ISSN: 2320-5407 Int. J. Adv. Res.

12(05), 1118-1137

Journal Homepage: - www.journalijar.com

Article DOI: 10.21474/IJAR01/18827

DOI URL: http://dx.doi.org/10.21474/IJAR01/18827

RESEARCH ARTICLE
EVALUATING SPEECH ANALYSIS TECHNIQUES FOR PARKINSON'S DISEASE DETECTION: A
COMPARISON OF MACHINE LEARNING AND DEEP LEARNING ALGORITHMS

Anand Ratnakar
M.Tech Robotics and Artificial Intelligence, Dept. of Manufacturing Engineering and Industrial Management COEP
Technological University Pune, India.
……………………………………………………………………………………………………....
Manuscript Info Abstract
……………………. ………………………………………………………………
Manuscript History Parkinson's disease (PD) presents a diagnostic challenge due to its often
Received: 31 March 2024 subtle and gradual onset. Speech analysis offers a promising avenue for
Final Accepted: 30 April 2024 early detection, enabling intervention before the disease significantly
Published: May 2024 progresses. This study investigates the efficacy of supervised machine
learning algorithms in identifying PD using speech features. We
Key words:-
Parkinson's Disease (PD) Diagnosis, compared the performance of Logistic Regression (LR), Decision Tree
Speech Analysis, Artificial Neural (DT), Random Forest (RF), Support Vector Machine (SVM), Naive
Networks (ANN), Support Vector Bayes (NB), K-Nearest Neighbors (KNN), and Artificial Neural
Machine (SVM), Naive Bayes (NB), K
Nearest Neighbors (KNN), Logistic
Networks (ANN) for PD classification. Our findings demonstrate the
Regression (LR), Decision Tree (DT), superiority of ANNs, achieving a test accuracy of 97.44%, which
Random Forest (RF), Machine Learning surpasses existing benchmarks and highlights their potential for PD
(ML), Deep Learning (DL) diagnosis. This approach leverages readily available speech data,
potentially reducing reliance on expensive and time-consuming clinical
procedures. This research contributes to the development of non-
invasive, speech-based diagnostic tools for PD, paving the way for
earlier intervention and improved patient management.

Copy Right, IJAR, 2024,. All rights reserved.

……………………………………………………………………………………………………....
Introduction:-
Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by a decline in dopamine levels
in the brain. This deficiency manifests in tremors, rigidity, and difficulties with balance and coordination. As the
second-most common neurodegenerative disease after Alzheimer's disease, PD affects millions globally, with its
prevalence expected to rise due to an aging population.

Early diagnosis of PD is crucial for optimizing patient outcomes and management. However, traditional diagnostic
methods rely heavily on a patient's medical history and neurological examinations, which can lack sensitivity,
especially in the early stages. Additionally, definitive diagnostic tests for PD are not currently available.
This necessitates the exploration of new and potentially more objective methods for PD detection. Machine learning
(ML) and deep learning (DL) offer promising avenues to address this challenge. ML algorithms can learn complex
relationships between features extracted from data (e.g., voice, gait) and disease status. Deep learning, a subfield of
ML, utilizes artificial neural networks with multiple layers to automatically discover these relationships from raw
data without the need for extensive feature engineering.

This research investigates the potential of both ML and DL techniques to analyze voice and potentially other
relevant data modalities for improved PD diagnosis. We compare the performance of various ML algorithms with a

Corresponding Author:- Anand Ratnakar 1118

Address:- M.Tech Robotics and Artificial Intelligence, Dept. of Manufacturing
Engineering and Industrial Management COEP Technological University Pune, India.
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

deep learning model called Artificial Neural Networks (ANN) to identify the most effective approach for accurate
PD classification. By leveraging the strengths of both ML and DL, we aim to develop a more objective and
potentially earlier detection method for PD compared to traditional approaches.

Leveraging Machine Learning for Improved Diagnosis

Parkinson's disease (PD) diagnosis is most effective in its early stages, offering patients better treatment outcomes
and improved quality of life. Traditionally, diagnosis relies on neurological history and motor assessments, which
can lack sensitivity. Machine learning (ML), a subfield of Artificial Intelligence (AI), offers promising avenues for
improved detection. By combining traditional methods with ML-based analysis, clinicians may achieve a more
comprehensive understanding of the disease in patients.

One readily observable aspect of PD is gait (walking pattern). As walking is a fundamental part of daily life, gait
analysis has emerged as a potential non-intrusive tool for PD detection, with the advantage of being deployable in
home settings. Researchers have explored various gait analysis approaches, with some focusing on combining ML
techniques for autonomous and offline operation.

Speech patterns can also be indicative of PD, particularly in early stages. Speech problems associated with PD
include dysphonia (weak vocal quality), diplophonia (repetitive echoes), and hypophonia (impaired vocal muscle
coordination). These subtle changes in speech can be detected and analyzed using computational methods to aid in
PD diagnosis.

Research Motivation and Proposed Approach

This research investigates the potential of multivariate data analysis (MVDA) combined with ML techniques for
early and accurate PD detection. Current research in this area often focuses on single-source data (text, speech,
video, or images). This study highlights the limitations of such an approach and proposes MVDA for more
comprehensive multimodal data processing. By analyzing a wider range of data points, including gait, speech, and
potentially other relevant information, MVDA has the potential to improve disease detection accuracy.

This work specifically investigates the effectiveness of MVDA powered by ML in processing multimodal data for
PD diagnosis. Existing research utilizes various ML algorithms like Support Vector Machines (SVM), Naive Bayes,
K-Nearest Neighbors (KNN), and Artificial Neural Networks (ANN) for PD detection based on vocal features. This
study builds upon these advancements by leveraging large datasets and diverse ML approaches for improved disease
identification. The proposed MVDA framework encourages the inclusion of a wide range of data points, such as
multivariate acoustic characteristics, from a large patient population. This objective approach, aided by ML
techniques, aims to achieve a more accurate and reliable diagnosis of PD compared to traditional subjective
methods.

