age of validation accu-
racy of 96.1, showing its outstanding efficacy in diagnosing and localising mon-
keypox diseases. This suggests that, when tested on a different validation dataset,
Mobilenetv2 outperformed the other algorithms in properly recognizing and
localizing objects within pictures. It demonstrates that Mobilenetv2’s design
and training procedure were successful in obtaining higher performance in object
detection tests. These findings shed light on the efficacy of various algorithms for
detecting monkeypox diseases, leading to the selection of appropriate models to
improve agricultural output and disease control techniques.
Figures 5 display the model loss, model accuracy, and confusion matrix of the
MobileNetV2 model, respectively.
The classification report and confusion matrix are thoroughly examined, and
it is clear that the model performs better than others in properly forecast-
ing instances of monkeypox. The model has a better sensitivity, indicating an
improvement in its ability to detect cases of monkeypox. This suggests that the
354 I. Ahmed et al.

Fig. 5. Confusion Matrix

model has a strong capacity to identify instances of monkeypox. We may eval-

uate the model’s ability to predict outcomes across several classes by using the
classification report and confusion matrix. Notably, it excels in correctly identi-
fying cases of monkeypox, demonstrating its effectiveness in this particular area
of categorization. Due to the increased sensitivity, there is a lower chance that
cases of monkeypox may be mistakenly labelled as negative cases. As a result,
the model successfully diagnoses monkeypox, increasing the possibility of quick
and efficient intervention in such situations.
In summary, the model’s performance analysis, based on the classification
report and confusion matrix, highlights its exceptional accuracy and capability
to accurately forecast incidences of monkeypox, suggesting its strength in this
particular area of classification.

6 Discussion
According to our study’s findings, the model’s validation accuracy was about
95.2%. This means that, on average, 95.2 out of 100 incidents were properly
classified by the model as cases of monkeypox. The program also demonstrated
a 93.6% sensitivity for identifying positive instances of monkeypox. We did see
difficulties, though, when misclassifications occurred because the illness resem-
bled measles or chickenpox. These challenges are caused by the similarity of
the symptoms and the individual variation in monkeypox presentation. We sug-
gest enlarging the training dataset and adding further picture augmentations to
improve the model’s performance. The algorithm may be trained to recognize
more variants if the training data contains a wider variety of instances of mon-
keypox. Furthermore, our results demonstrate the potential of computer vision-
based models like MobileNetV2 for locating cases of monkeypox. The model’s
accuracy and robustness might yet be improved, especially when it comes to
instances where Monkeypox imitates other illnesses.
 Enhancing Monkeypox Disease Detection 355

7 Conclusion
This research embarks on a mission to enhance our capabilities in identify-
ing monkeypoxes through the fusion of image processing and deep learning
approaches. Imagine a scenario where technology plays a pivotal role in early
diagnosis, potentially saving lives and minimizing the impact of this ailment on
affected communities. Our project aims to evaluate the effectiveness of cutting-
edge technologies, including computer vision-based approaches and CNN-based
models. We’re delving into a diverse array of algorithms - Googlenet, Resnet50,
VGG16, VGG19, Darknet53, Mobilenetv2, Inceptionv3, Inceptionresnetv2, and
Yolov7 - to understand how well they can spot instances of monkeypox. A star
performer in our evaluation is Mobilenetv2, boasting an exceptional validation
accuracy of 95.2%. This not only underscores its effectiveness in pinpointing
and identifying the disease but also raises hopes for a tech-driven leap in dis-
ease detection. We’re not just stopping at monkeypox; our study recognizes the
broader implications for monkeypox disease diagnosis, emphasizing the signifi-
cance of deep learning optimizers in refining image classification outcomes.

References
1. Altun, M., Gürüler, H., Özkaraca, O., Khan, F., Khan, J., Lee, Y.: Monkeypox
detection using CNN with transfer learning. Sensors 23(4), 1783 (2023)
2. Jaradat, A.S., et al.: Automated Monkeypox skin lesion detection using deep learn-
ing and transfer learning techniques. Int. J. Environ. Res. Public Health 20(5), 4422
(2023)
3. Ahsan, M.M., Uddin, M.R., Farjana, M., Sakib, A.N., Momin, K.A., Luna,
S.A.: The image data collection and implementation of the deep learning-based
model in detecting Monkeypox disease using modified VGG16. arXiv preprint
arXiv:2206.01862 (2022)
4. Chadaga, K., et al.: Application of artificial intelligence techniques for Monkeypox:
a systematic review. Diagnostics 13(5), 824 (2023)
5. Almufareh, M.F., Tehsin, S., Humayun, M., Kausar, S.: A transfer learning app-
roach for clinical detection support of Monkeypox skin lesions. Diagnostics 13(8),
1503 (2023)
6. Alakus, T.B., Baykara, M.: Comparison of Monkeypox and wart DNA sequences
with deep learning model. Appl. Sci. 12(20), 10216 (2022)
7. Ahsan, M.M., et al.: Deep transfer learning approaches for Monkeypox disease
diagnosis. Expert Syst. Appl. 216, 119483 (2023)
8. Abdelhamid, A.A., et al.: Classification of Monkeypox images based on trans-
fer learning and the Al-Biruni earth radius optimization algorithm. Mathematics
10(19), 3614 (2022)
9. Uzun Ozsahin, D., et al.: Computer-aided detection and classification of Mon-
keypox and chickenpox lesion in human subjects using deep learning framework.
Diagnostics 13(2), 292 (2023)
10. Marin, M., Lopez, A.: Varicella (Chickenpox). https://wwwnc.cdc.gov/travel/
yellowbook/2020/travel-related-infectious-diseases/varicella-chickenpox. Accessed
14 Oct 2022
356 I. Ahmed et al.

11. Abed Alah, M., Abdeen, S., Tayar, E., Bougmiza, I.: The story behind the first
few cases of Monkeypox infection in non-endemic countries, 2022. J. Infect. Public
Health 15, 970–974 (2022)
12. Ali, S.N., et al.: Monkeypox skin lesion detection using deep learning models. A
preliminary feasibility study. arXiv arXiv:2207.03342 (2022)
13. Chickenpox Stock Photos, Pictures & Royalty-Free Images-iStock. https://www.
istockphoto.com/search/2/image?page=15&phrase=chickenpox. Accessed 14 Oct
2022
14. Nafisa Ali, S., et al.: Monkeypox skin lesion detection using deep learning models:
a feasibility study. arXiv, 13. https://arxiv.org/pdf/2207.03342.pdf. Accessed 24
Aug 2022
 Tomato Leaf Disease Prediction Based on Deep
Learning Techniques

Anirudh Singh , Satyam Kumar , and Deepjyoti Choudhury(B)

Department of AIML, School of CSE, Manipal University Jaipur, Dehmi Kalan, Near GVK Toll
Plaza, Jaipur 303007, Rajasthan, India
[email protected]

Abstract. Tomatoes, globally cherished and pivotal to diverse cuisines, face sub-
stantial threats to their quality and quantity due to various diseases. This research
paper applies a deep learning approach for the precise detection of tomato leaf
diseases. Our methodology incorporates a range of classifiers, including Random
Forest (R.F), Inception V3, DenseNet, ResNet50, Xception, and MobileNet. Our
results showcase the proficiency of these classifiers in distinguishing among nine
distinct disease classes and one healthy class. Random Forest classifier exhibits an
accuracy of 68.00%. Inception V3 excels with an impressive accuracy of 97%, cou-
pled with high precision (98%), recall (96%), and an overall F1-Score of 97.50%.
DenseNet demonstrates robust performance with an accuracy of 94.00%, preci-
sion of 98%, recall of 89%, and an F1-Score of 93%. ResNet50 closely follows
with an accuracy of 93.30%, precision of 97%, recall of 91%, and an F1-Score of
94%. Xception maintains a well-balanced performance, achieving an accuracy of
95%, precision of 93%, recall of 99%, and F1-Score of 96%. Lastly, MobileNet,
while achieving a moderate accuracy of 76%, demonstrates precision, recall, and
F1-Score of 66%, 90%, and 76%, respectively. These experimental findings not
only underscore the varied performance of each classifier concerning different
disease classes but also emphasize the significant potential of deep learning to
enhance the accuracy and efficiency of tomato leaf disease detection. This contri-
bution supports ongoing endeavors to improve crop yield and quality through the
application of advanced machine learning techniques in agriculture.

Keywords: disease · prediction · tomato · leaf · agriculture

1 Introduction
India, ranking as the third-largest global producer and exporter of tomatoes, commands a
vast agricultural landscape spanning over 3,50,000 hectares, yielding a substantial annual
harvest of approximately 53,00,000 tons [1]. The significance of tomatoes in the coun-
try’s agricultural milieu is not only measured by sheer volume but also by the economic
and nutritional impact they impart. However, the grandeur of this vital crop is persistently
challenged by the looming specter of diseases that can exert a profound influence on
both the quantity and quality of tomato yields. These diseases, adept at traversing from
leaves to fruits, necessitate a vigilant and preemptive approach to ensure early detection

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 357–375, 2025.
https://doi.org/10.1007/978-3-031-71481-8_28
358 A. Singh et al.

on the plant’s foliage. Traditional disease detection methods, relying on manual obser-
vation by expert teams, present inherent challenges, manifesting as labor-intensive and
time-consuming processes that often lead to delayed detection and response [2]. Rec-
ognizing the imperative for a more efficient and timely approach, researchers have been
fervently engaged in the development of automated methods. These approaches, which
include advanced methods like deep learning, machine learning, and digital image pro-
cessing, have emerged as game-changing instruments with the potential to completely
change how diseases are detected in agriculture. Beyond streamlining detection pro-
cesses, these technological advancements hold the promise of reducing pesticide usage,
thereby fostering sustainable agricultural practices and elevating overall crop production
and quality. Against this backdrop, our research endeavors to address the critical chal-
lenge of predicting tomato leaf diseases through the innovative application of machine
learning techniques. Specifically, Convolutional Neural Network (CNN) models, with
ResNet (Residual Net), Inception V3, DenseNet, Xception, and MobileNet as prominent
exemplars, have been strategically employed. The integration of machine learning into
disease prediction protocols is motivated by its inherent ability to efficiently analyze
extensive datasets, extracting intricate patterns that may elude manual inspection.
Central to our study is a meticulously curated dataset of tomato leaf images sourced
from the Plant Village Dataset. This comprehensive dataset encompasses six distinct
disease categories afflicting tomato leaves: Bacterial Spot, Late Blight, Leaf Mold, Sep-
toria Leaf Spot, Tomato Mosaic, Yellow Curved, in addition to a category representing
healthy leaves [3]. Our approach pivots on the strategic utilization of a CNN model,
engineered to meticulously scrutinize these images with the overarching goal of not
only classifying but predicting the presence of diseases with a high degree of accuracy.
The selected CNN models, ResNet, Inception V3, DenseNet, Xception, and MobileNet,
have proven to be robust in the context of image classification and have demonstrated
notable success in various applications, including plant disease detection. These models
leverage the power of deep learning to decipher complex patterns within the dataset,
enabling accurate and efficient disease identification.
The subsequent sections of this research paper unfold as a journey through the
nuances of our methodology and the intricacies of our findings. Section 2 initiates the
discourse with an in-depth background study, providing a nuanced understanding of
the prevailing challenges in tomato cultivation, considering both the historical and con-
temporary perspectives. Section 3 meticulously delineates the proposed methodology,
offering a comprehensive exploration of the theoretical underpinnings and practical
considerations guiding our approach towards disease prediction. Section 4 delves into
the nuanced experimental settings and implementation strategies, shedding light on the
datasets employed, the intricacies of model development, and the deliberate choices
made in our research design. In conclusion, Sect. 5 provides a thorough and in-depth
examination of the outcomes obtained from our research.
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 359

