EARLY-STAGE PARKINSON'S DISEASE DETECTION

USING MACHINE LEARNING

A PROJECT REPORT

SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE AWARD OF THE DEGREE
OF

BACHELOR OF TECHNOLOGY
IN
SOFTWARE ENGINEERING
Submitted by:

AMIT BATRA
(2K20/SE/014)

LAKSHAY
(2K20/SE/077)

UNDER THE SUPERVISION OF

Dr. Abhilasha Sharma

DEPARTMENT OF SOFTWARE ENGINEERING

DELHI TECHNOLOGICAL UNIVERSITY
(Formally Delhi College of Engineering)
Bawana Road, Delhi-110042
MAY 2024
 DELHI TECHNOLOGICAL UNIVERSITY
(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

CANDIDATE’S DECLARATION

We, Amit Batra, 2K20/SE/14, and Lakshay, 2K20/SE/77 of B. Tech. (Software Engineering),

hereby declare that the project Dissertation titled "Early Stage Parkinson Disease Detection

using Machine Learning" which is submitted by us to the Department of Software

Engineering, Delhi Technological University, Delhi in partial fulfillment of the requirement

for the award of the degree of Bachelor of Technology, is original and not copied from any

source without proper citation. This work has not previously formed the basis for the award

of any Degree, Diploma Associateship, Fellowship, or other similar title or recognition.

Place: Delhi LAKSHAY

Date: 24 May 2024

AMIT BATRA

2
 DELHI TECHNOLOGICAL UNIVERSITY
(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

CERTIFICATE

I hereby certify that the Project Dissertation titled "Early Stage Parkinson Disease Detection

using Machine Learning" which is submitted by Amit Batra, 2K20/SE/14, and Lakshay,

2K20/SE/77, Software Engineering, Delhi Technological University, Delhi in partial

fulfillment of the requirement for the award of the degree of Bachelor of Technology, is a

record of the project work carried out by the students under my supervision.

Place: Delhi Dr. Abhilasha Sharma

Date: 24 May 2024 SUPERVISOR

Assistant Professor
Dept. of Software Engineering
Delhi Technological University, Delhi

3
 ABSTRACT

Current knowledge of the physio-pathological aspects of Parkinson's Disease suggests a

strong association with significant changes in the subcortical brain portions. Detecting the
structural alterations in these smaller regions, particularly during the quick stages of the
disease in de novo Parkinson's Disease (PD) patients, poses a challenge for neuro-
radiologists. The difficulty arises due to the intricate nature of these alterations, making them
elusive to traditional detection methods.

The lack of a trustworthy level of truth at the voxel level makes using traditional supervised
deep learning methods even more challenging. Our work uses an anomaly detection approach
to overcome this constraint. We investigate how well autoencoders (AE) perform in terms of
offering a productive anomaly scoring system using quantitative Magnetic Resonance
Imaging (MRI) data. According to our research, AE models present a useful tool for
Parkinson's disease early detection and treatment by providing a viable means of
differentiating de novo PD patients based on minute abnormalities in structural patterns.

4
 DELHI TECHNOLOGICAL UNIVERSITY
(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

ACKNOWLEDGMENT

“The successful completion of the task would not have been possible without the help of
people who supported us in our endeavor and whose encouragement and constant support
helped us throughout.”

Firstly, we are indebted to Dr. Abhilasha Sharma Assistant Professor (Department of

Software Engineering), and Ms. Roshni (Department of Software Engineering), Delhi
Technological University (Formerly Delhi College of Engineering), Delhi, for being our
guides and incorporating into us the idea behind the project assisting us in the undertaking of
the project, and for being ever available to guide us. Our future careers will be highly
influenced by the skills and knowledge we learned through this project. We also express our
thanks to all the faculties of the Software Engineering department for their impeccable
guidance, persistent upliftment, and generous support for this project work. We express our
wholehearted gratitude to our batchmates, who have directly or indirectly helped and
supported us in this undertaking.

AMIT BATRA

LAKSHAY

Place: Delhi
Date: 24 May 2024

5
 CONTENTS

Candidate’s Declaration i

Certificate ii

Acknowledgment iii

Abstract iv

Contents vi

List of Tables viii

List of Figures ix

List of Symbols/ abbreviations x

CHAPTER 1 INTRODUCTION 10

1.1 Background 10

1.2 Motivation 11

1.3 Problem Statement 11

CHAPTER 2: RELATED WORKS 12

2.1 Fuzzy C-Means (FCM) Clustering 12

2.2 Feature Subset Selection (FSS) 13

Techniques
2.3 Classification 16
CHAPTER 3: MATERIALS AND 17
METHODOLOGY

3.1 Datasets 17
3.2 Applied Methods 17
3.2.1 Artificial Neural Network (ANN) 17
3.2.2 Convolutional Neural Network (CNN) 19
3.2.2.1. General Model Of Convolution 20
Neural Network

6
3.2.3 K-Nearest Neighbors (K-NN) 20
3.2.4 Naive Bayes 21
CHAPTER 4: RESULTS AND 22
CONCLUSION

4.1 Results 22
4.1.1. Logistic Regression 22
4.1.2. K-Nearest Neighbors 22
4.1.3. Support Vector Machine 23
4.1.4. Random Forest: 23
4.1.5. Adaboost: 23
4.1.6. Multilayer Perceptron 24
4.1.7. Stacking 24
4.2 Conclusion 25
4.3 Future Works 26
REFERENCES 27