Research Contribution
This research explores various machine learning algorithms employed in speech analysis for PD diagnosis. The
strengths and limitations of these algorithms for PD detection are evaluated, while also considering potential
shortcomings in existing comparative studies. Artificial Neural Networks (ANN) have demonstrated promising
accuracy in speech analysis for PD diagnosis compared to other classifiers.

The key contributions of this paper are as follows:

1. Comparative Analysis of Machine Learning Algorithms: This research aims to identify which ML
algorithms, including SVM, KNN, Random Forest, Naive Bayes, and ANN, offer the most accurate
classifications for PD diagnosis.
2. Statistical Evaluation for Improved Diagnosis: This study proposes the development of statistical evaluations
for PD diagnosis. These evaluations aim to identify the optimal training and testing parameters, ultimately
contributing to future research efforts.
3. Comprehensive Machine Learning Model Exploration: The proposed system utilizes seven different
machine learning and deep learning models, including Logistic Regression (LR), Decision Tree (DT), Random
Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), K-Nearest Neighbors (KNN), and Artificial
Neural Networks (ANN). This comprehensive approach allows for identifying the model that performs best for
PD diagnosis based on training and testing results.

1119
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

4. Feature Selection Methodology: This study employs a comprehensive methodology to explore the
effectiveness and efficiency of various feature selection approaches to improve PD prediction accuracy.
5. Benchmarking Model Performance: By training all seven machine learning and deep learning models on the
same dataset, this research facilitates a direct comparison of their performance in PD diagnosis.

Related Works:-
In order to distinguish PD cases from healthy controls, a variety of modern machine learning algorithms, including
support vector machines, artificial neural networks, logistic regression, naïve Bayes, etc., have been successfully
used. In this study, numerous databases, including Web of Science, Elsevier, MDPI, Scopus, Science Direct, IEEE
Xplore, Springer, and Google Scholar, were utilized to survey relevant papers on Parkinson’s disease. The Table
Below shows the details about the previous work.
Machine
Learning Source of No. of
Reference Feature Objective Tools Used Outcomes
Algorithms Data Subjects
Used
Naïve Bayes,
JupyterLab
SVM (RBF and Classification Collected 252 (188 Highest accuracy
Sakar et al., With Python
Speech Linear), KNN, of PD from from PD + 64 obtained from SVM
2019 Programming
Random HC participants HC) (RBF)-86%
Language
Forest,MLP
Classification Collected 120 (40
Yasar A. et Artificial Accuracy of ANN-
Speech of PD from MATLAB from HC + 80
al., 2019 Neural Network 94.93%
HC participants PD+)
JupyterLab UCI
Avuçlu, E., KNN, Random Classification Accuracy from
With Python machine 31 (23
Elen, A, Speech Forest, Naïve of PD from Naïve Bayes-
Programming learning PD+8 HC)
2020 Bayes, SVM HC 70.26%
Language repository
Naïve
Bayes,ANN,
Classification Collected Highest accuracy
Marar et al., KNN, Random R 31 (23
Speech of PD from from obtained from
2018 Forest, SVM, programming PD+8 HC)
HC participants ANN-94.87%
Logistic
Regression
JupyterLab UCI
Classification Accuracy obtained
Sheibani R et Ensemble With Python machine 31 (23
Speech of PD from from ensemble
al., 2019 Based Method Programming learning PD+8 HC)
HC learning-90.6%
Language repository
Logistic Highest accuracy
Regression (L2- obtained from
John M. Classification 2289 (246
regularized), mPower gradient boosted
Tracy et al., Speech of PD from Python PD + 2023
Random Forest, database trees: Recall-79.7%,
2020 HC HC)
Gradient Precision-90.1%,
Boosted Trees F1-score-83.6%
Classification error
Classification Collected 270 (150 for rs11240569,
Cibulka et Handwriting Not
Random Forest of PD from from PD + 120 rs708727, rs823156
al., 2019 Patterns Mentioned
HC participants HC) is 49.6%, 44.8%,
49.3% respectively

1120
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Highest accuracy
202,94 obtained from
SVM with RBF Classification
Hsu S-Y et Handwriting Severe PD SVM-RBF 83.2%
Kernel, Logistic of PD from Weka PACS
al., 2019 Patterns +102 mild having sensitivity
Regression HC
PD + 6 HC 82.8%, specificity
100%
K-NN,
Ensemble
Classification Python
Drotár, P et Handwriting Classifier PaHaw 37 PD and
of PD from [scikit-learn Accuracy-81.3%
al., 2016 Patterns (AdaBoost), database 38 HC
HC library]
Support Vector
Machine
157,82 PD
+ 68 HC
UCI
Fabian Classification + 7 Normal
Handwriting machine Sensitivity-80%,
Maass et al., SVM of PD from Weka Pressure
Patterns learning and specificity-83%
2020 HC Hydroceph
repository
alus
(NPH))
accuracy-90% with
Classification
J. Mucha et Handwriting PaHaw 69, 33 PD sensitivity 89%, and
Random Forest of PD from Python
al., 2018 Patterns database + 36 HC specificity 91%
HC

