Parkinsons Disease Prediction Using ML

Technique

1st Varun Gupta 4th Swati Sharma

Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering
KIET Group of Institutions KIET Group of Institutions
Ghaziabad, India Ghaziabad, India
varunshivhare1729@[Link] [Link]@[Link]
2nd Yash Jain
Dept. of Computer Science and Engineering
KIET Group of Institutions
Ghaziabad, India
3rd Sumit Pal
Dept. of Computer Science and Engineering
KIET Group of Institutions
Ghaziabad, India

Abstract— Parkinson's Disease (PD) is the second most I. INTRODUCTION

prevalent neurological disorder associated with aging, marked
by various motor and cognitive impairments. Diagnosing PD is Millions of individuals worldwide are affected by Parkinson’s
challenging due to its resemblance to other conditions like Disease (PD), a progressively deteriorating disorder in which
normal aging and essential tremor. By the age of 50, prominent symptoms appear gradually over time. While visible symptoms
symptoms such as difficulties with walking and communication occur in people over the age of 50, roughly one in every ten
start to appear. Although there is no cure for PD, certain people shows signs of this disease before the age of 40
medications can alleviate symptoms, enabling patients to
(Marton, 2019). Parkinson’s disease causes the death of
maintain their quality of life by managing complications. Early
detection is crucial in slowing the disease’s progression. specific nerve cells in the brain’s substantia nigra, which
Numerous studies have focused on diagnosing PD, and in our generate chemical dopamine for directing bodily movements.
project, we aim to detect PD using a variety of Machine Dopamine deficiency causes additional progressive symptoms
Learning (ML) and Deep Learning (DL) models, including to emerge gradually over time. Typically, PD symptoms begin
Support Vector Machine (SVM), Random Forest (RF), Decision with tremors or stiffness on one side of the body, such as the
Tree (DT), K-Nearest Neighbor (KNN), and Multi-Layer hand or arm. Individuals with PD may acquire dementia at later
Perceptron (MLP). These models are trained to differentiate
stages (Tolosa et al., 2006). From 1996 to 2016, the global
between healthy individuals and PD patients based on voice
signal data. The dataset, sourced from the UCI machine prevalence of PD more than quadrupled, from 2.5 million to
learning repository, contains 195 voice recordings from 31 6.1 million individuals. Increased life expectancy has resulted
patients. To improve model performance, we applied in an older population, which explains the substantial rise
techniques such as Synthetic Minority Over-sampling (Fothergill-Misbah et al., 2020). The brain is the body’s
Technique (SMOTE), Feature Selection, and hyperparameter controlling organ. Trauma or sickness to any portion of the
tuning using GridSearchCV. Feature selection helped reduce brain will manifest in a variety of ways in numerous other
the dimensionality of the dataset, improving accuracy and sections of the body. PD causes a range of symptoms, including
reducing overfitting. Our results indicated that MLP and SVM,
trained using a 70:30 train/test split with GridSearchCV and partial or complete loss of motor reflexes, speech problems
SMOTE, produced the best outcomes. MLP achieved an and eventual failure, odd behavior, loss of mental thinking, and
accuracy of 98.31%, with a recall of 98%, precision of 100%, other critical skills. It is difficult to distinguish between typical
and an f1-score of 99%. Likewise, SVM attained an accuracy of cognitive function losses associated with aging and early PD
95%, a recall of 96%, precision of 98%, and an f1-score of 97%. symptoms. In the United States, the overall economic impact
These findings suggest that the proposed method is reliable for in 2017 was predicted to be $51.9 billion, including an indirect
predicting PD and could be integrated into healthcare systems cost of $14.2 billion, non-medical expenditures of $7.5 billion,
for diagnostic purposes.
and $4.8 billion accruing to disability income for owner’s public
works. The majority of Parkinson’s disease patients are over
the age of 65, and the overall economic burden is expected to disease diagnosis. The paper’s objective is to efficiently identify
approach $79 billion by 2037 (Yang et al., 2020).The diagnosis healthy individuals. A comparative study was carried out.
of PD in National Collaborating Centre for Chronic Conditions There were four different classification schemes used. These
(2006) is typically based on a few invasive techniques as well are, in order, Decision Trees, Regression, Neural Networks, and
as empirical testing and examinations. Invasive diagnostic DMneural. The performance score of the classifiers was