7
 LIST OF TABLES

Table 1 A literature review on applying ML techniques to diagnose 14

Parkinson's illness
Table 2 Results obtained using the Logistic Regression Model 21
Table 3 Results obtained using the K-Nearest Neighbors Classification 22
Model
Table 4 Results obtained using the Support Vector Machine Model 22
Table 5 Results obtained using Random Forest Model 22
Table 6 Results obtained using the Adaboost Model 23
Table 7 Results obtained using the Multilayer Perceptron Model 23
Table 8 Results obtained using Stacking 24

8
 LIST OF FIGURES

Figure 1 Normal Movement vs Movement disorders [media.net] 10

Figure 2 A schematic flow of fuzzy c-means (FCM) clustering algorithm 13
[researchgate.net]
Figure 3 Algorithm for medical image processing (MIP) using ML techniques 15
Figure 4 Methods applied to classify Parkinson’s Disease 16
Figure 5 A schematic diagram of the artificial neural network [23] 18
Figure 6 Elementary constituents of CNN 20
Figure 7 The Accuracy of the Models 24
Figure 8 F1 Scoring of the Models 24

9
 LIST OF SYMBOLS/ABBREVIATIONS

MLP Multilayer Perceptron 14

Table 2 Results obtained using the Logistic Regression Model 21
Table 3 Results obtained using the K-Nearest Neighbors Classification 22
Model
Table 4 Results obtained using the Support Vector Machine Model 22
Table 5 Results obtained using Random Forest Model 22
Table 6 Results obtained using the Adaboost Model 23
Table 7 Results obtained using the Multilayer Perceptron Model 23
Table 8 Results obtained using Stacking 24

10
 CHAPTER 1

INTRODUCTION

1.1 Background
Parkinson's disease is a neurodegenerative illness affecting the overall nervous system
and impairs speech and other motor activities. It is gradually more devastating. Since its
inception in 1817 [1], when Dr. James Parkinson published his groundbreaking research
and initially referred to the illness as "Shaking Palsy," Parkinson's disease continues to
be a significant global health issue. Classified among neurodegenerative disorders
characterized by the gradual breakdown of the nervous system, PD stands as the second
most common ailment in this group, following closely behind Alzheimer's disease [2].

Parkinson's disease is primarily caused by the progressive degradation of dopamine

neurons in the substantia nigra, a crucial midbrain region known as the "movement
control center" (Figure 1). Uncontrolled movements resulting from this neuronal loss
cause hypo-kinetic movement disorders. Even with its widespread occurrence and
significant influence on people's lives, Parkinson's disease (PD) is especially difficult to
manage due to the lack of a permanent cure or therapeutic intervention.

Figure 1: Normal Movement vs Movement disorders [media.net]

Despite decades of intensive study, the exact causes of Parkinson's disease remain
unknown. The onset of Parkinson's disease (PD) is thought to be influenced by a
complex combination of genetic and environmental variables [3], including exposure to
toxins, head traumas, living in rural areas, and pesticide exposure. Individual
differences in how symptoms present complicate diagnosis and therapy. The main
motor symptoms include limb stiffness, tremors, slowness of movement, and decreased
balance. The clinical picture is further complicated by non-motor symptoms such as
sadness and memory loss [4].

11
 Detecting Parkinson's disease (PD) in its early phases is crucial for successful
intervention. However, current diagnostic techniques have shortcomings. Traditional
assessment tools like the “Hoehn and Yahr scale” and “Unified Parkinson Disease
Rating Scale (UPDRS)” demand expertise and prolonged patient cooperation, posing
practical difficulties. To address the imperative for early and precise diagnosis, there is
a growing interest in exploring alternative approaches, specifically those rooted in
machine learning. These methods aim to improve diagnostic precision and facilitate
timely medical decisions.

Here, we start a thorough investigation of Parkinson's disease identification,

emphasizing the use of machine learning methods. The objective of our study is to
make a scholarly contribution by introducing a new methodology that utilizes multi-
modal data sources to improve diagnostic precision. By addressing the shortcomings of
existing diagnostic techniques, we want to enable earlier detection of the disease,
enabling people to live better lives.

1.2 Motivation
This research is motivated by the urgent need to address the hurdles associated with
quick detection and treatment of this disease (PD). Since this disease (PD) is a
neurological state that many people have, early intervention is essential to improve
patient outcomes. The diagnostic methods that are now in use have limitations in terms
of subjectivity, dependence on skilled personnel, and length of patient participation.
These methods mostly involve clinical assessments and rating scales. Given the wide
range of symptoms related to this disease (PD) and the lack of definitive therapy, our
study attempts to solve these problems through the use of ML techniques. We aim to
develop a more objective and efficient diagnostic model that can identify Parkinson's
disease (PD) at quick stage, even in the presence of mild symptoms, when therapy can
be most effective.

Our research aims to address the complexity of the intrinsic variability in Parkinson's
disease progression among individuals, which presents a difficulty for fast and accurate
diagnosis.

Furthermore, the need of creating a quick and efficient diagnostic tool is highlighted by
the worldwide influence Parkinson's disease has on public health and the heavy load it
takes on healthcare systems. Our research intends to utilize the power of these
algorithms to evaluate complicated datasets, perhaps uncovering patterns and
correlations that may defy standard diagnostic procedures. We do this by using the
rapid breakthroughs in machine learning. In an effort to close the gap between medical
knowledge and state-of-the-art technology, neurology, data science, and technology
have collaborated in an interdisciplinary manner on this project.