Classification 645, 438

Wenzel et Handwriting PPMI
CNN of PD from MATLAB PD + 207 accuracy-97.20%
al., 2019 Patterns database
HC HC
Virgen De
SVM with 10 Classification La Victoria
Segovia, F. Handwriting 189, 95 PD
Cross of PD from Python Hospital, accuracy-94.25%
et al., 2019 Patterns + 94 HC
Validation HC Malaga,
Spain
64, 15 PD
+ 16 HC
+13
Accuracy to
Neurology [Amyotrop
diagnose PD from
Outpatient hic lateral
Least Square HC- 90.32%,
Clinic at sclerosis
(LS)-SVM, Classification Accuracy to
Ye, Q. et al., Not Massachuset Disease
Gait Particle Swarm of PD, ALS, diagnose HD from
2018 mentioned ts General (ALS)] +
Optimization HD from HC HC-94.44%,
Hospital, 20
(PSO) Accuracy to
Boston, MA, [Huntingto
diagnose ALS from
USA n’s
HC- 93.10%
Disease
(HD)]

1121
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Accuracy to
Neurology
diagnose PD from
Outpatient
HC- 96.43%,
Clinic at 64, 15 PD
Classification Accuracy to
Klomsae, A Not Massachuset + 20 HD
Gait Fuzzy KNN of PD, ALS, diagnose HD from
et al., 2018 mentioned ts General +13 ALS +
HD from HC HC-97.22%,
Hospital, 16 HC
Accuracy to
Boston, MA,
diagnose ALS from
USA
HC-96.88%
Neurology
Outpatient
Highest Accuracy
SVM, KNN, Clinic at
Classification obtained from
J. P. Félix et Decision Tree, MATLAB Massachuset 31, 15 PD+
Gait of PD from SVM,KNN and
al., 2019] Naïve Bayes, R2017a ts General 16 HC
HC Decision Tree -
LDA Hospital,
96.80%
Boston, MA,
USA
Laboratory
Classification 166, 93
Andrei et al., Not for Gait and
Gait SVM of PD from PD+ 73 Accuracy- 100%
2019 mentioned Neurodynam
HC HC
ics
Laboratory
Classification 166 ,93
Priya SJ et MATLAB for Gait and
Gait ANN of PD from PD+ 73 Accuracy- 96.28%
al., 2021 R2018b Neurodynam
HC HC
ics
Laboratory
Classification 166 ,93
Oğul, et al., for Gait and Classification
Gait ANN of PD from MATLAB PD+ 73
2020 Neurodynam accuracy - 98.3%
HC HC
ics
Classification Collected
Li B et al., Not 20, 10 PD
Gait Deep CNN of PD from from Accuracy- 91.9%
2020 mentioned + 10 HC
HC participants
Table 1:- Comparative Studies of Machine Learning Approaches to diagnose Parkinson’s Disease.

Proposed System:-
This research investigates the potential of machine learning (ML) and deep learning (DL) algorithms for classifying
Parkinson's disease (PD) and healthy controls (HC) using voice analysis.

System Architecture:
The proposed work involves methods with several key components that work together to achieve PD classification:

Data Acquisition:
1. We retrieved voice recordings from the publicly available Max Little dataset.
2. This dataset contains recordings from individuals diagnosed with PD and healthy controls, along with 22 pre-
extracted features related to various aspects of the speaker's voice.

Data Preprocessing:
Depending on initial data exploration, preprocessing steps might have been applied to the data, including:
1. Handling missing values using techniques like mean/median imputation or more sophisticated methods.
2. Scaling features (normalization or standardization) to ensure all features contribute equally during model
training (applicable for specific algorithms).
3. Encoding categorical variables (if present) through appropriate techniques.

1122
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Model Training:
We implemented and trained various ML and DL models on the preprocessed data. This involved:
1. Selecting a diverse range of algorithms, such as Logistic Regression, Decision Tree, Random Forest, Support
Vector Machine, Naive Bayes, K-Nearest Neighbors, and potentially an Artificial Neural Network architecture.
2. Splitting the data into training and testing sets. The training set is used to train the models, allowing them to
learn the underlying patterns that differentiate PD from HC recordings in the voice data.
3. Training each chosen model on the training set.
4. Employing hyperparameter tuning (optional) to optimize the models' performance by adjusting their internal
settings.

Figure 1:- Flowchart of the proposed work.

Model Evaluation:
1. The performance of each trained model was rigorously evaluated on unseen data using the testing set.
2. Established classification metrics like accuracy, precision, recall, and F1-score were used to assess the models'
ability to correctly classify PD recordings.

Model Selection:
By comparing the evaluation metrics across all models, we identified the model that demonstrated the most robust
and accurate performance in classifying PD recordings from HC recordings within the Max Little dataset.

Methodology:-
Dataset
The dataset was created by Max Little of the University of Oxford, in collaboration with the National Centre for
Voice and Speech, Denver, Colorado, who recorded the speech signals. The original study published the feature
extraction methods for general voice disorders.

Dataset Information:
This dataset is composed of a range of biomedical voice measurements from 31 people, 23 with Parkinson's disease
(PD). Each column in the table is a particular voice measure, and each row corresponds to one of 195 voice
recordings from these individuals ("name" column). The main aim of the data is to discriminate healthy people from
those with PD, according to the "status" column which is set to 0 for healthy and 1 for PD.

1123
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Dataset Characteristic Multivariate

No. of Instances 197
Attributes Characteristic Real
No. of Attributes 23
Missing Values N/A
Made by Max Little of the University of Oxford
Associated Tasks Classification
Types of Classification Binary {0 for healthy and 1 for PD patient}
Table 2:- Detail of Parkinson’s Dataset.

