Developing an Explainable AI Framework for

Accurate ICU Length of Stay Predictions

to Optimize Hospital Resource Management
A PROJECT REPORT
Submitted by

VUYYURU AKSHITHA[RA2111047010172]
RAGHU VAMSI UPPULURI [RA2111047010180]
Under the Guidance of

Dr. T. R. SARAVANAN
(Associate Professor, Department of computational Intelligence)

in partial fulfillment of the requirements for the degree of

BACHELOR OF TECHNOLOGY
in
ARTIFICIAL INTELLIGENCE

DEPARTMENT OF COMPUTATIONAL INTELLIGENCE

COLLEGE OF ENGINEERING AND TECHNOLOGY
SRM INSTITUTE OF SCIENCE AND TECHNOLOGY
KATTANKULATHUR- 603 203

MAY 2025
 Department of Computational Intelligence
SRM Institute of Science & Technology
Own Work* Declaration Form

To be completed by the student for all assessments

Degree/ Course : B. Tech in Artificial Intelligence

Student Name : Vuyyuru Akshitha, Raghu Vamsi Uppuluri
Registration Number : RA2111047010172, RA2111047010180
Title of Work : Developing an Explainable AI Framework for Accurate ICU Length of Stay
Predictions to Optimize Hospital Resource Management
We hereby certify that this assessment compiles with the University’s Rules and Regulations relating to
Academic misconduct and plagiarism**, as listed in the University Website, Regulations, and the
Education Committee guidelines.
We confirm that all the work contained in this assessment is our own except where indicated, and that
we have met the following conditions:
 Clearly referenced / listed all sources as appropriate
 Referenced and put in inverted commas all quoted text (from books, web, etc.)
 Given the sources of all pictures, data etc. that are not my own
 Not made any use of the report(s) or essay(s) of any other student(s) either past or present
 Acknowledged in appropriate places any help that I have received from others (e.g.fellow
students, technicians, statisticians, external sources)
 Compiled with any other plagiarism criteria specified in the Course handbook /University
website
I understand that any false claim for this work will be penalized in accordance with theUniversity
policies and regulations.

DECLARATION:
I am aware of and understand the University’s policy on Academic misconduct and plagiarism and I certify
that this assessment is our own work, except where indicated by referring, and that I have followed the
good academic practices noted above.

Raghu Vamsi Uppuluri Vuyyuru Akshitha

[RA2111047010180] [RA2111047010172]

ii
SRM INSTITUTE OF SCIENCE AND TECHNOLOGY
KATTANKULATHUR – 603 203

BONAFIDE CERTIFICATE

Certified that 18AIP109L - Major Project report titled “Developing

an Explainable AI Framework for Accurate ICU Length of Stay
Predictions to Optimize Hospital Resource Management” is the
bonafide work of “Vuyyuru Akshitha [RA2111047010172], Raghu
Vamsi Uppuluri [RA2111047010180]” who carried out the project
work under my supervision. Certified further, that to the best of my
knowledge the work reported here in does not form any other project
report or dissertation on the basis of which a degree or award was
conferred on an earlier occasion on this or any other candidate.

SIGNATURE SIGNATURE

Dr. T. R. SARAVANAN Dr. R. ANNIE UTHRA

SUPERVISOR PROFESSOR & HEAD
ASSOCIATE PROFESSOR DEPARTMENT OF
DEPARTMENT OF COMPUTATIONAL
COMPUTATIONAL INTELLIGENCE
INTELLIGENCE SRM INSTITUTE OF SCIENCE
SRM INSTITUTE OF SCIENCE AND TECNOLOGY
AND TECNOLOGY KATTANKULATHUR
KATTANKULATHUR

EXAMINER 1 EXAMINER 2

iii
 ACKNOWLEDGEMENTS

We express our humble gratitude to Dr. C. Muthamizhchelvan, Vice-Chancellor, SRM Institute

of Science and Technology, for the facilities extended for the project work and his continued support.
We extend our sincere thanks to Dr. Leenus Jesu Martin, Dean-CET, SRM Institute of Science
andTechnology, for his invaluable support.
We wish to thank Dr. Revathi Venkataraman, Professor and Chairperson, School of Com putting,
SRM Institute of Science and Technology, for her support throughout the project work.
We encompass our sincere thanks to, Dr. M. Pushpalatha, Professor and Associate Chairperson,
School of Computing and Dr. C. Lakshmi, Professor and Associate Chairperson, School of Computing,
SRM Institute of Science and Technology, for their invaluable support.
We are incredibly grateful to our Head of the Department, Dr. R. Annie Uthra, Professor,
Department of Computational Intelligence, SRM Institute of Science and Technology, for her
suggestions and encouragement at all the stages of the project work.
We want to convey our thanks to our Panel Members, Dr. T. R. Saravanan, Dr. A. Sheryl Oliver,
Dr. R. Usharani, Dr. M. Kiruthika, Dr. M. Anousouya Devi, Department of Computational
Intelligence, SRM Institute of Science and Technology, for their inputs during the project reviews and
support.
We register our immeasurable thanks to our Faculty Advisor, Dr. B. Pitchaimanickam, Assistant
Professor, Department of Computational Intelligence, SRM Institute of Science and Technology, for
leading and helping us to complete our course
Our inexpressible respect and thanks to our guide and panel head, Dr. T. R.Saravanan, Associate
Professor, Department of Computational Intelligence, SRM Institute of Science and Technology, for
providing us with an opportunity to pursue our project under his mentorship. He provided us with the
freedom and support to explore theresearch topics of our interest. His passion for solving problems and
making a difference in the world has always been inspiring.
We sincerely thank all the staff and students of Computational Intelligence, School of Computing,
S.R.M Institute of Science and Technology, for their help during our project. Finally, we would like to
thank our parents, family members, and friends for their unconditional love, constant support and
encouragement

iv
 ABSTRACT

The efficient utilization of the available hospital resources, especially ICU beds, is critical for
the quality of health care services. This is one of the major challenges for the hospital to estimate the
duration of time the patient should stay in the ICU. This accurate prediction prevents inessential delays
in treatment, facilitates the smooth transfer of patients, and aids in the proper planning and availability
of the beds. This project involves designing a system that is able to predict the duration of patients' ICU
stay during their admission, based on their clinical details and medical histories.
The system was implemented through a hospital data set that provided different patient records,
such as their health and the actual lengths of their stays in the ICU. The gathered data was accurately
prepared, cleaned, and processed before it was utilized in the system. The system was developed to
determine whether a patient would need a short or long ICU stay, from the trends and patterns established
from previous patient histories. The strategy adopted ensures the ease of use of the system with results
that can help doctors and hospital administrators make timely and informed decisions.
Apart from making accurate predictions, the system also provides the critical drivers that impact
the length of stay of a patient. This information assists clinicians to better understand the potential causes
behind each outcome and gain greater insight into planning care for patients. The end system is
implemented in a simple web interface and aims to enable real-time prediction capabilities. With
continued enhancements, this project can be extended to other hospital wards and bigger data sets,
rendering it a useful tool for overall hospital resource management.

The goal of this project is to forecast the duration of ICU stays upon admission based on patient
health records. Accurate forecasts help in improved bed management, reducing hospital costs, and
improved patient care efficiency.

v
 TABLE OF CONTENTS

ABSTRACT v
TABLE OF CONTENTS vi
LIST OF FIGURES viii
ABBREVIATIONS ix
CHAPTER TITLE PAGE
NO. NO.
1 INTRODUCTION 1
1.1 Overview 1
1.2 Sustainable Development Goal of the Project 3
2 LITERATURE SURVEY 6
2.1 Machine learning in healthcare and disease prediction 6
2.2 Key research contributions in outbreak prediction 8
2.3 Limitations Identified from Literature Survey 10
2.4 Research Objectives 11
2.5 Key user stories with Desired outcomes 13
2.6 Proposed work 15
2.7 Plan of Action 16
3 SPRINT PLANNING AND EXECUTION METHODOLOGY 19
3.1 SPRINT I 19
3.1.1 Introduction 19
3.1.2 Data Augmentation and Preprocessing 20
3.1.3 Model Architecture 21
3.1.4 Training Procedure 21
3.1.5 Result Analysis 21
3.2 SPRINT II 22
3.2.1 Introduction 22
3.2.2 Backend and Web Framework 22
3.2.3 Frontend Development and Data Entry Interface 23
3.2.4 Output Generation 23
3.2.5 Result Analysis 24

vi
 3.3 SPRINT III 24
3.3.1 Introduction 24
3.3.2 Process Steps for Model Integration and System workflow 25
3.3.3 Model and System Architecture 25
3.3.4 Result Analysis 26
3.3.5 Limitations and Future Work 27
4 RESULTS AND DISCUSSIONS 28
4.1 Project Outcomes Performance Evaluation, Comparisons, Testing Results 28
4.2 Challenges and Improvements 32
5 CONCLUSION AND FUTURE ENHANCEMENT 35
5.1 Conclusion 35
5.2 Future Enhancement 37
6 REFERENCES 41
APPENDIX
A CODING 43
B CONFERENCE PRESENTATION 48
C PLAGIARISM REPORT 50

vii
 LIST OF FIGURES

CHAPTER TITLE PAGE

NO. NO.

Fig 3.1 Architecture diagram for ICU LOS prediction 26

Fig 4.1 Classification performance Accuracy 29

Fig 4.2 Classification performance Precision 30

Fig 4.3 Classification performance F1 score 31

Fig 4.4 ROC AUC Curve for the Extension CatBoost 31
Fig 4.5 Confusion matrix of CatBoost Model 32
Fig 5.1 Dashboard of ICUStay website 36
Fig 5.2 Result of testing data 39

viii
 ABBREVIATIONS

XAI Explainable Artificial intelligence

EHR Electronic health records
ICU Intensive Care Unit
ML Machine Learning
AI Artificial Intelligence
LOS Length of Stay
XAI Explainable Artificial Intelligence
XGBoost Extreme Gradient Boosting
CNN Convolutional Neural Network
RNN Recurrent Neural Network
LSTM Long Short-Term Memory
ANN Artificial Neural Network
SHAP SHapley Additive exPlanations
LIME Local Interpretable Model-Agnostic Explanations
CatBoost Categorical Boosting
RF Random Forest
MIMIC Medical Information Mart for Intensive Care
ROC Receiver Operating Characteristic
AUC Area Under Curve
RMSE Root Mean Squared Error
MAE Mean Absolute Error
SMOTE Synthetic Minority Over-sampling Technique
API Application Programming Interface
KNN K-Nearest Neighbor
SVM Support Vector Machine
DT Decision Tree

ix
 CHAPTER 1

INTRODUCTION

1.1 Overview
Effective resource allocation is an essential area of hospital administration, especially in the
Intensive Care Unit (ICU) where equipment and staff are intensive. Unpredictability of patient length
of stay (LOS) is one of the main ICU resource planning challenges. Patients can take different periods
in ICU beds depending on clinical condition, response to treatment, and post-treatment recovery
requirements.
This inequality makes it harder to forecast beds, allocate resources, and plan admissions. The
uncontrollable character of LOS typically creates crowding, delay in treatments, postponed elective
surgeries, and even accelerated deaths. The issue is addressed in this project through the creation of
a data-driven model that forecasts ICU LOS at patient admission. The answer is constructed on patient
history data, which include factors of health like age, gender, diagnostic data, admission type, and
various clinical variables. Through pattern identification in the data, the system tries to categorize
each patient's ICU stay as short or long.

The classification helps hospital managers in advance planning of bed utilization and transfer of
patients at the right time, ultimately yielding effective hospital operation and better patient care In
order to apply this framework, different algorithms were tried and analyzed on a healthcare dataset
downloaded from Kaggle that contained over 100,000 patient records. The dataset was preprocessed
by steps including missing value handling, normalization, and label encoding to represent non-
numeric data in a format that can be analyzed.

This project suggests an intelligent and integrated student transport management system to
improve safety, transparency, and operational efficiency in educational institutions. The system uses
GPS tracking, IoT sensors, and cloud-based software to provide real-time bus location updates and
student activity monitoring. With increasing concerns regarding student safety and the inadequacies
of traditional transport monitoring techniques, this solution offers a centralized and automated
platform for all the stakeholders such as parents, school authorities, and transport authorities.