1.3 Problem statement

Given the critical importance of early intervention and the lack of conclusive cures, the
goal of this paper is to improve the current framework for the detection of Parkinson's
disease by overcoming the subjectivity and limitations inherent in traditional
assessments and utilizing deep learning techniques to develop an objective, efficient,
and accessible tool for the accurate detection of early-stage PD.

12
 CHAPTER 2

RELATED WORKS

Researchers in the academic community have utilized various characteristics and datasets to
formulate predictions regarding Parkinson's disease (PD). One prominent feature, as
demonstrated by the work of Indira et al. [5], is the analysis of the human biomedical voice.
The researchers concentrated on analyzing individuals' voices in their investigation, aiming to
construct a model with the ability to automatically detect the existence of Parkinson's disease.

2.1 Fuzzy C-Means (FCM) Clustering

The researchers used pattern recognition and fuzzy c-means (FCM) clustering
algorithms on the dataset to accomplish this goal. They achieved noteworthy
performance numbers with these approaches. In particular, the model's impressive
68.04% accuracy rate showed that it could properly categorize people as having or not
having Parkinson's disease. Furthermore, the model's sensitivity rate—which gauges its
capacity to accurately identify genuine positive cases—reached 75.34%, highlighting its
efficacy in identifying Parkinson's disease patients. The model's accuracy in
recognizing real negatives is measured by the specificity rate, which was 45.83%.

This work provides encouraging findings by using sophisticated clustering and pattern
recognition algorithms to the biomedical voice dataset, highlighting the importance of
using voice-related aspects in Parkinson's disease prediction. The results add to the
increasing amount of information that is being gathered to improve diagnostic skills and
comprehend the possibilities of voice-based biomarkers in the context of Parkinson's
disease diagnosis.

Using the L2 norm metric in combination with a support vector machine (SVM), novel
method (Amit et al.'s)—described in their paper [6]—classifies PD patients according to
their postural instability. As described in their study, the authors used two widely used
“scent identification tests to evaluate olfactory function among the Brazilian population:
the 16-item Sniffin Sticks identification test and the 40-item University of Pennsylvania
smell identification test (UPSIT-40)” [7].

In their investigation, the authors applied logistic regression independently to each of

the mentioned characteristics. The results revealed noteworthy findings regarding the
diagnostic performance of the two olfactory tests. Specifically, when utilizing the
Sniffin Sticks test, they reported a sensitivity of 81.1%, indicating the test's ability to
correctly identify individuals with PD, and a specificity of 89.0%, signifying its
accuracy in correctly identifying individuals without PD.

13
Similarly, the UPSIT-40 demonstrated robust diagnostic capabilities with a sensitivity
of 82.1% and a specificity of 83.5%. These metrics underscore the effectiveness of the
UPSIT-40 in correctly identifying individuals with PD and those without. The logistic
regression analysis allowed for a comprehensive examination of each olfactory test's
contribution to PD diagnosis.

Figure 2: A schematic flow of fuzzy c-means (FCM) clustering algorithm

[researchgate.net]

This study clarifies the importance of olfactory evaluations, particularly the UPSIT-40
and Sniffin Stick tests, in classifying people according to their postural instability as an
indication of Parkinson's disease. The results highlight the promise of olfactory testing
as a non-invasive and trustworthy technique for diagnosing Parkinson's disease patients,
and they offer significant new insights into the field of PD diagnosis.

Prashant et al. [8] employed features related to olfactory loss from the 40-item
University of Pennsylvania Smell Identification Test (UPSIT) and sleep behavior
disorder, assessed through the Rapid Eye Movement Sleep Behavior Disorder
Screening Questionnaire (RBDSQ). In their study, they utilized classification tree and
support vector machine (SVM) techniques to train their algorithms, achieving notable
performance metrics with an accuracy of 85.48% and sensitivity of 90.55%.

Expanding upon their earlier research, as described in [9], the authors added new
features to their investigation. This included imaging indicators from Single Photon
Emission Computed Tomography (SPECT) and measures of the cerebrospinal fluid
(CSF). With the addition of these additional features, diagnostic performance was
enhanced; reported sensitivity increased to an astounding 97.03%, and overall accuracy
increased to 96.40%. This improvement emphasizes how important it is to incorporate

14
 cutting-edge diagnostic markers into the Parkinson's disease prediction models, like
CSF measurements and SPECT imaging.

The findings from this study encourage further exploration in the field. The authors
emphasize the importance of continuing the investigation, indicating a pathway for
future research in the diagnosis of Parkinson's disease. The incorporation of
sophisticated machine learning models in the current research reflects an attempt to
further enhance the accuracy of predictive algorithms, showcasing a commitment to
advancing the capabilities of diagnostic tools for Parkinson's disease detection.

2.2 Feature Subset Selection (FSS) Techniques

The following are among the neurodegenerative disease diagnoses made possible by
machine learning techniques:
(1) Compiling information (e.g., local field potential, walking, speech, and brain MRI
images).
(2) Feature extraction, or the process of extracting features that are suitable for testing
and training a classifier.
(3) Selecting a few feature subsets to minimize the number of duplicated features.
(4) Creating, and assessing the overall performance of the chosen classifier.

The steps that are involved in using machine learning techniques for medical image
processing (MIP) are depicted in Figure 3.