Dataset Attributes:
Attribute Name Description
name Unique identifier for each subject recording (e.g., "subject1_recording02")
MDVP:Fo(Hz) Average vocal fundamental frequency (perceived pitch) in Hertz (Hz)
MDVP:Fhi(Hz) Maximum vocal fundamental frequency (Hz)
MDVP:Flo(Hz) Minimum vocal fundamental frequency (Hz)
MDVP:Jitter(%) Variation in fundamental frequency over time (%)
MDVP:Jitter(Abs) Absolute variation in fundamental frequency
MDVP:RAP Ratio of Average Period to Average Amplitude variation
MDVP:PPQ Normalized logarithmic measure of variation in fundamental frequency
Jitter:DDP Local detrended fluctuation in fundamental frequency
MDVP:Shimmer Variation in amplitude of the voice signal over time
MDVP:Shimmer(dB) Amplitude variation in decibels (dB)
Shimmer:APQ3 Amplitude variation based on the 3rd quartile
Shimmer:APQ5 Amplitude variation based on the 5th quartile
MDVP:APQ Amplitude variation measure
Shimmer:DDA Local detrended fluctuation in amplitude variation
NHR Ratio of noise to tonal components in the voice
HNR Harmonic-to-Noise Ratio
status Health status (1: Parkinson's, 0: Healthy)
RPDE Nonlinear complexity measure 1
D2 Nonlinear complexity measure 2
DFA Signal complexity measure (fractal scaling exponent)
spread1 Nonlinear measure of fundamental frequency variation 1
spread2 Nonlinear measure of fundamental frequency variation 2
PPE Normalized log-area variation measure of fundamental frequency variation
Table 3:- Detail of Parkinson’s Dataset Attributes.

1124
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Figure 2:- Distribution plot displays a distribution and range of a set of numeric values plotted against a dimension.

Machine Learning And Deep Learning Classification Algorithms

This section explores the application of various machine learning algorithms for the diagnosis of Parkinson's
Disease (PD) based on voice analysis data. We investigate the following algorithms and their potential suitability for
this task:

Logistic Regression (LR):

1. Strengths: LR is a well-established linear classification algorithm. It excels at interpreting the coefficients of
the model, providing insights into the features that most significantly contribute to PD classification.
Additionally, LR offers efficient training and is relatively less prone to overfitting compared to more complex
models.
2. Limitations: LR assumes a linear relationship between features and the target variable (presence/absence of
PD). If the underlying relationships are non-linear, LR might not achieve optimal performance.

Decision Tree (DT):

1. Strengths: DT is a flexible and interpretable algorithm that can handle both continuous and categorical features
without extensive data preprocessing. It builds a tree-like structure where each node represents a decision based
on a specific feature. This structure allows for easy visualization and understanding of the decision-making
process.
2. Limitations: DTs can be susceptible to overfitting, particularly with deep trees and high dimensionality.
Additionally, they may be sensitive to small variations in the training data.

Random Forest (RF):

1. Strengths: RF addresses the overfitting limitations of DTs by constructing an ensemble of multiple decision
trees trained on random subsets of features and data points. This approach reduces variance and enhances the
model's generalizability to unseen data. RF also offers robustness to outliers and missing values.
2. Limitations: While interpretability is lower compared to individual DTs, techniques like feature importance
analysis can still provide insights into the most influential features. RF can be computationally expensive to
train compared to simpler models like LR.

1125
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Naive Bayes (NB):

1. Strengths: NB is a probabilistic classifier based on Bayes' theorem. It assumes conditional independence
between features, which can be a simplifying assumption but may be suitable for certain types of voice data. NB
is efficient for training and can handle high dimensionality.
2. Limitations: The conditional independence assumption may not always hold true for real-world data,
potentially impacting classification accuracy. Additionally, NB may struggle with imbalanced datasets where
one class (e.g., PD) has significantly fewer samples compared to the other (healthy controls).

K-Nearest Neighbors (KNN):

1. Strengths: KNN is a simple and intuitive non-parametric classification algorithm. It classifies a new data point
based on the majority vote of its K nearest neighbors in the training data. KNN requires minimal training and
can handle various feature types.
2. Limitations: KNN's performance is highly dependent on the choice of the K parameter (number of neighbors)
and the distance metric used. Additionally, KNN can be computationally expensive for large datasets due to the
need to compare new data points with all data points in the training set.

Support Vector Machine (SVM):

1. Strengths: SVMs are powerful algorithms that excel in high-dimensional feature spaces and can handle non-
linear relationships through the use of kernel functions. They are also robust to outliers and efficient in terms of
memory usage during training.
2. Limitations: SVMs can be challenging to tune due to the presence of hyperparameters (kernel type,
regularization parameter). Additionally, they may not provide clear interpretability of the model's decision-
making process.

Artificial Neural Networks (ANN):

1. Strengths: ANNs are powerful learning models inspired by the structure and function of the human brain. They
consist of interconnected nodes (artificial neurons) arranged in layers. ANNs can learn complex non-linear
relationships between features and the target variable, potentially achieving high accuracy on classification
tasks.
2. Limitations: ANNs are often considered "black boxes" due to their complex internal structure. This can make
interpretability challenging. Additionally, training ANNs can be computationally expensive and requires careful
hyperparameter tuning to avoid overfitting.

Machine Learning Classification for Parkinson's Disease

This section explores the use of machine learning (ML) classifiers for PD classification. We begin by identifying the
target variable, which in this case is the health status of the patient (presence or absence of PD). We then analyze the
distribution of health statuses within the dataset and visualize this data graphically. A common approach involves
splitting the data into two sets: a training set (typically 80%) used to train the ML model, and a testing set (20%)
used to evaluate the model's performance on unseen data.

Figure 3 depicts the distribution of health statuses in our dataset. A value of "0" represents healthy individuals, with
a count of 48. A value of "1" represents patients diagnosed with PD, with a count of 147. This translates to a
prevalence of PD in the dataset of 75.38% (147 out of 195) and a healthy control group of 24.62% (48 out of 195).

1126
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Figure 3:- Health Status of PD Patient.

Evaluation of Machine Learning Models for Parkinson's Disease Diagnosis

This section explores the evaluation of machine learning models employed for Parkinson's disease (PD) diagnosis.