1
 The architecture of the system includes three main elements: GPS modules for tracking locations,
RFID modules for tagging students, and microcontroller units (such as NodeMCU) for data handling
and transmission. The system compares various prediction models, and upon comparison, it was
determined that the CatBoost algorithm had the best accuracy (98.25%). This is primarily because it
can easily deal with categorical variables and avoid overfitting, and thus it is a good fit for healthcare-
related classification issues. Aside from prediction, the system is also concerned with explainability.
Perhaps one of the biggest hurdles in deploying such systems in healthcare is the "black box" tendency
of many prediction models. In order to mitigate this, the framework combines explainability
techniques that enable it to determine the most important factors that drive the predictions. Aside
from increasing the transparency of the system, it also increases trust among medical doctors and
practitioners, thereby making it more acceptable for actual deployment in real-world environments.
It is implemented as a web-based platform with Python and Flask.
The system is endowed with an intuitive interface to load patient information and get the
prediction in real-time. It ensures rapid storage and retrieval of data using SQLite as the backend
database, appropriate for small and medium-scale hospital environments. With both high accuracy of
prediction and ease of design, this solution is best suited for implementation into hospital management
systems. It has the potential to be used as a decision-support tool, making hospitals more efficient,
lessening operational stress, and delivering care in a timely manner.

The implementation process included creating a prototype that effectively illustrated the
feasibility of the solution proposed. GPS information was efficiently captured and forwarded to the
Firebase database, which displayed it dynamically on the dashboard. RFID readers were employed in
authenticating students boarding and alighting from the vehicle, hence automatic marking of
attendance. The dashboard interface was found intuitive and informative to suit multiple roles of
users. These results reflect the system's preparedness to scale up into larger learning environments
that necessitate real-time transportation monitoring and enhanced student safety procedures.

The architecture paves the way for enhancements like the integration of deep learning algorithms
(such as CNN and LSTM), inclusion of real-time sensor readings, and extending prediction features
to other hospital departments. This project not only solves a particular problem but also lays the
ground for larger hospital analytics systems in the future

2
 In further, this project addresses a vital need in hospital management by offering a practical,
accurate, and interpretable solution for ICU LOS prediction. It brings together structured data
processing, model evaluation, and system deployment to deliver a reliable tool for improving patient
care and resource optimization in critical care environments.

Summing up, the project responds to a real demand for educational transportation logistics with
an enterprise-grade, scalable IoT-solution. Such features as route optimization via AI, biometric
authentication, and compatibility with pollution sensing equipment in the future would reinforce the
role of the system as part of the smart campus concept and offer further improvement for users and
operators.

1.2 Sustainable Development Goal of the Project

The suggested project closely resonates with the United Nations Sustainable Development Goal
3 SDG 3: Good Health and Well-being, where the main target is ensuring healthy lives and enabling
well-being for all, at all ages. One of the main objectives of this goal is to make healthcare systems
efficient and accessible throughout the world. The project, which predicts ICU length of stay (LOS)
from patient health data, serves directly towards improving the health infrastructure through enhanced
critical care facility management. Hospitals' overcrowding, delayed hospital admissions, and
unavailability of ICU beds are common issues taking a heavy toll on the quality of healthcare services
Such problems not only undermine patient care but also generate more stress for healthcare workers.

The intended student transport monitoring system is consistent with the United Nations
Sustainable Development Goal SDG 11: Sustainable Cities and Communities. The goal is focused on
making cities safe, inclusive, resilient, and sustainable. With the introduction of intelligent transport
technology in schools, the project supports the development of safer and more efficient urban
infrastructure. Real-time tracking and RFID-based attendance enhance the safety and accountability
of student commuting, together with the reduction of the reliance on manual processes and archaic
systems. The centralized monitoring function allows the tracking of students' movement for
authorities and parents, hence reducing risks and improving response times during emergencies.
Furthermore, the utilization of digital resources and automation facilitates more environmentally
friendly management of resources, reducing paper-based attendance and manual reporting needs. The
employment of cloud-based platforms makes the system scalable and expandable to other institutions
without significant infrastructure modification. Through the provision of real-time bus location and
automated attendance capabilities, the system minimizes uncertainty for parents and students alike.
This promotes consistent attendance and provides peace of mind, enabling students to concentrate

3
more on learning than transport issues. It also allows school authorities to more effectively manage
transport logistics, hence indirectly contributing to a more structured and inclusive learning
environment.

Through estimating how many days a patient is going to be in the ICU on the day of admission,
this project is a good way of preplanning bed availability. This ensures improved hospital workflow,
reduces the waiting time of patients, and allows timely access to care—all of which conform to SDG
3.8 that is concerned with universal health coverage and access to health services. Another significant
factor is the decrease in mortality and morbidity rates due to critical care delays.in cases of
emergencies, when seconds count, being able to use a free ICU bed can mean life and death.
Proper prediction of ICU stay can enable hospitals to plan admissions and discharges better and
minimize the likelihood of avoidable deaths. This supports SDG 3.2 and 3.4, which are designed to
eliminate premature mortality and non-communicable diseases by increasing readiness and
management in health systems. The project indirectly supports SDG 9: Industry, Innovation, and
Infrastructure by ensuring digital solutions and data-driven interventions are made accessible for use
in health centers. Increased use of such types of forecasting models in hospital operations is an
indication of innovation in health infrastructure. It shows the potential of intelligent solutions to
enhance service provision in an industry historically under strain to perform with limited resources.
By stimulating technological advancement in health, the project enhances resilience and
efficiency in hospitals. Additionally, the system facilitates the achievement of SDG 10: Reduced
Inequalities by providing equitable access to ICU treatment. The healthcare system in most
developing and resource-constrained countries tends to be over-stretched. Better bed allocation of
ICU beds by a predictive system can help rank patients according to need and urgency, thus reducing
inequalities in care access. By removing delays and bottlenecks, all patients—regardless of socio-
economic status—are bound to receive treatment in a timely and quality fashion. From the
sustainability perspective, optimization of resources in hospitals reduces unnecessary operational
costs.
Reducing long stays in ICU, which otherwise can be avoided with proper planning, not only
reduces the cost for hospitals but also the environmental impact due to energy-consuming ICU wards.
This is in relation to SDG 12: Responsible Consumption and Production, where optimal utilization
of medical infrastructure reflects against the idea of minimizing wastage and maximizing utilization
of available infrastructure. Finally, the project opens up long-term hospital operational reforms and
can be replicated to support national and global healthcare programs.

4
 Through the integration of data analysis, clinical expertise, and operation strategy, it results in
the development of a healthcare model that is not only reactive but proactive. Therefore, it is an
example of how technology and innovation can be utilized to attain sustainable development goals in
public health and well-being.

Aside from health impact, the project also advances SDG 11: Sustainable Cities and
Communities, particularly for city hospitals with continuous patient flow. By optimizing ICU space
design and mitigating bottlenecks in treatment, the project contributes indirectly towards constructing
superior urban health systems. They are better able to respond to emergencies and can provide quality
care even with high-density settings.

Besides, the project is in line with SDG 17: Partnerships for the Goals, as it brings together
collaboration among healthcare professionals, technologists, and research scholars. It requires
successful implementation of such a system to be often a collaboration among universities, hospitals,
software developers, and data scientists. Such collaborations open up avenues for knowledge sharing,
capacity building, and development of integrated systems deployable at scale for public good.

Finally, the project responds to the need for technology-driven, sustainable health reform
necessary in the context of increasing global health issues. Whether it is coping with looming
pandemics or coping with an aging population, the capacity to forecast and plan ICU resource
consumption becomes progressively more important. This book, despite being written with a
particular hospital function in mind, can be used as a starting block in the collective effort towards
achieving health equity, resilience, and sustainability for everyone.

5
 CHAPTER 2

LITERATURE SURVEY

2.1 Machine Learning in Healthcare and Disease Prediction

Machine learning (ML) has transformed the healthcare industry by allowing predictive analytics
and decision support systems to improve patient care and operational effectiveness, as shown by
Alsinglawi et al. [1] and Su et al. [4]. In the era of data-driven medicine, ML models are able to dig
out information from huge amounts of electronic health records (EHR), reveal latent patterns and
associations that can be used to inform clinical decisions, as demonstrated in Alghatani et al. [5].
Disease prediction and hospital outcome forecasting, such as predicting the length of stay (LOS) in
intensive care units (ICU), have been particularly facilitated by the use of supervised machine
learning algorithms, including ensemble techniques, support vector machines, and neural networks,
as discussed in Staziaki et al. [3]b .

Machine Learning (ML) has emerged as a transformative force in the healthcare sector, for it
enables data-driven decisions, early disease detection, together with personalized treatment strategies,
which is also evident in the findings of Alsinglawi [2]. ML algorithms are able to uncover patterns
which are often too complex for customary statistical approaches through analyzing huge amounts of
medical data such as electronic health records (EHRs) and diagnostic images, a challenge addressed
effectively by Rocheteau et al. [7]. ML supports the development of personalized medicine also.
Treatment plans are customized for an individual’s specific genetic makeup, lifestyle, and health
history there. Techniques such as clustering as well as classification help to segment patients into
groups that have similar traits. This segmentation does make it easier for one to prescribe the most
effective treatment. For drug dosage optimization and therapy scheduling, reinforcement learning is
also especially explored in chronic disease management like HIV and cancer, as noted in Islam et al.
[15]. Healthcare is going from a reactive state to a proactive one with these advances. ML applies
powerfully for predicting disease and early diagnosis. Decision Trees, Support Vector Machines
(SVM), and Neural Networks are algorithms used successfully to predict diseases like diabetes,
cancer, heart conditions, and neurological disorders, as supported by Yeh et al. [13].

6
 This identification allows for preventive care or treatment early on. By a minimizing of late-
stage intervention needs, this then reduces the healthcare infrastructure burden and improves survival
rates. The prediction of ICU LOS is a complex task influenced by numerous factors, including patient
demographics, comorbidities, admission source, laboratory values, and treatment interventions,
according to Su et al. [4] and Blom et al. [6]. Traditional statistical models, although useful, often fall
short in capturing non-linear relationships and interactions among features. Machine learning
algorithms, on the other hand, offer the flexibility to model these complexities with greater accuracy,
which is supported by Li et al. [14]. For example, Random Forest and Gradient Boosting Machines
have been found to perform well in classifying patients according to their predicted LOS, particularly
when used in conjunction with feature engineering methods that identify temporal and aggregate
patterns from time-series data, as demonstrated in Rocheteau et al. [7].

Explainable AI (XAI) has become an essential complement to ML in healthcare, responding to

the requirement for transparency and trust in model predictions, as discussed by Alsinglawi et al. [1]
and Keegan et al. [12]. Methods such as SHAP (SHapley Additive exPlanations) and LIME (Local
Interpretable Model-agnostic Explanations) enable clinicians to see which features had the greatest
impact on a prediction, thus building confidence in the recommendations made by the system, as
pointed out in Alsinglawi [2]. The incorporation of ML and XAI into healthcare procedures is hence
not just about attaining high accuracy but also maintaining accountability and user trust, which has
been emphasized in studies by Knaus et al. [10], [11]. Interpretability is critical in clinical settings
where decisions have direct implications on patient outcomes.

The incorporation of ML and XAI into healthcare procedures is hence not just about attaining
high accuracy but also maintaining accountability and user trust, which has been emphasized in
studies by Knaus et al. [10], [11]. In ICU administration, prediction of LOS early on through ML can
have a substantial influence on the availability of beds, staff, and treatment, as seen in Blom et al. [6]
and Alghatani et al. [5]. Hospitals operating with limited capacity, particularly during pandemics, can
use these predictions to make data-driven choices on admissions, discharge, and transfers, as shown
in Su et al. [4]. In addition, ML models can facilitate individualized treatment regimens by finding
subgroups of patients who share similar clinical profiles, thus enabling precision medicine strategies,
as highlighted by Alsinglawi et al. [1]. Overall, the contribution of machine learning in healthcare
continues to expand, providing predictive ability as well as operational knowledge to enhance care
delivery, as also discussed in Islam et al. [15].

7
2.2 Key Research Contributions in Outbreak Prediction
In the last decade, a large body of work has been directed toward applying machine learning to
predict clinical outcomes, specifically ICU length of stay. Different models and frameworks have
been developed, each building on a more nuanced understanding of how clinical and operational
factors drive patient trajectories in critical care environments, as explored by Alsinglawi et al. [1] and
Staziaki et al. [3]. Among these advances, ensemble learning techniques like XGBoost, CatBoost,
and Random Forest have become popular because they can handle intricate feature interactions and
yield high predictive power, as supported by Alghatani et al. [5].