Numerous machine-learning algorithms have been documented in scholarly works

including instance-based (IBL) and induction-based algorithms such as (ID3, and
CART) for medical image categorization. Nevertheless, these methods decrease the
prediction accuracy due to the abundance of features that are not even necessary to
predict the outcome. Consequently, there is a need for FSS techniques that improve
classification accuracy by selecting the right subset to maximize the number of features.
An FSS generally consists of four basic steps:
1. Subset Generation,
2. Subset Evaluation,
3. Stopping Criterion, and
4. Result Validation [10].

Using preset parameters, a search method called the "subset generation procedure"
generates feature subsets for evaluation [11]. Based on a variety of different evaluation
criteria, FSS algorithms are known to divide into three groups of models: (1) the
wrapper [10]; (2) the hybrid; and (3) the filter model. More distinction is made possible
by the way algorithms in each category investigate the space of feature subsets and the
details of their evaluation function.

Table 1 Literature review on applying ML techniques to diagnose Parkinson illness [12]

15
The filter model assesses and chooses the feature subsets based on the overall
characteristics of the data, without utilizing any kind of learning technique.
Occasionally, however, the basic filter approach fails to identify the relevant subset of
all features if the relevant criterion being applied differs from the one being trained
with. Moreover, there is a possibility that the filter approach will not be able to identify
a feature subset that will optimize the criterion, hence decreasing the learning process'
efficacy. model (13, 14). By contrast, the performance of the learning algorithm is used
as the assessment criterion in the wrapper technique.

If wrappers consider prediction accuracy, they can achieve even better outcomes than
others. However, wrapper models are less versatile and more computationally costly
than filter models since they need additional computing power and a particular learning
technique [15, 16]. Because filters function several times quicker than wrappers, they
have a considerably greater chance of scaling large databases with many features.
Moreover, filters for different learning algorithms don't need to be performed
repeatedly. As a result, filters and wrappers both have the potential to support learning.

Figure 3: Algorithm for medical image processing (MIP) using ML techniques

The typical hybrid model incorporates the advantages of the various filter and the
wrapper models by using different assessment criteria at various stages of the search
process [17]. Although hybrid approaches are more and more efficient than the filter
and wrapper approaches, only specific learning machines can use them due to their
increased complexity [18, 19]. This topic has also been the subject of several studies

16
 [20], with different researchers emphasizing the right advantages and disadvantages of
the defined filter and wrapper approaches. The FSS has the dimensionality reduction
methods for PD detection and classification utilizing machine learning techniques that
are currently available in the literature and are included in Table 3. These methods seek
to remove aspects that are superfluous or unnecessary or to decrease the dimensionality.

2.3 Classification
The classification stage comes after feature extraction and subset selection. This is an
example of supervised learning, which is characterized as the challenge of determining
which category a new observation will fall into.

Statistical algorithms, learning-based algorithms for pattern recognition, search

heuristics, and algorithm combinations are the three categories into which different
classification techniques fall.
The mean and standard deviation of the features in the template are computed in
statistical approaches.

Figure 4: Methods applied to classify Parkinson’s Disease

Methods to calculate distance such as “weighted Euclidean”, “Manhattan”, and

“Euclidean distance” are used to compare the training and testing data. Pattern
recognition is the act of collecting raw data and classifying it using clustering
techniques like K-means [21], artificial neural networks (ANN) [21], SVM, Bayes
classifier, K-NN rule, and others. Numerous evolutionary techniques, such as ant
colony optimization and particle swarm optimization, can be used for classification
[22]. One advantage of these transformative algorithms is their capacity to handle larger
databases. Figure 4 displays the ways to categorize Parkinson's Disease in this
investigation.

CHAPTER 3

MATERIALS AND METHODOLOGY

17
3.1 Datasets
The Parkinson's Progression Labels Initiative (PPMI) stands as a groundbreaking
experimental clinical study designed to comprehensively assess diverse cohorts through
behavioral and clinical evaluations, with the primary objective of identifying
biomarkers indicative of the progression of Parkinson's disease (PD). In the realm of
neuroimaging, biomarkers are measurements derived from images that signify the
presence, extent, or severity of an illness, thereby facilitating the early detection of
symptoms.

The dataset utilized in this study was generously provided by the PPMI, a research
initiative dedicated to uncovering precise biomarkers and advancing the early detection
of Parkinson's disease. The Siemens scanner was utilized to generate T1-weighted
(T1w) MRI images, which were made available in both DICOM and NIFTI formats for
three-dimensional MRI imaging.

This dataset, accessible through the Parkinson Parkinson Imaging Initiative (PPMI),
comprises a substantial collection of natural samples, image processing data, and
clinical information. It plays a pivotal role in furthering our understanding of
Parkinson's disease by providing a valuable resource for researchers. The Digital
Imaging and Communications in Medicine (DICOM) format encompasses 458 T2-
weighted MRI examinations, with 229 of these MRI datasets pertaining to individuals
with Parkinson's disease (PD), while the remaining datasets are from healthy controls
(HC). The repository is readily accessible at http://www.ppmiinfo.org.

Furthermore, the dataset has been organized into three subsets, each allocated for
testing, validation, and training with proportions of 10%, 15%, and 75% respectively.
This division ensures the robustness and reliability of the dataset for various stages of
machine learning model development and evaluation.