Confusion Matrix
A confusion matrix is a visualization tool that summarizes a classification model's performance on a set of test data.
It allows us to understand how well the model distinguishes between different classes (healthy vs. PD in this case).
The matrix displays the number of correctly and incorrectly classified instances based on the model's predictions.

Figure 4:- Confusion Matrix for Model Evaluation.

Key Metrics Derived from the Confusion Matrix:

1. True Positives (TP): These represent instances where the model correctly identifies a patient with PD (positive
class).
2. True Negatives (TN): These represent instances where the model correctly identifies a healthy individual
(negative class).

1127
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

3. False Positives (FP): These represent instances where the model incorrectly classifies a healthy individual as
having PD (incorrect positive prediction).
4. False Negatives (FN): These represent instances where the model incorrectly classifies a patient with PD as
healthy (incorrect negative prediction).

By analyzing these values within the confusion matrix, we have calculated various performance metrics to assess the
effectiveness of the machine learning models for PD diagnosis.

Performance metrics for PD diagnosis:

1. Accuracy: Overall proportion of correctly classified cases
Accuracy = (TP + TN) / (Total samples)
2. Precision: Proportion of true positives among all predicted positive cases
Precision = (TP / (TP + FP))
3. Recall: Proportion of true positives among all actual positive cases
Recall = (TP / (TP + FN))
4. F1-Score: Harmonic mean of precision and recall, providing a balanced view of model performance
F1-score = 2 * (Precision * Recall) / (Precision + Recall)

By evaluating these metrics for different machine learning models, we have identified the model that achieves the
most accurate and reliable classification for PD diagnosis based on the chosen features.

Kappa Statistic for Evaluating Inter-Rater Reliability in PD Diagnosis

While the confusion matrix provides valuable insights into a machine learning model's performance for PD
diagnosis, it doesn't necessarily address the question of agreement between the model's predictions and a potential
"gold standard" diagnosis by a human expert. Here, the Kappa statistic (κ) emerges as a valuable tool for assessing
inter-rater reliability.

Understanding Kappa:
Kappa is a statistical measure that goes beyond simple agreement between two raters (model and human expert in
this case). It considers the agreement that occurs by chance and focuses on the agreement beyond this random
chance. Unlike the percentage agreement, which can be misleading, Kappa provides a more robust measure of
agreement, ranging from -1 to 1.

Interpreting Kappa Values:

1. κ < 0: Indicates disagreement worse than chance.
2. 0 ≤ κ ≤ 0.20: Represents slight agreement.
3. 0.21 ≤ κ ≤ 0.40: Indicates fair agreement.
4. 0.41 ≤ κ ≤ 0.60: Represents moderate agreement.
5. 0.61 ≤ κ ≤ 0.80: Suggests substantial agreement.
6. 0.81 ≤ κ ≤ 1.00: Indicates almost perfect agreement.

Formula for Kappa Score:

The Kappa statistic is calculated using the following formula:
κ = (P(A) - P(E)) / (1 - P(E))
Where:
1. P(A): Represents the observed agreement between the model and the human expert. This is calculated as the
sum of the diagonal elements of the confusion matrix divided by the total number of samples.
2. P(E): Represents the expected agreement by chance. This is calculated by summing the product of row and
column totals in the confusion matrix (excluding diagonal elements) and then dividing by the total number of
samples squared.

Experiments and Results:-

The proposed work, The Machine Learning algorithms including Logistic Regression (LR), Decision Tree (DT),
Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), K-Nearest Neighbors (KNN) are
implemented in Python 3.11.4: Jupyter Notebook And Deep Learning Algorithm Artificial Neural Networks (ANN)

1128
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

is implemented in Python 3.10.12: Google Colab. Here we detail the experimental setup and the results of the Total
Seven machine learning and Deep Learning classification methods.

Logistic Regression (LR):

Title Results
Training Accuracy 87.18%
Testing Accuracy 84.62%
Precision 80.65%
Recall 78.13%
F1-Score 79.37%
Kappa Score 0.3546
Table 4:- Performance Analysis for Logistic Regression (LR) Classifier.

Figure 5:- Confusion Matrix and Heatmap for Logistic Regression (LR) Classifier.

Decision Tree (DT):

Title Results
Training Accuracy 100%
Testing Accuracy 100%
Precision 77.42%
Recall 75.00%
F1-Score 76.16%
Kappa Score 1.00
Table 5:- Performance Analysis for Decision Tree (DT) Classifier.

1129
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Figure 6:- Confusion Matrix and Heatmap for Decision Tree (DT) Classifier.

Random Forest (RF):

Title Results
Training Accuracy 100%
Testing Accuracy 89.74%
Precision 81.25%
Recall 81.25%
F1-Score 81.25%
Kappa Score 0.6855
Table 6:- Performance Analysis for Random Forest (RF) Classifier.

Figure 7:- Confusion Matrix and Heatmap for Random Forest (RF) Classifier.

1130
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Naive Bayes (NB):

Title Results
Training Accuracy 71.79%
Testing Accuracy 69.23%
Precision 84.21%
Recall 50.00%
F1-Score 62.75%
Kappa Score 0.0414
Table 7:- Performance Analysis for Naive Bayes (NB) Classifier.

Figure 8:- Confusion Matrix and Heatmap for Naive Bayes (NB) Classifier.

K-Nearest Neighbors (KNN):

Title Results
Training Accuracy 89.74%
Testing Accuracy 82.05%
Precision 87.88%
Recall 90.63%
F1-Score 89.23%
Kappa Score 0.3546
Table 8:- Performance Analysis for K-Nearest Neighbors (KNN) Classifier.

1131
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Figure 9:- Confusion Matrix and Heatmap for K-Nearest Neighbors (KNN) Classifier.

