Enhanced Parkinson's Disease Prediction Using

Feature Selection, SMOTE, and Machine Learning

with Deep Learning Models
Al-Amain Rajon Bardhan
Dept. of Computer Science and Mirza Nadim Saad Esrat Jahan Dept. of Computer & Cyber
Engineering Dept. of Computer Science and Dept. of Computer Science and Science
Southeast University Engineering Engineering Augusta University
Dhaka, Bangladesh Southeast University Southeast University Georgia, United States
[email protected] Dhaka, Bangladesh Dhaka, Bangladesh [email protected]
[email protected] [email protected]

features of Parkinson's is necessary to better understand and

Abstract— This research leverages the UCL Parkinson's
predict its progression [6,7]. In clinical settings, the Unified
Disease clinical dataset to predict Parkinson's disease using both
machine learning and deep learning approaches. To address class
imbalance, SMOTE was applied, and feature selection was
conducted using the sci2 tool. The study assessed six deep learning Parkinson’s Disease Rating Scale (UPDRS) has become the
algorithms, such as LSTM, PLSTM, and MLP, as well as six gold standard for assessing the severity of symptoms and
machine learning classifiers, including SVM, KNN, Logistic tracking the progression of the disease [8-11]. Yet, while this
Regression, Naive Bayes, Decision Tree, and AdaBoost. MLP scale helps to monitor the physical manifestations of the disease,
outperformed the others in terms of accuracy for deep learning there is a growing call for more comprehensive approaches that
models. With 97.20% accuracy, 97.44% precision, 97.20% recall, incorporate genetic and molecular data to predict how the
and 97.32% F1-score, KNN was the most successful overall, disease might evolve. The National Human Genome Research
nonetheless. These findings show how well KNN detects Institute envisions a future where epigenetics and
Parkinson's disease and emphasize the technology's potential for transcriptomics will be used routinely in predictive models,
early diagnosis and monitoring. shedding light on the link between genotype and phenotype by
the year 2030 [12]. Although Parkinson's affects both men and
Keywords— Parkinson's Disease, SMOTE, Scaling, Machine women, research indicates that men are more susceptible to the
Learning, Deep Learning, KNN
disease [13]. This growing burden, particularly in countries like
I. INTRODUCTION India, exacerbates economic challenges. Studies suggest that
non-communicable diseases, including Parkinson's, contribute
Parkinson's disease, a neurodegenerative disorder, silently to significant economic losses, slowing down GDP growth and
alters the lives of millions worldwide [1]. It creeps in slowly, pushing families into poverty. At its core, Parkinson's disease
targeting the brain's dopamine-producing cells, impairing motor primarily impacts the neurons located in a region of the midbrain
functions and mobility [2]. The gradual loss of dopamine, a vital known as the substantia nigra. This critical area is responsible
neurotransmitter, disrupts communication between synapses, for producing dopamine, and as the disease progresses, it leads
leading to a cascade of symptoms such as tremors, muscle to a shortage of this neurotransmitter, exacerbating the disorder's
rigidity, bradykinesia, and postural instability [3]. These devastating effects on movement and coordination [14]. As we
changes are not merely physical; Parkinson's disease also brings learn more about Parkinson's disease, its complexity and far-
with it a host of psychological symptoms including depression, reaching impacts become more evident. By embracing new
behavioral disorders, dementia, and mood changes [2]. The scientific approaches and understanding the intricate web of
disease manifests in many ways, from shaking and difficulty in symptoms and underlying causes, we may one day find a path to
movement to cognitive and behavioral disturbances. One early more effective treatments and perhaps even a cure.
and detectable sign can often be observed in changes to a
patient's voice. The voice, once strong and steady, may become  This study employs sci2 for feature selection and
tremulous, soft, or slurred as the disease progresses [1]. SMOTE for balancing the dataset, integrating both
Parkinson's disease is not just a battle for the elderly. It can machine learning (SVM, KNN, Logistic
present as early-onset or late-onset, moving through slow- Regression, Naive Bayes, Decision Tree,
progressing "benign" forms or fast-progressing "malignant" AdaBoost) and deep learning (LSTM, PLSTM,
types. Some patients may suffer from dementia, while others do MLP) models. We evaluated performance using
not. The disease can also be categorized into tremor-dominant accuracy, precision, recall, and F1-score.
subtypes or those characterized by postural instability and gait
disorders [4,5]. However, these traditional categories often fail  The data were analyzed using confusion matrices
to capture the full complexity of the disease. A more nuanced, and ROC curves, and our suggested KNN model
data-driven approach that accounts for the many interconnected was compared with current studies, showcasing its