3.2 Applied Methods

3.2.1 Artificial Neural Network (ANN)

Artificial Neural Networks (ANNs) constitute a crucial component of artificial

intelligence, mirroring the structure and functions of the human brain to process and
interpret information. ANNs excel at unraveling intricate patterns within data due to
their composition of interconnected nodes or neurons. To leverage an ANN for tasks
such as classification or prediction, artificial neurons undergo a sequence of operations,
including accepting input, applying weighted sums, and passing the result through an
activation function.
The adaptability and versatility of ANNs make them invaluable in a myriad of
applications, spanning predictive modeling, natural language processing, and image

18
recognition. These networks are particularly adept at handling complex and non-linear
relationships within data, allowing them to discern patterns that might be challenging
for traditional algorithms.
The three fundamental layers of an ANN—namely the input layer, hidden layers, and
output layer—constitute the architecture responsible for its learning and decision-
making processes. The input layer receives and processes the initial data, while the
hidden layers iteratively refine and transform the information through weighted
connections and activation functions. The output layer eventually produces the
network's final result, whether it be a classification, prediction, or any other desired
output.
1. Input Layer:
○ The input layer receives the raw data, and each node represents a feature of the
input.
2. Hidden Layers:
○ By using activation functions and weighted connections, hidden layers process
the incoming data. The network can learn and retrieve complicated patterns
thanks to these layers.
3. Output Layer:
○ The output layer generates the ultimate forecast or outcome. Each node in a
classification task may represent a class, and the anticipated class is indicated by
the node with the highest activation.
4. Training:
○ The network modifies its weights during training by utilizing techniques like
backpropagation. This entails computing the error, comparing the expected and
actual outputs, and adjusting the weights to minimize the error.

Figure 5: A schematic diagram of the artificial neural network [23]

3.2.2 Convolutional Neural Network (CNN)

19
Convolutional Neural Networks (CNNs) represent a specialized class of deep neural
networks designed specifically for interpreting structured grid data, akin to the grid-like
structure found in images. Originating from the groundbreaking work of researchers
like Yann LeCun, CNNs have proven highly effective in revolutionizing computer
vision tasks.

At the core of CNN architecture are convolutional and pooling layers, which play
pivotal roles in capturing hierarchical features. The convolutional subcaste involves the
systematic movement of filters or kernels across input data, enabling the identification
of intricate patterns and the learning of hierarchical representations. This process is
fundamental to the superior performance of CNNs, particularly in tasks such as image
classification, object detection, and facial recognition.

Following the convolutional layers, pooling layers contribute to the efficiency of CNNs
by reducing spatial dimensions and enhancing computational effectiveness. This step
helps in retaining essential information while discarding less relevant details,
contributing to the network's ability to handle complex visual data.

The prowess of CNNs extends beyond computer vision applications. These networks
have become indispensable in various fields, including natural language processing and
medical image analysis. In natural language processing, CNNs excel at tasks like text
classification and sentiment analysis, leveraging their capacity to automatically extract
hierarchical features from structured grid-like data, such as word embeddings in
sentences.

In medical image analysis, CNNs have demonstrated remarkable success in tasks like
tumor detection and medical image segmentation. Their ability to automatically discern
hierarchical features within medical images has significantly contributed to the
advancement of diagnostic capabilities and treatment planning in healthcare.

3.2.2.1. General Model Of Convolution Neural Network

Artificial Neural Network (ANN) conventional architecture consists of one input and
output layer plus several hidden layers [24]. Equation (1) expresses how each neuron in
the network uses a particular function F to analyze an input vector X and produce an
output vector Y [25].

F(X,W) = Y equation (1)

The weight vector (W) in this equation denotes the strength of connections between
neurons in successive levels. After that, the weight vector that is produced is used to
categorize images. Although a large amount of research has been done on pixel-based
picture categorization, more recent work—most notably [26]—has shown that better
results are obtained when contextual information—like the image's shape—is taken into

20
account. Convolutional Neural Networks (CNNs) have gained popularity due to their
ability to classify images according to context.

Figure 6 demonstrates the core model of a CNN, having four main components: (a)
convolution layer, (b) pooling layer, (c) activation function, and (d) fully linked layer.
Each part contributes differently to improving the network's capacity to identify
complex relationships and patterns in the data.

Figure 6: Elementary constituents of CNN

3.2.3 K-Nearest Neighbors (K-NN)

The K-Nearest Neighbors (K-NN) algorithm is a popular and successful supervised

machine learning classifier that can be utilized for both regression and classification
tasks. This algorithm, as highlighted by Sarker et al. (2019), follows a lazy or case-
based learning approach. Unlike other algorithms, K-NN doesn't have a distinct training
phase; instead, the entire training dataset is employed for categorization.
The fundamental idea behind the K-NN algorithm is rooted in determining the class or
category of a new data point based on the classes of its k-nearest neighbors, as
measured by the Euclidean distance metric. In simpler terms, if we have a dataset with
two distinct classes, when a new case is presented, its class is established by examining
the classes of its k-nearest neighbors. These neighbors are identified based on their
proximity to the new case in the feature space, typically using Euclidean distance as a
measure of similarity.
The lazy learning approach of K-NN has its advantages, such as adaptability to changes
in the dataset without requiring a retraining phase. However, it also has its challenges,
such as potential inefficiency with large datasets and sensitivity to the choice of the
distance metric. The methodology of K-NN has been explored and validated by various
researchers. For example, Rani et al. (2020) and Uddin et al. (2019) have contributed to
the understanding of the algorithm's functioning and its application in different
domains.