Support Vector Machine (SVM):

Title Results
Training Accuracy 86.54%
Testing Accuracy 87.18%
Precision 88.57%
Recall 96.88%
F1-Score 92.54%
Kappa Score 0.4772
Table 9:- Performance Analysis for Support Vector Machine (SVM) Classifier.

Figure 10:- Confusion Matrix and Heatmap for Support Vector Machine (SVM) Classifier.

1132
 ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Artificial Neural Networks (ANN):

Title Results
Training Accuracy 100%
Testing Accuracy 97.44%
Precision 96.67%
Recall 100%
F1-Score 98.31%
Kappa Score 0.9305
Table 10:- Performance Analysis for Artificial Neural Networks (ANN) Classifier.

Figure 11:- Confusion Matrix and Heatmap for Artificial Neural Networks (ANN) Classifier.

Comparative Study of Machine Learning Algorithms Used in Proposed Work

Training Testing
Classifier Precision Recall F1-Score Kappa Score
Accuracy Accuracy
Logistic
Regression 87.18% 84.62% 80.65% 78.13% 79.37% 0.3546
(LR)
Decision Tree
100% 100% 77.42% 75.00% 76.19% 1.00
(DT)
Random Forest
100% 89.74% 81.25% 81.25% 81.25%
(RF) 0.6855
Naive Bayes
71.79% 69.23% 84.21% 50.00% 62.75% 0.0414
(NB)
K-Nearest
Neighbors 89.74% 82.05% 87.88% 90.63% 89.23% 0.3546
(KNN)
Support Vector
Machine 86.54% 87.18% 88.57% 96.88% 92.54% 0.4772
(SVM)

1133
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Artificial
Neural
100% 97.44% 96.67 100% 98.31% 0.9305
Networks
(ANN)
Table 11:- An overview of evaluation results and Performance Analysis for all Classifiers used in Proposed Work.

Figure 12:- Graphical Representation of Comparison of Training And Testing Accuracy for all Classifiers.

Figure 13:- Graphical Representation of Comparison of Kappa Score for all Classifiers.

1134
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

Conclusion:-
This research investigated the potential of automated machine learning (ML) and deep learning (DL) techniques to
classify Parkinson's disease (PD) from healthy controls (HC) based on non-invasive speech biomarkers. Our study
focused on comparing the performance of various classifiers in handling the challenges of noisy and high-
dimensional speech data, common in real-world applications. The findings demonstrate that achieving clinical-level
accuracy for PD detection is feasible with careful feature selection and appropriate model selection.

Among the evaluated algorithms, Logistic Regression (LR) achieved an accuracy of 84.62%, Decision Tree (DT)
achieved 100% accuracy, Random Forest (RF) achieved 89.74% accuracy, Naive Bayes (NB) achieved 69.23%
accuracy, K-Nearest Neighbors (KNN) achieved 82.05% accuracy, Support Vector Machine (SVM) achieved
87.18% accuracy, and Artificial Neural Networks (ANNs) achieved the highest accuracy of 97.44%. It's important
to note that while the Decision Tree classifier achieved the highest reported accuracy, it is susceptible to overfitting,
which can lead to poor performance on unseen data.

This research highlights the significant advantage of Artificial Neural Networks (ANNs) for PD classification using
speech analysis. The deep learning model achieved an impressive accuracy of 97.44%, significantly outperforming
other methods. ANNs' inherent ability to learn complex, non-linear relationships within the data offers a clear
advantage for this specific task.

References:-
[1] DeMaagd, G., & Philip, A. (2015). Parkinson’s Disease and Its Management: Part 1: Disease Entity, Risk
Factors, Pathophysiology, Clinical Presentation, and Diagnosis. Pharmacotherapy, 40, 504–532.
[2] Rizek, P., Kumar, N., & Jog, M. S. (2016). An update on the diagnosis and treatment of Parkinson disease.
Canadian Medical Association Journal, 188(16), 1157–1165.
[3] World Health Organization. (n.d.). Parkinson's disease. Retrieved October 30, 2022, from
https://www.who.int/news-room/fact-sheets/detail/parkinson-disease
[4] de Rijk, M. C., Launer, L. J., Berger, K., Breteler, M. M., Dartigues, J. F., Baldereschi, M., ... & Trenkwalder, C.
(2000). Prevalence of Parkinson’s disease in Europe: A collaborative study of population-based cohorts. Neurology,
54(Suppl. 5), S21–S23.
[5] Cantürk, İ., & Karabiber, F. (2016). A machine learning system for the diagnosis of Parkinson’s disease from
speech signals and its application to multiple speech signal types. Arabian Journal for Science and Engineering, 41,
5049–5059.
[6] Singh, N., Pillay, V., & Choonara, Y. E. (2007). Advances in the treatment of Parkinson’s disease. Progress in
Neurobiology, 81, 29–44.
[7] Rana, A., Rawat, A. S., Bijalwan, A., & Bahuguna, H. (2018). Application of multi-layer (perceptron) artificial
neural network in the diagnosis system: A systematic review. In Proceedings of the 2018 International Conference
on Research in Intelligent and Computing in Engineering (RICE) (pp. 1–6). San Salvador, El Salvador.
[8] Lakany, H. (2008). Extracting a diagnostic gait signature. Pattern Recognition, 41, 1627–1637.
[9] Figueiredo, J., Santos, C. P., & Moreno, J. C. (2018). Automatic recognition of gait patterns in human motor
disorders using machine learning: A review. Medical Engineering & Physics, 53, 1–12.
https://doi.org/10.1016/j.medengphy.2017.12.006.
[10] Hazan, H., Hilu, D., Manevitz, L., Ramig, L. O., & Sapir, S. (2012). Early diagnosis of Parkinson’s disease via
machine learning on speech data. In Proceedings of the 2012 IEEE 27th Convention of Electrical and Electronics
Engineers in Israel (pp. 1–4). Eilat, Israel. https://doi.org/10.1109/eeei.2012.6377065.
[11] Karan, B., Sahu, S. S., & Mahto, K. (2019). Parkinson disease prediction using intrinsic mode function based
features from speech signal. Biocybernetics and Biomedical Engineering, 40, 249–264.
https://doi.org/10.1016/j.bbe.2019.05.005.
[12] Frid, A.; Safra, E.J.; Hazan, H.; Lokey, L.L.; Hilu, D.; Manevitz, L.; Ramig, L.O.; Sapir, S. Computational
diagnosis of Parkinson’s Disease directly from natural speech using machine learning techniques. In Proceedings of
the 2014 IEEE International Conference on Software Science, Technology and Engineering, Washington, DC, USA,
11–12 June 2014; pp. 50–53.
[13] Rawat, A. S., Rana, A., Kumar, A., & Bagwari, A. (2018). Application of multi layer artificial neural network
in the diagnosis system: a systematic review. IAES International Journal of Artificial Intelligence, 7(3), 138.
[14] Little, M.A., McSharry, P.E., Roberts, S.J. et al. Exploiting Nonlinear Recurrence and Fractal Scaling
Properties for Voice Disorder Detection. BioMed Eng OnLine 6, 23 (2007). https://doi.org/10.1186/1475-925X-6-