Research has indicated that the inclusion of heterogeneous data types, such as laboratory tests,
vital signs, demographic features, and past hospital utilization patterns, greatly enhances LOS
prediction model performance, which is consistent with findings from Su et al. [4] and Rocheteau et
al. [7]. Such a study utilized CatBoost, which is a categorical data-optimized algorithm, to predict
LOS on real and simulated hospital data. The outcome revealed that CatBoost performed better
compared to common algorithms in classification as well as regression scenarios, having high
accuracy with high robustness against missing and imbalanced data, as reported by Alsinglawi [2].
Its potential to handle big feature sets with minimal preprocessing was especially useful for clinical
uses.

Recent studies have also highlighted hybrid approaches that merge epidemiological models with
machine learning approaches to address the prediction of outbreaks. Epidemiological models like the
SEIR (Susceptible-Exposed-Infectious-Recovered) model are usually limited with their staticity.
However, a hybrid model consisting of SEIR and machine learning models such as Random Forest
or Gradient Boosting is better able to adjust and capture changing parameters from data in the real
world, as reflected in Alsinglawi et al. [1]. Hybrid approaches possess the capability to analyze
nonlinear relationships between factors affecting disease transmission, including behavior of humans,
environmental conditions, and responses of the healthcare sector on the spread of the disease.

These models have been an important part of the decision-making process during emergencies,
they help the decision makers understand best policies for lockdowns, vaccine roll out, and
distribution of health resources. Further, there has been interest in community-based participatory
surveillance, in which data collected from populations through mobile applications, wearable devices,
or by self-report surveys, is then used to track public health trends.

The evidence is building that crowdsourced data obtained in this manner, and processed using
machine-learning models, can serve as an early warning system to identify asymptomatic cases and
underreported areas, as discussed by Islam et al. [15].

8
 This option gives the population a sense of responsibility to help contribute to public health
safety and enhances surveillance in remote or marginalized areas. Together, these innovations
represent a shift towards a more decentralized, data-driven, and responsive public health surveillance
paradigm. Another significant contribution is from the incorporation of explainable AI methods into
LOS prediction models. SHAP values, for example, have been utilized to order the feature importance
of patient age, admission type, and lab results in extended stay prediction, as demonstrated in
Alsinglawi et al. [1], [2]. Such insights have been extremely useful for clinicians who wish to
comprehend and verify the rationale behind model predictions. In one instance, SHAP analysis
identified high creatinine levels and lower Glasgow Coma Scale scores as robust predictors of
extended ICU stays, consistent with established clinical risk factors, as supported by Su et al. [4].

The design of predictive systems has also changed to accommodate modularity and scalability.
Microservices-based architectures enable the separation of data preprocessing, model training,
inference, separate components. This modular approach enables easier updates, debugging, and
integration with hospital information systems, as discussed by Islam et al. [15]. Additionally, real-
time deployment pipelines have been constructed to handle incoming patient data in real-time and
supply dynamic predictions that evolve as new data emerges. Such a system illustrates predictive
modeling's role not as an intellectual exercise but as an empirical tool poised for clinical deployment.

In addition, academic and healthcare researchers have supported the efforts by developing open-
access data collections and simulation platforms for outbreak modeling. Sites such as HealthMap and
World Health Organization (WHO) and Centers for Disease Control and Prevention (CDC) datasets
have tremendously boosted research through the availability of quality data used in training and
testing models, as mentioned by Alsinglawi [2]. All these have been instrumental in facilitating the
international community of researchers to share, innovate, and respond swiftly to new public health
threats.

Last, hybrid models incorporating structured EHR data and unstructured data from clinical notes
and imaging data mark the horizon for LOS prediction studies. Natural language processing (NLP)
algorithms and computer vision algorithms have been applied to extracting useful features from
discharge summaries and CT scans, respectively, and enriching data available to predictive models,
as explored by Staziaki et al. [3]. Such inter-disciplinary endeavors reflect the diversity and
broadening applications of machine learning for ICU outcome prediction.

9
2.3 Limitations Identified from Literature Survey
In spite of tremendous progress in machine learning-based ICU LOS prediction, there are still
some limitations and gaps in research. One of the most enduring issues is the completeness and quality
of electronic health records. Most datasets have missing values, variable data entry, and non-
standardization across institutions, as pointed out by Alghatani et al. [5] and Rocheteau et al. [7].
These problems not only impact the performance of predictive models but also their generalizability.
Though methods such as data imputation and normalization assist to a certain degree, they are unable
to completely make up for the lack of important clinical variables, which are capable of biasing model
training and evaluation, as also mentioned by Li et al. [14].

A further limitation is that most studies have a narrow scope of binary classification of LOS,
usually defined as short or long stays. This simplification does not capture the entire range of patient
experiences and might not be sufficient for subtle clinical decision-making. Regression-based
methods, while more informative, are less commonly investigated because of the increased
complexity in assessment and interpretation, as shown in Alsinglawi et al. [1] and Su et al. [4].

In addition, most current models are trained and validated on data from a single institution or
within a particular geographic area, restricting their transferability to larger settings. The absence of
external validation makes the models' robustness and ability to transfer to other populations or
healthcare systems uncertain, as noted by Staziaki et al. [3]. While there has been tremendous progress
in outbreak prediction utilizing machine learning and data analytics, several limitations have been
identified in the current literature. A major issue is that there is often a lack of sufficiently high-
quality, real-time data. Most models rely on historic datasets that do not reflect the ever-changing
reality of outbreaks, a challenge emphasized by Islam et al. [15].

Additionally, the black-box construction of some deep-learning models can complicate the
interpretation of predictions; any rationale behind predictions is a significant consideration for
healthcare practitioners and policymakers, who will only be able to engage with the model if adequate
explanations are exhibited to justify a decision, as addressed by Alsinglawi [2].
In addition complexities arise in managing multiple data sources (e.g., clinical data,
environmental conditions, mobility trends, and behavioral trends). Integrating structured and
unstructured data continues to be an issue because of different formats, privacy, and data standards,
which, along with other reasons, raises many technical challenges, as observed by Rocheteau et al.
[7] and Li et al. [14]. While some models perform satisfactorily in a fixed research environment, they
tend to not generalize to rapidly changing real-world contexts like pandemics, where variants or
behavior changes can change the kinetics of disease transmission.

10
 The limitations described above highlight the necessity of more research, improving data
infrastructure, and the need for specific collaboration across disciplines to help develop the next
generation of surveillance and predictive systems that are more resilient and adaptable, as concluded
by Islam et al. [15].
Without extensive testing in diverse cohorts, it is challenging to know whether a model
developed in one dataset will perform consistently elsewhere. This task highlights the imperative for
cross-collaboration data-sharing initiatives and multi-institution studies, as recommended by Blom et
al. [6]. Interpretability is another key gap issue. Although certain investigations include explainable
AI approaches such as SHAP and LIME, there are many high-performance models, especially deep
learning ones, that are still black boxes. Clinicians understandably will not turn to predictions that
they do not understand or challenge, particularly in high-stakes settings such as the ICU, as described
by Alsinglawi [2].
Filling the gap between performance and interpretability is needed to create trust and uptake in
clinical practice. Although certain investigations include explainable AI approaches such as SHAP
and LIME, there are many high-performance models, especially deep learning ones, that are still black
boxes. Clinicians understandably will not turn to predictions that they do not understand or challenge,
particularly in high-stakes settings such as the ICU, as described by Alsinglawi [2]. Lastly, real-time
deployment and integration of ML models are underexploited. Most studies rely on retrospective
analysis without considering the infrastructural and technical requirements for real-time inference.
Embedding predictive systems into hospital workflows requires solid data pipelines, easy-to-use
interfaces, and continuous model monitoring to identify performance drift, as emphasized by Islam
et al. [15] and Keegan et al. [12].

2.4 Research Objectives

The key goal of this study is to create an accurate and interpretable machine learning model for
forecasting ICU length of stay upon patient admission. Leaning on electronic health records and
advanced ML techniques, the project will provide clinicians and hospital administrators with
actionable, real-time intelligence that empowers effective bed management, staff planning, and
treatment planning, as discussed by Alsinglawi et al. [1], [2] and Alghatani et al. [5]. The model will
be designed to take in a wide variety of patient information such as demographic, clinical, and
procedural features, and make accurate and timely LOS predictions, a strategy shown to be effective
by Su et al. [4].

11
 A secondary objective is to integrate explainable AI techniques such as SHAP into the prediction
pipeline for giving transparent model decision-making. Clinical adoption mandates interpretability,
which allows physicians to understand the rationale behind every prediction, as recommended by
Alsinglawi et al. [1] and Keegan et al. [12]. Through highlighting significant traits that affect LOS
estimates, the model will make users capable of making informed decisions and have trust in the
suggestions of the system, supporting arguments made by Knaus et al. [10], [11]. This dual emphasis
on accuracy and interpretability is a testament to the project's dedication to technical sophistication
and practicality.
The project aims to compare various machine learning techniques, such as Logistic Regression,
Random Forest, Gradient Boosting, XGBoost, and CatBoost, to determine the best method of ICU
LOS prediction. The performances will be benchmarked against various metrics including accuracy,
precision, recall, F1-score, and area under the curve (AUC), following evaluation frameworks similar
to those in Rocheteau et al. [7] and Li et al. [14].
Finally, the project will emphasize ethical concerns and data privacy. The methodology will
adhere to best practices in data anonymization and model validation to facilitate both ethical
compliance and scientific integrity, a concern underscored by Islam et al. [15]. Through these goals,
the project seeks to provide a substantive contribution to the healthcare analytics domain and enable
the continued shift towards data-driven clinical decision-making, aligning with the predictive
direction outlined in Staziaki et al. [3]. The comparative assessment will give detailed insights into
model trade-offs as well as serve as the input for the model selection process. Besides, the project
will establish a user-friendly web application implemented under the Flask framework to provide
real-time predictions for end-users. A user-friendly visualization of results interface, data entry
interface, and support for interpretation will be included in the application, echoing the modular and
scalable approach advocated by Alsinglawi [2] and Islam et al. [15].
By integrating the prediction engine with a frontend interface, the project ensures that the outputs
become accessible and usable by non-technical users like clinicians and hospital administrators. The
eventual vision is to design a system that can seamlessly integrate with existing hospital workflows
and enhance operation effectiveness, consistent with suggestions from Blom et al. [6]. The process
starts with a comprehensive requirement analysis involving collaborative efforts from healthcare
professionals, data specialists, and software engineers. During this first phase, project objectives like
prediction accuracy, real-time response, transparency, and ethical appropriateness will be defined in
terms of stakeholder expectations.

12
2.5 Project Backlog
The project backlog for this research is a formalized roadmap that captures the key development
stages needed to produce a predictive system that can predict ICU length of stay through explainable
machine learning. The process starts with a comprehensive requirement analysis involving
collaborative efforts from healthcare professionals, data specialists, and software engineers. During
this first phase, project objectives like prediction accuracy, real-time response, transparency, and
ethical appropriateness will be defined in terms of stakeholder expectations. Having a good
understanding of clinical workflows and what data can be made available, the team will define
realistic deliverables and performance metrics for success measurement.

The development of an interpretable ICU LOS prediction system is guided by specific user
stories that reflect the practical needs of different healthcare stakeholders. For instance, as a clinician,
the goal is to quickly understand whether a patient is likely to require a prolonged ICU stay at the
time of admission. This enables better early care planning, staffing allocation, and prioritization of
resources, as emphasized by Alghatani et al. [5] and Su et al. [4]. If LOS can be estimated at the
outset, clinicians can intervene earlier or order specialized care pathways, as discussed by Alsinglawi
et al. [1].

Once a neat and clean dataset is prepared, feature engineering will be done to enhance the
predictive power of the model. This will involve constructing derived variables incorporating time-
series patterns, severity scoring, and clinical variable interaction effects. Feature selection algorithms
will be used to find the most explanatory features, and dimensionality reduction can be employed to
eliminate duplicative information. By developing a quality feature set, the task of training the model
will be better equipped to detect meaningful patterns and relationships.