This method allows for flexibility and adaptation to the features of the dataset by
guaranteeing that the categorization choice for a new instance is impacted by the
agreement of its closest neighbors. K-NN is a flexible and broadly applicable algorithm
in many different disciplines, where it shines in situations where the underlying data
distribution is non-linear or lacks a clear pattern. This is due to its simplicity and
intuitive methodology.

3.2.4 Naive Bayes

21
 The Naive Bayes (NB) algorithm is a probabilistic classification method based on
Bayes' theorem (Sarker et al., 2019). Well-known for its simplicity, NB has been
applied to a wide range of fields, including spam filtering, illness diagnosis or
prediction, and document categorization. Class conditional independence, according to
Murthy and Koolagudi (2018), is the core premise of NB and explains why a particular
characteristic inside a class is considered independent of all other features.

The chance that an item with given characteristics will belong to a specific class is
known as the posterior probability (P(m|n)), and it may be calculated using the
following formula in the Naive Bayes framework. The notation for the probability that
an attribute belongs to a class is by P(n|m) in this equation, whilst the prior class
probability and the attribute's prior probability are represented by P(m) and P(n),
respectively. Because of its straightforward methodology and foundation in
probabilistic reasoning, Naive Bayes is a useful and extensively used tool. It is
especially appropriate in situations where the assumption of feature independence
remains true or yields a good approximation.

CHAPTER 4

RESULTS AND CONCLUSION

4.1 Results
As discussed, we applied several machine learning models to the dataset, and using these
models the following results were obtained:
4.1.1. Logistic Regression:
An event, observation, or outcome's probability is predicted using this supervised
machine learning technique, which produces a dichotomous or binary result that can
only be true or false, 0 or 1, or yes or no. By applying the Logistic Regression model to
the dataset, the following results were obtained:
Table 2: Results obtained using the Logistic Regression Model

Metric Score

22
 Accuracy 73.3%
F1 Score 82.6%
Recall Rate 79.1%
Precision 86.3%
ROC AUC Score 64.5%

4.1.2. K-Nearest Neighbors:

A non-parametric supervised learning classifier, the K-Nearest Neighbours (KNN)
method saves every instance that is already available and uses a similarity measure to
classify incoming information or cases. A data point is primarily categorized by the
classification of its neighbors. By applying the K-Nearest Neighbors classification
model to the dataset, the following results were obtained:
Table 3: Results obtained using the K-Nearest Neighbors Classification Model

Metric Score
Accuracy 96.6%
F1 Score 97.9%
Recall Rate 100%
Precision 96%
ROC AUC Score 91.6%

4.1.3. Support Vector Machine:

A Support Vector Machine (SVM) is a supervised machine learning algorithm that is
used to solve complex classification and regression tasks by finding optimal
lines/hyperplanes having maximum distance between different labels/ outputs. By
applying the Support Vector Machine model to the dataset, the following results were
obtained:
Table 4: Results obtained using the Support Vector Machine Model

Metric Score
Accuracy 86.6%
F1 Score 92.3%
Recall Rate 100%
Precision 85.7%
ROC AUC Score 66.6%

4.1.4. Random Forest:

23
Random forest is a machine learning algorithm that creates an ensemble of multiple
decision trees to reach a singular, more accurate prediction or result. By applying the
Random Forest model to the dataset, the following results were obtained:
Table 5: Results obtained using Random Forest Model

Metric Score
Accuracy 90%
F1 Score 93.8%
Recall Rate 95.8%
Precision 92%
ROC AUC Score 81.2%

4.1.5. Adaboost:
The objective of Adaboost, a supervised learning classification method, is to enhance
overall prediction accuracy by iteratively training weak classifiers to predict the
likelihood of a target variable with binary outcomes. The following are the findings of
the above Adaboost Model:
Table 6: Results obtained using the Adaboost Model

Metric Score
Accuracy 83.3%
F1 Score 89.3%
Recall Rate 87.5%
Precision 91.3%
ROC AUC Score 77%

4.1.6. Multilayer Perceptron:

An MLP, or Multilayer Perceptron, represents a neural network utilized within
supervised learning contexts for classification tasks, where it predicts the probability of
a binary target variable with two possible outcomes. Using the Multilayer perceptron
model, the following results were obtained:
Table 7: Results obtained using the Multilayer Perceptron Model

Metric Score

Accuracy 93.3%

24
 F1 Score 96%

Recall Rate 100%

Precision 92.3%

ROC AUC Score 83.3%

4.1.7. Stacking:
Stacking is a machine learning ensemble method that combines predictions from
multiple base algorithms to create a new meta-learning algorithm. While applying to a
stack of KNN, SVM, Random Forest, AdaBoost, and MLP, the following results were
obtained:
Table 8: Results obtained using Stacking

Metric Score
Accuracy 93.3%
F1 Score 96%
Recall Rate 100%
Precision 92.3%
ROC AUC Score 83.3%

After evaluating all the models discussed above, the following observations were made
regarding their accuracy scores.

Figure 7: The Accuracy of the Models.

