J.
Electrical Systems 20-2 (2024): 2255-2266
Early Detection of Parkinson’s
1
Dr.Shweta S.Salunkhe
²Samita Ganveer Disease Using Machine Learning
³Himani Bire
⁴Rutuja Deshmukh
Abstract: - A hallmark of Parkinson's disease is the degeneration of dopaminergic neurons in the midbrain's substantia nigra pars
compacta. However, machine learning is needed to design and implement early Parkinson's disease detection. by using machine
learning methods such as CNN and SVM, which can reliably identify voice signals and spiral images to identify early indications of
Parkinson's disease. Using machine learning on handwriting, tremor, and gait datasets, the method addresses the shortcomings of
individual analyses for a more complete diagnosis solution by investigating relationships between symptoms. This increases accuracy.
When voice and spiral drawing data were combined, Parkinson's disease diagnosis accuracy showed promise, with the machine learning
model successfully differentiating between unaffected patients and those who were affected. This observation suggests a viable path
for precise Parkinson's disease diagnosis: an integrated method that combines machine learning techniques with data from spiral
drawings and voice.
Keywords: Machine learning, Python, voice dataset, spiral dataset, Support Vector Machine (SVM), Convolutional
Neural Network (CNN).
I. INTRODUCTION
Diseases affecting the nervous system are a major health concern due to their high mortality rates. Nervous system
diseases are projected to cause more deaths than cancer soon [2]. Parkinson's disease is one such illness that causes
specific motor symptoms due to the loss of dopaminergic neurons in the midbrain's substantia nigra pars compacta.
Early detection is crucial for timely intervention and improved patient outcomes, but current diagnostic methods
often lack the necessary sensitivity to identify the disease in its early stages.
While machine learning techniques such as support vector machines (SVM) and convolutional neural networks
(CNN) have shown promise in various medical applications, their specific application to voice datasets and spiral
images for early detection of Parkinson's disease remains an underexplored area. It is known that dopamine
deficiency in the striatum is responsible for motor skills in Parkinson's disease [1]. While clinical evaluations are
currently the primary method for diagnosing the disease, they may not be able to identify subtle early indicators.
By utilizing cutting-edge machine learning algorithms like SVMs and CNNs to analyze voice sounds and spiral
drawings, a more comprehensive diagnostic strategy can be developed, overcoming the limitations of current
approaches and improving sensitivity and specificity. This work aims to bridge the gap in the early diagnosis of
Parkinson's disease by evaluating voice samples and spiral imagery using SVM and CNN algorithms. It is
anticipated that this approach will significantly improve the effectiveness and accuracy of early-stage Parkinson's
disease identification.
By examining the connections between speech signals and motor function as represented in spiral drawings, this
research seeks to develop a distinct and reliable diagnostic paradigm. The creation of a trustworthy and accessible
early detection tool could potentially lead to better patient outcomes in the treatment of Parkinson's disease. The
ultimate goal of this work is to improve the early diagnosis of Parkinson's disease by utilizing CNN and SVM
Copyright © JES 2024 on-line: journal.esrgroups.org
2255
J. Electrical Systems 20-2 (2024): 2255-2266
algorithms with spiral and audio datasets. By correlating speech signals and motor function in spiral images, it is
expected that the accuracy of the diagnosis will increase. Successful validation of this approach could lead to the
development of a reliable machine-learning-based diagnostic tool for improved treatment of Parkinson's disease.
II. LITERATURE REVIEW
The neurodegenerative condition known as Parkinson's disease (PD) is characterized by both motor and non-
motor symptoms.
Early detection is crucial for effective disease management, leading researchers to explore the use of machine
learning (ML) techniques for timely diagnosis. Studies have shown that ML systems utilizing speech signals have
the potential for non-invasive and early detection methods [3].
By analyzing multiple types of speech signals, these systems can contribute to a more accurate and timely
diagnosis. The understanding of Parkinson's genetics has undergone significant changes, as highlighted by M.J.