procedures for PD are exceedingly expensive, inefficient, and determined using a variety of evaluation techniques. The
require extremely complex equipment with poor accuracy. neural network classifier produces the best outcomes, as
New techniques are needed to diagnose PD. Therefore, less determined by the application scores. The neural network’s
expensive, simplified, and reliable methods should be adapted overall classification performance is 92.9%. A deep belief
to diagnose disease and ensure treatments. However, network (DBN) has been used as a successful method to
noninvasive diagnosis techniques for PD require being identify Parkinson’s disease in the paper by Al-Fatlawi et al.
investigated. Machine learning techniques are used to classify (2016). The deep belief network (DBN), which is used to
people with PD and healthy people. It has been determined produce a template match of the voices, has been configured
that disorders’ vocal issues can be assessed for early PD to accept input from a feature extraction procedure. Using two
detection (Harel et al., 2004). So, this study attempts to identify stacked Restricted Boltzmann Machines (RBMs) and one
Parkinson’s disease (PD) by utilizing Machine Learning (ML) output layer, DBN is employed in this study to categorize
and Deep Learning (DL) models to discriminate between Parkinson’s illness. To maximize the networks’ parameters, two
healthy and PD patients based on voice signal features, stages of learning must be used. Unsupervised learning, the
perhaps lowering some of these expenditures. first stage, uses RBMs to address the issue that can arise from
the initial weights’ unpredictable initial value. Secondly, the
A variety of ML models are trained and evaluated, including backpropagation technique is employed for the fine tuning as
Decision Tree, Random Forest, Logistic Regression, Support a supervised learning approach. The experimental results are
Vector Machine (SVM), Naive Bayes, K-Nearest Neighbors contrasted with various strategies and related work to
(KNN), and XGBoost classifiers. Each model is assessed based demonstrate the efficacy of the suggested system. The
on accuracy, F1 Score, and R² Score to determine which proposed approach outperforms all other methods in
algorithm delivers the highest performance in predicting comparison with its 94% total testing accuracy. Rasheed et al.
Parkinson’s Disease. We also perform detailed comparisons of (2020) proposed two classification schemes to improve the
these models, analyzing the trade-offs between simplicity, accuracy of PD case identification from voice measurements.
interpretability, and accuracy. They began by applying a variable adaptive moment-based
backpropagation algorithm to BPVAM, an artificial neural
II. RELATED WORK network. The researchers then investigated the use of
dimensionality reduction methods such as principal
Several researchers have classified Parkinson’s disease using component analysis (PCA) in conjunction with BPVAM to
various methods. These studies provide a solid foundation for classify the same dataset. The main goal was to improve PD
how machine learning can be applied to neurodegenerative prediction in the early stages by increasing the system’s
diseases in the face of current challenges in Parkinson’s disease sensitivity to dealing with fine-grained data. The best results
subclassification, risk assessment, and prognosis using voice were obtained by BPVAM and BPVAM PCA (97.50%), followed
signal features. Selection and classification procedures are by ANN with Levenberg-Marquardt (95.89%). In their (Kadam
used in the (Senturk, 2020) diagnosis technique. The feature and Jadhav, 2019) study, they proposed a feature ensemble
selection task took into consideration the methodologies of learning method based on sparse autoencoders to classify
Feature Importance and Recursive Feature Elimination. healthy people and people with Parkinson’s disease using
Artificial neural networks, support vector machines, and proper representation of vocal and speech datasets. Feature
classification and regression trees were all utilized in the trials ensemble learning based on the Sparse Autoencoders method
to categorize Parkinson’s patients. Performance comparisons achieves the highest sensitivity and specificity of 97.28% and
of the different techniques revealed that Support Vector 90%, respectively. The DNN method achieves the highest
Machines with Recursive Feature Elimination outperformed sensitivity and specificity of 93.59% and 90%, respectively.
them. With the fewest vocal features necessary to diagnose
Parkinson’s, 93.84% accuracy was attained. The results of the Our research follows a structured process that involves
methods provided by Gil and Manuel (2009) based on artificial multiple stages, including Data Collection, Data Preprocessing,