25
 Figure 3: F1 Scoring of the Models.

4.2 Conclusion

Studies show that 90% of people with Parkinson's disease (PD) have speech
impairment, which is typified by a forceful, strained, or raspy voice. Voice datasets
have been investigated in a number of research initiatives to automate Parkinson's
disease diagnosis. This study assessed the total classification, sensitivity, specificity,
geometric mean, and accuracy of Artificial Neural Networks (ANN), k-nearest
Neighbors (KNN), and Support Vector Machine (SVM) classifiers on a speech
database.
Using the Levenberg–Marquardt method, the ANN proved to be the most successful
classifier for speech datasets in this study, with an astounding 95.89% classification
accuracy. This research highlights the advantages that physicians may have when using
machine learning methods to diagnose Parkinson's disease, especially when it comes to
voice-related evaluations.
Although there are many different approaches available, there are recognized
limitations in the ways that Parkinson's disease is currently diagnosed. Therefore, more
advancements are needed to raise the accuracy of algorithms for computer-aided
diagnosis (CAD). Subsequent investigations will focus on incorporating other
evolutionary algorithms, including the Genetic algorithm, to enhance and optimize
machine learning systems for Parkinson's disease identification and classification. The
goal of this continuing study is to overcome current barriers and further advance
Parkinson's disease diagnostic technologies.

4.2 Future Works

Future possibilities for machine learning-based Parkinson's disease (PD) diagnosis
encompass a variety of techniques to improve the usefulness and accuracy of models.
Neural network models should undergo large-scale validation studies to confirm their
effectiveness, with an exploration into generalization across various Parkinsonian
disorders. Incorporating additional data sources, such as Inertial Measurement Unit
(IMU) data, could improve model accuracy, and longitudinal monitoring studies could
assess the models' ability to track disease progression. In addition, studying the gut-
brain axis and molecular pathways may offer a more thorough knowledge of
Parkinson's disease (PD). For solid findings, deep learning models must be verified

26
 using larger sample sizes and investigate a range of MRI measurements, including
supervised techniques for increased accuracy in various phases of Parkinson's disease.
The interpretability of models in real-world clinical settings should be a focus. Hybrid
LSTM-GRU models can benefit from advanced feature selection methods, addressing
biases from sampling techniques, and exploring real-world applications. Improved
interpretability can enhance clinical utility. For PD diagnosis systems, feasibility in
real-time clinical settings and integration with existing workflows are crucial, alongside
studies exploring longitudinal data analysis and interpretability.
Machine learning method improvements should explore diverse methods and increase
sample sizes for better generalizability, while also considering real-time application
feasibility. New machine-learning algorithms need comprehensive comparisons,
longitudinal validation, and practical implementation exploration. For efficient
detection using deep learning, integration of multimodal data, interpretability
improvements, and ethical considerations are essential. An improved approach for
prediction must delve deeper into non-motor symptoms, longitudinal data, and
integration of biomarkers. In voice-based diagnosis, extensive data collection,
differential diagnosis exploration, and benchmark establishment are needed. High-
accuracy detection through multimodal features requires larger and more diverse
samples, longitudinal data integration, exploration of non-motor symptoms, and
comparisons with existing diagnostic tools. These future works collectively aim to
refine machine learning models, making them more reliable and applicable for early PD
detection in clinical settings.

REFERENCES

[1] J. Parkinson, An Essay on Shaking Palsy. London: Whitting- ham and Rowland Printing,
1817.
[2] D. B. Calne, “Is idiopathic parkinsonism the consequence of an event or a process,”
Neurology, Vol. 44, no. 15, pp. 5–5, 1994.
[3] A Samii, J. G. Nutt, and B. R. Ransom, “Parkinson’s disease,” Lancet, Vol. 363, no. 9423,
pp. 1783–93, 2004.
[4] Wu, S. L., Liscic, R. M., Kim, S., Sorbi, S., & Yang, Y. H. (2017). Nonmotor Symptoms of
Parkinson’s Disease. Parkinson’s Disease, 2017, 1–2. https://doi.org/10.1155/2017/4382518
[5] Rustempasic, Indira, & Can, M. (2013). Diagnosis of Parkinson’s disease using Fuzzy C-
Means Clustering and Pattern Recognition. SouthEast Europe Journal of Soft Computing,
2(1).