Farrer in a comprehensive review [1]. Identifying genetic factors is essential for a nuanced understanding of the
disease. Deep learning techniques, specifically convolutional neural networks (CNNs), have been applied to both
voice and spiral datasets, demonstrating their potential for early PD diagnosis [10] and [12]. The intricate patterns
detected by CNNs showcase their ability to identify subtle signs of the disease. Datasets such as Hand PD [4] and
Parkinson's Drawings [5] provide a visual dimension for ML applications, opening up new possibilities for image-
based diagnostic tools. While traditional methods for Parkinson's diagnosis have relied on clinical assessments,
they have limitations in accuracy and early detection [2]. Therefore, it is essential to augment these methods with
advanced technologies like ML. While SVMs have been used for PD diagnosis [6], a comparative analysis with
CNN models reveals the strengths and weaknesses of each approach. SVMs, with their interpretability, contrast
with the intricate pattern recognition capabilities of CNNs. Emerging trends, such as the integration of diverse
datasets [13], hold promise for refining diagnostic models. However, challenges still exist, and continuous
interdisciplinary collaboration is necessary to address ethical concerns.
In conclusion, the integration of ML, particularly deep learning techniques, into Parkinson's disease diagnosis
presents a transformative approach. As technology advances and datasets expand, the potential for early detection
and effective management becomes increasingly viable.
III. METHODOLOGY
"Machine learning techniques can be effectively utilized to solve problems with the lowest possible error rate. In
this study, we use a speech dataset related to Parkinson's disease, obtained from the UCI Machine Learning library.
By combining spiral drawing inputs from both normal individuals and those with Parkinson's disease, our system
can produce accurate results. Our proposed hybrid approach incorporates both spiral drawing and speech data
from patients, resulting in reliable findings. This integrated approach allows physicians to determine the presence
and severity of the illness and provide appropriate treatment. Additionally, we suggest including handwriting
samples, tremors, and gait as part of the dataset for Parkinson's disease diagnosis. By identifying correlations
between these symptoms, we can improve the accuracy of our diagnosis. While analyzing each symptom
separately may have limitations, such as the complexity of handwriting and the potential influence of other factors
on motor movement, incorporating multiple symptoms can help mitigate these issues. Furthermore, previous
studies have shown that using breath samples alone may not yield clinically meaningful results for speech
recognition.
Therefore, we have also included additional steps, such as noise removal and speech segmentation, to improve
the accuracy of our approach. By incorporating multiple symptoms, we aim to address the limitations of relying
on a single symptom for diagnosis.
A. Proposed System
2256
J. Electrical Systems 20-2 (2024): 2255-2266
Figure 1 Proposed system
B. Data sets
a. Voice
To diagnose diseases, data is collected from Kaggle and then processed.
• The dataset consists of measurements from 31 individuals' biological voices, 23 of whom have been
diagnosed with Parkinson's disease (PD).
• Each row in the table represents one of the 181 voice recordings from these individuals, with each column
representing a specific voice measure (referred to as the "name" column).
• The main purpose of this data is to distinguish between individuals with PD and those who are healthy,
as indicated by the "status" column, where a value of 0 represents a healthy individual and a value of 1 represents
someone with PD.
• The data is stored in CSV ASCII format, with each row in the CSV file containing one instance of a
voice recording. Each patient has about six recordings available, with the patient's name listed in the first column.
b. Spiral
The second data set was retrieved from Kaggle and then processed.
• There is a spiral pattern in the data set.
• The format of each one is PNG. There are training and testing data sets inside the data set.
• Personal details like age and gender are not disclosed. The training data includes 72 spiral drawings,
with 36 created by Parkinson's sufferers and 36 by healthy individuals.
• The testing data include thirty spiral drawings in total, fifteen of which were created by patients with
Parkinson’s disease and fifteen by Healthy individuals.
C. Data-Pre-Processing
A. Voice data
Clean and preprocess the data, normalize the features, and handle missing values if necessary. This step is
crucial to ensure the quality of the dataset.
a. Normalization
2257
J. Electrical Systems 20-2 (2024): 2255-2266
• Min-max scaling, often referred to as normalizing, is a preprocessing method for data that is used to convert
numerical variables or characteristics to a range, usually between 0 and 1.