979-8-3315-1909-4/24/$31.00 ©2024 IEEE

 higher efficacy in Parkinson's disease older adults. Meanwhile, Cordella et al. [22] used audio data to
identification. classify individuals with Parkinson's, relying on MATLAB
There are four components to this study: Previous research models. This particular study stands out by utilizing Python-
is summarized in the Literature Review section. The study's based, open-source models that are both faster and more
methodology and data are described in depth in the Materials memory-efficient. A consistent theme across recent literature is
and Methods section. The analysis of the results of the the emphasis on deep learning approaches. Ali et al. [23]
experiment is done in the Result Analysis section, and the employed ensemble deep learning models on phonation data to
conclusion includes important Conclusions and suggestions for track PD progression. While their results were impressive, the
more study. lack of feature selection meant that their deep learning model
(DNN) performance could have been further improved. To
II. LITERATURE REVIEW address this gap, this paper applied Principal Component
In recent years, the quest to predict Parkinson’s disease (PD) Analysis (PCA) to identify seven key voice modalities from a
and its severity has garnered immense attention. Various larger dataset of 22 attributes, enabling more effective PD
studies have explored how machine learning and deep learning detection.Huang et al. [24] sought to move away from wearable
models can help forecast PD, but much of the research has technology by training a decision tree model using speech
focused solely on detection. What’s been left behind, however, features from the MDVR-KCL [25] dataset, achieving notable
is the critical task of predicting the severity of Parkinsonism. success. On a different front, Wodzinski et al. [26] took a
One such endeavor led by Das and colleagues [15] dove into creative approach by training a ResNet model on audio data
the world of neural networks, machine learning regression images rather than on the frequency nuances of the audio itself.
algorithms, and decision trees to assess their abilities in Despite their unique methods, their results fell short, reaching
diagnosing PD. Their findings were intriguing, with neural an 85% accuracy ceiling. In another study, Wang et al. [27]
networks emerging as the most consistent and systematic applied 12 machine learning models to a vast voice biomarker
classifier, outperforming other models. The growing body of dataset, with their custom deep learning model (DEEP) pushing
work has largely concentrated on the potential of voice signal the boundaries with an impressive 96.45% classification
features to predict the severity of PD. For instance, Genain et accuracy. Other researchers have explored gait patterns and
al. [16] delved into this area by using Bagged Decision Trees to their connection to PD. For example, Alkhatib et al. [28]
estimate PD severity based on audio recordings. Their model implemented a linear classification model to characterize the
saw a small yet significant improvement in accuracy, gaining a shuffling gait of PD patients, achieving 95% accuracy. In
2% edge. Similarly, Malek et al. [17] zeroed in on a 40-feature comparison, Ghaderyan and Fathi et al. [29] developed a
dataset, identifying the nine most essential features with the technique based on signal separation and interlimb divergence,
help of LLBFS (Local Learning-Based Feature Selection). achieving an impressive accuracy range of 95.59% Yet, audio
Their goal was to divide PD subjects into four classes (Healthy, data seems to consistently outperform gait analysis in terms of
Early, Intermediate, and Advanced) based on their Unified predictive power. A standout study by Ashour, El-Attar, Dey,
Parkinson's Disease Rating Scale (UPDRS) scores. Other Abd El-Kader, and Abd El-Naby [30] (2020) delved into the
researchers like Seeja K.R. et al. [18] constructed a two-class world of LSTM, focusing on Freezing of Gait (FOG) signals
neural network, trained on acoustic data, to distinguish between from sensors worn by PD patients. Although their model
Parkinson’s patients and healthy individuals. While the model's achieved an 83% accuracy, the limited dataset size impacted the
classification accuracy reached 79%, it was considered a overall generalizability of the findings. Similarly, Paragliola