2 Related Work

Pivotal contribution to the literature came from Thangaraj et al. [4], who conducted a
comprehensive review and discussion on artificial intelligence in tomato leaf disease
detection. In 2019, Kaur and Bhatia [5] proposed an improved method for disease detec-
tion and classification, contributing to the precision agriculture landscape. The subse-
quent year witnessed Kibriya et al. [6] extending the capabilities of CNNs, further show-
casing their effectiveness in disease detection. Chowdhury et al. [7], in 2021, focused
on deep learning techniques for tomato leaf disease detection, emphasizing the applica-
bility of these methods in agriculture. This work demonstrated a pivotal shift towards
leveraging advanced technologies for automating disease identification processes in the
agricultural sector.
Hasan et al. [8] explored precision farming and introduced transfer learning to
enhance tomato leaf disease detection. Kaushik et al. [9], in 2020, contributed to the
field by emphasizing the importance of data augmentation for improved disease detec-
tion. Durmuş et al. [10] proposed a disease detection system on tomato leaves using
deep learning, underscoring the importance of advanced technologies in agriculture. The
subsequent year witnessed Karthik et al. [11] introducing attention-embedded residual
CNNs for disease detection in tomato leaves, demonstrating the potential of attention
mechanisms in refining model accuracy.
Anandhakrishnan and Jaisakthi [12] delved into image-based disease detection, con-
tributing to sustainable agriculture. Their work highlighted the continuous exploration
of innovative technologies for addressing agricultural challenges. Chowdhury et al. [13]
explored automatic and reliable leaf disease detection using deep learning techniques,
showcasing the ongoing efforts to automate and streamline disease identification pro-
cesses. David et al. [14] conducted a literature review in 2021, offering a comprehensive
overview of disease detection in tomato leaves using deep learning.
Mim et al. [15] explored image processing techniques in 2019, presenting an alter-
native avenue for disease detection. Their work contributed to the diversification of
methodologies employed in tomato leaf disease detection. Irmak and Saygili [16]
extended the use of CNNs for tomato leaf disease detection, contributing to the ongoing
advancements. Their work showcased the continued exploration and refinement of deep
learning techniques for agriculture applications. The application of transfer learning
for mobile applications was introduced by Bir et al. [17] in 2020, demonstrating the
adaptability of advanced techniques to diverse contexts. Rahman et al. [18] proposed an
image processing-based system for detection, identification, and treatment of tomato leaf
diseases in 2023, showcasing the potential for holistic solutions in disease management.
360 A. Singh et al.

Nagamani and Sarojadevi [19] explored deep learning techniques for disease detec-
tion, further diversifying the methodologies. Their work highlighted the ongoing explo-
ration of innovative technologies to improve disease identification accuracy. Hong et al.
[20] focused on disease detection and classification in 2020, emphasizing the role of deep
learning in addressing complex agricultural challenges. Zhao et al. [21] introduced an
improved CNN with an attention module for disease diagnosis, showcasing the ongoing
efforts to enhance model architectures. Batool et al. [22] extended the classification and
identification of tomato leaf diseases using deep neural networks. Their work contributed
to the ongoing exploration of neural network architectures for disease detection.
Zhang et al. [23] worked on improving object detection for tomato diseases using
deep learning in 2020, showcasing the potential for enhancing disease identification
through advanced computer vision techniques. Wu et al. [24] explored tomato leaf disease
identification based on deep CNNs in 2021, contributing to the refinement of deep
learning models for agricultural applications. Paymode et al. [25] contributed to CNN-
based disease detection, showcasing the continued emphasis on leveraging convolutional
neural networks for accurate disease identification. Alhaj Ali et al. [26] delved into deep
learning for tomato leaf diseases, contributing to the ongoing efforts to improve disease
detection accuracy. Anandhakrishnan and Murugaiyan [27] focused on pretrained deep
CNN models for disease detection, showcasing the potential for leveraging pre-existing
knowledge in deep learning applications. Jasim and Al-Tuwaijari [28] explored plant leaf
diseases detection using image processing and deep learning techniques, contributing
to the diversification of methodologies. In 2021, Kodali and Gudala [29] specifically
addressed tomato leaf disease detection using CNNs, providing a targeted approach to
disease identification. Their work contributes to the specificity required in agricultural
applications.
Adhikari et al. [30] developed a system for tomato plant diseases detection, showcas-
ing the application of machine learning in addressing broader agricultural challenges.
Gehlot et al. [31] introduced the “Tomato-Village” dataset in 2023, providing a real-
world environment for end-to-end tomato disease detection. This dataset serves as a
valuable resource for benchmarking and advancing the development of robust disease
detection models. This comprehensive review encapsulates the evolution of tomato leaf
disease detection, highlighting the diverse methodologies and approaches that collec-
tively contribute to the advancement of precision agriculture. The continuous explo-
ration of advanced technologies and the development of innovative datasets underscore
the dynamic nature of research in this field.
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 361

3 Design and Implementation

The workflow diagram is shown in Fig. 1.

Fig. 1. Workflow diagram of our working model

3.1 Dataset Description

These datasets collectively form the foundation for our research, enabling comprehensive
evaluation and comparison of deep learning models for.
• Benchmark Dataset (Kaggle): The Benchmark dataset serves as the primary dataset
for evaluating the performance of our machine learning models in tomato leaf disease
detection. The images in this dataset come from two different sources: the Tomato-
Village dataset from Kaggle and the Tomato Leaf Disease dataset, resulting in a
significant and diverse collection of data for robust model testing.
362 A. Singh et al.

• Tomato-Village Dataset (Kaggle): This dataset emerges in response to the critical

demand for authentic and varied datasets reflecting real-world scenarios of tomato
diseases. Notably, existing datasets, notably Plant Village, predominantly originate
from controlled laboratory environments, constraining their relevance in addressing
real-world challenges. Noteworthy for its focus on diseases prevalent in the Jodhpur
and Jaipur districts of Rajasthan, India, including Leaf Miner, spotted wilt virus, and
nutrition deficiency diseases.
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 363

• Tomato Leaf Disease (Kaggle): This dataset focuses on the detection of diseases in
tomato plant leaves using deep learning techniques. With a specific emphasis on the
diverse diseases that afflict plant leaves, the project aims to create models capable
of accurately identifying and classifying these diseases. Deep learning algorithms
are applied to train models for disease detection. Additionally, the dataset explores
the impact of transfer learning on disease detection, providing insights into the effi-
cacy of leveraging pre-trained models for improved accuracy in tomato leaf disease
identification.
364 A. Singh et al.

3.2 Data Vectorization and Feature Selection

3.2.1 Data Vectorization

The process of data vectorization for the purpose of predicting tomato leaf disease
through deep learning begins with the gathering of tomato leaf images. Utilizing the
OpenCV library, images are loaded, and a structured Data Frame is constructed to manage
file paths and corresponding labels. To augment the dataset’s diversity and improve
model generalization, an ImageDataGenerator is employed, applying transformations
such as horizontal flipping, rotation, and shift during training. This not only introduces
variability but also prepares the model for real-world scenarios where leaves may exhibit
different orientations.
The dimensions of the images play a crucial role in understanding the dataset’s
characteristics. A subset of training images is randomly selected, and their average
height, width, and aspect ratio are calculated. This information provides insights into
the inherent variability of leaf sizes and shapes within the dataset. The final step in data
vectorization involves configuring data generators for training, validation, and testing,
ensuring consistent image sizes (256 × 256 pixels) and appropriate batch sizes (30).
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 365

These generators efficiently handle the flow of data to the model during training and
evaluation, setting the stage for the subsequent machine learning process.

3.2.2 Feature Selection

Feature selection is a critical aspect of building an effective tomato leaf disease pre-
diction model. In this research, a strategic trimming process is implemented to balance
the number of samples per class. The ‘trim’ function ensures that each class is rep-
resented by a specified number of images (both maximum and minimum), promoting
model robustness and preventing bias towards classes with larger sample sizes. Con-
volutional Neural Network (CNN) architecture, specifically the EfficientNetB3 model,
forms the backbone of feature extraction. During training, the CNN layers automatically
learn to discern relevant patterns, textures, and local spatial information from the input
tomato leaf images. The subsequent layers, including Batch Normalization, Dense, and
Dropout layers, contribute to further feature refinement and model adaptability. The
chosen model, EfficientNetB3, is pre-trained on ImageNet, allowing it to capture both
low-level details and high-level features during training. The fine-tuning process aligns
the learned features with the specifics of tomato leaf disease patterns. The model’s ability
to automatically select discriminative features is enhanced through regularization tech-
niques, such as L1 and L2 regularization in the Dense layer. In summary, the combination
of meticulous data vectorization and thoughtful feature selection techniques forms the
foundation for a robust machine learning model capable of accurately predicting tomato
leaf diseases. The insights gained from image dimensions and the pre-trained Efficient-
NetB3 model contribute to a holistic understanding of the dataset, paving the way for
effective disease classification.