• ensuring all the variables are on the same scale without misrepresenting their relative differences is the aim
of min-max scaling.
• The following formula may be used to apply min-max scaling to a feature:
(X – X_min) / (X_max - X_min) equals X_scaled.
B. Spiral Image
• Data augmentation: This involves making various adjustments to the training image to increase the
diversity of the training process; This will help increase the generality and robustness of the model.
• Rescale: This parameter scales the pixel values of the image. In the case, it rescales the pixel values to [0,1]
by dividing each pixel’s value by 255. This is a first step to enable equipment to emulate a neural network, which
can help improve convergence during education.
• Rotation_range: This parameter specifies the range of the image that can be rotated. In this case, the image
can be rotated at any angle within the range of [-40,40] degrees. Width_shift_ range and height_shift_range.
Parameters control the horizontal and vertical range of the image respectively. In the case, the image can be moved
horizontally up to 20% of the total width and vertically up to 20% of the total height.
• Shear_range: shear moves part of the image in a fixed direction while preserving other parts. This
parameter controls where the crop is applied to the image.
• Zoom_range: This parameter specifies the range of images to be randomly zooed. In this case, the image
can be expanded up to 20%.
• Horizontal_flip: This parameter specifies whether the image will be allowed to flip horizontally. Randomly
flipping horizontally will mirror the images horizontally, which will help make a difference in the study material.
• Fill_mode: This parameter specifies the strategy for filling new pixels that will appear after various changes
such as rotation or replacement. In this case “nearest” is used; This means that the closest available pixel value
will be used to write the new pixel.
a. Instruction of Models
Data splitting is a term used to describe the process of training, validating, and testing models.
i. Training Set: A machine learning model is trained using the training set, which is a subset of the dataset.
It is used to train the model to identify patterns and correlations within the data and comprises most of the data.
ii. Test Set: This entirely separate subset is used to assess how well the trained model performs in the end.
The test set acts as a stand-in for hidden, real-world data. It offers an assessment of the model's generalization
performance and aids in evaluating how effectively the model functions in real-world situations
IV. FEATURE DESCRIPTION
A. Voice dataset
Table 1 Voice Attributes
Attributes Description
Name ASCII subject name and recording
number
2258
J. Electrical Systems 20-2 (2024): 2255-2266
Table 1 Voice Attributes
MDVP: Fo(Hz) Average vocal fundamental
frequency
MDVP: Fhi(Hz) Maximum vocal fundamental
frequency
MDVP: Flo(Hz) Minimum vocal fundamental
frequency
MDVP; Jitter(%),MDVP: Several
Jitter(Abs),MDVP: RAP,MDVP: measures of variation in fundamental
PPQ ,Jitter: DDP frequency
MDVP: Shimmer, MDVP: Several measures of variation in
Shimmer(dB),Shimmer:APQ3,Shim amplitude
mer:APQ5,MDVP:APQ,Shimmer:D
DA
NHR,HNR Two measures of ratio of noise to
tonal components in the voice
status Health status of the subject (one) -
Parkinson's, (zero) – healthy
RPDE,D2 Two nonlinear dynamical
complexity measures
DFA Signal fractal scaling exponent
spread1,spread2,PPE Three nonlinear measures of
fundamental frequency variation
B. Spiral dataset
Extracting statistical features from an image using the Gray Level Co-occurrence Matrix (GLCM) technique.
GLCM is a technique used to describe the spatial relationship between pixels in an image. It quantifies how often
pairs of pixel intensities with specific values occur about each other at a specified distance and direction within
an image. After computing the GLCM, it extracts four statistical features: Contrast, Energy, Correlation, and
Homogeneity.
i.Contrast:
Contrast is the measure of the contrast between a pixel and its neighbouring pixels throughout an image.
Calculates local changes in the grey level correlation matrix(GLCM). More contrast compared to the difference
between adjacent pixels makes the image more beautiful.