neural networks and support vector machines to aid specialists Exploratory Data Analysis (EDA), Dataset Balancing and
in the diagnosis of Parkinson’s disease indicate a high accuracy Scaling, and finally Machine Learning Model Training and
of about 90%. Das (2010) compared various classification Evaluation. The feature extraction process is particularly
techniques for the purpose of making an accurate Parkinson’s critical in this study, as the quality and relevance of the
extracted features greatly influence the performance of the is crucial to achieve a balance that minimizes errors while
machine learning models. The importance of selecting key maximizing model performance.
features that are most predictive of PD is comparable to how
feature extraction has been used in credit card fraud detection Furthermore, future research could explore more
[4], where prioritizing the most informative features improved advanced techniques, such as deep learning models or hybrid
classifier performance. approaches that combine multiple classifiers to further
Data preprocessing plays a pivotal role in ensuring the improve diagnostic accuracy. Other studies have shown that
success of the machine learning models. It involves handling Convolutional Neural Networks (CNNs) and Autoencoders
missing values, dealing with outliers, normalizing the data, and can provide excellent results in handling large, complex
scaling the features, much like the methods used in [8] to datasets [10], suggesting potential for application in
handle transaction data in fraud detection. In our project, we Parkinson’s Disease prediction. Additionally, exploring
faced similar challenges in ensuring the dataset was ready for different feature extraction and selection methods, such as
training the models. Additionally, balancing the dataset is Genetic Algorithms [12] or Principal Component Analysis
crucial to address the problem of class imbalance, where there (PCA), could further enhance model performance.
might be more data points for healthy individuals than those
with Parkinson’s. Techniques like oversampling and This project illustrates the versatility of machine learning in
undersampling, similar to methods used in fraud detection addressing challenges across various domains, from healthcare
studies, are implemented to mitigate these imbalances, diagnostics . By comparing different machine learning
ensuring that the models do not become biased toward the algorithms and utilizing advanced techniques like feature
majority class. extraction, dataset balancing, and performance evaluation,
Once the data is preprocessed and ready, we train and this study not only contributes to the ongoing research in
evaluate several machine learning models, including Decision Parkinson's Disease prediction but also demonstrates the
Tree, Random Forest, Logistic Regression, Support Vector broader applicability of machine learning in solving complex
Machine (SVM), Naive Bayes, K-Nearest Neighbors (KNN), real-world problems. Further exploration of techniques such as
and XGBoost. Each model is assessed based on its accuracy, F1 federated learning and blockchain technology could enhance
score, precision, and recall, allowing for a comprehensive data privacy and security in future models, providing a more
evaluation of their effectiveness in predicting Parkinson’s robust framework for sensitive medical data applications.
Disease. Random Forest, in particular, was found to be the
best-performing model, achieving high accuracy and an
excellent balance between precision and recall. This aligns with III. Material and methods
findings in fraud detection research [7][12], where ensemble
A. Dataset
models like Random Forest and XGBoost were shown to
provide superior performance in classification tasks. The dataset utilized in the research was obtained from the
The high accuracy and F1 score achieved by the Random University of Oxford (UO) repository with collaboration from
Forest Classifier (accuracy: 99.61%, F1 score: 0.961) suggest the National Center for Voice, established by Little et al. (2007,
that ensemble models are well-suited for handling complex 2009), and is available at the UCI Machine Learning Repository
datasets with many features, as also noted in studies on fraud (Little, 2008). The original study presented feature extraction
detection. The success of Random Forest in our study methods for general voice disorders The study included voice
highlights the importance of using models that can effectively recordings from 31 people, including 23 people with
capture non-linear relationships between features, a factor Parkinson’s Disease (PD) (16 males and 7 females) and eight
that often influences the performance of classifiers in Healthy Controls (HC) (males = 3 and females = 5). The dataset