stepping stone toward more advanced methods. Drawing and Coronato [31] utilized a hybrid approach that combined
inspiration from these efforts, a groundbreaking study [19] CNNs and LSTMs for diagnosing PD based on gait patterns,
aimed to use keyboard typing tests as a predictor for PD. resulting in a 95% accuracy. However, their method was
Researchers found this method to be a promising detection tool. limited to binary classification, leaving the severity aspect
Combining the typing data with patients' voice impairments unexplored. Yang et al. (2022) [32] took a leap forward by
allowed them to predict not just the presence of Parkinson’s but developing PD-ResNet, a cutting-edge model rooted in residual
its severity as well. They referenced a study that achieved high network architecture. They focused on discerning gait
accuracy in detecting early PD based on finger movement differences not only between individuals with PD and healthy
characteristics during typing. While MRI scans, gait analysis, controls but also among PD patients at various severity levels.
and genetic data have been the traditional avenues for Using a combination of SMOTE, data augmentation, and early
predicting PD, the use of voice signals for early detection is stopping, their model reached remarkable accuracy rates. For
relatively novel. Bilal et al. [20] examined genetic data to instance, when classifying early PD from healthy controls, they
predict PD in elderly patients using an SVM model, achieving achieved a precision of 94.20%, recall of 90.28%, and an
an accuracy of 88.9%. In contrast, the current research overall accuracy of 92.03%. When analyzing the severity of
demonstrates that SVM, when applied to audio data, could PD, their model performed similarly well, boasting a precision
reach an accuracy of 91.83%, further validating the advantage of 94.29% and an F1 score of 92.31%. By examining previous
of audio-based classification. Raundale, Thosar, and Rane [21] studies and integrating new methodologies, the field continues
explored keystroke data from the UCI Telemonitoring dataset, to evolve, with each model contributing a piece to the puzzle of
training a Random Forest classifier to predict PD severity in
understanding and predicting Parkinson’s disease more B. Feature Selections
accurately and efficiently. In this study, feature selection was performed using the sci2
tool, resulting in the selection of 15 key features crucial for
III. MATERIAL AND METHODS diagnosing Parkinson's disease. These characteristics
The present study's approach, as seen in Fig.1, commences with encompass a variety of voice metrics, including non-linear
data preparation, which encompasses the use of imputation dynamics (spread1, PPE), shimmer (e.g., MDVP, Shimmer),
techniques to address missing data and the elimination of jitter (e.g., MDVP(%), MDVP), and frequency measurements
duplicate entries to guarantee data quality. In order to overcome (e.g., MDVP(Hz), MDVP(Hz)). Fig.2 illustrates the selected
class imbalance, SMOTE is used for data balancing. Then, to traits and emphasizes their significance in differentiating
keep the most pertinent features feature selection is performed between healthy persons and those diagnosed with Parkinson's
using only sci2. After scaling the data using normalization or disease. The inclusion of the most essential features for precise
standards, it is divided into two parts: 25% is used for testing model training is ensured by this selection procedure.
and 75% is used for training. Models for deep learning and
machine learning are both used. The machine learning models
include Support Vector Machine (SVM), K-Nearest Neighbors
(KNN), Logistic Regression, Naive Bayes, Decision Tree, and
AdaBoost, while the deep learning models include Long Short-
Term Memory (LSTM), Parallel LSTM (PLSTM), and
Multilayer Perceptron (MLP). Based on measures like F1 score,
accuracy, precision, recall, and area under the ROC curve
(AUC), these models are assessed. Confusion matrices and
ROC curves are used to better depict the evaluation. The best-
performing model is chosen for additional examination and
optimization once all models' performances have been
compared. With close consideration to data quality, model
selection, and performance evaluation, this all-inclusive
method guarantees the choice of the best model for the assigned
job.
Fig. 2. Selected 15 Features Visualization