1135
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

23
-[15] Little, M.; McSharry, P.; Hunter, E.; Spielman, J.; Ramig, L. Suitability of dysphonia measurements for
telemonitoring of Parkinson’s disease. Nat. Preced. 2008, 1-1. https://doi.org/10.1038/npre.2008.2298.1.
[16] Lichman, M. UCI Machine Learning Repository; University of California, School of Information and
Computer Science: Irvine, CA, USA. Available online: http://archive.ics.uci.edu/ml (accessed on 25 September
2022).
[17] Rana, Arti & Dumka, Ankur & Singh, Rajesh & Rashid, Mamoon & Ahmad, Nazir & Panda, Manoj. (2022).
An Efficient Machine Learning Approach for Diagnosing Parkinson's Disease by Utilizing Voice Features.
Electronics. 11. 3782. 10.3390/electronics11223782.
[18] Sakar, B.E.; Isenkul, M.E.; Sakar, C.O.; Sertbas, A.; Gurgen, F.; Delil, S.; Apaydin, H.; Kursun, O. Collection
and Analysis of a Parkinson Speech Dataset With Multiple Types of Sound Recordings. IEEE J. Biomed. Health
Inform. 2013, 17, 828–834. https://doi.org/10.1109/jbhi.2013.2245674.
[19] Vadovský, M.; Paralič, J. Parkinson's disease patients classification based on the speech signals. In Proceedings
of the 2017 IEEE 15th International Symposium on Applied Machine Intelligence and Informatics (SAMI),
Herlany, Slovakia, 26–28 January 2017; pp. 000321–000326.
[20] Ouhmida, A., Raihani, A., Cherradi, B., & Terrada, O. (2021). A Novel Approach for Parkinson’s Disease
Detection Based on Voice Classification and Features Selection Techniques. International Journal of Online &
Biomedical Engineering, 17(10).
[21] Mabrouk, R.; Chikhaoui, B.; Bentabet, L. Machine Learning Based Classification Using Clinical and
DaTSCAN SPECT Imaging Features: A Study on Parkinson’s Disease and SWEDD. IEEE Trans. Radiat. Plasma
Med. Sci. 2018, 3, 170–177. https://doi.org/10.1109/trpms.2018.2877754.
[22] Benba, A.; Jilbab, A.; Hammouch, A. Using Human Factor Cepstral Coefficient on Multiple Types of Voice
Recordings for Detecting Patients with Parkinson’s Disease. Irbm 2017, 38, 346–351.
https://doi.org/10.1016/j.irbm.2017.10.002.
[23] Sakar, C.O.; Serbes, G.; Gunduz, A.; Tunc, H.C.; Nizam, H.; Sakar, B.E.; Tutuncu, M.; Aydin, T.; Isenkul,
M.E.; Apaydin, H. A comparative analysis of speech signal processing algorithms for Parkinson’s disease
classification and the use of the tunable factor wavelet transform. Appl. Soft Comput. 2019, 74, 255–263.
[24] Yasar, A.; Saritas, I.; Sahman, M.A.; Cinar, A.C. Classification of Parkinson disease data with artificial neural
networks. In Proceedings of the IOP Conference Series: Materials Science and Engineering, Wuhan, China, 10–12
October 2019; Volume 675, p.012031.
[25] Avuçlu, E.; Elen, A. Evaluation of train and test performance of machine learning algorithms and Parkinson
diagnosis with statistical measurements. Med. Biol. Eng. Comput. 2020, 58, 2775–2788.
https://doi.org/10.1007/s11517-020-02260-3.
[26] Marar, S.; Swain, D.; Hiwarkar, V.; Motwani, N.; Awari, A. Predicting the occurrence of Parkinson’s Disease
using various Classification Models. In Proceedings of the 2018 International Conference on Advanced
Computation and Telecommunication (ICACAT), Bhopal, India, 28–29 December 2018; pp. 1–5.
[27] Nikookar, E.; Sheibani, R.; Alavi, S.E. An ensemble method for diagnosis of Parkinson’s disease based on
voice measurements. J. Med. Signals Sens. 2019, 9, 221–226. https://doi.org/10.4103/jmss.jmss_57_18.
[28] Tracy, J.M.; Özkanca, Y.; Atkins, D.C.; Ghomi, R.H. Investigating voice as a biomarker: Deep phenotyping
methods for early detection of Parkinson’s disease. J. Biomed. Inform. 2019, 104, 103362.
https://doi.org/10.1016/j.jbi.2019.103362.
[29] Cibulka, M.; Brodnanova, M.; Grendar, M.; Grofik, M.; Kurca, E.; Pilchova, I.; Osina, O.; Tatarkova, Z.;
Dobrota, D.; Kolisek, M. SNPs rs11240569, rs708727, and rs823156 in SLC41A1 Do Not Discriminate Between
Slovak Patients with Idiopathic Parkinson’s Disease and Healthy Controls: Statistics and Machine-Learning
Evidence. Int. J. Mol. Sci. 2019, 20, 4688. https://doi.org/10.3390/ijms20194688. for Parkinson
[30]. Hsu, S.-Y.; Lin, H.-C.; Chen, T.-B.; Du, W.-C.; Hsu, Y.-H.; Wu, Y.-C.; Tu, P.-W.; Huang, Y.-H.; Chen, H.-Y.
Feasible Classified Models Disease https://doi.org/10.3390/s19071740. from 99mTc-TRODAT-1 SPECT Imaging.
Sensors 2019, 19, 1740.
[31] Drotár, P., Mekyska, J., Rektorová, I., Masarová, L., Smékal, Z., & Faundez-Zanuy, M. (2016). Evaluation of
handwriting kinematics and pressure for differential diagnosis of Parkinson's disease. Artificial intelligence in
Medicine, 67, 39-46.
[32] Maass, F.; Michalke, B.; Willkommen, D.; Leha, A.; Schulte, C.; Tönges, L.; Mollenhauer, B.; Trenkwalder,
C.; Rückamp, D.; Börger, M.; et al. Elemental fingerprint: Reassessment of a cerebrospinal fluid biomarker for
Parkinson’s disease. Neurobiol. Dis. 2019, 134, 104677. https://doi.org/10.1016/j.nbd.2019.104677.
[33] Mucha, J.; Mekyska, J.; Faundez-Zanuy, M.; Lopez-De-Ipina, K.; Zvoncak, V.; Galaz, Z.; Kiska, T.; Smekal,
Z.; Brabenec, L.; Rektorova, I. Advanced Parkinson’s Disease Dysgraphia Analysis Based on Fractional Derivatives