As a hospital administrator, the prediction system will support capacity planning by anticipating
bed turnover rates. During high patient load periods or crises such as pandemics, it allows informed
decisions on patient transfers or scheduling of elective surgeries, in line with the findings of Blom et
al. [6] and Rocheteau et al. [7]. As an ICU nurse manager, knowing anticipated LOS helps determine
workload distribution and shift scheduling, improving staff efficiency and patient care, a workflow
impact also noted in the outcomes from Hanson et al. [8].

As a policymaker or quality assurance officer, access to system-wide LOS data and its predictors
supports long-term decisions like ICU expansion, training investment, and identification of
bottlenecks. These insights provide the basis for reforms and better resource utilization, which is
consistent with suggestions from Alsinglawi [2] and Islam et al. [15].

13
 As a data analyst or AI researcher, having a transparent system with SHAP-based feature
explanations allows for validation, model improvement, and publication of interpretable clinical AI
tools. This kind of explainability and modularity has been discussed by Keegan et al. [12] and Li et
al. [14].

After the requirement gathering, data acquiring and preprocessing will be the following step.
Because of the privacy limitations involved in real-world hospital data, the project will work on a mix
of public data sets and techniques of creating synthetic data. These datasets will simulate actual
clinical situations, including different patient populations, vital signs, diagnosis codes, and treatment
histories. Preprocessing steps will involve handling missing values, normalization of features, and
categorical encoding. This stage will also focus on generating features known to be associated with
ICU stays, such as past admissions, comorbidities, and early laboratory results.

The outcomes from these user stories are not only improved individual patient care but also an
overall more efficient and responsive ICU management system. The goal is to meet clinical and
administrative needs while ensuring ethical AI deployment, patient data safety, and continuous
monitoring—challenges frequently encountered in real-world healthcare systems, as highlighted by
Staziaki et al. [3].

Model development will be the next step in the project, where multiple machine learning
algorithms—Logistic Regression, Random Forest, XGBoost, and CatBoost—will be utilized and
fine-tuned. All models will be trained on cross-validation procedures and tested against a holdout test
set. Accuracy, precision, recall, and F1-score metrics will be used for performance comparison. Care
will be taken to ensure that the model can generalize to unseen data, making it reliable and robust.

Later on, explainability will be incorporated into the pipeline through SHAP, which will give
qualitative and quantitative explanations of how each feature is contributing to the output of the
model. These explanations will be used as a foundation for creating an interactive web interface
designed with Flask. This program will be a real-time dashboard where input is taken from the users
in the form of patient data, predictions are made, and each result's justification is provided.

The web application will be subjected to intense usability testing with real end-users to further
hone the interface and confirm that it satisfies clinical requirements. Ultimately, the complete
integrated system will be tested in simulated real-world environments. This will involve stress-testing
the backend with heavy data loads and confirming the accuracy and responsiveness of the prediction
engine. Once tested, the system will be ready to deploy, with regular monitoring and periodic updates
ensuring that the model is always accurate as new clinical trends arise. The organized backlog
guarantees user-focused solution for ICU LOS forecasting.

14
2.6 Proposed work
The proposed research seeks to build an end-to-end, interpretable machine learning model for
forecasting the length of ICU stay from structured electronic health record (EHR) data. The model
begins by deriving relevant features from patients' history, including vital signs, laboratory tests,
admitting diagnoses, and demographics. As confidentiality of patients is of utmost importance,
synthetic data will be used in the early phases of the project so that actual hospital data can be
replicated without any compromise on confidentiality. Synthetic data will replicate closely the form
and nature of actual ICU records and offer a safe environment where modeling methods can be
experimented with.

Data preprocessing will involve some of the most crucial steps such as imputation of missing
values, encoding categorical features, and scaling continuous features. Furthermore, the process at
this stage will also implement outlier detection methods and test data distributions to determine any
anomalies that might impact model training. After preprocessing, the data will be partitioned into
training, validation, and test sets such that performance measures are run on unseen data to ensure
unbiased measurement.

Model development will proceed with the deployment of a range of machine learning models.
These will include Logistic Regression as a baseline comparator, followed by ensembling techniques
such as Random Forest, Gradient Boosting, and the more advanced XGBoost and CatBoost
algorithms. Every model will be hyperparameter-tuned using methods such as grid search or random
search to realize optimal predictive performance.

Evaluation metrics will include accuracy, precision, recall, F1-score, and AUC to represent a full
understanding of model performance at different classification thresholds. Explainable AI (XAI) will
be central to the framework, with SHAP being the primary feature importance explanation tool. SHAP
will offer visualizations of the impact of individual features on each prediction, offering transparency
and explainability to healthcare clinicians who will be using the system. This is crucial to guarantee
clinical acceptability of the model and to detect biases or omitted risk factors.

To realize the model's predictive power into real-world utility, a web application will be built
based on the Flask framework. The application will have an easy-to-use interface to input patient data
and observe LOS predictions accompanied by SHAP visualizations explaining each prediction. The
application will be made to fit well into current hospital IT infrastructure as well as workflows. The
application will be built to fit in seamlessly with current hospital IT infrastructure and processes.
Clinicians and hospital administrators will, through this interface, be able to make timely and
informed decisions about patient care and resource utilization.

15
 In summary, the work in consideration combines data preprocessing, model building,
explainability, and user interface design within one deployable system for predicting ICU LOS. The
project focuses as much on technical accuracy as on ethical responsibility and usability in the actual
clinical setting. This ensures that the predictive system not only works but also is trusted and widely
used by medical professionals.

2.7 Plan of Action (Project Road Map)

This project will identify its objectives and scope through set success criteria during the start
phase. Within this phase, the team will have engagements with vital stakeholders, comprising hospital
managers, clinicians, as well as data scientists, so that detailed requirements regarding the ICU LOS
prediction system are obtained. The goals for the project shall be created accordingly based on
stakeholders' needs and encompass predictive quality, interpretability, clarity in user interface, as well
as compliance with laws governing healthcare data privacy like HIPAA (Alsinglawi et al. [2]; Islam
et al. [15]). A detailed schedule will be prepared, emphasizing major milestones, projected
deliverables, and the use of both computational and human resources required for keeping the project
within time and within scope.

After the planning phase, the project will proceed to the data preparation and collection phase.
Because true patient records can be accessed only with the constraint of privacy, a synthetic dataset
that mirrors hospital EHRs will be created (Alsinglawi et al. [1]; Alsinglawi et al. [2]). The synthetic
data will mimic an extensive array of patient factors—demographics, lab test results, and treatment
history—that affect ICU LOS. Collected data will be preprocessed with tasks like missing value
imputation, normalization, and one-hot encoding categorical features (Su et al. [4]; Alghatani et al.
[5]). This will lead to a tidy and clean dataset ready for machine learning model training. Feature
engineering and selection will follow, which would uncover the most effective variables on LOS by
exploring patient age, vital signs, comorbidities, admission type, and lab results.

Feature engineering techniques will include lag variable generation and statistical aggregations
in order to identify temporal patterns of patient status and hospital response times (Rocheteau et al.
[7]). Interaction terms between features, such as age and disease severity or oxygen saturation and
diagnosis type, will also be explored. The designed characteristics must increase the model's
sensitivity and accuracy with which it can forecast LOS.

This ensures that the feature set encompasses not only static values but dynamic patterns
accountable for patient outcomes in intensive care settings. The project will then advance to the model
architecture and design stage. During this stage, the team will architect various machine learning
models, with ensemble methods such as XGBoost and CatBoost being prioritized due to performance

16
during prior LOS prediction tasks (Staziaki et al. [3]; Su et al. [4]; Alghatani et al. [5]). These models
will be designed to work on high-dimensional, complex health care data. Tree depth, learning rate,
methods of class balancing, and regularization of overfitting are some of the major architectural
choices. These architectural choices will make the model efficient enough to detect nonlinear
relationships between the data and render both classification and regression techniques feasible to
predict ICU LOS.

Having laid down the model architecture, the project will then advance to the training phase. All
the models will be trained on the synthetic ICU dataset. The hyperparameters like the number of
estimators, tree depth, and learning rate will be tuned using cross-validation and grid search
algorithms (Li et al. [14]). Training will be done in iterative cycles with accuracy, recall, and F1-score
checkpoints to monitor progress. This stage will make sure the models generalize well and learn
patterns that map well to unseen patient cases. Convergence of the model will be checked to ensure
that underfitting and overfitting don't occur for different patient subgroups.

After training is complete, test and evaluation will begin. Trained models will be compared with
the held-out validation set to measure predictive performance. Performance will be quantified using
performance measures such as accuracy, AUC, precision, recall, and F1-score. In addition, SHAP
(SHapley Additive exPlanations) will be employed to provide explanations of the impact of each
input feature on the decisions made by the model (Alsinglawi et al. [1]; Alsinglawi et al. [2]).
This explainability layer is such that clinical stakeholders are aware of why some predictions are
made, and hence trust in using the system. Any indication of bias or abnormal feature dominance will
be corrected in this phase in order to calibrate the models for deployment. The second phase will
incorporate the top-performing model into a real-time prediction and system. This module will review
patient data at ICU admission and repeatedly scan for patterns of high LOS risk. Upon detecting a
risk, alerts will be sent to care teams with suggestions such as early discharge planning, additional
monitoring, or resource redistribution. This system, as a backend service, will make predictive
analytics dynamically assist clinical workflows and aid hospitals in managing patient loads and
resources more effectively.

In addition to the system, a patient-facing and admin-accessible web application will be built
using Flask. The UI will allow clinicians and administrators to input patient data, view forecasted
LOS results, and view feature contributions using interactive SHAP plots (Alsinglawi et al. [1]). The
UI will be user-friendly, with clean and clear visualizations of predictions and confidences. This real-
time decision support system will enable better communication and permit frontline staff to make
time-effective action, hence maximizing the efficiency of ICU workflow.

17
 The project will next move into the user testing and feedback stage. The project will then proceed
into the user testing and feedback phase. This includes beta testing among hospital staff and clinical
practitioners for usability testing, accuracy, and responsiveness of web application. The feedback will
target UI design, alert clarity, response time, and compliance with hospital decision-making
requirements. According to such feedback, the system will be optimized to maximize user interaction
and remove any technological or interpretability barriers during the application process prior to its
mass deployment.

Finally, the project will move into deployment and upkeep. The combined platform—model,
web app. The system will be designed to scale with growing numbers of patients and datasets. Upkeep
will involve periodic updates to retrain the model using fresher datasets, deploy new research results,
and keep up with changing healthcare data standards (Islam et al. [15]). Real-world performance will
be monitored continuously, and feedback loops will be used for iterative improvements. This roadmap
includes all required phases for successful execution, focusing on data quality, sound model
development, and seamless real-time integration. The system will be a sophisticated ICU
management tool that will enable hospitals to make better resource allocation decisions and allow
clinicians to provide more informed and individualized care. By technology alignment with actual
clinical needs, the project aims to develop a practical, precise and interpretable LOS forecasting
solution that can easily fit into hospital procedures on a daily basis.

18
 CHAPTER 3

SPRINT PLANNING AND EXECTION METHODOLOGY

In this project of forecasting ICU length of stay (LOS) employing explainable machine learning
models, the team followed an Agile development methodology spread over three sprints. Each
sprint emphasized distinct phases of development, ranging from data processing and model training
to system integration and ultimate deployment. This iterative methodology permitted ongoing
improvement of the system in accordance with testing, feedback, and performance assessment. The
sprints were designed to provide a progressive and additive build-up of the project so that each stage
contributed functional and architectural value to the ultimate deployment. This chapter outlines the
objective, implementation plan, technological foundation, and results analysis of each sprint so as to
provide transparency in the workflow and effectiveness in execution.

3.1 Sprint I
3.1.1 Introduction
Sprint I set the tone for the rest of the project and centered around building a strong, interpretable
machine learning model that would forecast the length of stay (LOS) for ICU patients. Since ICU
care is so critical and resource-hungry, precise LOS forecasts are important for maximizing bed usage,
treatment planning, and general hospital efficiency. The overall goal of this sprint was to create a
model for making predictions based on structured healthcare information that would classify ICU
stays into either short or extended categories during patient admission. Such early classification
makes possible better clinical decision-making and operational planning, both critical functions in
high-stress ICU environments.

Early on, the team realized predictive accuracy would not be enough. For this model to be reliable
and used in actual clinical settings, it had to be interpretable and transparent. Therefore, the sprint not
only addressed creating high-performance algorithms, but also selecting models and techniques
allowing for explainable artificial intelligence (XAI). Traditional methods such as linear regression
or rule-based systems were felt to be too limiting in managing the non-linear, multi-faceted form of
ICU data.