3.3 Machine Learning Models

This research employs a comprehensive suite of state-of-the-art models, including Ran-
dom Forest (RF), Inception V3, DenseNet, ResNet50, Xception, and MobileNet, to
develop a robust and versatile predictive system. Each model brings its unique strengths
and characteristics to the forefront, contributing to the overall efficacy of disease
classification.
• Random Forest (RF):
As an ensemble learning method, RF excels in handling high-dimensional data with
complex relationships. It leverages a multitude of decision trees, each trained on a ran-
dom subset of the data. The collective wisdom of these trees culminates in a powerful
predictive model capable of capturing intricate patterns and nuances present in the tomato
leaf images. RF’s ability to handle diverse feature sets makes it a valuable component
in the ensemble of classifiers.
• Inception v3:
This CNN architecture is renowned for its depth and efficiency in extracting hierar-
chical features. Inception V3 introduces inception modules that enable the simultaneous
processing of multiple receptive fields. This results in the model’s capability to capture
366 A. Singh et al.

both local and global features, making it well-suited for discerning the intricate patterns
indicative of various tomato leaf diseases.
• DenseNet:
Densely Connected Convolutional Networks, or DenseNet, introduces a unique con-
nectivity pattern where each layer receives direct input from all preceding layers. This
architecture promotes feature reuse and enhances gradient flow throughout the network.
In the context of tomato leaf disease prediction, DenseNet’s dense connections enable
effective information propagation.
• ResNet50:
Residual Networks, or ResNets, address the challenge of training deep networks by
introducing residual connections. ResNet50, a variant of ResNet, comprises 50 layers
and is renowned for its depth. This architecture facilitates the training of deeper models
without succumbing to issues like vanishing gradients. ResNet50’s skip connections
enhance the flow of information, enabling the model to learn intricate disease-related
features.
• Xception:
Inspired by the idea of extreme depth, Xception (Extreme Inception) explores a novel
architecture by replacing the standard inception modules with depthwise separable con-
volutions. This modification significantly reduces the number of parameters, making the
model more computationally efficient. In the context of tomato leaf disease prediction,
Xception’s depthwise separable convolutions contribute to the model’s ability to capture
detailed spatial information efficiently.
• MobileNet:
In the realm of tomato leaf disease prediction, MobileNet’s lightweight design allows
for efficient inference without compromising on classification accuracy.

3.4 Confusion Matrix and Accuracy Curve

• Confusion matrix:

The Confusion Matrix is a tabular representation that delineates the classifier’s pre-
dictions against the ground truth labels. The four quadrants of the matrix are populated
with values indicating true positives, true negatives, false positives, and false negatives.
– True Positives (TP): Instances where the classifier correctly predicts the presence of
a specific disease.
– True Negatives (TN): Instances correctly identified as not having the predicted
disease.
– False Positives (FP): Instances where the classifier incorrectly predicts the presence
of a disease (Type I error).
– False Negatives (FN): Instances incorrectly identified as not having the predicted
disease (Type II error).
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 367

In our research on tomato leaf disease prediction utilizing advanced machine learning
classifiers, the Confusion Matrix and Accuracy Curve play pivotal roles in assessing and
interpreting the performance of our models.
• Accuracy Curve:
The Accuracy Curve is a graphical representation of the classifier’s performance
across different epochs or iterations during the training process. In our research, we
employ machine learning models such as Random Forest, Inception V3, DenseNet,
ResNet50, Xception, and MobileNet. The Accuracy Curve allows us to track how well
these models are learning and adapting over time. Training Accuracy: Represents the
accuracy of the model on the training dataset. Validation Accuracy: Reflects the accuracy
of the model on a separate validation dataset not used for training. The trajectory of the
Accuracy Curve provides valuable insights into the learning dynamics of our models.
Observing steady increases in both training and validation accuracy indicates effective
learning and generalization. However, significant gaps between training and valida-
tion accuracy curves could signal overfitting, necessitating adjustments to the model
architecture or training strategy.

3.5 Evaluation Parameters

• Accuracy:

Accuracy assesses the correctness of predictions by comparing the sum of true

positives and true negatives to the total instances.

Accuracy = (TP + TN)/(TP + TN + FP + FN) (1)

• Precision:

Precision gauges the precision of positive predictions, the formula is given below:

Precision = TP/(TP + FP) (2)

• Recall:

Recall measures the ability to capture positive instances, The formula is given below:

Recall = TP/(TP + FN) (3)

• F1-Score:

The F1-score offers a balanced assessment, The formula is given below:

F1-Score = 2 ∗ (Precision ∗ Recall)/(Precision + Recall) (4)

368 A. Singh et al.

Table 1. Benchmark Dataset

Classifier Accuracy (%) Precision (%) Recall (%) F1-Score (%)

R.F 68 64 71 68
Inception v3 97 98 96 97.5
DenseNet 94 98 89 93
ResNet50 93.3 97 91 94
Xception 95 93 99 96
MobileNet 76 66 90 76

4 Results Analysis

Across all datasets, Inception V3 consistently exhibits outstanding performance, boast-

ing high accuracy, precision, recall, and F1-Score. DenseNet also demonstrates com-
mendable results, particularly excelling in accuracy and precision. On the other hand,
Random Forest and MobileNet exhibit varying degrees of performance across datasets,
indicating sensitivity to dataset characteristics (Table 1).
The analysis of the BenchMark Dataset for Tomato Leaf Disease Detection from
Kaggle reveals distinct performance characteristics among various classifiers, each eval-
uated based on key metrics such as Accuracy, Precision, Recall, and F1-Score. In this
dataset, Inception V3 emerges as the standout performer, showcasing exceptional accu-
racy at 97%, precise disease identification with a precision of 98%, robust recall at
96%, and an overall F1-Score of 97.50%. DenseNet also demonstrates commendable
results with a high accuracy of 94%, noteworthy precision of 98%, and balanced recall
at 89%, resulting in an impressive F1-Score of 93%. ResNet50 and Xception exhibit a
balanced performance, with ResNet50 achieving accuracy of 93.30% while Xception
attains 95% accuracy, 93% precision, and an impressive 99% recall. The Random Forest
(R.F) model, with a moderate accuracy of 68%, shows balanced precision and recall,
indicating its competence in disease detection. However, MobileNet, while achieving a
reasonable accuracy of 76%, demonstrates sensitivity to certain dataset characteristics
with comparatively lower precision (66%) and high recall (90%). This comprehensive
analysis underscores the efficacy of deep learning models, particularly Inception V3 and
DenseNet, in accurately identifying tomato leaf diseases, and emphasizes the need for
careful model selection based on dataset characteristics to optimize disease detection
outcomes. The Confusion Matrix and Accuracy Curves for Inception V3 model is given
below (Fig. 2 and Table 2).
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 369

Fig. 2. Confusion Matrix and Accuracy Curve on Benchmark Dataset

Random Forest, displaying a moderate accuracy of 61%, demonstrates a balanced

performance in precision and recall, suggesting its capability for correct predictions.
Inception V3 emerges as a robust performer with an accuracy of 69%, demonstrating
balanced precision and recall, highlighting its reliability in accurately identifying and
classifying tomato leaf diseases. DenseNet achieves commendable accuracy at 74%,
primarily driven by its high precision rate, while the slightly lower recall indicates
370 A. Singh et al.

Table 2. Kaggle Dataset (Tomato Village)

Classifier Accuracy (%) Precision (%) Recall (%) F1-Score (%)

R.F 61 57 73 64
Inception v3 69 69 75 72
DenseNet 74 80 69 74
ResNet50 64 61 65 64
Xception 70 66 80 72
MobileNet 63 80 52 63

potential challenges in capturing all instances of leaf diseases. ResNet50 exhibits a bal-
anced and consistent performance with an accuracy of 64%, showcasing reliability in
disease detection scenarios. Xception displays a balanced triad of accuracy, precision,
and recall at 70%, 66%, and 80%, respectively, showcasing its versatility in effectively
identifying tomato leaf diseases with a nuanced understanding of the dataset’s intricacies.
MobileNet, with an accuracy of 63%, shows sensitivity, particularly evident in preci-
sion and recall, implying potential challenges in correctly identifying specific disease
characteristics within the Tomato Village dataset. In summary, the results underscore
the nuanced performance of different classifiers in the context of tomato leaf disease
detection. Inception V3 and DenseNet emerge as reliable choices, while MobileNet’s
sensitivity emphasizes the crucial need for meticulous model selection and parameter
tuning to address dataset-specific challenges. These findings provide valuable insights
for refining machine learning models and enhancing disease detection capabilities in
diverse agricultural settings. The Confusion Matrix and Accuracy Curves for DenseNet
Model is given below (Fig. 3 and Table 3).
Random Forest, with an accuracy of 83%, demonstrates a robust performance, bal-
ancing precision, recall, and F1-Score at 82%, 84%, and 83%, respectively. Inception V3
emerges as a standout performer with an impressive accuracy of 97.5%, showcasing high
precision, recall, and F1-Score at 96%, 100%, and 98%, respectively. DenseNet achieves
near-perfect accuracy at 99%, with exceptional precision, recall, and F1-Score all at 98%
or higher, indicating its capability to effectively identify and classify tomato leaf dis-
eases. ResNet50 maintains a high level of accuracy at 95%. Xception, with an accuracy
of 98.6%, demonstrates exceptional precision and recall at 99.42% and 98.92%, respec-
tively, resulting in a robust F1-Score of 98.68%. MobileNet, with an accuracy of 88%,
exhibits balanced precision, recall, and F1-Score metrics, all hovering around 87–89%.
Inception V3, DenseNet, and Xception emerge as top-performing models, showcas-
ing their potential for accurate identification and classification of tomato leaf diseases.
These findings contribute valuable insights for the selection of classifiers based on spe-
cific requirements and performance expectations in agricultural settings. The Confusion
Matrix and Accuracy Curve for DenseNet Model is Given Below (Fig. 4).
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 371

Fig. 3. Confusion Matrix and Accuracy Curve on Kaggle Dataset (Tomato Village)

Table 3. Kaggle Dataset (Tomato Leaf Disease)

Classifier Accuracy (%) Precision (%) Recall (%) F1-Score (%)

R.F 83 82 84 83
Inception v3 97.5 96 100 98
(continued)
372 A. Singh et al.

Table 3. (continued)

Classifier Accuracy (%) Precision (%) Recall (%) F1-Score (%)

DenseNet 99 98 100 99
ResNet50 95 93 97 95
Xception 98.6 99.42 98.92 98.68
MobileNet 88 87 89 88

Fig. 4. Confusion Matrix and Accuracy Curve on Kaggle Dataset (Tomato Leaf Disease)
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 373

5 Conclusion and Future Work

In the culmination of our extensive exploration into the realm of tomato leaf disease
prediction using deep learning, a thorough analysis of diverse classifiers – including
Random Forest, Inception V3, DenseNet, ResNet50, Xception, and MobileNet – across
three distinct datasets has provided nuanced insights. The BenchMark Dataset (Tomato
Leaf Disease Detection Kaggle), Kaggle Dataset (Tomato Village), and Kaggle Dataset
(Tomato Leaf Disease) have served as robust testing grounds, enabling a meticulous
examination of the classifiers’ performances.The discerning evaluation has illuminated
the efficacy of specific models, with Inception V3, DenseNet, and Xception emerging as
stalwart contenders, showcasing high accuracy and commendable balance in precision,
recall, and F1-Score. The utilization of sophisticated machine learning methods, specif-
ically deep learning models, holds potential to improve the precision and effectiveness
of disease identification in tomato plants. Yet, the dynamic landscape of agricultural
technology demands continuous refinement. Subsequent endeavors should be dedicated
to fine-tuning and optimizing the proposed models, exploring the potential synergies
offered by ensemble methods, and delving into transfer learning approaches to enhance
model generalization across diverse datasets.
The scalability and real-world applicability of these models should be subject to thor-
ough investigation, considering the practical constraints and exigencies of agricultural
settings. Moreover, the integration of cutting-edge technologies such as edge computing
and the Internet of Things (IoT) stands as a promising avenue for crafting more effi-
cient and real-time disease detection systems tailored for tomato crops. Collaborative
efforts involving domain experts, agronomists, and farmers will be pivotal in refining
models to align seamlessly with the intricate needs of agricultural practices. The con-
tinuous expansion and diversification of annotated datasets, incorporating geographical
variations and diverse weather conditions, will contribute substantially to fortifying the
robustness and generalization capabilities of machine learning models. In summation,
this research not only establishes a foundation for the application of machine learning
in tomato leaf disease prediction but also underscores the potential of advanced models
for precise and timely detection. The identified directions for future research encapsu-
late exciting prospects, promising to elevate the impact of these models on agricultural
practices.