C. Preprocessing and Data Balancing

To ensure data integrity, we first performed a missing data and
duplicate data check during the preprocessing stage. We used
the SMOTE technique to remedy class imbalance, and the
balanced dataset that emerged is shown in Fig.3. In order to
standardize the characteristics and improve model
performance, we also scaled the data.

Fig. 1. Working Procedure of the Proposed Parkinson's Disease Detection

System.

A. Dataset
The dataset [33] used in this research comprises 195 phonation
samples aimed at diagnosing Parkinson's disease through vocal
features. It consists of 24 attributes: 23 features that characterize
different aspects of voice, and a binary target attribute (status) Fig. 3. Data Balancing Using SMOTE
that indicates if Parkinson's disease is present (1) or absent (0).
These characteristics include non-linear dynamics like RPDE,
DFA, spread1, spread2, D2, and PPE, as well as frequency Fig.4 shows the correlation of Selected features, shedding light
measurements like MDVP, jitter, shimmer, and noise ratios on feature relationships. In order to ensure a thorough
(NHR, HNR). The dataset is well-rounded and comprehensive, assessment of the model, the dataset was finally divided into
which makes it perfect for creating machine learning models that training and testing sets, with 75% of the data designated for
identify Parkinson's disease. training and 25% for testing.
 judgment, respectively. For example, TP can be used to
describe the total number of accurately recognized healthy data
points. Equation (3) uses calculations to show the correctness
overall. Recall evaluates a model's ability to identify each
successful instance. Recall is defined mathematically in
Equation (4). Equation (5) has been used to calculate the ratio
of consistently expected positive results among all positive
forecasts. Formula (6) is used to determine the F1-Score
measure, which assesses the harmonic mean across recall and
precision.
𝑻𝑷 + 𝑻𝑵
Fig. 4. Correlation Matrix of Selected Features 𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = (𝟑)
𝑻𝑷 + 𝑻𝑵 + 𝑭𝑷 + 𝑭𝑵
𝑻𝑷
𝑹𝒆𝒄𝒂𝒍𝒍 = (𝟒)
𝑻𝑷 + 𝑭𝑵
D. Machine Learning Models 𝑻𝑷
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 = (𝟓)
𝑻𝑷 + 𝑭𝑷
𝟐 ∗ 𝑻𝑷
In this study, six machine learning models—SVM, KNN, 𝑭𝟏 𝑺𝒄𝒐𝒓𝒆 = (𝟔)
𝟐 ∗ 𝑻𝑷 + 𝑭𝑷 + 𝑭𝑵
Logistic Regression, Naive Bayes, Decision Tree, and
AdaBoost—were employed, with KNN achieving the highest
accuracy of 97.20%. LSTM, PLSTM, and MLP models were IV. RESULT ANALYSIS
used in the deep learning section; MLP performed the best,
obtaining an accuracy of 94.59%.
The performance of machine learning models and deep
KNN: learning for Parkinson's disease prediction was assessed in our
The k-Nearest Neighbors (KNN) approach is one of the most study, and the findings are compiled in Table I. K-Nearest
fundamental and straightforward models in supervised machine Neighbors (KNN) method has the greatest accuracy of 97.20%
learning. Given that KNN is a non-parametric technique and among machine learning models, making it stand out. Overall
does not assume anything about the distribution of the reliability and efficacy were demonstrated by its excellent
underlying data, it can be applied freely in a wide range of performance in precision (97.44%), recall (97.20%), and F1-
scenarios. The ability of KNN to make predictions based on a score (97.20%). KNN performs exceptionally well, with very
comparison of the similarity between each dataset data point is few misclassifications, as evidenced by the confusion matrix
its basic concept. This technique is utilized for both regression (Fig. 5).
and classification applications because of its adaptability.
TABLE I. SUMMARY OF PERFORMANCE METRICS FOR MACHINE
Additionally, KNN has demonstrated potential in applications LEARNING AND DEEP LEARNING CLASSIFICATION MODELS
including text classification, image recognition, and clustering.
The performance of the algorithm is usually determined by the Machine Learning Models
selection of distance measurements, of which there are three Model Accuracy Precision Recall F1-Score
primary categories that are commonly utilized [34].With the use KNN 97.20 97.44 97.20 97.20
of these equations, we may degree separate within the KNN
computation. These two types of criteria are for two- SVM 89.19 89.32 89.19 89.19
dimensional data and multidimensional data, respectively. The Logistic 86.49 86.57 86.49 86.49
number of measurements is 'n'. '𝑋𝑖 ' might contain information regression
from the prepared dataset. The highlight for information point Naïve Bayes 83.78 86.53 83.78 83.54
'x' is denoted by ' 𝑋𝑗 '. The value of the ‘j’ highlight for Decision 94.59 94.72 94.59 94.55
information point '𝑋𝑖 ' is represented by '𝑋𝑖𝑗 '. Equations (1) and Tree
(2) deal with the two- and three-dimensional data. AdaBoost 89.19 89.32 89.13 89.19
Deep Learning Models
√(𝑋1 − 𝑋𝑖2 )2 + (𝑋2 − 𝑋𝑖2 )2 (1)
Model Accuracy Precision Recall F1-Score