19
 Therefore, the more sophisticated models such as tree-based ensembles were favored as they
were capable of learning complex feature interactions while providing feature importance
interpretation paths. This first sprint laid the groundwork for all following developments, securing a
solid technical basis and answering clinical requirements.

3.1.2 Data Augmentation and Preprocessing

Largely because of privacy limitations and restricted access to actual ICU patient records, the
initial dataset only contained 300 records—too few to support training a performing machine learning
algorithm. Synthetic data generation methods were used to alleviate this, resulting in a much larger
dataset with 3,000 records. These synthetic records were generated by observed patterns on open-
source ICU datasets like MIMIC-III and hospital datasets from Kaggle. The created data covered a
good variety of clinically useful features: patient information (age, gender), admission information
(type, place), vital signs, laboratory test results, and comorbidity scores. Data augmentation by this
created data allowed simulation of a greater diversity of clinical cases and enhanced the model's
power to generalize across diverse patient categories.

After data generation, an extensive preprocessing step was performed to condition the dataset
for modeling. Missing values were filled with median imputation methods for numerical and mode
for categorical variables to have no gaps impact model training. Outliers were detected and dealt with
via IQR-based capping. Age, blood pressure, and temperature numerical features were normalized
via z-score standardization to bring them onto a comparable scale. Categorical features like diagnosis
codes, ICU types, and gender were label encoded and one-hot encoded where required. Lastly, the
entire dataset was split into 80:20 training-to-test ratio with preservation of class distribution for
unbiased estimation of model performance. These preprocessing steps ensured clean, well-structured
training data ready to be fed into machine learning models.

3.1.3 Model Architecture

The design of the architecture started by benchmarking a number of baseline and state-of-the-art
models to determine their appropriateness for LOS classification tasks. Logistic Regression was used
as a baseline because it is simple and easy to interpret, but its performance constraints were quickly
evident in dealing with the complex, non-linear relationships found in ICU data. The group then
turned their attention to more advanced models, such as Random Forest, XGBoost, and CatBoost.
These ensemble-based algorithms using trees work well on structured healthcare data and provide
integrated management of missing values and categories. CatBoost emerged especially because it
performed extremely well with low preprocessing, hence efficient and easy to use.

20
 The final architecture was made modular and scalable. Each model generated a continuous
probability score reflecting the probability of a long ICU stay, which may be thresholded to activate
alarm systems or help clinicians with risk stratification. Feature importance was computed during
early training iterations using mutual information and Gini importance, which assisted in the
identification of notable variables like age, comorbidity score, and serum creatinine. By embracing
an architecture that struck a balance between predictive capability and interpretability, the team
guaranteed the model would be clinically useful and poised for integration into downstream systems
such as real-time alert modules and patient-facing dashboards.

3.1.4 Training Procedure

Model training for the selected models entailed a series of carefully planned experiments
designed to maximize performance and generalizability. Grid search and random search were utilized
to optimize hyperparameters to iterate over sets of tree depth, learning rate, regularization strength,
and number of estimators. The team also experimented with performance on different train-test splits
and stratified sampling techniques to check robustness of models across heterogeneous patient
subgroups. The training was done using mini-batches of size 32, and validation accuracy was
monitored every couple of epochs to see if it converged as well as overfitted.

Early stopping was utilized in order to cut training when validation loss of the model did not get
better past a certain epoch so as not to incur undue computation as well as model degradation. The
Adam optimizer was utilized wherever possible so as to facilitate better convergence, particularly
when utilized with neural models which were put into consideration by concurrent experiments.
Accuracy, recall, precision, and F1-score metrics were calculated at every epoch and returned for
examination on gains. CatBoost model ranked best with strong generalization over the test set and
little variance across training and validation results. This confirmed its choice for deployment in
future development stages and underscored the need for stringent training methodology in clinical
machine learning use.

3.1.5 Result Analysis

Following training, model performance was evaluated using an extensive evaluation framework.
Precision, recall, F1-score, and accuracy were computed to measure the predictive power of the
model. Besides the above measures, confusion matrices were created to analyze the true positives,
false positives, false negatives, and true negatives distribution. This was important to detect edge
cases where the model could be classifying patients wrongly, for instance, ascribing a high-risk
patient a low-risk label that would have dangerous clinical implications.

SHAP (SHapley Additive exPlanations) values were used to improve the interpretability of the
predictions. These gave a decomposition of how each feature contributed towards each prediction.
21
SHAP plots also pointed out the most impactful variables like age, creatinine levels, type of ICU
admission, and comorbidity numbers. This enabled clinicians and developers to verify the rationale
behind each prediction, thus building greater confidence in the system. The end CatBoost model was
over 96% accurate on the test set, and its high recall and precision levels indicated that it was both
reliable and clinically relevant. These findings showed the model ready for its integration into real-
time systems and constituted the technical foundation for the subsequent sprint into user interaction
and alerting mechanisms.

3.2 Sprint II
3.2.1 Introduction
Sprint II focused on transitioning from model construction to system integration by building an
operative predictive model constructed in Sprint I. This capability was vital for anticipatory resource
planning, especially in situations of high tension where ICU space and critical staff resources were
limited. Usability of the system was also given importance during this sprint, wherein the alerts
generated were easily accessible, understandable, and actionable. The aim was to have an system in
real-time that could present model predictions in a clear and understandable format through an
interactive interface. With the incorporation of backend predictive logic and interaction points on the
frontend, the system could alert the respective users about patients whose ICU stay was expected to
be longer. In achieving this, Sprint II laid down the foundation for developing a clinical decision
support tool that was both technologically feasible and user-centric. Building such a system required
cooperation between multiple technical layers, including backend development, API design, data
routing, and user interface building.

3.2.2 Backend and Web Framework

The backend for the system was built using Flask, a high-level Python web framework based on
its high scalability, security, and reliability. Flask was selected since it natively supports user
authentication, URL routing, and database management, simplifying the process. The first task was
to set up a PostgreSQL database where patient records, prediction outputs, and user profiles would
be stored and maintained.

The schema was designed such that the system would efficiently support increasingly more users
as well as data points and yet provide high reliability and data integrity. Input validation was also
handled in the backend, and critical operations were secured by role-based access control.
Restful APIs were subsequently used to allow communication between frontend and machine
learning model.

22
 The APIs allowed user input age, vitals, and lab tests to be passed from the server to the frontend,
where the latter then forwarded the data using the CatBoost model that was trained. The model
output—a probability score and classification result—was returned through the API to be displayed
to the user. It was ensured that the APIs were secure, lightweight, and well-optimized for processing
concurrent requests. The API layer was employed to test the response time, error management, and
overall stability of the APIs using tools like Postman. The API layer became imperative in ensuring
that the predictive model could be deployed into a working web environment, a grand system
integration achievement.

3.2.3 Frontend Development and Data Entry Interface

The frontend of the application was made to possess a simple yet effective user interface for
clinicians as well as administrative users. The interface was made using different tools like HTML,
CSS, Bootstrap, and JavaScript and was made with ease of use and accessibility in mind. The focal
point of the frontend was the data entry form where users could enter patient data including
demographic data, vitals, and comorbidities. All areas of input were client- and server-side validated
to maintain input errors as minimal as possible and to standardize data formatting. Visual hints and
tooltips were included to inform users about what type of data is being requested from which field
and to reduce the learning curve for non-technical users.

The frontend not only consisted of input fields but also visual outputs for showing prediction
output. When a prediction was triggered, the system presented a summary of the results with the
probability score of a long ICU stay and an explanation of the most important contributing features
based on SHAP-derived insights. This feedback was intended to enable clinicians to comprehend the
reasoning behind the prediction, which is essential for clinical adoption. The UI also facilitated alert
logging, allowing users to see past predictions and associated patient outcomes. This provided
constant monitoring and backward analysis, again adding strength to the value of the system as a
decision-support tool.

3.2.4 Output Generation

The process of output generation was focused on converting raw model outputs into useful
clinical alerts. After the model produced a probability score for ICU LOS, the backend compared the
score with a predetermined threshold to determine whether the patient was high or low risk. If the
score was higher than the threshold, the system initiated an alert message, which was sent directly to
the frontend. These notifications were made to be brief but descriptive, emphasizing both the type
(short vs. long stay) and the key contributing features. Real-time processing was used to guarantee
that the users received their predictions immediately upon entering patient data.

23
3.2.5 Result Analysis
Sprint II testing was done with mock inputs to determine the efficacy of the system in processing
inputs and providing actionable information. The backend API responsiveness, the precision of the
prediction output, and the readability of the alerts were all quantified and examined. Response times
were on average less than two seconds, which was considered acceptable for clinical use. In addition,
users indicated that the prediction summaries and alerts were simple to comprehend and consistent
with their expectations of what a clinical support tool should provide.
In addition, a formal usability test was conducted in which participants used the system and
offered qualitative feedback. This feedback emphasized improvement areas including spacing
between layouts, feature description tooltips, and customizing alerts. Measurable indicators like user
satisfaction, perceived accuracy, and probability of future use were high, reflecting successful
deployment. Real-time system logging under load revealed well-behaved system behavior, few API
errors, and no system crashes. These results showed the successful completion of Sprint II and
established a useful benchmark for improving user experience and increasing system features in
Sprint III, where mobile responsiveness and richer user interaction would take precedence.

3.3 Sprint III

3.3.1 Introduction
Sprint III was the last and most integrative sprint of the project, which was intended to integrate
the predictive model and alert logic into a single patient-facing and administrator-support system.
The main objective of this sprint was to develop a working interface and communication mechanism
by which patients and hospital personnel could receive actionable insights in real-time. The attention
turned from strictly technical deployment to end-user experience, with a focus on intuitive design,
timely alerting, and smooth interaction between the machine learning backend and the user interface.
Since the model had already been shown to be accurate and robust in Sprint I, and its integration
tested in Sprint II.

One of the most important parts of this sprint was making sure the system was able to provide
sound information to the user based on real-time or near-real-time data. This involved improving
threshold logic, performance under load, and adding more user interaction points. Preparing the
platform for wider deployment, ensuring stability, accessibility, and responsiveness were also part of
this sprint. To make this a reality, there were improvements on both frontend and backend processes
such that the exchange of information between the predictive engine and the end-user dashboard
continued in a regular, uninterrupted flow. By the completion of Sprint III, the system was no longer
just predicting accurate ICU LOS, but in presenting those predictions in a coherent, understandable
form to end users.

24
3.3.2 Process Steps for Model Integration and System Workflow
In order to provide guaranteed functionality, the development team first charted the pipeline of
data flow and interaction between predictive model, backend server, and user interface. The process
began with the creation of input parameters for the model and defining how real-time patient data
were to be gathered, validated, and passed through the system. Secondly, logic was added to define
the conditions upon which an alert was to be created. These included establishing a classification
threshold for the probability score produced by the model, above which the system would trigger a
risk alert for extended ICU stay. Careful selection of the threshold was required to achieve sensitivity
and specificity balance, minimizing false positives while avoiding high-risk patients being missed.

The integration process also included extensive end-to-end testing with synthetic patient
scenarios to mimic varied clinical cases. Each subsystem of the system—from data upload and
prediction computation to alert visualisation—was tested for working and stability. RESTful APIs
formed the framework of communication such that data were transferred reliably from one module
to another.

Security aspects were reinforced to guarantee confidentiality of data, particularly during API
transactions. Once functional stability was attained, the system was rolled out to a controlled test
environment and was monitored for bottlenecks in performance, response times, and user behavior
patterns. These process steps enabled a unified, integrated experience where predictions and alerts
could be processed in real-time with low latency.

3.3.3 Model and System Architecture

The final architecture was a multi-layered one with the balance of predictive accuracy and user
accessibility. The central theme of the architecture was the CatBoost model that was trained on
structured clinical data to make predictions for ICU LOS. This model was incorporated in the backend
infrastructure developed using Flask such that inputs from users in real-time could be processed
instantly.

The system employed the use of modular design principles in such a way that the pieces of
prediction model, database management, and handling of alerts were independently but cohesive.
This also provided scalability in a manner whereby new additions like new models or new data
sources could be inserted with little future refactoring.

25
 Fig 3.1 Architecture Diagram for ICU LOS prediction

This fig 3.1 architecture diagram outlines the end-to-end pipeline for ICU LOS prediction, beginning
with data collection, visualization, label encoding, and normalization. The processed data is split into
training and test sets, used to train various machine learning models including CatBoost. Model
performance is evaluated using accuracy, precision, recall, and F1-score.