Calculated as the sum of the squares of the differences between the pixel values of all possible pixel pairs in the
GLCM and their averages. Similar results indicate more texture or roughness in the image.
ii.Energy:
Energy, also known as angular second time, represents the uniformity or smoothness of the beautiful image.
Indicates the distribution decision of pixel pairs by evaluating the sum of squares of the point in GLCM. Higher
2259
J. Electrical Systems 20-2 (2024): 2255-2266
energy values indicate a smoother texture where pixel values are similar across the images. For example, lower
energy values indicate more tissue. The results show that there is a lot of complexity and a large number of textures
with different patterns.
iii. Correlation:
Correlation measures the line between the grey levels of adjacent pixels in an image.
Shows the correlation or relationship between pixel values in GLCM. Values close to 1 indicate good correlation,
meaning adjacent pixels will have similar values. A value close to -1 indicates a negative correlation; This means
that neighbouring pixels will have similar values. Pixels have different values. Correlation values around 0
indicate poor or no correlation between pixel intensities.
iv. Homogeneity:
Homogeneity, also known as contrast moment, measures the closeness of pixel intensities and distribution of
elements in a GLCM along the GLCM diagonal. Measures the local uniformity of the beautiful images. The higher
the consistency value, the closer the GLCM points are to the diagonals, which means the texture becomes more
and more regular. Lower homogeneity values indicate disorganization of GLCM elements, indicating diversity
and heterogeneity.
v. Stroke thickness:
This function calculates the stroke thickness of object in the binary image: Multiplies the distance between the
skeleton to get the contour images; where each pixel value represents the distance from the pixel to the stroke
weight, according to the skeleton.
V. MODELS
A. Support Vector Machine
A popular supervised learning technique for classification and regression problems is support vector machines
(SVMs). An algorithm for supervised learning is a support vector machine. An N-dimensional hyperplane is
formed and classification is performed using an SVM, which models the data into k categories. Train the SVM
model using the training data. The SVM algorithm aims to find the optimal hyperplane (or decision boundary)
that maximizes the margin between different classes in the feature space.
Mathematically, the linear kernel between two feature vectors x and y is given by:
K (x, y) = x ^T. y
Figure 2 SVM Model
2260
J. Electrical Systems 20-2 (2024): 2255-2266
B. Convolutional Neural Network
One kind of deep neural network that is specially made for tasks requiring visual input, such as image and video
recognition, is the convolutional neural network (CNN). CNNs have proven to be highly effective in a wide range
of computer vision applications, including image classification, object detection, facial recognition, and more.
They are inspired by the visual processing that occurs in the human brain and are particularly well-suited for
handling grid-like data, such as pixels in an image.
Figure 3 CNN Model
VI. RESULT
The proposed work has considered a limited data set for both. The existing work has expressed the probability of
using the voice dataset of humans as an effective tool for PD diagnosis when we use the SVM model as it gives
an accuracy of 87% and the CNN model gives an accuracy of 82.05%. It also expressed the probability of using
the spiral dataset as an effective tool for PD diagnosis when we use the CNN model as it gives an accuracy of
93% and the SVM model it gives an accuracy of 60% However when machine learning and deep learning are the
emerging techniques to be applied for healthcare sector, it requires significant execution and accurate results to
demonstrate the technique. This work has suggested machine learning methods for the classification of spiral
drawing images of healthy control and PD patients to aid the medical practitioner in diagnosing the diseases at an
early stage.
Table 2 Result Evaluation table
Accuracy
Dataset Classification model
SVM 87%
Voice dataset CNN 82.05%
CNN 93%
Spiral dataset SVM 60%
A.Metrics
tp:true positive, tn: true negative
fp: false positive, fn: false negative.
Accuracy: Classification accuracy is what we usually mean when we use the word accuracy. It is the ratio of the
number of correct predictions to the total number of input samples.