detecting both diseases and fraudulent activities. contains 195 records, 24 columns, and as presented in Table 1,
In addition to model performance, this study also a series of biomedical voice measurements. Table 1 is divided
emphasizes the significance of evaluating the trade-offs into columns that represent each of the voice measurements
between various metrics such as accuracy, precision, recall, and rows which represent vocal recordings from individuals
and F1 score. In real-world applications, a high accuracy alone (the “name” column). An average of six recordings were made
may not be sufficient if the model fails to generalize well to for each patient; six recordings were taken from 22 patients,
unseen data or if it produces an unacceptable number of false and seven recordings were taken from nine patients. The
negatives or false positives. For example, a misclassified case patients’ ages ranged from 46 to 85 years (mean 65.8, standard
of Parkinson’s could delay treatment, while a false positive deviation 9.8), and the time since diagnosis ranged from 0 to
could cause unnecessary anxiety and medical tests. Hence, like 28 years. Each row corresponds to one voice recording for 36
in fraud detection systems, where false negatives are costly, it s. The voice was recorded in an industrial acoustic company
sound-treated booth by a microphone placed 8 cm from the
mouth and calibrated according to Little et al. (2009). In the
dataset, the “status” column is set to 0 for HC and 1 for those
1. Feature selection (FS)
with PD, to distinguish healthy individuals from those with PD.
In this phase, SelectKBest was applied to select the eight
B. Methods
best features of the dataset. SelectKBest has been found as the
The proposed method is designed to classify whether the second most commonly used dimensionality reduction
patient has PD or not by using the Google Colab environment technique, accounting for 29.1% of total usage (Bilgen et al.,
and Python language. The methodology of the proposed 2020). This technique chooses features based on the highest k
model is structured into six steps: data preprocessing, features score, aiding in the removal of less essential data and reducing
selection, Synthetic Minority Over-sampling Technique training time.
(SMOTE), hyperparameter tuning (GridSearchCV), machine The eight features used were: MDVP:Fo(Hz), MDVP:Flo(Hz),
and deep learning classification models, and performance MDVP:Shimmer, MDVP:APQ, HNR, spred1 spread2, and PPE.
evaluation. These steps of the proposed model are shown in
Figure 1. 2. Synthetic Minority Over-sampling Technique (SMOTE)
In our dataset, there are fewer HC samples than PD samples.
C. Data Preprocessing Oversampling samples in the minority class is one way of
resolving imbalanced classes. Duplicate instances from the
Preprocessing is the most important aspect of data processing,
minority class in the training dataset can be used to accomplish
which helps the model learn the features of the data
this. This may equalize the distribution of classes, but it
effectively and remove unnecessary information (Singh, 2020).
provides no extra information. SMOTE, or Synthetic Minority
The dataset was imported into the Google Colab platform as a
Oversampling Technique, is another method for improving
CSV file using the Pandas package. After we screened for any
minority data based on previous samples. The SMOTE
duplicates or null entries, we used the “status” column and
approach builds a linear connection using close features, then
found that the dataset was imbalanced with 147 for PD and 48
selects a new sample from the minority class along that line
for HC, which is equivalent to 25% for HC and 75% for PD. In
(Brownlee, 2020).
order to avoid under-fitting and over-fitting, we split our
dataset into a ratio of 70:30 train/test split. The training set
3. Hyperparameter tuning (GridSearchCV)
includes known outputs, and what the model learns from it
The hyperparameters are variables that the user normally
may be extended to other data sets. By computing the relevant
specifies when building the machine learning model. To get the
statistics on the samples in the training set, each feature is
best results from the model, we need to use GridSearchCV to
scaled individually. The mean and standard deviation are then
discover the optimum hyperparameter values. Grid search is
saved and utilized on later data using the transform in
the most basic search algorithm that produces the most
StandardScaler (Teo, 2021). Equation (1) express the
accurate predictions. Grid search is simple to conduct in
mathematical form of StandardScaler normalization. For this
parallel since each trial runs independently without regard for