References
1. Agarwal, M., et al.: ToLeD: tomato leaf disease detection using convolution neural network.
Proc. Comput. Sci. 167, 293–301 (2020)
2. Tm, P., et al.: Tomato leaf disease detection using convolutional neural networks. In: 2018
Eleventh International Conference on Contemporary Computing (IC3). IEEE (2018)
3. Ashok, S., et al.: Tomato leaf disease detection using deep learning techniques. In: 2020 5th
International Conference on Communication and Electronics Systems (ICCES). IEEE (2020)
4. Thangaraj, R., et al.: Artificial intelligence in tomato leaf disease detection: a comprehensive
review and discussion. J. Plant Dis. Prot. 129(3), 469–488 (2022)
5. Kaur, M., Bhatia, R.: Development of an improved tomato leaf disease detection and classifi-
cation method. In: 2019 IEEE Conference on Information and Communication Technology.
IEEE (2019)
374 A. Singh et al.

6. Kibriya, H., et al.: Tomato leaf disease detection using convolution neural network. In: 2021
International Bhurban Conference on Applied Sciences and Technologies (IBCAST). IEEE
(2021)
7. Chowdhury, M.E.H., et al.: Tomato leaf diseases detection using deep learning technique.
Technol. Agric. 453 (2021)
8. Hasan, M., Tanawala, B., Patel, K.J.: Deep learning precision farming: tomato leaf disease
detection by transfer learning. In: Proceedings of 2nd International Conference on Advanced
Computing and Software Engineering (ICACSE) (2019)
9. Kaushik, M., et al.: Tomato leaf disease detection using convolutional neural network with
data augmentation. In: 2020 5th International Conference on Communication and Electronics
Systems (ICCES). IEEE (2020)
10. Durmuş, H., Güneş, E.O., Kırcı, M.: Disease detection on the leaves of the tomato plants by
using deep learning. In: 2017 6th International Conference on Agro-Geoinformatics. IEEE
(2017)
11. Karthik, R., et al.: Attention embedded residual CNN for disease detection in tomato leaves.
Appl. Soft Comput. 86, 105933 (2020)
12. Anandhakrishnan, T., Jaisakthi, S.M.: Deep Convolutional Neural Networks for image based
tomato leaf disease detection. Sustain. Chem. Pharm. 30, 100793 (2022)
13. Chowdhury, M.E.H., et al.: Automatic and reliable leaf disease detection using deep learning
techniques. AgriEngineering 3(2), 294–312 (2021)
14. David, H.E., et al.: Literature review of disease detection in tomato leaf using deep
learning techniques. In: 2021 7th International Conference on Advanced Computing and
Communication Systems (ICACCS), vol. 1. IEEE (2021)
15. Mim, T.T., et al.: Leaves diseases detection of tomato using image processing. In: 2019 8th
International Conference System Modeling and Advancement in Research Trends (SMART).
IEEE (2019)
16. Irmak, G., Saygili, A.: Tomato leaf disease detection and classification using convolutional
neural networks. In: 2020 Innovations in Intelligent Systems and Applications Conference
(ASYU). IEEE (2020)
17. Bir, P., Kumar, R., Singh, G.: Transfer learning based tomato leaf disease detection for
mobile applications. In: 2020 IEEE International Conference on Computing, Power and
Communication Technologies (GUCON). IEEE (2020)
18. Rahman, S.U., et al.: Image processing based system for the detection, identification and
treatment of tomato leaf diseases. Multimed. Tools Appl. 82(6), 9431–9445 (2023)
19. Nagamani, H.S., Sarojadevi, H.: Tomato leaf disease detection using deep learning techniques.
Int. J. Adv. Comput. Sci. Appl. 13(1) (2022)
20. Hong, H., Lin, J., Huang, F.: Tomato disease detection and classification by deep learning.
In: 2020 International Conference on Big Data, Artificial Intelligence and Internet of Things
Engineering (ICBAIE). IEEE (2020)
21. Zhao, S., et al.: Tomato leaf disease diagnosis based on improved convolution neural network
by attention module. Agriculture 11(7), 651 (2021)
22. Batool, A., et al.: Classification and identification of tomato leaf disease using deep neu-
ral network. In: 2020 International Conference on Engineering and Emerging Technologies
(ICEET). IEEE (2020)
23. Zhang, Y., Song, C., Zhang, D.: Deep learning-based object detection improvement for tomato
disease. IEEE Access 8, 56607–56614 (2020)
24. Wu, Y., Xu, L., Goodman, E.D.: Tomato leaf disease identification and detection based on
deep convolutional neural network. Intell. Autom. Soft Comput. 28(2) (2021)
25. Paymode, A.S., Magar, S.P., Malode, V.B.: Tomato leaf disease detection and classification
using convolution neural network. In: 2021 International Conference on Emerging Smart
Computing and Informatics (ESCI). IEEE (2021)
 Tomato Leaf Disease Prediction Based on Deep Learning Techniques 375

26. Alhaj Ali, A., Chramcov, B., Jasek, R., Katta, R., Krayem, S., Awwama, E.: Tomato leaf
diseases detection using deep learning. In: Silhavy, R., Silhavy, P., Prokopova, Z. (eds.)
CoMeSySo 2021. LNNS, vol. 231, pp. 199–208. Springer, Cham (2021). https://doi.org/
10.1007/978-3-030-90321-3_18
27. Anandhakrishnan, T., Murugaiyan, J.S.M.: Identification of tomato leaf disease detection
using pretrained deep convolutional neural network models. Scalable Comput.: Pract. Exp.
21(4), 625–635 (2020)
28. Jasim, M.A., Al-Tuwaijari, J.M.: Plant leaf diseases detection and classification using image
processing and deep learning techniques. In: 2020 International Conference on Computer
Science and Software Engineering (CSASE). IEEE (2020)
29. Kodali, R.K., Gudala, P.: Tomato plant leaf disease detection using CNN. In: 2021 IEEE 9th
Region 10 Humanitarian Technology Conference (R10-HTC). IEEE (2021)
30. Adhikari, S., et al.: Tomato plant diseases detection system, vol. I, pp. 81–86 (2018)
31. Gehlot, M., Saxena, R.K., Gandhi, G.C.: “Tomato-Village”: a dataset for end-to-end tomato
disease detection in a real-world environment. Multimed. Syst. 1–24 (2023)
 Predicting Stock Market Trends: Machine
Learning Approaches of a Possible Uptrend
or Downtrend

Raima Joseph and Mausumi Goswami(B)

Department of Computer Science and Engineering, Christ (Deemed to Be) University, Kengeri,
Bangalore, Karnataka 560074, India
[email protected]

Abstract. This paper delves into a statistical analysis of the stock market, empha-
sizing the significance of accuracy in stock predictions. Large data sets can be han-
dled by machine learning algorithms, which can also forecast outcomes based on
past data and spot intricate patterns in financial data. They assist control risks, auto-
mate decision-making procedures, and adjust to changing circumstances. Multi-
source data can be combined by ML models to provide a comprehensive picture of
market circumstances. They can manage intricate, nonlinear interactions, provide
impartial analysis, and lessen human bias. Models are able to adjust to shifting
market conditions through ongoing learning and retraining. They must, however,
exercise caution when deploying models in real-world situations and ensure that
they are validated. Although machine learning has advantages for stock market
analysis, it must be carefully evaluated for dangers and validated before being used
in practical situations. The traditional machine learning model, Logistic Regres-
sion has been used in order to predict stock prices. It focuses on binary classifica-
tion based on the trend of the stock. Through the model training and evaluation and
additional analysis done on the results, this research contributes towards obtain-
ing predictions and studying reasons of a possible uptrend or downtrend to further
assist companies.

Keywords: Stock Market · Machine Learning · Logistic Regression

1 Introduction
Stock prices represent ownership or a part of a company bought, or rather invested
in, and sold in the finance market. In order to benefit from investing in stocks, many
researches and studies have shed a light on Machine Learning (ML) techniques to fore-
see stock prices. The performance of stock trends and prices is hard to predict as they
are uncertain. As a lot of factors incorporate accurate analysis and prediction of the
stock data, it is necessary to work forward in stock prediction using various efficient ML
techniques. Stock market (SM) analysis essentially refers to the process of interpreting
the factors affecting the performance of stock in order to make informed decisions. The

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 376–387, 2025.
https://doi.org/10.1007/978-3-031-71481-8_29
 Predicting Stock Market Trends: Machine Learning Approaches 377

impact of SM on the global business is profound in risk management for making strate-
gic choices when it comes to investing. It also creates investor confidence by creating
a well-analysed prediction. Stock price prediction consists a lot of scope for accuracy
including its challenges. As mentioned within the background, while a company’s finan-
cial performance is required for fundamental analysis, this paper focuses on technical
analysis using machine learning. The global financial stability increases as the analysis
of future status of the stocks can prevent financial crises.

Fig. 1. Applications of ML in SMA

Various papers suggest that, [1] there can be two main approaches to stock predic-
tion: Fundamental and technical analysis. Fundamental analysis includes evaluating a
company’s finances and its industry, while technical analysis relies on historical data
and predictive data to forecast price trends. While these methods have been traditionally
used, there are limitations. The data and the trends followed can be subjective which can
lead to different interpretations by different investors. Machine learning models can iden-
tify market inefficiencies that traditional analysis methods may not account for. Another
limitation is that handling large amounts of financial data manually can be very time
consuming and challenging. A subset of artificial intelligence, machine learning, uses
algorithms as models to analyse data. In the context of SM, machine learning concepts
can be applied in different factors. The main applications include predictive modelling,
risk assessment, algorithmic trading and sentiment analysis (see Fig. 1). Here, we focus
on predictive modelling [2]. Various techniques in machine learning include:
1. Traditional ML analysis: Algorithms such as regression or classification when given
sequential data, such as that of stock prices. These models prove to be efficient in
terms of performance.
2. Neural-Network based analysis: Deep learning models that capture intricate patterns
in price data.
3. Time-series based analysis: These models analyse the historical data and identify
trends over time.
378 R. Joseph and M. Goswami

4. Graph-based analysis: Financial networks and creating relationships between stock

prices and technical indicators in order to branch out options.
The main idea would be to analyse historical data of each company by its prices and
identify patterns in order to predict what the trend would be the next day. The question
this paper highlights is: Will tomorrow’s price fall or rise?
Using big data techniques, stock market prediction combines fundamental and tech-
nical analysis. Whereas fundamental analysis employs sentiment analysis to analyze
social media data, technical analysis uses machine learning to analyze historical stock
price data [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]. Online user-generated textual material
predicts stock prices and offers insightful analysis of investor psychology. Support vec-
tor machines are used to include sentiment analysis into machine learning techniques.
Few studies use deep neural networks or other advanced machine learning techniques
to forecast daily stock market returns. Support vector machines and artificial neural net-
works are the most often utilized algorithms for stock market prediction, according to
a systematic analysis of 122 academic publications. For predicting the bitcoin market,
a hybrid strategy combining natural language processing, deep learning, and machine
learning is suggested group methods such as bagging, boosting.