27
[6] Amit S., Ashutosh M., A. Bhattacharya, F. Revilla, (2014, March).Understanding Postural
Response of Parkinson’s Subjects Using Nonlinear Dynamics and Support Vector Machines.
Austin J, Biomed Eng 1(1): id1005.
[7] L. Silveira-Moriyama, J. Carvalho Mde, R. Katzenschlager, A. Petrie, R. Ranvaud, E.R.
Barbosa, A.J. Lees, The use of smell identification tests in the diagnosis of Parkinson’s
disease in Brazil, Mov Disord, 23 (2008) 2328-2334.
[8] Parkinson’s disease detection using olfactory loss and REM sleep disorder features. R.
Prashanth-EMBS Member, S. Dutta Roy-IEEE Member, and P. K. Mandal, S. Ghosh.
[9] High Accuracy Detection of Early Parkinson’s Disease through Multi-modal Features and
Machine Learning. R Prashant, Sumantra Dutta Roy, Pravat K. Mandal, Shantanu Ghosh.
[10] M. Dash, and H. Liu, “Feature selection for classification,” Intell. Data Anal., Vol. 1,
pp. 131–56, 1997.
[11] L. Huan and H. Motoda, Feature Selection for Knowledge Discovery and Data
Mining. Boston, MA: Kluwer Aca-demic, 1998.
[12] Gunjan Pahuja & T. N. Nagabhushan (2018): A Comparative Study of Existing
Machine Learning Approaches for Parkinson's Disease Detection, IETE Journal of Research,
DOI: 10.1080/03772063.2018.1531730
[13] M. E. ElAlami, “A filter model for feature subset selection based on genetic
algorithm,” Knowledge-Based Syst., Vol. 22, pp. 356–62, 2009.
[14] H. Yoon, C-S. Park, J. S. Kim, and J-G. Baek, “Algorithm learning based neural
network integrating feature selection and classification,” Expert Syst. Appl., Vol. 40, pp.
231–41, 2013.
[15] R. Kohavi and G. H. John, “Wrappers for feature subset selection,” Artif. Intell., Vol.
97, pp. 273–324, 1997.
[16] M. Hall, “Correlation based feature selection for machine learning,” Ph.D.
dissertation, Dept. of Computer Science, University of Waikato, 1999.
[17] A. H. Hadjahmadi and T. J. Askari, “A decision support system for Parkinson’s
disease diagnosis using classification and regression tree,” J. Math. Comp. Sci., Vol. 4, pp.
257–63, 2012.
[18] O. Uncu and I. B. Turksen, “A novel feature selection approach: combining feature
wrappers and filters,” Inf. Sci., Vol. 177, pp. 449–66, 2007.
[19] J. Huang, Y. Cai, and X. Xu, “A hybrid genetic algorithm for feature selection
wrapper based on mutual information,” Pattern Recog. Lett., Vol. 28, pp. 1825–1844, 2007.
[20] D. Guan, W. Yuan, Y-K. Lee, K. Najeebullah, and M. K. Rasel, “A review of
ensemble learning based feature selection,” IETE Tech. Rev., Vol. 31, pp. 190–198, 2014.
[21] R. Das, “A comparison of multiple classification methods for diagnosis of Parkinson's
disease,” Expert Syst. Appl., Vol. 37, pp. 1568–72, 2010.
[22] G. Sateesh Babu, and S. Suresh, “Parkinson’s disease pre-diction using gene
expression – a projection based learning meta-cognitive neural classifier approach,” Expert
Syst. Appl., Vol. 40, pp. 1519–29, 2013.

28
[23] Lahiri, S., & Ghanta, K. (2009). Artificial neural network model with the parameter
tuning assisted by a differential evolution technique: The study of the hold up of the slurry
flow in a pipeline. Chemical Industry and Chemical Engineering Quarterly, 15(2), 103–117.
https://doi.org/10.2298/ciceq0902103l
[24] Lee, K. B., Cheon, S., & Kim, C. O. (2017) “A convolutional neural network for fault
classification and diagnosis in semiconductor manufacturing processes.” IEEE Transactions
on Semiconductor Manufacturing 30 (2): 135-142.Lima, E., Sun, X.
[25] LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998) “Gradient-based learning
applied to document recognition.” Proceedings of the IEEE 86 (11): 2278-2324
[26] Maggiori, E., Tarabalka, Y., Charpiat, G., and Alliez, P. (2017) “Convolutional neural
networks for large-scale remote-sensing image classification.” IEEE Transactions on
Geoscience and Remote Sensing 55 (2): 645-657.
[27] Lin, C. H., Wang, F. C., Kuo, T. Y., Huang, P. W., Chen, S. F., & Fu, L. C. (2022).
Early detection of Parkinson’s disease by neural network models. IEEE Access, 10, 19033-
19044.
[28] Prashanth, R., Roy, S. D., Mandal, P. K., & Ghosh, S. (2016). High-accuracy
detection of early Parkinson's disease through multimodal features and machine learning.
International journal of medical informatics, 90, 13-21.
[29] Rehman, A., Saba, T., Mujahid, M., Alamri, F. S., & ElHakim, N. (2023). Parkinson’s
disease detection using hybrid lstm-gru deep learning model. Electronics, 12(13), 2856.
[30] Lamba, R., Gulati, T., Alharbi, H. F., & Jain, A. (2021). A hybrid system for
Parkinson’s disease diagnosis using machine learning techniques. International Journal of
Speech Technology, 1-11.
[31] Celik, E., & Omurca, S. I. (2019, April). Improving Parkinson's disease diagnosis
with machine learning methods. In 2019 Scientific Meeting on Electrical-Electronics &
Biomedical Engineering and Computer Science (EBBT) (pp. 1-4). Ieee.
[32] Mandal, I., & Sairam, N. (2014). New machine-learning algorithms for prediction of
Parkinson's disease. International Journal of Systems Science, 45(3), 647-666.
[33] Sahu, L., Sharma, R., Sahu, I., Das, M., Sahu, B., & Kumar, R. (2022). Efficient
detection of Parkinson's disease using deep learning techniques over medical data. Expert
Systems, 39(3), e12787.
[34] Challa, K. N. R., Pagolu, V. S., Panda, G., & Majhi, B. (2016, October). An improved
approach for prediction of Parkinson's disease using machine learning techniques. In 2016
international conference on signal processing, communication, power and embedded system
(SCOPES) (pp. 1446-1451). IEEE.
[35] Wroge, T. J., Özkanca, Y., Demiroglu, C., Si, D., Atkins, D. C., & Ghomi, R. H.
(2018, December). Parkinson’s disease diagnosis using machine learning and voice. In 2018
IEEE signal processing in medicine and biology symposium (SPMB) (pp. 1-7). IEEE.
[36] Prashanth, R., Roy, S. D., Mandal, P. K., & Ghosh, S. (2016). High-accuracy
detection of early Parkinson's disease through multimodal features and machine learning.
International journal of medical informatics, 90, 13-21.

29
30