study, we employed a variety of libraries, including NumPy,
time sequence (Yu and Zhu, 2020). Primarily, it takes
Pandas, Matplotlib, Seaborn, and Sickit-learn (Sklearn). Numpy
arguments i.e., estimator, param grid, cv. Each of the
is Python’s fundamental package for scientific computation. It
arguments is described as follows: • Estimator: the estimator
is used to insert any form of mathematical operation into the
object being used. • Param grid: a list of parameter values and
code. Also, it allows you to include large multidimensional
their names. • cv: an integer represents the folds for a K-fold
arrays and matrices in your code. The Pandas library is
cross-validation.
excellent for data manipulation and analysis; it is extensively
used for importing and organizing datasets. Matplotlib and
D. Classification Models
Seaborn are the foundations of Python data visualization.
Following the preceding stages, the desired classifiers were
Matplotlib is a Python library that can be used to plot 2D
chosen and applied. Deep Learning (DL) and various Machine
graphs with the help of other libraries such as Numpy and
Learning (ML) algorithms were explored, including K-Nearest
Pandas. Seaborn is used to plot graphs using Matplotlib,
Neighbors (KNN), Support Vector Machine (SVM), Decision
Pandas, and Numpy. The last one is Sklearn, the most usable
Tree (DT), Random Forest (RF), and Multi-Layer Perceptron
and robust machine learning package in Python. It provides a
(MLP). [Link]. K-nearest neighbors (KNN) The supervised
Python-based consistency interface as well as tools for
machine learning algorithm KNN is a simple and
classification, regression, clustering, and dimensionality
straightforward technique. The KNN algorithm assumes that
reduction (Desai, 2019).
related things are located in close proximity. In other words,
Standard Scaler = xi − mean(x) stdev(x) (1) comparable objects are close to each other (Hossain et al.,
2019). [Link]. Support vector machine (SVM) Based on recent
advances in statistical learning theory, SVM is part of a new
generation of learning systems. It is a linear and non-linear
data algorithm. It converts the original data into a higher
dimension, from which it may create a hyperplane for data
separation using support vectors, which are crucial training
tuples (Bind et al., 2015). [Link]. Decision tree (DT) The DT
belongs to the supervised learning algorithm family. Unlike
other supervised learning algorithms, the decision tree
approach may also be used for regression and classification.
Because of its resilience to noise, tolerance for missing
information, management of irrelevant redundant predictive
attribute values, low processing cost, interpretability, and
robust predictors, the DT is one of the most popular and widely
used machine learning algorithms (Charbuty and Abdulazeez,
2021). [Link]. Random forest (RF) RF is a collection of
classifiers based on decision trees. Each tree is built using a
bootstrap sample from the data and a candidate set of features
chosen at random. It employs both bagging and random
variable selection for tree construction. Once the forest has
been built, test instances are percolated down each tree, and
the trees give class predictions for their particular classes. A
random forest’s error rate is determined by the strength of
each tree and the correlation between any two trees. It may be
used to naturally order the relevance of variables in a 2) XGBoost (Extreme Gradient Boosting: XGBoost
regression or classification task (Bind et al., 2015). constructs an ensemble of decision trees in a sequential
[Link]. Multilayer perceptron (MLP) MLP is a feed-forward manner, with each tree aiming to correct the errors of the
artificial neural network having three nodes: an input layer, a preceding ones. XGBoost leverages gradient descent
hidden layer, and an output layer. The input signal that will be optimization to minimize a differentiable loss function,
processed is received by the input layer. The output layer is resulting in improved model performance [17]. Objective
responsible for tasks like prediction and classification. The Function:
MLP’s true computational engine is an arbitrary number of
hidden layers sandwiched between the input and output layers. Objective(XGBoost) = L(θ) + Ω(f)(2)
Data in an MLP moves forward from the input to the output
layer, comparable to a feedforward network. The neurons of
the MLP are trained using the back propagation learning
approach (Abirami and Chitra, 2020).
1) Decision Tree Classifier: Decision trees are versatile
supervised machine learning algorithms commonly used for
classification and regression tasks. They make predictions by
recursively partitioning the data set based on feature
attributes. Information Gain (IG) Formula (for classification):