2 Background
Accurate stock price predictions are invaluable for investors and financial organisations,
as they can help make investment decisions, manage risks, and come up with better
strategies. Firstly, we have some terms related to analysis and machine learning. Stock
market indicators are some data points derived from historical data of the company. An
example of such indicators is Moving Average, a mean price value of a particular set of
days. These are very useful for identifying patterns in trends. Historical data refers to
the past stock prices, volumes traded, opening and closing prices and the high as well as
low prices of the stock for a specific period. This data serves as the foundation dataset
for analysing trends. Machine learning models are trained using this very historical data
(Fig. 2).
The methodology brings out a broader understanding of the features or factors used
in the training of the ML algorithms. The algorithms that can be used for SM analysis
include Regression, Classification and Deep Learning models. Regressive models bring
out a relationship between the input variables (indicators and features) and the target
variable (stock prices). Classification prediction involves categorising data into groups
or classes. For SM, the classification required is the trend pattern. An uptrend means that
the price of the stock has been increasing steadily, and downtrend indicates a decline
in the trend of the stocks. This brings about the conclusion that the task would be a
binary classification [3, 4]; uptrend or downtrend. Hence, we choose traditional ML
analysis by using logistic regression to identify the relationship between input features
and the days to predict an uptrend or a downtrend. In the context of SM analysis and
prediction, accuracy refers to the performance of the model and the ability for it to make
accurate forecasts, when compared to the actual prices. The accuracy of the predictive
model decides the validity of the predictions which impacts the development of strategy
highly. It also serves as a competitive advantage. Highly accurate predictions also help
 Predicting Stock Market Trends: Machine Learning Approaches 379

Fig. 2. Historical Quotes for AMZN from NASDAQ

in assessing the chances of loss or gain, hence risk management. We continue with
technical analysis and use Machine Learning to use the pointers and the past data to
predict prices. The companies chosen for this analysis are Microsoft [MSFT], Amazon
[AMZN], Google [GOOG], Apple [AAPL] and Tesla [TSLA] from NASDAQ Historical
quotes for Logistic Regression to determine next day predicted price trends.

3 Methodology

Data collection: Collected historical stock price data for Microsoft (MSFT), Google
(GOOG), Amazon (AMZN), Tesla (TSLA), and Apple (AAPL) from the NASDAQ
stock exchange. The dataset includes daily Open, High, Low, Volume, and Closing
(OHLCV) prices. The dataset size ranges from August 2018–2023. Training and testing
data is initially set to 80 and 20% respectively. The size of each dataset comes out to be
1005 and 251 each.

3.1 Features

These are common indicators in financial analysis and stand for:

1. Open: At the beginning of a trading session, the price at which a stock opens for
trading.
2. High: The highest price at which the stock was traded during a trading session.
3. Low: The lowest price at which the stock was traded during a trading session.
380 R. Joseph and M. Goswami

4. Close: The price at which a stock closes at the end of a trading session.
5. Volume: During a given period (market activity), the number of shares or contracts
traded.
These terms are fundamental in financial analysis and are widely used by traders,
investors, and analysts to assess market trends.

3.2 Feature Selection

1. Next-Day Close: This feature represents the next trading day close price. It is essen-
tially the value of the “Close” data with the next day values. It maps the data to ‘1’ if
the next day’s price is greater than the previous day, else it maps ‘0’.
2. Open-Close: This is the variance between the opening price and the closing price of
a stock during a trading period. It represents the change between the opening and
closing prices and can indicate pre-market activity, i.e., the few minutes right before
the trading session commences.
3. Daily Return: Daily return measures the change in the price percentage of a stock
from one trading day to the next. Calculated as the following equation.

((Close − Previous Close)/Previous Close) × 100 (1)

4. Open-Open: This feature represents the difference between the open price of a trading
day and the open price of the previous trading day. It can indicate how the market
opens compared to the previous day, reflecting financial market conditions within the
closed hours of trading sessions.
5. Moving Average: Moving averages are designed to classify the trend direction of a
stock. An increasing moving average show that the stock is in an uptrend, while a
falling moving average shows that it is in a downtrend. The “rolling mean” function
has been used to find the simple moving average over varying ranges.

3.3 Challenges

Relevance of Open-Open vs. Open-Close. The relevance of using Open-Open over

Open-Close was assessed by calculating trends along with close prices for a specific
range. Results showed that a higher difference in Open-Close results in a downtrend
of the close price, whereas the Open-Open gap fluctuates. Broadly, there is very little
difference between the features, and Open-Open does not prove to be very relevant
(Fig. 3).
Observations
Whenever there is a significant fluctuation with Open-Close, a similar spike or fall occurs
for the same duration in the return prices. Graphs for Open-Close and Open-Open are
similar. The presence of both features together does not seem to hinder the performance
of the model. This made it necessary to verify the relevance of each with respect to the
accuracy shift that occurs whenever one of the features is not present (Tables 1 and 2).
It is observed that the accuracy does not decline by a lot when Open-Open is omitted.
However, the accuracy significantly reduces with removal of Open-Close, indicating that
 Predicting Stock Market Trends: Machine Learning Approaches 381

Fig. 3. Open-Open vs. Open-Close with Returns

Table 1. Results without Open-Open

Open-Close is much more relevant than Open-Open. To generalize the data to relevant
features and inputs and reduce noise, Open-Open feature is removed. This also helps
to reduce the chances of over-fitting, where training accuracy is higher than testing
accuracy.
Appropriate Range for Moving Average. For calculating the simple moving average,
the rolling mean function is used to find the average close prices over a window of trading
days. The range of the window plays a crucial role in generating an apt moving average.
Depending on the type of trader, the time or range of the window can be adjusted. Long-
term traders prefer a weekly average and short-term traders would select a daily moving
average.
Results
382 R. Joseph and M. Goswami

Table 2. Results without Open-Close

Due to these results, a window of 5 days gives the highest accuracy, and hence, we move
forward with a window range or period within 5 days (Figs. 4, 5 and 6).

Fig. 4. Accuracy with a window size of 3 days: 88.44%

Predicting Stock Market Trends: Machine Learning Approaches 383

Fig. 5. Accuracy with a window size of 5 days: 88.87%

Fig. 6. Accuracy with a window size of 7 days: 88.4%

384 R. Joseph and M. Goswami

3.4 Logistic Regression

An ML algorithm that is used for binary classification tasks. For the context of SM
analysis and prediction, logistic regression can be utilised to forecast the uptrend or the
downtrend of the stock. It effectively makes a connection between the input and output
variables, in this case the features of the stock data-set and the predicted values, and
converts it into two classes; 0 or 1 (downtrend or uptrend).
A sigmoid function is used to reduce the outputs to 0 or 1 (see Fig. 7).
Here, ‘x’ is the linear combination calculated by the model as:
X = c0 + c1 ∗ feature1 + c2 ∗ feature2 + . . . + cN ∗ featureN

Feature = OHCLV + Added features

where c0, c1… cN are the coefficients that are learned by the logistic regression model.
Simply put, X here gives the linear regression output and this X value is used in the
above-mentioned Sigmoid function to find the binary outputs between 0 and 1. f(x)
mentioned gives the value between 0 and 1 with the threshold value as 0.5. This means
that the values above 0.5 are termed as uptrend (1) and below 0.5 are termed as downtrend
(0)

Fig. 7. Sigmoid function and plot.

3.5 Splitting and Scaling the Data

We need to split the data to train the model using a given data as well as testing it with
hidden data. Hidden and given data-sets are taken as the training and testing data, which
 Predicting Stock Market Trends: Machine Learning Approaches 385

is retrieved from one data-set split in a ratio of 7:3 or 8:2. This is done to that the model
created generalises to unseen data as well as it generalises to the given data.
This model takes in 80% data as training and 20% as testing. Scaling the data refers
to the normalisation of features in a data-set in order to achieve consistency. Having
dissimilar data or high differences between values in a data-set would affect the perfor-
mance of the model. Scaling leads to more accurate predictions as it makes data uniform.
For instance, the Volume in the data-set is in millions while the Close and Open prices
are in hundreds. Larger values may dominate the learning princess of the model.
Min-Max scaler [5] is a commonly used scaler for the purpose of logistic regression.
This scaler does not change the relationship between data and retains the relevance
of the features. The mathematical formula for scaling the data is:

Xscaled = (X − Xmin)/(Xmax − Xmin) (2)

where Xscaled is the value, we get after scaling the feature X using Minmax Scaler.
Xmin and Xmax are the minimum and maximum values of the X feature in the dataset.
Finally, the scaled dataset is fed into the Logistic Regression model from Python’s Sci-kit
learn library.

4 Predictions and Accuracy

The model is trained using the training dataset, and the mentioned mathematical calcu-
lations are performed. The model calculates X values or the linear regression output and
feeds it to the sigmoid function to get corresponding outputs for the predicted price of
each day. The coefficients (model parameters) to be found for the optimal values are iter-
atively found by the function using optimization techniques. Model evaluation is needed
for any machine learning model to be assessed. Certain performance metrics are used to
evaluate the accurateness of the model. The first technique employed is to calculate the
confusion matrix. It basically tells us the number of values the model predicts as true
positives, false positives, true negatives, and false negatives. True values are predicted
correctly, and the false values are predicted as the class which was incorrect. A classi-
fication report generates a summary of metrics that determine the accurate predictions
and inaccuracies of the model (Table 3).

Table 3. Accuracy Results for each Company Stocks.