Accuracy= correctly classified sequence/ all samples
2261
J. Electrical Systems 20-2 (2024): 2255-2266
Accuracy: Also called positive predictive value, is the proportion of relevant events in the sample. The sample
was returned. Intuitively, precision is the classifier’s ability to not write good example when it is not good. The
higher the value, the better the classifier.
Return= tp/(tp+tn+fp+fn)
Reflection(return) is the fraction of all relevant events returned.
Recall is the ability of a classifier to intuitively find all good patterns. The higher the number, the more accurate
it is.
Recall= tp/(tp+fn)
Specificity (also called negative value): Specificity related to the ability of the class to detect negative
consequences.
F1 score: F1 score*((sensitivity*return) / (sensitivity + return)). Also called F score or F-measure.
When F1, it shows the balance between accuracy and return.
F1= (2*(tp/(tp+fp+tn+fn))*(tp/(tp+fn)))/((tp/tp+fp+tn+fn))+(tp/(tp+fn))
Figure 4 Confusion matrix of voice data using SVM
Figure 5 Confusion matrix of voice data using CNN
2262
J. Electrical Systems 20-2 (2024): 2255-2266
Figure 6 Confusion matrix of spiral image using SVM
Figure 7 Confusion matrix of spiral image using CNN
Figure 8 Model accuracy histogram for voice data set
Figure 9 Model accuracy histogram for spiral data set
2263
J. Electrical Systems 20-2 (2024): 2255-2266
Figure 10 Healthy Image
Figure 12 Parkinson’s Image Figure 13 Feature Extracted Image
Among traditional machine learning models, naive Bayes and logistic regression outperform decision trees in
terms of voice data classification accuracy. SVM performs better on voice data than on spiral data, showing better
sensitivity to data features. CNN outperforms traditional machine learning models, especially when processing
complex data such as voice and spiral data. It is especially effective in voice data classification, with an accuracy
2264
J. Electrical Systems 20-2 (2024): 2255-2266
of 82.03% compared to CNN, SVM and traditional machine learning models. Spiral object seems difficult for
SVM as it only achieves 60% accuracy while CNN's accuracy is higher at 93%.
VII. DISCUSSION
The performance of SVM and CNN in many areas, such as voice data and spiral image data, is affected by the
characteristics of the data and the algorithm itself. For voice data, SVMs are generally effective because they have
high feature values and can handle data relationships. Voice data are often represented by high-dimensional
feature vectors; where each feature corresponds to a different part of the speech signal, such as frequency
components or spectrogram representations. SVM can divide the height of the surface into different classes,
making them suitable for tasks such as speech recognition and classification. Additionally, voice data often has
arbitrary decision boundaries, and SVM can effectively capture these boundaries using techniques such as kernel
techniques. In contrast, CNNs may not perform well on speech objects because they are designed to learn
representations of objects (e.g., images). Although speech data can be represented as spectrograms or other
visualizations, it does not have the spatial structure that good CNNs use. Therefore, CNN may not be able to
utilize its full hierarchical feature extraction capabilities for speech data, resulting in fewer features compared to
SVM. For spiral data, CNNs perform well because they are specifically designed to learn hierarchical properties
of spatial data. Spiral images, like other visual objects, have a relationship between pixels that captures important
patterns and patterns. CNNs use convolutional layers to extract hierarchical features from image data, capturing
low-level features such as edges and high-level features such as shape and texture. This algorithm makes CNNs
efficient for tasks such as image classification and pattern recognition, including line detection in images. SVMs,
on the other hand, may have problems handling convoluted data because they rely on identifying classes at a
particular location. If objects have irregular relationships (such as spirals), SVM will have difficulty finding well-
defined boundaries separating classes. SVM often requires architecture or the use of kernel methods to handle
nonlinear data; this may not be as good as CNN's ability to learn hierarchical features directly from raw data. In
summary, SVM is suitable for voice data due to its efficiency. CNN is good at processing highly spatial and
temporal data, while CNN is good at processing convoluted data because they can learn hierarchical features of
spatial data, which are important to capture. The patterns and patterns found in the spiral image are very important.