√∑𝑛𝑗=1(𝑋𝑗 − 𝑋𝑖𝑗 )2 (2) LSTM 83.78 82.50 86.84 84.61

PLSTM 85.14 82.93 89.47 86.08
E. Performance Metrics MLP 94.59 97.22 92.11 94.59
The effectiveness of the sorting techniques used for the
Parkinson diseases categorization was assessed using a variety
of measures, including sensitivity, F-score, accuracy, and With constant accuracy, recall, and F1-scores within the
precision. The terms false, positive, negative, and T, F, P, and same range, and accuracies of 89.19% and 86.49%, respectively,
N stand for unfavorable facts and the confusion matrix's other models, like Support Vector Machine (SVM) and Logistic
Regression, demonstrated modest performance. Six machine ResNet (85% accuracy) and Pd-ResNet (92.03% accuracy)
learning classifiers are shown with their ROC curves in Fig.6. models developed by Wodzinski et al. and Yang et al. The
While KNN has the maximum area under the curve (AUC), accuracy of 95.20% for our KNN model is higher than even that
other models, such as Decision Tree and AdaBoost, also of Paragliola et al.'s hybrid CNN and LSTM model, which is
function well, albeit not as well as KNN. The Multilayer only 95% accurate. These results show that our KNN model
Perceptron (MLP) model was the best performance in the deep performs better than other models in Parkinson's disease
learning category, achieving the maximum accuracy of 94.59%, prediction, indicating its efficacy across many assessment
precision of 97.22%, recall of 92.11%, and F1-score of 94.59%. metrics.
Fig.7 and 8, which show the MLP's confusion matrix and ROC
curve, respectively, provide evidence of its efficacy.

MLP
Fig. 8. ROC Curve of MLP Algorithm
Fig. 5. Confusion Matrix of KNN Classifier
TABLE II. COMPARING VARIOUS MACHINE LEARNING AND DEEP
LEARNING ALGORITHMS FOR PREDICTING PARKINSON'S DISEASE.
Author Name Methods Accuracy
Seeja K.R. et al. [18] Two-class Neural Network 79%
Bilal et al. [20] SVM on Genetic Data 88.9%
Bilal et al. [20] SVM on Audio Data 91.83%
Wodzinski et al. [26] ResNet on Audio Data Images 85%
Wang et al. [27] 12 Machine Learning Models 96.45%
Alkhatib et al. [28] Linear Classification on Gait 95%
Patterns
Ghaderyan et al. [29] Signal Separation, Interlimb 95.59%
Divergence
Ashour et al. [30] LSTM on Freezing of Gait 83%
(FOG) Signals
Paragliola et al. [31] Hybrid CNN and LSTM, Gait 95%
Patterns
Yang et al. [32] Pd-ResNet (Residual Network 92.03%
Fig. 6. Performance Evaluation: ROC Analysis of Multiple Machine Architecture)
Learning Classifiers Proposed Model KNN 97.20%