Ticker Company LR Accuracy

AMZN Amazon 92.43
TSLA Tesla 88.84
MSFT Microsoft 86.85
GOOGLE Alphabet Inc (Google) 90.03
AAPL Apple 88.04
386 R. Joseph and M. Goswami

5 Conclusion

This paper navigated through the relevance of stock analysis and the application of
machine learning in the prediction and analysis of certain stock markets with five chosen
companies. The results of the Logistic Regression prediction model showed that after
fine-tuning the parameters and feature engineering, accurate predictions could be made
with the test datasets. Among the five companies amazon showed highest accuracy. The
second highest is Google. Tesla and Apple are very close. Microsoft dataset reflects 86.85
percentage of accuracy. This research can help investors to make a better management
of their portfolios.

References
1. Obthong, M., Tantisantiwong, N., Jeamwatthanachai, W., Wills, G.: A Survey on Machine
Learning for Stock Price Prediction: Algorithms and Techniques, 06 May 2020. https://epr
ints.soton.ac.uk/437785/
2. Mintarya, L.N., Halim, J.N.M., Angie, C., Achmad, S., Kurniawan, A.: Machine learning
approaches in stock market prediction: a systematic literature review. Proc. Comput. Sci. 216,
96–102 (2023). https://doi.org/10.1016/j.procs.2022.12.115
3. Ali, S.S., Mubeen, M., Hussain, I.L.: Prediction of stock performance by using logistic regres-
sion model: evidence from Pakistan stock exchange (PSX). Asian J. Empir. Res. 8(7), 247–258
(2018)
4. Zaidi, M., Amirat, A.: Forecasting stock market trends by logistic regression and neural net-
works. Int. J. Econ. Commer. Manag. (2016). https://ijecm.co.uk/wp-content/uploads/2016/
06/4614.pdf. Evidence From Ksa Stock Market
5. Chatterjee, A., Bhowmick, H., Sen, J.: Stock price prediction using time series, econometric,
machine learning, and deep learning models. In: 2021 IEEE Mysore Sub Section International
Conference (MysuruCon), pp. 289–296, October 2021. https://doi.org/10.1109/MysuruCon
52639.2021.9641610
6. Attigeri, G.V., Manohara Pai, M.M., Pai, R.M., Nayak, A.: Stock market prediction: a big
data approach. In: TENCON 2015 - 2015 IEEE Region 10 Conference (2015). (IF:3)
7. Sharma, A., Bhuriya, D., Singh, U.: Survey of stock market prediction using machine learning
approach. In: 2017 International Conference of Electronics, Communication (2017). (IF: 4)
8. Ren, R., Wu, D.D., Liu, T.: Forecasting stock market movement direction using sentiment
analysis and support vector machine. IEEE Syst. J. (2019). (IF:4)
9. Zhong, X., Enke, D.: Predicting the daily return direction of the stock market using hybrid
machine learning algorithms. Financ. Innov. (2019). (IF: 4)
10. Nti, I.K., Adekoya, A.F., Weyori, B.A.: A systematic review of fundamental and technical
analysis of stock market predictions. Artif. Intell. Rev. (2019). (IF: 4)
11. Gandhmal, D.P., Kumar, K.: Systematic analysis and review of stock market prediction
techniques. Comput. Sci. Rev. (2019). (IF: 4)
12. Mehtab, S., Sen, J.: A robust predictive model for stock price prediction using deep learning
and natural language processing. Econom. Model.: Cap. Mark. - Forecast. eJ. (2019). (IF:3)
13. Valencia, F., Gómez-Espinosa, A., Valdés-Aguirre, B.: Price movement prediction of cryp-
tocurrencies using sentiment analysis and machine learning. Entropy (Basel, Switzerland)
(2019). (IF:4)
 Predicting Stock Market Trends: Machine Learning Approaches 387

14. Nti, I.K., Adekoya, A.F., Weyori, B.A.: A comprehensive evaluation of ensemble learning for
stock-market prediction. J. Big Data (2020). (IF:4)
15. Nabipour, M., Nayyeri, P., Jabani, H., Shahab, S., Mosavi, A.: Predicting stock market trends
using machine learning and deep learning algorithms via continuous and binary data; a
comparative analysis. IEEE Access (2020). (IF: 4)
 A Scientific Classification-Based Survey
of Intrusion Detection System Using Machine
Learning Approach

Neha Srivastava(B) and R. K. Singh

Kamla Nehru Institute of Technology, Sultanpur, Uttar Pradesh, India

[email protected], [email protected]

Abstract. In this paper, the prime focus is to find and explore the field of intrusion
detection with supervised ML approaches. The objective is to create and publish
a scientific classification for combined IDS and supervised ML techniques. To do
the same, examination of the fundamentals of IDS, ML algorithms (supervised),
and cyber attacks is done. In addition, the work has been explored that has recently
been done in the sphere of supervised learning intrusion detection. Then, using
these works, a taxonomy is provided using different publicly available datasets
to classify the high and accurate performance of supervised machine learning
algorithms.

Keywords: Intrusion Detection System · Cyber Attacks · Machine Learning

1 Introduction
The possibility of cyberattacks is increasing with the rise in internet usage. Now, there
are new classifications of attacks, and it is necessary to recognize these types of new
attacks using intelligent systems. An IDS can be described as software that examines
network traffic to find malicious activity. It can be distributed into two types: one is
signature-based and the other is anomaly-based. Signature-based is used to search for
already available attacks and compares new attacks with old ones; if it matches, an alarm
is raised. While anomaly-based detection searches the new type of traffic to see if it is
showing unusual behavior and reports it as an anomaly. This detection is very critical
for detecting zero-day attacks [17].
There is a need for huge amounts of data to create a model for detecting normal and
abnormal behavior. Thus, it is required to create a model using supervised ML algorithms
to efficiently comprehend the data and to create a prediction model for the identification
of new attacks. The use of feature reduction along with feature selection helps improve
classification to overcome the problem of high dimensionality in data. It also studied
the best strategies for supervised machine learning applications in intrusion detection
systems, which can further be used for the efficient detection of new cyberattacks.
The major aim of this investigation work is to create a survey based on supervised
ML approaches, and IDS. The paper begins with the introduction of IDS and focus on its

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025

A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 388–396, 2025.
https://doi.org/10.1007/978-3-031-71481-8_30
 A Scientific Classification-Based Survey of Intrusion Detection System 389

significance, types, and definition. Then, it discuss the well-known supervised learning
methods, the publicly available datasets in this field, and the idea of dimensionality
reduction. It also discuss various types of cyber-attacks. Finally, it provides a taxonomy
to find out which methods are most appropriate to use with various IDS available datasets
and if feature selection could improve classification results or not.
The paper is presented in five sections which are as follows: Sect. 2 covers the
information related to intrusion detection systems, supervised ML, and cyberattacks.
Section 3 comes with the study of supervised ML methods for intrusion detection sys-
tems. Section 4 discusses the scientific classification provided based on the studies.
Finally, Sect. 5 presents the summary and findings of this research paper.

2 Background

This particular section provides information on the following topics:

• Intrusion Detection System
• Supervised ML (Machine Learning)
• Cyber Attack

2.1 Intrusion Detection System

Computer networking is crucial for information sharing and communication. Computers,

however, can be vulnerable to outside dangers that require constant monitoring and
detection [1]. Such risks have the potential to breach the computer security’s three pier
that are: confidentiality, integrity, and availability (CIA).
An intrusion can be defined as an attempt to bypass a computer or network’s security
measures or breach the CIA. It is a procedure of tracking and examining network or
system activity for indicators of intrusion [2]. Thus, the four basic goals of IDS is to (1)
monitor the host and network (2) analyze the behavior of the user computer network,
(3) raise alerts and (4) react to apprehensive behavior.
IDS systems are widely discussed in the literature. However, two issues plague a lot
of IDS:
1. High false alarm rate: Many significant violations go unreported while alarms are
activated for non-threatening infractions.
2. Because new attacks are crucial to spot, attention has been given to the employment
of cutting-edge ML algorithms for IDS.
IDS is often be categorized in two ways: 1) based upon methodology, and 2) based
upon platform being monitored.
With respect to the approach, it can divide Intrusion detection system into detection
of anomaly-based and misused-based. Anomaly-based detection techniques create a
model of standard behavior based on system activity.
This methodology attempts to make a model that closely resembles a system’s typical
behavior by monitoring the activity of a user with time. The IDS is suited for encoun-
tering zero-day attacks since it shows every event that diverges from typical behavior as
malicious activity and raises an alarm. There were two problems identified as significant
390 N. Srivastava and R. K. Singh

drawbacks: 1) unable to handle the changes in typical user behavior, and 2) a significant
FPR i.e. false positive rate [3].
On the other side signature based detection compares captured events with known
assaults and threats in order to identify potential intrusions. Comparing this method with
the anomaly based detection method, it is much more accurate and generates minimum
false alarms. However, it cannot be used to identify fresh attacks [4].
There are two categories: one is Host-based IDS (HIDS) and other one is network-
based IDS (NIDS) into which IDS is classified based on the platform being monitored.
The tasks on the user’s system where Host-based IDS is installed are tracked and
analyzed [5]. It is possible to keep track of some of the system’s dynamic conduct and
state. On the other side, Network-based IDS keeps an eye on network traffic to find
assaults that are transmitted via a network connection remotely [6]. It is a gadget that
is dispersed among networks to examine traffic passing through the ones it is seated
on. It could consist of h/w or s/w. NIDS often has two network interfaces -: one is for
controlling and reporting, the other one to listen network communications [7].

2.2 Supervised Machine Learning Approaches

For automatic identification of intrusions, there are typically two methods of machine
learning: supervised learning and unsupervised learning [8–11]. This survey’s major
objective is to use supervised machine learning methods for IDS. Labeled data must
be there to train a model for detection using supervised learning or classification. The
following step-by-step action can be used to summarize the categorization process:
Collection of Data , the process of gathering data to supply the knowledge needed to
train the categorization model. Basically, a set of features that can differentiate between
the classes is used to define the datasets. Since collecting data is a crucial task, there
are numerous benchmark publicly available data sets available, including KDD’99 [14],
NSL-KDD [14], and CICIDS2017 [14]. A summary of mentioned data collection is
included in Table 1.

Table 1. Description of IDS datasets

Dataset Description
CICIDS2017 It took five days to develop in a realistic environment. It has both bidirectional
and packet-based traffic. There are 80 extracted features. There are various
types of attacks which includes DDoS, DoS, Web attacks, BF (bruteforce)
attacks as FTP and SSH, infiltration, heartbleed, botnet and many more
NSL-KDD This data is an improvement of the KDD’99 dataset that addresses issues with
duplication, redundancy, and data imbalance
KDD-99 It is produced by simulating regular and hostile traffic in a military setting (US
AirForce LAN). Rat tcpdump files representing nine weeks of simulation are
included. 41 intrinsic, content, and traffic-related factors are used to describe
the dataset. Simulated assaults include DoS, Prob, U2R, and R2L
 A Scientific Classification-Based Survey of Intrusion Detection System 391

Reduction of Data: High dimensionality in feature spaces may be difficult and can
cause problems like the “curse of dimensionality” when there is a presence of little
amount of data in big-dimensional [13]. Therefore, techniques like PCA and LDA may
be used to convert the data into a low-dimensional space [8]. Feature selection techniques
such as the best-first algorithm [9] can also be used to select the best-performing features
in the opposite way.
Classification: At this phase, one portion of the data has to be used for developing
the model known as training data, while the other portion has to be used for testing the
performance of the classification model known as testing data. The data divided into
training set and testing set is defined as hold-out test. One other common test type is the
N-fold cross-validation test, in which the data is split into N-folds, with the tenth being
used to test the model and the other nine being used for training. A fresh set of 9-folds
for training and a 10-fold for testing are then used to repeat the process. The average
value across all folds is used to calculate accuracy.
Evaluation of Performance: The performance of classification model is evaluated with
the help of accuracy and the false positive rate (FPR).