VIII. CONCLUSION
Parkinson's disease cannot be diagnosed directly; That is, a specific test such as a blood test or electrocardiogram
cannot determine whether a person has Parkinson's disease. The doctor will take a detailed medical history of the
patient and then perform a psychological examination. They identified at least two significant symptoms in the
subjects and then predicted whether the patients had Parkinson's disease. The prognosis of PD is poor because
there is no definitive diagnosis. In this case, it will help us assist doctors by providing machine learning models.
Predictive models were developed for these key features using machine learning methods of support vector
machines and convolutional neural networks. This article draws some conclusions. First of all, Parkinson's disease
can be diagnosed using patterns and recording systems. Looking at the results, it can be seen that SVM and CNN
models provide different accuracies for different data. On speech data, SVM is more accurate than CNN, while
on spiral data, CNN is more accurate than SVM. From these results, we can conclude that CNN is more effective
for images and SVM is better for registration.
ACKNOWLEDGMENT
I feel profound happiness in forwarding this paper and indebted to all who generously helped by sharing valuable
experience and devoting their precious time, without which this project
paper would have never been accomplished. I extend my heart felt gratitude to my project guide. We avail this
opportunity to express my sincere gratitude to all those people who contributed to this paper I am grateful to Prof.
Dr. S. R. Patil, Principal and head of department for their continuous encouragement and support. At last, I would
like to thank my parents for rendering their continuous love and care during the course of this work, also to my
friends to help me out to carry out my work directly or indirectly.
2265
J. Electrical Systems 20-2 (2024): 2255-2266
REFERENCES
[1] M.J. Farrer, “Genetics of Parkinson disease: paradigm shifts and prospects”, Nature Reviews Genetics, vol. 7, no. 4, pp.
306-318, 2006.
[2] M. Lozano and A. E. Lang, “Parkinson’s disease,” New England Journal of Medicine, vol. 339, no. 15, pp. 1044–1053,
1998.
[3] Cantürk and F. Karabiber, ‘‘A machine learning system for the diagnosis of Parkinson’s disease from speech signals and
its application to multiple speech signal types,” Arabian J. Sci. Eng., vol. 41, pp. 5049– 5059, Dec. 2016
[4] Pereira, "Welcome to the HandPD dataset," HandPD, 2017. [Online]. Available:
https://www.p.fc.unesp.br/~papa/pub/datasets/Handpd. [Accessed 3rd September 2022].
[5] K. Mader, "Parkinson's Drawings," Kaggle.com, 2019. [Online]. Available:
[6] https://www.kaggle.com/kmader/parkinsons-drawings. [Accessed 27 September 2021].
[7] Huseyn, "Deep Learning Based Early Diagnosis of Parkinson’s Disease," Cornell University arXiv.org, August 2020.
[8] Figure 2 https://www.javatpoint.com/machine-lerning-support-vector-machine-algorithm.
[9] Figure 3. http://images.app.goo.gl/wJ1RnN6crcRhsCmk7.
[10] Editorial Team, "Diagnosis-Early Symptoms & Early Diagnosis," parkinsonsdisease.net, 8 March 2017. [Online].
Available:
[11] https://parkinsonsdisease.net/diagnosis/early-symptoms-signs. [Accessed 3rd September 2022].
[12] Pereira, "Convolutional neural networks applied for Parkinson's disease identification,” Machine Learning for Health
Informatics, 2017.
[13] C. Pereira, "Deep Learning-Aided Parkinson's Disease Diagnosis from Handwritten Dynamics.,” Graphics. Patterns and
Images (SIBGRAPI) 2016 29th SIBGRAPI Conference IEEE, 2016.
[14] M. Gil-Martin, "Parkinson's Disease Detection from Drawing Movements using convolutional neural networks," in
Electronics (MDPI), 2019.
[15] DraganaMiljkovic et al, “Machine Learningand Data Mining Methods for Managing Parkinson’s Disease” LNAI 9605,
pp 209-220, 2016.
[16] International Conference on Advanced Computation and Telecommunication (ICACAT) held in Bhopal, India, in 2018.
2266