V. CONCLUSION AND FUTURE WORK

Our study's machine learning method for predicting the
identification of Parkinson's disease using the KNN algorithm
performed very well, with 97.20% accuracy, 97.44% precision,
97.20% recall, and 97.32% F1-score. This high degree of
accuracy demonstrates how well KNN classifies voice
characteristics suggestive of Parkinson's illness. Our findings
demonstrate KNN's promise as a reliable tool for early illness
detection and surveillance. To improve and expand on this
knowledge, future work will concentrate on a number of
important areas. To further improve the model, we intend to add
Fig. 7. Confusion Matrix of MLP Algorithm
more features and investigate cutting-edge feature selection
strategies like SHAPE. Enhancing the dataset by include a wider
As we can see from Table II, our proposed result with 97.20% range of samples would enhance the generalizability. Further
accuracy are superior to those of earlier research, including the insights may also be obtained by looking into hybrid models that
incorporate KNN with other algorithms and by using
sophisticated deep learning methods. To determine the 18. Seeja K.R., Srishti Grover, Saloni Bhartia, Akshama, and Abhilasha
suggested model's usefulness and practical application, it must Yadav. “Predicting Severity Of Parkinson’s Disease Using Deep
Learning”. Procedia Computer Science, 132:1788–1794.
be put to the test and validated in clinical settings.
19. Warwick R. Adams. “High-accuracy detection of early Parkinson’s
Disease using multiple characteristics of finger movement while
typing”. 2017.
REFERENCES 20. Alatas Bilal, Moradi Shadi, Tapak Leili, Afshar Saeid (2022),
“Identification of Novel Noninvasive Diagnostics Biomarkers in the
Parkinson’s Diseases and Improving the Disease Classification
1. P. Krishna, D. StalinDavid. "An Effective Parkinson's Disease Using Support Vector Machine”, BioMed Research International,
Prediction Using Logistic Decision Regression and Machine Hindawi
Learning with Big Data," in Turkish Journal of Physiotherapy and 21. P. Raundale, C. Thosar and S. Rane (2021), "Prediction of
Rehabilitation, vol. 32, no. 3, pp. 778–786, 2021. Parkinson’s disease and severity of the disease using Machine Learn
2. P. Raundale, C. Thosar, S. Rane, "Prediction of Parkinson’s disease ing and Deep Learning algorithm," 2021 2nd International
and severity of the disease using Machine Learning and Deep Conference for Emerging Technology (INCET), pp. 1-5, doi:
Learning algorithm," in 2021 2nd International Conference for 10.1109/INCET51464.2021.9456292.
Emerging Technology (INCET), 2021, pp. 1–5. 22. F. Cordella, A. Paffi and A. Pallotti (2021) "Classification-based
3. Prabhavathi, K., Patil, S. (2022). “Tremors and Bradykinesia. In: screening of Parkinson’s disease patients through voice signal,"
Arjunan, S.P., Kumar, D.K. (eds) Techniques for Assessment of 2021 IEEE International Symposium on Medical Measurements and
Parkinsonism for Diagnosis and Rehabilitation”. Series in Applications (MeMeA), pp. 1-6, doi:
BioEngineering. Springer. 135–149 https://doi.org/10.1007/978- 10.1109/MeMeA52024.2021.9478683.
981-16-3056-9_9 23. Ali, L., Chakraborty, C., He, Z. et al. (2022) “A novel sample and
4. Jankovic, J. et al. Variable expression of Parkinson’s disease: a feature dependent ensemble approach for Parkinson’s disease
base‐line analysis of the DAT ATOP cohort. Neurology 40, 1529– detection”. Neural Comput & Applic.
1529 (1990). https://doi.org/10.1007/s00521-022-07046-2
5. Zetusky, W. J., Jankovic, J. & Pirozzolo, F. J. The heterogeneity of 24. F. Huang, H. Xu, T. Shen and L. Jin (2021), "Recognition of
Parkinson’s disease: clinical and prognostic implications. Parkinson's Disease Based on Residual Neural Network and Voice
Neurology 35, 522–526 (1985). Diagnosis," 2021 IEEE 5th Information Technology, Networking,
6. Faghri, F. et al. Identifying and predicting amyotrophic lateral Electronic and Automation Control Conference (ITNEC), pp. 381-
sclerosis clinical subgroups: a population-based machine-learning 386, doi: 10.1109/ITNEC52019.2021.9586915.
study. Lancet Digit Health 4, e359–e369 (2022). 25. D. Trivedi H. Jaeger and M. Stadtschnitzer. (2019) “Mobile Device