where H is the entropy [16].

 3) Random Forest Classifier: Random Forest developed
by Leo Breiman [20] is an ensemble learning technique that
combines multiple decision trees to enhance classification
accuracy. It introduces randomness by creating bootstrapped
samples from the data set and using random feature selection
when building each tree. The final prediction is determined
through a voting mechanism for classification tasks or
averaging for regression tasks, where each tree’s prediction
contributes to the overall decision.

4) SVM
Support Vector Machine (SVM) is a supervised machine
learning algorithm used primarily for classification tasks but
can also be extended to regression (SVR). SVM works by
finding the best hyperplane that divides the data points into
different classes in a high-dimensional space. The objective is
to maximize the margin, which is the distance between the
hyperplane and the closest data points from each class, called
support vectors. This maximization helps ensure better
classification and generalization to new data.

6) Logistic Regression

Logistic regression is a statistical method used for binary

classification problems where the outcome is a categorical
variable, typically 0 or 1. Unlike linear regression, which
predicts continuous outcomes, logistic regression predicts the
probability that a given input belongs to a specific class. The
algorithm models the relationship between one or more
independent variables (features) and the dependent variable
(the class) using the logistic function, also known as the
sigmoid function.
p(X)=i/1+e−(β0 +β1 X1 +β2 X2 +⋯+βn Xn )1
 5) Naive Bayes
Naive Bayes is a probabilistic classifier based on Bayes’
8)KNN
Theorem, with the assumption that features are conditionally
independent given the class label. Despite the “naive”
assumption of independence, it often performs well in real- KNN is a non-parametric, instance-based learning algorithm
world applications, especially in text classification. that classifies data points based on the majority class of their
nearest neighbors in the feature space. The algorithm assumes
P(C∣X)=P(X)P(X∣C)⋅P(C) (2) that similar instances exist in close proximity to one another.

f(x)=k1 i=1∑k yi
 F. Performance Metrics
For the proposed work, the three standard performance
metrics are used - Precision, Recall and F1 Score.
1) Precision: Precision measures the proportion of true
positive predictions among all positive predictions made.
True Positives
Precision = (5) True
Positives + False Positives