Frontend UI, built using HTML, Bootstrap, and JavaScript, enabled users to enter health-related
details and see real-time predictions. Predictions were provided with SHAP-based explainability
outcomes showing the most important features in determining the decision. The backend also had a
PostgreSQL database to retain historical user input, predictions, and alert responses for longitudinal
analysis and system auditability. Seamless synchronization across data input, model inference, and
alert rendering was among the key architectural accomplishments of this sprint. Every action by a
user—from form-filling to glance at an alert—triggered backend activities that executed without
glitches, thus realizing the underlying vision of the project of real-time, interpretable clinical decision
support.

3.3.4 Result Analysis

The efficacy and performance of the completely integrated system were thoroughly examined at
the conclusion of Sprint III. Technical as well as user-oriented metrics were taken into consideration
to test the result. Technically, latency readings showed that the delay between the submission of user
input to the delivery of alerts was consistently below two seconds—well within the thresholds for
real-time clinical use. The alert generation system was also highly consistent in threshold accuracy,
with most alerts properly aligned with the model's output probabilities. Backend logs validated few
API failures, and server load tests demonstrated the system's ability to support multiple users at once
without degrading response time.

26
 User testing was also a major area of focus in this sprint. A clinical user group and students were
sampled and engaged with the deployed system and gave systematic feedback in the form of surveys
and interviews. The participants appreciated the clear alerts, the ease of the interface, and the
explainability provided by the SHAP-based feature explanations. Feedback also pointed out areas of
improvement, like providing more contextual assistance and risk stratification categories instead of
simple binary results. On the basis of this feedback, small adjustments were implemented to the user
interface and alert display logic. In general, the outcomes of Sprint III confirmed the system's
readiness for general application and proved that the combination of predictive accuracy and
explainable output could lead to significant interaction in clinical environments.

3.3.5 Limitations and Future Work

Although the system was successful in reaching its key goals, a number of limitations were noted
that should be addressed in future development rounds. The use of synthetic data, while convenient
for initial model training, will not necessarily recreate the richness of real-world ICU patient
populations. While synthetic datasets were modeled to reflect natural patient distributions, they
necessarily lack the unexpected quirks and unusual edge cases inherent in actual hospital data. This
constraint can impair the model's generalizability when implemented in real-world clinical settings.
Also, demographic bias in the synthetic data can generate biased predictions in underrepresented
populations, a problem that needs to be addressed by real-world verification and ongoing model
improvement. Ongoing work will include integrating anonymized real-world patient data by
collaborating with health care providers to improve model training and testing. The team also intends
to add the system's functionality by implementing mobile accessibility, supporting push notifications
and providing notification of alerts to clinicians on handheld platforms. Other features of proposals
include user role-specific threshold customizability, multilingual user interfaces, and integration with
electronic health record (EHR) systems in hospitals. In addition, a continuous learning pipeline can
be adopted to enable the model to retrain from new data over time and, in turn, enhance accuracy and
adaptability. Breaking through these constraints and integrating real-world feedback, the system can
develop into an all-encompassing, trusted mechanism for handling ICU processes and enhancing
patient care outcomes.

27
 CHAPTER 4

RESULTS AND DISCUSSIONS

4.1 Project Outcomes

The ultimate aim of this project was to predict ICU length of stay (LOS) in patients by using
various machine learning models and comparing them with the help of conventional classification
methods evaluation metrics. The trained models are Logistic Regression, Multi-Layer Perceptron
(MLP), Random Forest (RF), Gradient Boosting (GB), XGBoost, and an advanced version, CatBoost.
With a stringent data preprocessing, training, and test performance sequence, the project unearthed
substantial trends and achieved impressive predictive performances everywhere. Preprocessing
pipeline ensured that the dataset was strong, clean, and model-ready. Exploratory visualization of
data, label encoding categorical features, scaling continuous variables, and missing values handling
were involved. The data was split into training and testing sets to ensure fair evaluation and prevent
overfitting.

Both XGBoost and CatBoost among the tested models performed better than the common models
like Logistic Regression and MLP in terms of accuracy and generalization ability. XGBoost had the
best overall accuracy, proving its strength in dealing with structured data with non-linear
relationships. CatBoost which is also known to deal with categorical data effectively, gave
competitive results and was especially beneficial when working with features like admission types,
comorbidities, and demographic variables. The performance metrics employed to compare models
were Accuracy, Precision, Recall, and F1-Score. These measures enabled a thorough assessment of
the quality of short, medium, and long ICU stays prediction by the models. For instance, CatBoost
scored 0.87 on F1-score of the test set, which is an excellent balance of precision and recall. Random
Forest exhibited stable performance with 84% accuracy, while Logistic Regression trailed by
moderate performance, as predicted from its linear nature. MLP, although less accurate than
XGBoost, was useful because it had the ability to fit complex, non-linear patterns

28
 Fig 4.1: Classification Performance (Accuracy Score)

Fig 4.1 shows The graph shows the accuracy of the different classification models, emphasizing
the better performance of the customized "Extension" model. In contrast to standard models such as
Logistic Regression and MLP, the Extension model produced the best accuracy.

What made the project effective was the use of explainable AI (XAI) methods. Through the use
of tools like SHAP (SHapley Additive exPlanations), we would be able to comprehend the
contribution of every feature to the model's predictions.For instance, age, comorbidities, severity
scores like APACHE II or SOFA, type of admission, and ventilation needs were the most significant
features in determining the duration of stay in the ICU. SHAP value visualizations gave a detailed
insight into how each feature pulled the prediction in the direction of a longer or shorter stay, thus
adding interpretation to otherwise black-box models such as XGBoost.

Another significant contribution of the research was the creation of a generalizable prediction
framework. The trained models were tested on unseen test datasets and produced consistent
performance, confirming that the model had not overfit the training dataset and could generalize to
new ICU cases. This is critical in clinical usage where model adaptability to novel data is an essential
requirement.

29
 Further, comparison between the models enabled us to establish trade-offs. For example, while
the highest F1-score was achieved by XGBoost, it was more computationally intensive to calculate
than Logistic Regression. That being said, when clinical stakes involve patient triage and resource
allocation in the ICU, prediction prowess over computational cost is paramount.CatBoost's lack of
need for extensive hyperparameter tuning and built-in support for categorical features rendered it both
powerful and easy.

This project also provided the groundwork for scalability. Once refined, the pipeline can be
applied to multiple hospitals or health systems, if the input features are aligned with what is necessary.
The model could also be scaled up to predict outcomes besides LOS, for example, the risk of ICU
readmission, mortality prediction, or ventilator dependency. All these wider use cases are supported
because of modular architecture and evident separation between layers of preprocessing, modeling,
and explainability.

Fig 4.2: Classification Performance (Precision Score)

Fig 4.2 This bar chart shows the precision scores of various classification models, with the tailor-
made "Extension" model having the best precision. It shows better sensitivity in detecting positive
cases than traditional models such as Logistic Regression and Random Forest.

30
 Fig 4.3: Classification performance (F1 Score)

Fig 4.3 shows the bar chart compares the classification performance of various machine
learning models based on their F1 scores. The custom "Extension" model outperformed all other
models, including XGBoost, Gradient Boosting, and Random Forest.

Fig 4.4: ROC AUC Curve for the Extension CatBoost model

Fig 4.4 shows the ROC AUC curve for the Extension CatBoost model indicates near-perfect
classification performance with an AUC close to 1.0.

31
4.2 Challenges and Improvements

While the project yielded acceptable outcomes, it also experienced a series of technical and
methodological issues which provide avenues for future enhancement. Among the earliest of these
were data imbalance problems where unequal higher numbers of short-stay ICU cases existed
compared to long-stay cases. Such imbalance may bias model performance, particularly if taken in
terms of accuracy alone. To counter this, techniques like SMOTE (Synthetic Minority Over-sampling
Technique) and weighting by class were tried but require further optimization to get full performance
on all LOS classes.The use of proxy variables was a limitation on prediction accuracy. More clinically
meaningful data collated—possibly from Electronic Health Record (EHR) systems—would
significantly enhance the model. For practical world advancements, hospital partnerships to obtain
de-identified real-world patient information can assist in overcoming this limitation as well as
facilitating external validity.

Computational complexity was also a consideration, especially when hyperparameter tuning for
such models as XGBoost and MLP. Grid search with cross-validation was time-consuming and
computationally expensive. Optimizing the process in the future could involve using methods such
as Bayesian Optimization or AutoML frameworks, which could automate model selection and tuning
without sacrificing performance.

Fig 4.5 Confusion matrix of CatBoost Model

Fig 4.5 shows Extension CatBoost model confusion matrix indicates that it wrongly classified a
significant number of 'Short Stay' records as 'Long Stay'. It made 136 accurate predictions of 'Long
Stay' cases but was unable to correctly identify any 'Short Stay'

32
 The other area of enhancement is in dynamic prediction ability. The model employing static
inputs at admission or shortly after is what is currently being used. Yet, patient status in the ICU
changes quickly. Adding time-series models that refine predictions as additional patient data is
received would dramatically improve prediction applicability. Techniques such as Temporal
Convolutional Networks (TCNs) or Recurrent Neural Networks (RNNs) could be investigated to
model ICU stay as a function of evolving patient data streams. Even hybrid models utilizing static
predictors in addition to dynamic time-series inputs could enhance accuracy and responsiveness.

From a deployment perspective, the current setup is more of a proof-of-concept and less an end-
to-end deployable solution. Further modules such as API integrations, security compliance
(HIPAA/GDPR), and clinician-friendly interfaces would need to be included for practical
deployment. Developing a web dashboard with visual explainability tools may enable easy integration
into hospital workflows. They might make it possible for clinicians to enter patient parameters, get
real-time predictions, and see contributing factors visually—thus taking model understanding and
turning it into clinical action. Perhaps the most crucial area for advancement is ethical management
of predictions. ICU length of stay predictions can have an impact on clinical judgment and resource
planning. As such, it is imperative to ensure that these tools remain decision support instruments and
not stand-alone decision makers. Human oversight, transparency, and clinical validation must remain
paramount to any future system created out of this model.

In summary, while the current project is a satisfactory step towards data-driven ICU
management, future growth must focus on enhancing data diversity, real-time adaptability,
computational efficiency, and clinical usability. A constant feedback loop between clinicians and
machine learning practitioners will be essential in bridging the gap between model development and
clinical implementation. With such enhancements, the system can grow into a powerful, scalable, and
ethical AI instrument for enhancing critical care provision.

From a deployment perspective, the existing system is more of a proof-of-concept than a mature,
production-capable solution. Though it poses a powerful predictor and an architecture with great
promise, a number of key pieces need to be integrated into advancing it as a viable tool that can be
usefully applied in actual healthcare settings. Some examples are robust API integrations into existing
hospital information systems (HIS), support for data security and privacy regulations such as HIPAA
and GDPR, and design of clinician-friendly user interfaces that align with actual medical workflows.
In particular, designing a simple-to-use web-based dashboard with built-in visual explainability
tools—such as SHAP or LIME plots—is most effective at boosting adoption.

33
 This would allow clinicians to just input patient parameters, receive instant predictions of ICU
length of stay, and view the features influencing them in plain sight. This transparency, and not just
confidence, but empowering clinicians to make informed decisions with greater certainty. But
perhaps most essential for development is the ethical use and management of predictive systems. ICU
length of stay predictions can greatly influence treatment prioritization, bed allocation, and clinical
decision-making. It is therefore vital to maintain the role of these models as decision support, not as
decision-makers. Maintaining human review, interpretability, and verifying results through clinical
trials are important to avoid over-reliance on algorithmic outcomes. In addition, bias and fairness
monitoring must be part of deployment to prevent disparate patient populations from receiving
unequal care.

In summary, though the model is an important advancement towards AI-assisted ICU treatment,
future refinement has to prioritize data diversity, real-time operation, scalability of computation, and
most importantly clinical utility and ease of use. Interdisciplinary collaboration among data scientists,
doctors, and administrators will continue to be critical to transform the system into a highly capable,
scalable, and ethically sustainable AI tool that maximizes the delivery of critical care.