7. 7. Satone, V. K. et al. Predicting Alzheimer’s disease progression Voice Recordings at King’s College London (MDVR -KCL) from
trajectory andclinical subtypes using machine learning. Preprint at both early and advanced Parkinson’s disease patients and healthy
https://www.biorxiv.org/ content/10.1101/792432v2 (2020). controls.” https://doi.org/10.5281/zenodo.2867216
8. O. Eskidere, F. Ertas ̧, and C. Hanilçi, “A comparison of r̈ egression 26. M. Wodzinski, A. Skalski, D. Hemmerling, J. R. Orozco-Arroyave
methods for remote tracking of Parkinson’s disease progression,” and E. Nöth, (2019) "Deep Learning Approach to Parkinson’s
Expert Systems with Applications, vol. 39, pp. 5523–5528, 2012. Disease Detection Using Voice Recordings and Convolutional
9. C. G. Goetz, W. Poewe, and O. Rascol, “Movement disorder society Neural Network Dedicated to Image Classification," 2019 41st
task force on rating scales for Parkinson’s disease: the unified Annual International Conference of the IEEE Engineering in
Parkinson’s disease rating scale (UPDRS): status and Medicine and Biology Society (EMBC), pp. 717-720, doi:
recommendations,” Movement Disorders, vol. 18, pp. 738–750, 10.1109/EMBC.2019.8856972.
2003. 27. W. Wang, J. Lee, F. Harrou and Y. Sun, "Early Detection of
10. G. Moguel-Cobos, C. Saldivar, P. W. Goslar, and H. A. Shill, “)e Parkinson’s Disease Using Deep Learning and Machin e Learning,"
relationship between social anxiety disorder and motor symptoms in IEEE Access, vol. 8, pp. 147635-147646, 2020, doi:
of Parkinson disease: a pilot study,” Psychoso- matics, vol. 61, pp. 10.1109/ACCESS.2020.3016062.
321–326, 2020. 28. R. Alkhatib, M. O. Diab, C. Corbier and M. E. Badaoui, "Machine
11. M. Nilashi, O. Ibrahim, S. Samad, H. Ahmadi, L. Shahmoradi, and Learning Algorithm for Gait Analysis and Classification on Early
E. Akbari, “An analytical method for measuring the Parkinson’s Detection of Parkinson," in IEEE Sensors Letters, vol. 4, no. 6, pp.
disease progression: a case on a Parkinson’s telemonitoring 1-4, June 2020, Art no. 6000604, doi:
dataset,” Measurement, vol. 136, pp. 545–557, 2019b. 10.1109/LSENS.2020.2994938.
12. Green, E. D. et al. Strategic vision for improving human health at 29. P. Ghaderyan and G. Fathi, “Inter-limb time-varying singular value:
The Forefront of Genomics. Nature 586, 683–692 (2020). a new gait feature for Parkinson’s disease detection and stage
classification,” Measurement, vol. 177, Article ID 109249, 2021.
13. Michael, J.: Parkinson Disease (2017).
https://www.nia.nih.gov/health/ 30. A. S. Ashour, A. El-Attar, N. Dey, H. Abd El-Kader, and M. M.
parkinsonsdisease#:∼:text=Parkinson%27s%20disease%20is%20a Abd El-Naby, “Long short term memory based patient- dependent
%20brain. Accessed 20 May 2021. model for FOG detection in Parkinson’s disease,” Pattern
Recognition Letters, vol. 131, pp. 23–29, 2020.
14. M. Fjodorova, E.M. Torres and S.B. Dunnett, “Transplantation site
influences the phenotypic differentiation of dopamine neurons in 31. G. Paragliola and A. Coronato, “Gait anomaly detection of subjects
ventral mesencephalic grafts in Parkinsonian rats," Experimental with Parkinson’s disease using a deep time series- based approach,”
Neurology, vol. 291, pp. 8-19, 2017. IEEE Access, vol. 6, pp. 73280–73292, 2018.
15. Das R. “Comparison of multiple classification methods for 32. Yang X, Ye Q, Cai G, Wang Y and Cai G, (2022). PD-ResNet for
diagnosis of Parkinson disease”. Expert Systems With Applications, Classification of Parkinson’s Disease from Gait, in IEEE Journal of
37:1568–1572, 2010. Translational Engineering in Health and Medicine, 10, 1-11, 2022,
Art no. 2200111, doi: 10.1109/JTEHM.2022.3180933.
16. Genain N, Huberth M, and Vidyashankar R. “Predicting
Parkinson’s Disease Severity from Patient Voice Features.” 2014. 33. M. Little, "Parkinsons," , 2008
17. Benmalek E, Elmhamdi J, and Jilbab A. “UPDRS tracking using 34. Tyagi A, Singh VP, Gore MM. An efficient automated detection of
linear regression and neural network for Parkinson’s disease schizophrenia using k-NN and bag of words features. SN Computer
prediction.” International Journal Of Emerging Trends and Science 2023;4(5):518.
Technology In Computer Science, 4:189–193, 2015.