Fig. 2. Proposed Architecture

2) Recall: Recall measures the proportion of true

positive predictions among all actual positive instances in the
data.
True Positives
Recall = (6) True
E. Resampling Positives + False Negatives
In this dataset, the target attribute has a majority class of 3) F1 Score: F1 score is the harmonic mean of precision
non-fraud cases so, the dataset is resampled using SMOTE over and recall, providing a balanced evaluation metric.
sampling technique. Credit card fraud detection is a
classification challenge where the aim is to achieve Precision × Recall
classification for a given instance in the data set whether it is a F1 = 2 × (7) Precision +
fraud or not. Balancing the data set is a critical step in building Recall
an effective fraud detection model. The data set typically These metrics are widely used in evaluating the performance
exhibits a severe class imbalance, with a vast majority of of classification models. They provide insights into the model’s
legitimate (non-fraudulent) transactions and a relatively small ability to correctly classify positive instances (precision), its
number of fraudulent ones. Research in the field of fraud ability to capture all positive instances (recall), and a balance
detection, such as studies by [5], highlights the importance of between the two (F1 score).
addressing class imbalance for robust model performance.
Balancing the data set involves addressing this imbalance by IV. PROPOSED ARCHITECTURE
either oversampling the minority class (fraudulent
In figure 2, firstly the data has been acquired, then on this
transactions) or under sampling the majority class (legitimate
data EDA was performed and the data was preprocessed for
transactions. Techniques such as Synthetic Minority Over-
further work where this dataset was resampled using
sampling Technique (SMOTE), introduced by [6], are commonly
balancing techniques and then the dataset is divided into
used to generate synthetic examples of the minority class,
training and test data and the models were trained and the
which helps create a more balanced distribution of data. In the
results were produced.
paper [7], it is found that the random oversampling method
has come out to be the best suited technique over the V. RESULTS AND DISCUSSIONS
imbalance data and yields 0.99 precision and 0.99 accuracy This paper attempts to develop an efficient method of
score when applied on XGBoost. In our proposed work, we detecting PD by using voice samples. We used the UCI dataset,
have however used a SMOTE over sampling method for which contains 195 records of voice signal features collected
balancing our data set which is suggested by most of the from 147 PD and 48 HC. We compared different techniques and
researchers around the globe. how they affected our models, although we used various
traditional machine learning and deep learning algorithms, Also, the voice diagnosis will be better in terms of low cost,
such as k-nearest neighbors (KNN), support vector machine simplicity, and it can be easily incorporated into healthcare.
(SVM), decision tree (DT), random forest (RF), and multilayer This research diagnoses PD by applying several classification
perceptron (MLP). We split the sample into two groups of data models and comparing their performance to choose the most
remaining samples. Our dataset appears to be unbalanced accurate one.
because there are much more PD than HC. To address this
issue, we used the Synthetic Minority Over-sampling Technique
(SMOTE) to balance the dataset. GridSearchCV and SelectKBest
were also performed to choose the best features in order to
determine the best hyperparameters for our models. The
following figures show the results with and without using the
SMOTE, as well as with and without GridSearchCV and
FeatureSelection. The results in Figures 2, 3 showed that using
SMOTE and GridSearch achieved the best performance.
However, due to the importance of each feature in the training,
we discovered that using FeatureSelection in our dataset did
not yield satisfactory results. As a result, we chose to use all
features. Fig. 3. Confusion Matrix for XGBoost
According to new findings (Ma et al., 2020), voice dysfunction is
the first indicator of motor impairment in PD. Because of the
complexity and precision required for vocalization, VI. CONCLUSION & FUTURE SCOPE
malfunctions may occur here before the limbs. In perceptual
and auditory studies, the voice in Parkinson's disease exhibits In conclusion, we proposed using machine learning and deep
distinct changes. So, we are optimistic about the use of voice as learning approaches to Identify Parkinson’s Disease by using voice
a dense biomarker for PD. Our approach exclusively employs signal features. These methods’ results (SVM 95% and MLP 98.31%)
voice measurements for clinical diagnosis, as opposed to the are more accurate than previous works. The proposed working
most generally acknowledged biomarkers for diagnosis, such as model can help in reducing treatment costs by providing initial
DaT scans or clinician-scored supervised motor assessments in diagnostics on time. This model can also be used as a teaching tool
the Unified Parkinson's Disease Rating Scale (UPDRS). Since the for medical students and as a soft diagnostic tool for physicians.
voice is one of the first visible signs, we believe that using it will Also, the accuracy and scalability of this prediction model can both
give a faster and more accurate diagnosis than traditional and be improved with numerous possible improvements.
harmful diagnostic methods, such as handwriting and MRI.