34
 CHAPTER 5

CONCLUSION AND FUTURE ENHANCEMENT

5.1 Conclusion

This ICU Length of Stay (LOS) forecasting project illustrates the potential application of
machine learning to contemporary healthcare, particularly within the critical care setting where time,
resources, and decision-making are critical to the survival and recovery of patients. Utilizing patient
Electronic Health Records (EHR), the system was designed to predict the length of stay for ICU
admissions, thus allowing hospitals to enhance patient care, maximize resource utilization, and
enhance overall treatment procedures.

One of the success stories of this research was being able to deploy and compare an array of
machine learning models, i.e., Logistic Regression, Multi-Layer Perceptron (MLP), Random Forest,
Gradient Boosting, XGBoost, and CatBoost. The models were trained as well as tested on real ICU
datasets that included rich clinical features like patient demographics, comorbidities, vital signs, types
of admission, and severity scores. The performance of each model was evaluated extensively in
parameters such as Accuracy, Precision, Recall, and F1-score so that a strong and meaningful
comparison could be ensured.

In the results section, the prediction model worked efficiently in classifying ICU patients into
"Short Stay" and "Long Stay" categories. Among all the compared models, Extension CatBoost
achieved the best recall score, suggesting that it worked best to predict the long-stay patients. The
confusion matrix revealed that the prediction for the long stays was good with only a few short stays
being wrongly classified. In addition, the ROC AUC curve for the Extension model nearly reached a
perfect score, thus confirming its excellent discriminatory capability between the two
classes.mAmong all the models experimented with, CatBoost was the best performing algorithm with
an excellent accuracy of 98.25%. This may be because CatBoost is natively suitable to deal with
categorical data with less preprocessing, and on top of that, gradient boosting on decision trees for
accurate and stable learning.

35
 In addition to predictive precision, this project also emphasized explainability since healthcare
professionals not only need precise predictions but also explanations of the predictions that are
understandable. To meet this need, SHAP (SHapley Additive exPlanations) was used as an
Explainable AI (XAI) technique to provide an open transparent window to the operation of the model.
SHAP identified the most significant features that made a difference in ICU LOS predictions—e.g.,
patient age, comorbidities (e.g., diabetes and hypertension), source of admission, type of ICU
intervention, APACHE II scores, ventilator status, and initial lab results. These visual explanations
helped clinicians to understand the logic of the model and establish trust in its results.
Use of SHAP not only enhanced interpretability but also equaled the objective of clinical decision
support by providing medical practitioners with actionable information instead of mere abstract
results. For example, SHAP plots identified how increased SOFA scores and longer ventilator stay
always biased predictions toward a longer stay category, whereas early admission from emergency
departments and low initial severity scores indicated shorter stays.

From a systems viewpoint, the project also set the stage for constructing an AI-driven ICU
management system that would be implemented in hospital information systems. The architecture
was structured to support modules that could integrate various machine learning models, data sources,
and visualization tools seamlessly. As a proof-of-concept, this configuration demonstrated that an
interpretable and scalable prediction pipeline could not only be used to support outcome prediction
but also real-time ICU workflow planning, discharge time prediction, and bed availability prediction.

Fig 5.1: Dashboard of ICUStay website

Fig 5.1 shows the dashboard provides a user-friendly interface for predicting hospital ICU
stay duration using explainable machine learning models.

36
 Its broader impact is to introduce AI-powdered insights to the frontlines of healthcare, crossing
the divide from raw patient information to clinical decision. By processing historical ICU datasets
into predictive wisdom, the project provides a glance into how hospitals can shift the paradigm from
reacting to proactive modes of care management. This AI-fueled paradigm shift, in turn, may
eventually make way for enhanced patient outcomes, fewer ICU readmissions, and more efficient
deployment of staff load.

In addition, the high performance of the CatBoost model and explainable AI integration
highlights the increasing value of interpretable and ethical machine learning in medicine. Clinical
decisions tend to carry high stakes, and this project recognizes that any prediction system has no
business being a black box. The explainability aspect is thus not merely a technical aspect but an
ethical necessity, serving to ensure accountability, transparency, and clinical utility.

5.2 Future Enhancement

Although the project has shown excellent performance, there is sufficient room for improvement
that can really enhance the system's accuracy, scalability, and clinical usage. One of the main areas
of future research includes working with deep learning models—mainly Convolutional Neural
Networks (CNNs) and Long Short-Term Memory (LSTM) networks. CNNs may be employed to
obtain high-level features from multi-dimensional clinical data such as lab images or waveforms,
whereas LSTMs may be used to extract temporal dynamics in patient data, e.g., trends in vitals,
medication schedules, and symptom progression during ICU admission. The fact that LSTM is
designed to handle sequences makes it especially well-suited for dealing with time-series data, which
is prevalent in ICU monitoring systems
Moreover, subsequent iterations of this system might play with hybrid models that use classical
tree-based learners and deep neural networks. Ensembling or stacking these traditional interpretable
models with sophisticated pattern recognizers might yield both accuracy and transparency. For
example, stacking over CatBoost, LSTM, and Random Forest model predictions might be able to
perform better than any one of them, particularly for the scenarios involving heterogeneous patient
data. Enhancing feature engineering is another important direction. Although the existing system had
a solid set of clinical features, subsequent models might gain from incorporating derived features
(e.g., heart rate or oxygen saturation rate of change), interaction terms, or even latent variables that
have been found through unsupervised learning. These upgrades might enhance model depth and
detect more subtle trends in patient trajectories.

37
 In order to manage high-dimensional patient data, dimensionality reduction methods like
Principal Component Analysis (PCA) or Autoencoders can be integrated. These can help diminish
computational costs, remove multicollinearity, and enhance model generalization capability,
particularly when working with large hospital data.
Another dimension of improvement is real-time integration of the system within hospital
workflows. Today, the model is a batch prediction engine, but in clinical environments, predictions
must frequently be refreshed dynamically as new patient information arrives at the data streams.
Adding a real-time prediction module leveraging streaming patient data (e.g., vitals that are hourly
updated) could make the system more responsive and actionable. This would involve serving the
model as an API using APIs and supporting secure communication with EHR systems.
From the user interface and experience point of view, upcoming iterations of the system may
have clinician dashboards to engage with predictions. The interfaces may have SHAP-based
visualizations, graphs of historical trends, and notifications when a patient is forecast to have an
exceptionally long or brief stay. These types of tools can assist in ICU planning, discharge planning,
and intervention.
To enhance generalizability, the model also needs to be tested on various datasets from various
hospitals or regions. This would make the model invariant to variations in clinical practices, patient
populations, and EHR formats. Partnerships with medical institutions would provide access to de-
identified patient data, which would increase the validity and scope of the project. Federated learning
techniques may be integrated to enable model training across institutions without sharing data, thus
maintaining privacy while enhancing performance.
Furthermore, as a matter of ethical AI practices, the model ought to have fairness checks in place
to avoid biased predictions against particular groups, e.g., older patients, minorities, or patients with
unusual diseases. There ought to be recurring audits, techniques for mitigating bias, as well as
clinician feedback loops incorporated into subsequent iterations of the system to facilitate trustworthy
AI deployment.
Lastly, the system may also be able to factor in environmental and operational context, including
ICU levels of occupancy, availability of staff, and seasonal fluctuations in admissions. Such
contextual information usually has an effect on LOS but is seldom factored into clinical ML models.
Adding such information to the dataset could allow the model to take operational limitations into
account, rendering the predictions even more practical and applicable.

38
 Fig 5.2: Results of the testing data
Fig 5.2 shows the output shows ICU stay predictions (Long or Short) for test data samples using
the trained model on a web-based interface.
The model comparison employing various algorithms like XGBoost, Gradient Boosting,
Random Forest, MLP, and Logistic Regression identified that ensemble-based models fared better on
average. The in-house "Extension" model, though, excelled in recall and ROC-AUC score compared
to others. This implies that adaptive models, when optimized according to the dataset features,
provide significant gains in predictive power.
The dashboard deployment delivers real-time prediction outcomes for each patient data entry.
The result is easily seen to indicate the model's choice of ICU stay length, adding clinical transparency
and decision-making. Every prediction is supported by the most important patient parameters,
showing how the system incorporates explainable AI to enable interpretability in healthcare
environments.
In summary, this project provides the groundwork for a machine learning and explainable AI-
powered predictive ICU management system. Not only does it have high predictive performance, but
it also maintains transparency and ethical deployment through SHAP. With ongoing research and
development, the system can be further developed into a full-fledged clinical support tool that can
enhance ICU outcomes, resource planning, and patient care at scale. With the healthcare sector
growing more and more reliant on AI, this project is a blueprint of how data science can transform
critical car.

39
 The deployed dashboard delivers real-time ICU LOS predictions for any patient input, enabling
quick assessment of outcomes without delay. The predictions are also presented in a visually easy-
to-understand format, providing a clear indication of the estimated ICU stay length. By adding
explainable AI (XAI) features such as SHAP visualizations, the system is more than output and
contains reasoning for each prediction. The most significant patient parameters responsible for each
of the predictions are highlighted, promoting readability and transparency within clinical practice.
The most important patient parameters that affect each outcome are emphasized, ensuring
transparency and interpretability in clinical settings. This method enables medical personnel to not
only see what the model predicts but also why, thereby building trust and facilitating adoption in
sensitive healthcare environments.
Briefly, the project establishes an underlying framework of an AI-facilitated ICU decision-
making system with priorities in both predictive performance and interpretability in clinics. Through
its use of machine learning and understandable algorithms, it ensures that moral deployment is core
to it. Continuous improvement and verification utilizing real-world data makes it poised to become
an end-to-end clinical decision tool that maximizes ICU outcomes, streamlines operational efficiency,
and enhances care for patients. As AI plays a more central role in healthcare, this initiative is an
innovative example of how data science can be brought into critical care procedures.

40
 CHAPTER 6

REFERENCES

[1] B. Alsinglawi, O. Alshari, M. Alorjani, O. Mubin, F. Alnajjar, M. Novoa, and O. Darwish,

‘‘An explainable machine learning framework for lung cancer hospital length of stay
prediction,’’ Sci. Rep., vol. 12, no. 1, pp. 1–10, Jan. 2022.
[2] B. Alsinglawi, ‘‘Predictive analytics framework for electronic health records with machine
learning advancements: Optimizing hospital resources utilization with predictive and
epidemiological models,’’ Ph.D. thesis, Western Sydney Univ., Penrith, NSW, Australia,
2022. [Online]. Available: islandora/object/uws:67523/
[3] P. V. Staziaki, D. Wu, J. C. Rayan, I. D. D. O. Santo, F. Nan, A. Maybury, N. Gangasani, I.
Benador, V. Saligrama, J. Scalera, and S. W. Anderson, ‘‘Machine learning combining CT
findings and clinical parameters improves prediction of length of stay and ICU admission in
torso trauma,’’ Eur. Radiol., vol. 31, no. 7, pp. 5434–5441, Jul. 2021.
[4] L. Su, Z. Xu, F. Chang, Y. Ma, S. Liu, H. Jiang, H. Wang, D. Li, H. Chen, X. Zhou, N.
Hong, W. Zhu, and Y. Long, ‘‘Early prediction of mortality, severity, and length of stay in
the intensive care unit of sepsis patients based on sepsis 3.0 by machine learning models,’’
Frontiers Med., vol. 8, p. 883, Jun. 2021
[5] K. Alghatani, N. Ammar, A. Rezgui, and A. Shaban-Nejad, ‘‘Predicting intensive care unit
length of stay and mortality using patient vital signs: Machine learning model development
and validation,’’ JMIR Med. Informat., vol. 9, no. 5, May 2021, Art. no. e21347.
[6] M. C. Blom, K. Erwander, L. Gustafsson, M. Landin-Olsson, F. Jonsson, and K. Ivarsson,
‘‘The probability of readmission within 30 days of hospital discharge is positively associated
with inpatient bed occupancy at discharge—A retrospective cohort study,’’ BMC Emergency
Med., vol. 15, no. 1, pp. 1–6, Dec. 2015
[7] E. Rocheteau, P. Liò, and S. Hyland, ‘‘Temporal pointwise convolutional networks for length
of stay prediction in the intensive care unit,’’ 2020, arXiv:2007.09483.
[8] C. W. Hanson, C. S. Deutschman, H. L. Anderson, P. M. Reilly, E. C. Behringer, C. W.
Schwab, and J. Price, ‘‘Effects of an organized critical care service on outcomes and resource
utilization: A cohort study,’’ Crit. Care Med., vol. 27, no. 2, pp. 270–274, Feb. 1999
41
 [9] S. Siddiqui, S. Ahmed, and R. Manasia, ‘‘Apache II score as a predictor of length of stay and
outcome in our ICUs,’’ J. Pakistan Med. Assoc., vol. 55, no. 6, p. 253, 2005.
[10] W. A. Knaus, J. E. Zimmerman, D. P. Wagner, E. A. Draper, and D. E. Lawrence,
‘‘APACHE—Acute physiology and chronic health eval uation: A physiologically based
classification system,’’ Crit. Care Med., vol. 9, no. 8, pp. 591–597, Aug. 1981.
[11] A. Knaus, D. P. Wagner, E. A. Draper, J. E. Zimmerman, M. Bergner, P. G. Bastos, C. A.
Sirio, D. J. Murphy, T. Lotring, A. Damiano, and F. E. Harrell Jr., ‘‘The APACHE III
prognostic system: Risk prediction of hospital mortality for critically III hospitalized adults,’’
Chest, vol. 100, no. 6, pp. 1619–1636, 1991.
[12] M. T. Keegan, O. Gajic, and B. Afessa, ‘‘Comparison of APACHE III, APACHE IV, SAPS
3, and MPM0III and influence of resuscitation status on model performance,’’ Chest, vol.
142, no. 4, pp. 851–858, Oct. 2012.
[13] C.-C. Yeh, Y.-A. Chen, C.-C. Hsu, J.-H. Chen, W.-L. Chen, C.-C. Huang, and J.-Y. Chung,
‘‘Quick-SOFA score ≥ 2 predicts prolonged hospital stay in geriatric patients with influenza
infection,’’ Amer. J. Emergency Med., vol. 38, no. 4, pp. 780–784, Apr. 2020.
[14] C. Li, L. Chen, J. Feng, D. Wu, Z. Wang, J. Liu, and W. Xu, ‘‘Prediction of length of stay on
the intensive care unit based on least absolute shrinkage and selection operator,’’ IEEE
Access, vol. 7, pp. 110710–110721, 2019.
[15] M. M. Islam, T. N. Poly, and Y.-C. Li, ‘‘Recent advancement of clinical information systems:
Opportunities and challenges,’’ Yearbook Med. Informat., vol. 27, no. 1, pp. 83–90, Aug.
2018.