2.3 Cybersecurity Attacks

Cyberattacks are deliberate, malicious attempts to breach another person’s or organiza-
tion’s information system. Usually, the attacker takes advantage of the victim’s network
outage. There are various types of cyber-attacks [16]:
Spyware, ransomware, viruses, and worms are examples of dangerous software that
falls under the umbrella term “malware.” Malware enters a network by taking advantage
of a weakness, usually caused by clicking on a link or some attachment that installs or
downloads software.
The practice of sending phony emails or other contact that appears to be from a
trustworthy source is known as phishing. The goal is to either use malware to infect
the victim’s computer or steal private information like credit card numbers and login
credentials. Phishing is a cyber-threat that is becoming more frequent.
Eavesdropping attacks, sometimes referred to as man-in-the-middle (MitM) attacks,
happen when an attacker interjects oneself into a two-party transaction. After severing
the connection, the attackers can filter and pilfer information.
In an effort to exhaust the bandwidth and resources of servers, networks, or systems,
a denial-of-service assault floods them with traffic. As such, legitimate requests cannot
be processed by the system. Attackers can even use many compromised devices to carry
out this attack.
An attack known as a “SQL injection” happens when a hacker introduces malicious
code into a SQL-using server, compelling it to divulge data that it would not typically. All
it takes to execute a SQL injection is for an attacker to insert malicious code anywhere
in the search bar of website which is vulnerable.
A zero-day exploit occurs when a network vulnerability is discovered just before a
patch and/or fix has been implemented. The Attacker target the widely known suscep-
tibility during the window of opportunity. Continuous monitoring is required for threat
identification resulting from zero-day vulnerabilities.
392 N. Srivastava and R. K. Singh

3 Literature Review
Automating intrusion detection with supervised ML techniques has been receiving
enough attention nowadays.
As per [12], the dataset NSL-KDD was utilized to address the intrusion classification
problem. In order to classify numerical attributes and address missing data issues, the
data set is first preprocessed. After that, the dataset is sub-divided into four groups and
categorized as train data and test data. The classification, accuracy, and False positive
rate are then calculated using a random forest classifier after the data has been input. In
a feature selection technique, the Symmetrical Uncertainty measure is also employed.
Using feature selection led to a little performance gain, according to the authors. Random
forests performed better than C4.5 when the results were compared [12]. Following
feature selection, the received accuracy is of 99.67 with an average FAR of 0.005.
Authors [17] compared the evaluation of 10 algorithms based on classification on
the dataset NSL-KDD, similar to the authors in [12], but applying a different and various
feature selection strategy. The technique of the selection of feature is dependent on the
use of attribute evaluators or filters. The authors used the following algorithms: NB,
Logistics, RT, RF, Bagging, J48, OneR, ZERO, PART. The random forest classification
system performs best having accuracy 99.9% and a FAR as 0.001.
Similarly, the author [18] used the dataset NSL-KDD and used a hold-out test strategy
with feature selection not applicable. RF, SVM, logistical regression, and the Gaussian
mixture model are the four classification techniques that are being tested. Having an
accuracy of 99%, the random forest performed as the best algorithm.
The authors of [19] recently stated the working of SVM and ANN for the NSLKDD
dataset. The data sample taken is the 20% of the whole dataset. The author used a chi-
squared-based and correlation-based approach with 35 features which were earlier 14.
The data is then sent to ANN and SVM after the selection of feature. Hence it is found
that ANN with correlation-based feature achieved the accuracy of 94%.
In other approach, the authors of [20] used the KDD’99 data set with the following
classes: DoS, normal, U2R, Prob and R2L. CFSSubSet Eval and Best First feature
selection algorithms are used along with four approaches: k-means, SVM, NB, and RF.
Based on the eight best attributes, it is found that the supervised approach (random
forest) performed well in comparison with k-means having accuracy of 99%.
The authors also used KDD’99 in [21] where an Algorithm selects a good represen-
tative sample of the original dataset having a record of 550. The [21] implemented a new
feature reduction method called GFR (Gradual Feature Removal) for reducing the space
of feature dimension to 19 features. For classifying, the reduced characteristics are then
combined using SVM. Before feature selection, the received accuracy was 98.67% with
no improvement and after the selection of feature, the accuracy is 98.62%.
In [22], the author reduced the dimensional space of KDD’99 to see its impact
on the overall performance of classification They looked specifically at the algorithm
for picking features, getting information by entropy (IG), and using BBNN with feature
vector (22). They found that the accuracy dropped to 91% when the feature was removed.
Another dataset that has generated interest is CICIDS2017. The authors of [23] and
[24] used sampling followed by feature reduction and then boosting strategy to create
an Intrusion detection system. This dataset has a difficult part due to its imbalance
 A Scientific Classification-Based Survey of Intrusion Detection System 393

feature. To overcome this issue, they preprocessed the data with the Synthetic Minority
Overamplification Technique (SMOT) to reduce the number of case in certain class.
They then used a reduced feature space (25) to implement a reduction strategy using
Principal Component Analysis (PCA) and ensemble feature selection (EFS). Finally,
they evaluated the findings using the hold-out approach based on the Adaboost algorithm
having accuracy of 81.83%.
In a recent paper [25], the authors discussed the topic of assault detection using
RF and the ANN approach on the dataset CICIDS2017. They used Boruta, a feature
selection software, to return the top ten most significant features. Then, they fed the
classifiers the feature set. Using random forest, they reported 96% accuracy, and using
ANN, 96.4%.
In order to review and brief the related research, the paper presents a classification
of IDS and supervised ML approach in the following section to better understand the
presented final findings and conclusions.

4 Scientific Classification of IDS and Supervised Machine Learning

Algorithms

Considering the useful findings presented in the mentioned part, it constructs a scientific
classification of IDS and ML in this section. The scientific classification also known as
taxonomy having following characteristics, which are listed in Table 2 dataset used; 2)
The usage of feature selection or not: Y or N, 3) Usefulness of feature selection: Y or
N, 4) The supervised ML used, 5) Method of validation, 6) The best-received accuracy
and FPR outcome. 7) year of research.
To comprehend the data shown in Table 2, it shall examine the datasets individually
before providing an complete summary:
1. NSL-KDD dataset: When feature selection is applied to this dataset, it improves
classification performance. Furthermore, the random forest technique performs well
on this dataset and produces high-quality result utilizing several methods of validation.
ANN appears to perform good on NSL-KDD; however, it has tested on only 20% of
entire data as sample.
2. KDD99 dataset: When feature selection is performed on this dataset, it does not
improve classification results. The data set is validated using several approaches, and
no immediate result can be taken regarding the top-behaving classification algorithm;
however, Random forest and Support Vector Machine do quite good on KDD’99
dataset.
3. CICIDS2017 dataset: It is not possible to directly determine the impact of feature
selection; however, it was useful in enhancing performance of the classification. This
dataset contains the significant imbalance issue that can be remedied through sampling
techniques. QDA is the most effective algorithm.
Overall, the paper can draw the following conclusions:
1. There has been a lot of discussion in the literature on IDS using machine learning
methodologies.
394 N. Srivastava and R. K. Singh

2. The results of the categorization are pretty good and strong, as per the review of three
separate publicly available datasets.
3. The selection of feature is crucial for performance enhancement and is often necessary.
4. Data imbalancy can be a thought of concern, and sampling procedures can assist in
resolving it.

Table 2. IDS Scientific Classification

Dataset Feature Algorithm Used Best Result Paper Year

Selection used Validation Method
Applied
(Y/N)
KDD’99 Y SVM TEN-FOLD BBNN Acc. 98.7% 21 2012
KDD’99 Y RF, SVM, Hold-out RF Acc-99% 20 2015
NB,
k-means
KDD’99 Y BBNN TEN-FOLD QDA Acc-91% 22 2015
NSL-KDD Y RF, J-48 TEN-FOLD RF Acc-99.7% 12 2016
FPR-.005%
NSL-KDD N RF, GMM, HOLD OUT RF Acc-99% 18 2016
LR, SVM
NSL-KDD Y NB TEN-FOLD RF Acc-99.9% 17 2019
20% of Y SVM, Hold-out RF Acc-94% 19 2019
NSL-KDD ANN
CICIDS2017 Y QDA, Hold-out QDA Acc = 24 2019
LDA, BN, 98.8%
RF FPR =
.001%
CICIDS2017 Y Adaboost Hold-out Adaboost Acc = 23 2019
81.83%
CICIDS2017 Y RF, ANN Hold-out RF Acc = 25 2020
96.4%

5 Conclusion
Because of the current boom in internet material, cybercrime is on the rise. The first and
immediate step for detecting and raising such attacks is to employ intrusion detection
systems (IDS). Anomaly detection relies on being able to identify new attacks, which
can be a tricky job. In this paper, it can be seen how IDS classification works, how
supervised learning works, and how cyber security attacks work.
It then presents all the work that’s been done in this area and compared it to each
other using three different data sets that were made available to the public: the KDD’99
 A Scientific Classification-Based Survey of Intrusion Detection System 395

dataset, the NSL-KDD dataset, and the CICIDS2017 dataset. A scientific classification
i.e., taxonomy is also provided having related studies. This classification suggests that
the subject of intrusion detection utilizing supervised ML approaches is gaining pop-
ularity. The study looked at three data sets and found that the performance of various
classification of the supervised learning algorithms was good and promising. It’s impor-
tant to pick the right features, and in many cases, it’s mandatory to improve performance.
Data mismatches can also be a problem, and sampling procedures can help with that.