References
[1] Al-Fatlawi A. H., Jabardi M. H., Ling S. H. (2016). “Efficient diagnosis system for
parkinson's disease using deep belief network,” in 2016 IEEE Congress on evolutionary
computation (CEC) (Vancouver, BC, Canada: IEEE; ), 1324–1330.
10.1109/CEC.2016.7743941 [CrossRef] [Google Scholar]
[2] Bilgen I., Guvercin G., Rekik I. (2020). Machine learning methods for brain network
classification: application to autism diagnosis using cortical morphological networks. J.
Neurosci. Meth. 343, 108799. 10.1016/[Link].2020.108799 [PubMed] [CrossRef]
[Google Scholar]
[3] Bind S., Tiwari A. K., Sahani A. K., Koulibaly P., Nobili F., Pagani M., et al.. (2015). A survey
of machine learning based approaches for parkinson disease prediction. Int. J. Comput.
Sci. Inf. Technol. 6, 1648–1655. [Google Scholar]
[4] Brownlee J. (2020). Imbalanced Classification with Python: Better Metrics, Balance
Skewed Classes, Cost-sensitive Learning. Machine Learning Mastery. Available online at:
[Link]
dq=Brownlee,+J.+(2020).+%E2%80%9CImbalanced+classification+with+Python:+be
tter+metrics,+balance+skewed+classes,+cost-
 sensitive+learning,%E2%80%9D+in+Machine+Learning+Mastery&ots=CfNF8NM2X
W&sig=6urQFaaAxqDDHzqTPTI9yjzr0rQ&redir_esc=y#v=onepage&q&f=false
[5] Charbuty B., Abdulazeez A. (2021). Classification based on decision tree algorithm for
machine learning. J. Appl. Sci. Technol. Trends. 2, 20–28. 10.38094/jastt20165
[CrossRef] [Google Scholar]
[6] Das R. (2010). A comparison of multiple classification methods for diagnosis of
parkinson disease. Expert Syst. Appl. 37, 1568–1572. 10.1016/[Link].2009.06.040
[CrossRef] [Google Scholar]
[7] Desai R. (2019). Top 10 Python Libraries for Data Science. Available online at:
[Link]
cd82294ec266 (accessed July 3, 2022).
[8] Fothergill-Misbah N., Maroo H., Hooker J., Kwasa J., Walker R. (2020). Parkinson's
disease medication in kenya–situation analysis. Pharmaceutica l J. Kenya. 24, 38–41.
[Google Scholar]
[9] Gil D., Manuel D. J. (2009). Diagnosing parkinson by using artificial neural networks and
support vector machines. Glob. J. Comput. Sci. Technol. 9, 63–71. Available online at:
[Link]
e=pdf
[10] Harel B., Cannizzaro M., Snyder P. J. (2004). Variability in fundamental frequency
during speech in prodromal and incipient parkinson's disease: a longitudinal case study.
Brain Cognit. 56, 24–29. 10.1016/[Link].2004.05.002 [PubMed] [CrossRef] [Google
Scholar]
[11] Hossain E., Hossain M. F., Rahaman M. A. (2019). “A color and texture based approach
for the detection and classification of plant leaf disease using knn classifier,” in 2019
International Conference on Electrical, Computer and Communication Engineering
(ECCE) (Cox's Bazar: IEEE; ), 1–6. 10.1109/ECACE.2019.8679247 [CrossRef] [Google
Scholar]
[12] Jayaswal V. (2020). Performance Metrics: Confusion Matrix, Precision, Recall, and f1
Score. Available online at: [Link]
confusion-matrix-precision-recall-and-f1-score (accessed December 6, 2021).
[13] Kadam V. J., Jadhav S. M. (2019). “Feature ensemble learning based on sparse
autoencoders for diagnosis of parkinson's disease,” in Computing, Communication and
Signal Processing. Advances in Intelligent Systems and Computing, Vol. 810, eds B. Iyer, S.
Nalbalwar, N. Pathak (Singapore: Springer; ), 567–581. 10.1007/978-981-13-1513-
8_58 [CrossRef] [Google Scholar]
[14] Little M. (2008). UCI Machine Learning Repository. Irvine, CA: University of California,
School of Information and Computer Science. Available online at:
[Link] (accessed March 17, 2023).