1136
ISSN: 2320-5407 Int. J. Adv. Res. 12(05), 1118-1137

of Online Handwriting. In Proceedings of the 2018 10th International Congress on Ultra Modern
Telecommunications and Control Systems and Workshops (ICUMT), Moscow, Russia, 5–9 November 2018; pp. 1–
6. https://doi.org/10.1109/icumt.2018.8631265.
[34] Wenzel, M.; Milletari, F.; Krüger, J.; Lange, C.; Schenk, M.; Apostolova, I.; Klutmann, S.; Ehrenburg, M.;
Buchert, R. Automatic classification of dopamine transporter SPECT: Deep convolutional neural networks can be
trained to be robust with respect to variable image characteristics. Eur. J. Pediatr. 2019, 46, 2800–2811.
https://doi.org/10.1007/s00259-019-04502-5.
[35] Segovia, F.; Gorriz, J.M.; Ramirez, J.; Martinez-Murcia, F.J.; Castillo-Barnes, D. Assisted Diagnosis of
Parkinsonism Based on the Striatal Morphology. Int. J. Neural Syst. 2019, 29, 1950011.
https://doi.org/10.1142/s0129065719500114.
[36] Ye, Q.; Xia, Y.; Yao, Z. Classification of Gait Patterns in Patients with Neurodegenerative Disease Using
Adaptive Neuro-Fuzzy Inference System. Comput. Math. Methods Med. 2018, 2018, 9831252.
https://doi.org/10.1155/2018/9831252.
[37] Klomsae, A.; Auephanwiriyakul, S.; Theera-Umpon, N. (2018). String grammar unsupervised possibilistic
fuzzy c-medians for gait pattern classification in patients with neurodegenerative diseases. Comput. Intell. Neurosci.
2018, 2018, 1869565.
[38] Felix, J.P.; Vieira, F.H.T.; Cardoso, A.A.; Ferreira, M.V.G.; Franco, R.A.P.; Ribeiro, M.A.; Araujo, S.G.;
Correa, H.P.; Carneiro, M.L. A Parkinson’s Disease Classification Method: An Approach Using Gait Dynamics and
Detrended Fluctuation Analysis. In Proceedings of the 2019 IEEE Canadian Conference of Electrical and Computer
Engineering (CCECE), Edmonton, AB, Canada, 5–8 May 2019. https://doi.org/10.1109/ccece.2019.8861759.
Bioengineering
[39] Andrei, A.-G.; Tautan, A.-M.; Ionescu, B. Parkinson’s Disease Detection from Gait Patterns. In Proceedings of
the 2019 E-Health and Conference (EHB), https://doi.org/10.1109/ehb47216.2019.8969942. Iasi, Romania, 21–23
November 2019; pp. 1–4.
[40] Priya, S.J.; Rani, A.J.; Subathra, M.S.P.; Mohammed, M.A.; Damaševičius, R.; Ubendran, N. Local Pattern
Transformation Based Feature Extraction for Recognition of Parkinson’s Disease Based on Gait Signals.
Diagnostics 2021, 11, 1395. https://doi.org/10.3390/diagnostics11081395.
[41] Yurdakul, O.C.; Subathra, M.; George, S.T. Detection of Parkinson’s Disease from gait using Neighborhood
Representation Local Binary Patterns. Biomed. Signal Process. Control 2020, 62, 102070.
https://doi.org/10.1016/j.bspc.2020.102070.
[42] Li, B.; Yao, Z.; Wang, J.; Wang, S.; Yang, X.; Sun, Y. Improved Deep Learning Technique to Detect Freezing
of Gait in Parkinson’s Disease Based on Wearable Sensors. Electronics 2020, 9, 1919.
https://doi.org/10.3390/electronics9111919.

1137