42
 APPENDIX A

CODING:

BACKEND:
# IMPORT MODULES
# TURN ON the GPU !

import os
from operator import itemgetter
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
get_ipython().magic(u'matplotlib inline')
plt.style.use('ggplot')

from sklearn.model_selection import StratifiedShuffleSplit

from sklearn.tree import DecisionTreeRegressor

from sklearn import preprocessing
from sklearn.preprocessing import Imputer
from pandas.tools.plotting import scatter_matrix
from sklearn.preprocessing import RobustScaler, StandardScaler, LabelEncoder, MinMaxScaler, OneHotEncoder,
LabelBinarizer
from sklearn.metrics import mean_squared_error, accuracy_score, mean_absolute_error
#from sklearn.cross_validation import KFold, cross_val_score
from sklearn.model_selection import cross_val_score, GridSearchCV, RandomizedSearchCV, KFold,
cross_val_predict, StratifiedKFold, train_test_split, learning_curve, ShuffleSplit
from sklearn.model_selection import train_test_split
from sklearn import model_selection, preprocessing
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
from sklearn.model_selection import cross_val_score, GridSearchCV, RandomizedSearchCV,KFold,
cross_val_predict, StratifiedKFold, train_test_split, learning_curve, ShuffleSplit
from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
from sklearn.naive_bayes import GaussianNB
from sklearn.svm import SVC
from sklearn.linear_model import SGDClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import GridSearchCV, ShuffleSplit
from sklearn.metrics import confusion_matrix, precision_score, recall_score, accuracy_score, f1_score
from sklearn.model_selection import cross_val_predict
from sklearn.metrics import confusion_matrix
from sklearn.metrics import precision_recall_curve, average_precision_score, auc
from sklearn.utils.fixes import signature
from sklearn.metrics import roc_curve

43
from sklearn.metrics import roc_auc_score

from mlxtend.plotting import plot_learning_curves

from mlxtend.preprocessing import shuffle_arrays_unison

import tensorflow as tf

from keras import models, regularizers, layers, optimizers, losses, metrics

from keras.models import Sequential
from keras.layers import Dense
from keras.utils import np_utils
from keras.utils import to_categorical

print(os.getcwd())
print("Modules imported \n")
import os
print(os.listdir("../input/"))
print(os.listdir("../input/"))
# Load MIMIC2 data

data = pd.read_csv('../input/mimic3d/mimic3d.csv')
print("With id", data.shape)
data_full = data.drop('hadm_id', 1)
print("No id",data_full.shape)
print(data_full.shape)
data_full.info()
data_full.describe()
data_full.head(10)
# Label = LOS

y = data_full['LOSgroupNum']
X = data_full.drop('LOSgroupNum', 1)
X = X.drop('LOSdays', 1)
X = X.drop('ExpiredHospital', 1)
X = X.drop('AdmitDiagnosis', 1)
X = X.drop('AdmitProcedure', 1)
X = X.drop('marital_status', 1)
X = X.drop('ethnicity', 1)
X = X.drop('religion', 1)
X = X.drop('insurance', 1)

print("y - Labels", y.shape)

print("X - No Label No id ", X.shape)
print(X.columns)
data_full.groupby('LOSgroupNum').size().plot.bar()
plt.show()
data_full.groupby('admit_type').size().plot.bar()
plt.show()
data_full.groupby('admit_location').size().plot.bar()
plt.show()
# Check that all X columns have no missing values
X.info()
X.describe()
# MAP Text to Numerical Data
# Use one-hot-encoding to convert categorical features to numerical
44
print(X.shape)
categorical_columns = [
'gender',
'admit_type',
'admit_location'
]

for col in categorical_columns:

#if the original column is present replace it with a one-hot
if col in X.columns:
one_hot_encoded = pd.get_dummies(X[col])
X = X.drop(col, axis=1)
X = X.join(one_hot_encoded, lsuffix='_left', rsuffix='_right')

print(X.shape)
print(data_full.shape)
print(X.shape)
#XnotNorm = np.array(X.copy())
XnotNorm = X.copy()
print('XnotNorm ', XnotNorm.shape)

ynotNorm = y.copy()
print('ynotNorm ', ynotNorm.shape)
# Normalize X

x = XnotNorm.values #returns a numpy array

scaler = preprocessing.StandardScaler()
x_scaled = scaler.fit_transform(x)
XNorm = pd.DataFrame(x_scaled, columns=XnotNorm.columns)
#print(XNorm)
#print(y)
print('X normalized')
# SPLIT into Train & Test

X_train, X_test, y_train, y_test = train_test_split(XNorm, y, test_size=0.2, random_state=7)

print ('X_train: ', X_train.shape)
print ('X_test: ', X_test.shape)
print ('y_train: ', y_train.shape)
print ('y_test: ', y_test.shape)
# Test Models and evaluation metric
seed = 42
scoring = 'accuracy'

# Spot Check Algorithms

Mymodels = []
#Mymodels.append(('LogReg', LogisticRegression()))
Mymodels.append(('RandomForestClassifier', RandomForestClassifier()))
Mymodels.append(('SGDclassifier', SGDClassifier()))
#Mymodels.append(('KNearestNeighbors', KNeighborsClassifier()))
Mymodels.append(('DecisionTreeClassifier', DecisionTreeClassifier()))
#Mymodels.append(('GaussianNB', GaussianNB()))
#Mymodels.append(('SVM', SVC()))

# Evaluate each model in turn

45
results = []
names = []
for name, model in Mymodels:
kfold = model_selection.KFold(n_splits=10, random_state=seed)
cv_results = model_selection.cross_val_score(model, X_train, y_train, cv=kfold, scoring=scoring)
results.append(cv_results)
names.append(name)
msg = "%s: %f (%f)" % (name, cv_results.mean(), cv_results.std())
print(msg)

# Optimize hyper params for one model

model = RandomForestClassifier()

param_grid = [{},]

grid_search = GridSearchCV(model, param_grid, cv=5, scoring=scoring)

grid_search.fit(XNorm, y)

print(grid_search.best_estimator_)
model = RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
max_depth=None, max_features='auto', max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, n_estimators=10, n_jobs=1,
oob_score=False, random_state=None, verbose=0,
warm_start=False)
# FEATURE IMPORTANCE - NORMALIZED - last model

trainFinalFI = XNorm
yFinalFI = y

model.fit(trainFinalFI,yFinalFI)

FI_model = pd.DataFrame({"Feature Importance":model.feature_importances_,}, index=trainFinalFI.columns)

FI_model[FI_model["Feature Importance"] > 0.005].sort_values("Feature
Importance").plot(kind="barh",figsize=(15,25))
plt.xticks(rotation=90)
plt.xticks(rotation=90)
plt.show()
# List of important features for model
FI_model = pd.DataFrame({"Feature Importance":model.feature_importances_,}, index=trainFinalFI.columns)
FI_model=FI_model.sort_values('Feature Importance', ascending = False)
print(FI_model[FI_model["Feature Importance"] > 0.001])
def plot_learning_curve(estimator, title, X, y, ylim=None, cv=None,
n_jobs=1, train_sizes=np.linspace(.1, 1.0, 10)):
plt.figure()
plt.title(title)
if ylim is not None:
plt.ylim(*ylim)
plt.xlabel("Training examples")
plt.ylabel("Error")
train_sizes, train_scores, test_scores = learning_curve(
estimator, X, y, cv=cv, n_jobs=n_jobs, train_sizes=train_sizes)
train_scores_mean = 1-np.mean(train_scores, axis=1)
46
 train_scores_std = np.std(train_scores, axis=1)
test_scores_mean = 1-np.mean(test_scores, axis=1)
test_scores_std = np.std(test_scores, axis=1)
plt.grid()

plt.fill_between(train_sizes, train_scores_mean - train_scores_std,

train_scores_mean + train_scores_std, alpha=0.1,
color="r")
plt.fill_between(train_sizes, test_scores_mean - test_scores_std,
test_scores_mean + test_scores_std, alpha=0.1, color="g")
plt.plot(train_sizes, train_scores_mean, 'o-', color="r",
label="Training score")
plt.plot(train_sizes, test_scores_mean, 'o-', color="g",
label="Cross-validation score")

plt.legend(loc="best")
return plt
# LEARNING CURVES Train / Validation

title = "Learning Curves "

cv = ShuffleSplit(n_splits=7, test_size=0.2)
plot_learning_curve(model, title, X_train, y_train, cv=cv, n_jobs=4)
#plot_learning_curve(model, title, XNorm, y, ylim=(0.01, 0.99), cv=cv, n_jobs=4)
# Split into Train & Test

X_train, X_test, y_train, y_test = train_test_split(XNorm, y, test_size=0.2, random_state=42)

print ('X_train: ', X_train.shape)
print ('X_test: ', X_test.shape)
print ('y_train: ', y_train.shape)
print ('y_test: ', y_test.shape)
# Model FINAL fit and evaluation on test

model.fit(X_train, y_train)
final_predictions = model.predict(X_test)

#final_acc = accuracy(y_test, final_predictions)

# Confusion matrix

conf_mx = confusion_matrix(y_test, final_predictions)

print('conf_mx ready')
def plot_confusion_matrix(cm,target_names,title='Confusion Matrix',cmap=None,
normalize=False):
import itertools
accuracy = np.trace(cm) / float(np.sum(cm))
misclass = 1 - accuracy

if cmap is None:
cmap = plt.get_cmap('Blues')

plt.figure(figsize=(8, 6))
plt.imshow(cm, interpolation='nearest', cmap=cmap)
plt.title(title)
plt.colorbar()

47
 APPENDIX B

CONFERENCE PRESENTATION

Our paper “Developing an Explainable AI Framework for Accurate ICU Length of Stay Predictions
to Optimize Hospital Resource Management”was presented at ICONDEEPCOM conference held at
SRM. 200+ shortlisted teams presented their papers on various fields in the conference. Our paper got

accepted as paper id : ICONDEEPCOM_154 with a plagiarism of just

On presenting the paper in this international conference held at SRM KTR campus, we received
positive remarks and suggestion from the judging panel.

48
Certificate from ICONDEEPCOM 2025

49
 APPENDIX C

PLAGIARISM REPORT

50
51