References
1. Graham, J., Olson, R., Howard, R. (eds.): Cyber Security Essentials. CRC Press (2011)
2. Liao, H.J., Lin, C.H.R., Lin, Y.C., Tung, K.Y.: Intrusion detection system: a comprehensive
review. J. Netw. Comput. Appl. 36(1), 16–24 (2013)
3. Liu, H., Lang, B.: Machine learning and deep learning methods for intrusion detection
systems: a survey. Appl. Sci. 9(20), 4396 (2019)
4. Hamid, Y., Sugumaran, M., Balasaraswathi, V.R.: IDS using machine learning-current state
of the art and future directions. Curr. J. Appl. Sci. Technol. 1–22 (2016)
5. Masdari, M., Khezri, H.: A survey and taxonomy of the fuzzy signature-based intrusion
detection systems. Appl. Soft Comput. 106301 (2020)
6. Milenkoski, A., Vieira, M., Kounev, S., Avritzer, A., Payne, B.D.: Evaluating computer intru-
sion detection systems: a survey of common practices. ACM Comput. Surv. (CSUR) 48(1),
1–41 (2015)
7. Conrad, E., Misenar, S., Feldman, J.: CISSP Study Guide (2012)
8. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer, New York (2006)
9. Witten, I.H., Frank, E.: Data mining: practical machine learning tools and techniques with
Java implementations. ACM SIGMOD Rec. 31(1), 76–77 (2002)
10. https://en.wikipedia.org/wiki/. Accessed 10 Aug 2023
11. Sahasrabuddhe, A., Naikade, S., Ramaswamy, A., Sadliwala, B., Futane, P.: Survey on intru-
sion detection system using data mining techniques. Int. Res. J. Eng. Technol. 4(5), 1780–1784
(2017)
12. Farnaaz, N., Jabbar, M.A.: Random forest modeling for network intrusion detection system.
Proc. Comput. Sci. 89, 213217 (2016)
13. Rust, J.: Using randomization to break the curse of dimensionality. Econometrica: J. Econom.
Soc. 487–516 (1997)
14. https://www.unb.ca/cic/datasets/index.html. Accessed 11 Aug 2023
15. www.kaggle.com. Accessed 03 Aug 2023
16. https://www.cisco.com/c/en_in/products/security/common-cyberattacks.html#~how-cyber-
attacks-work
17. Malhotra, H., Sharma, P.: Intrusion detection using machine learning and feature selection.
Int. J. Comput. Netw. Inf. Secur. 11(4) (2019)
18. Belavagi, M.C., Muniyal, B.: Performance evaluation of supervised machine learning
algorithms for intrusion detection. Proc. Comput. Sci. 89, 117–123 (2016)
19. Taher, K.A., Jisan, B.M.Y., Rahman, M.M.: Network intrusion detection using supervised
machine learning technique with feature selection. In: 2019 International Conference on
Robotics, Electrical and Signal Processing Techniques (ICREST), pp. 643–646. IEEE,
January 2019
20. El Mourabit, Y., Bouirden, A., Toumanari, A., Moussaid, N.E.: Intrusion detection techniques
in wireless sensor network using data mining algorithms: comparative evaluation based on
attacks detection. Int. J. Adv. Comput. Sci. Appl. 6(9), 164–172 (2015)
396 N. Srivastava and R. K. Singh

21. Li, Y., Xia, J., Zhang, S., Yan, J., Ai, X., Dai, K.: An efficient intrusion detection system based
on support vector machines and gradually feature removal method. Expert Syst. Appl. 39(1),
424–430 (2012)
22. Shah, B., Trivedi, B.H.: Reducing features of KDD CUP 1999 dataset for anomaly detection
using back propagation neural network. In: 2015 Fifth International Conference on Advanced
Computing & Communication Technologies, pp. 247–251. IEEE, February 2015
23. Yulianto, A., Sukarno, P., Suwastika, N.A.: Improving AdaBoost-based intrusion detection
system (IDS) performance on CIC IDS 2017 dataset. J. Phys.: Conf. Ser. 1192(1), 012018
(2019)
24. Abdulhammed, R., Faezipour, M., Musafer, H., Abuzneid, A.: Efficient network intrusion
detection using PCA-based dimensionality reduction of features. In: 2019 International
Symposium on Networks, Computers and Communications (ISNCC), pp. 1–6. IEEE, June
2019
25. Pelletier, Z., Abualkibash, M.: Evaluating the CIC IDS-2017 dataset using machine learning
methods and creating multiple predictive models in the statistical computing language R.
Science 5(2), 187–191 (2020)
 Author Index

A G
Abidin, Shafiqul I-136 Gandhi, Dhruv II-305
Agarwal, Gaurav II-259 Gandhi, Geeta Chhabra II-319
Ahmed, Imtiaj I-346 Gangrade, Jayesh I-86, I-239, I-328, II-161
Ahmed, Muqeem II-69 Garg, Amit II-329
Ali, Md. Sawkat I-174 Girisha, S. II-210, II-220, II-272
Anand, Aastha II-29 Gite, Shilpa II-1
Anjum II-190 Gnanaprakasam, Johnbenetic I-16, I-30
Arora, Chandani II-51 Goel, Rajat II-319
Arora, Palak II-102, II-170 Gopalakrishnan, E. A. I-47
Gopalani, Bhavya II-202
B Goswami, Aarti II-161
Balusamy, Balamurugan II-1 Goswami, Mausumi I-376
Bansal, Deepika II-29 Gudapati, Harshitha II-42
Bhaisare, Bhumika I-312 Gupta, Bhoomi II-29
Bharati, Rajesh I-312 Gupta, Kanav I-278
Bhargava, Lava II-113 Gupta, Mayuri I-239, I-328
Bhosale, Jyoti Dinkar I-205 Gupta, Pratik II-305
Bind, Aman I-290 Gupta, Sachin II-29
Bukya, Mahipal II-346 Gupta, Sunita I-1

H
C
Hasan, Mahamudul I-174
Chaudhari, Durvesh I-254
Hemlata I-184
Chaurasia, Sandeep II-161, II-356
Chhikara, Rita I-218
J
Choudhary, Sonia II-142
Jadon, Aryan II-281
Choudhury, Deepjyoti I-357
Jain, Jagrite II-319
Chowdary, Paleti Nikhil I-47
Jain, Nishant I-278
Chugh, Aarti II-77
Jain, Piyush I-124
Jangirala, Srinivas II-233
D Janu, Neha I-1
Damre, Suraj S. I-205 Jha, Aparna I-290
Delvadia, Om II-305 Joseph, Raima I-376
Dhanya, M. I-47 Joshi, Nisheeth II-102, II-170
Dikshit, Adrika I-254
Durga, Kbks II-42 K
Kamath, Ritesh V. II-272
F Kamboj, Akshit I-63
Ferdaus, Md. Hasanul I-174 Karia, Deepak I-254
Fifatin, François-Xavier II-128 Katarya, Rahul II-190

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2025
A. K. Bairwa et al. (Eds.): ICCAIML 2024, CCIS 2184, pp. 397–399, 2025.
https://doi.org/10.1007/978-3-031-71481-8
398 Author Index

Katyayan, Pragya II-102, II-170 O

Kedari, Bhavana II-272 Oloulade, Arouna II-128
Kumar, Abhishek I-86
Kumar, Ajay II-10 P
Kumar, Lokesh I-136 Palati, Madhu II-346
Kumar, Rajan II-329 Pandey, Pranjal Mohan II-90
Kumar, Rajesh I-113, II-346 Pandey, Swati I-328
Kumar, Rohan I-184 Pandey, Vedant II-90
Kumar, Satyam I-86, I-357 Parmar, Jitendra II-305
Kumar, Sunil I-150 Patil, Avinash II-281
Kushwaha, Satpal Singh II-329 Paul, Chirag I-278
Pavithra, N. II-244
Pawar, Omkar I-254
L Pawar, Rajkumar B. I-205
Lakhwani, Love I-150 Pingali, Sathvika I-47
Lamidi, Taohidi Alamou II-128 Praksh, Aditya I-265
Lawanya Shri, M. II-1 Preethi II-244
Lita, Mario Gracius Krishna II-179
Logeshwaran, J. II-329
Lourdusamy, Ravi I-16, I-30 R
Raghuwanshi, Pranshu II-202
Rai, Anmol I-164
M Rai, Bipin Kumar I-290, II-259
Maharani, I. Gusti Ayu Ngurah Stita II-179 Rani, Ashima II-77
Mahmud, Adnan I-346 Rashmi, R. II-210, II-220
Malik, Meenakshi I-218, I-226, I-302 Rathi, Vedansh I-164
Malik, Seema II-142 Rayan I-346
Mandal, Anandadeep II-233 Ritu I-226, I-302
Mardhiyyah, Alya Dhiya’ II-179 Rosni, Tanjela Rahman I-174
Mathur, Akhilesh II-346 Roy, Ritwick II-356
Maulana, Fairuz Iqbal II-179
Mayuri, A. V. R. II-90 S
Md. Gulzar II-69 Sagar, Aditya II-29
Mishra, Ashish I-239, I-328 Sahoo, Debajani I-74, I-99
Mishra, Ashutosh I-239, I-328 Salagrama, Shailaja II-329
Mitra, Sakalya II-90 Sandhu, Ramandeep II-51
Mittra, Tanni I-174 Sanjeev, Rohan I-47
Motwani, Deepak I-124 Santhi, K. II-1
Mukthala, Shilpa II-42 Sapna, R. II-244
Mulia, Angelina Patience II-179 Saroop, Shashank I-136
Mundra, Ankit I-63 Savitha, G. II-272
Mundra, Shikha I-63 Semassou, Guy Clarence II-128
Shah, Payal I-254
Shaik, Ayesha II-42
N Sharath Kumar, Y. H. II-244
Naajim, M. Mohamed II-77 Sharma, Amit Kumar II-356
Nagar, Anushka II-170 Sharma, Pankaj I-124
Nandal, Rainu I-218 Sharma, Paribhasha II-233
Nathani, Bharti II-102 Sharma, Sandeep Kumar I-164
Nounangnonhou, Cossi Télesphore II-128 Sharma, Santosh II-113
Author Index 399

Sharma, Shivani II-259 T

Sharma, Shrey I-164 Tayal, Prince I-184
Sharma, Surbhi II-202 Tihany, Sayma Akter I-346
Shreesha, S. II-272 Tognigban, Romaric II-128
Shrotriya, Neha I-1 Tulsyan, Rajiv II-60
Shukla, Pranjal II-60 Tyagi, Rajesh Kumar I-136
Singh, Anirudh I-86, I-357
Singh, Chandrakant Kumar I-113 U
Singh, Juhi I-265 Unnikrishnan, Pranav I-47
Singh, R. K. I-388 Urkude, Shubhangi V. I-74, I-99
Singh, Satyendra I-113
Singh, Vibhanshu I-150
Sinha, Shweta I-136 V
Sowmya, V. I-47 Vardhan, Kuruba Lakshmi Harsha II-305
Srivastava, Neha I-388 Vats, Subhash II-142
Srivastava, Shreyal I-290 Verma, Atul Kumar II-305
Sudhamsh, G. V. S. II-220 Vijarania, Meenu I-226, I-302
Sujigarasharma, K. II-1
Sumit II-29 Y
Sundari, M. Shanmuga II-42 Yadav, Narendra Singh II-142
Suryawanshi, Ujwala V. I-205 Yadav, Poonam I-226, I-302