Fine-Tuned Machine Learning Classifiers for Diagnosing Parkinson's Disease Using Vocal Characteristics: A Comparative...

Area of Science:

• Neurology
• Computational Linguistics
• Biomedical Engineering

Background:

• Parkinson's Disease (PD) diagnosis requires early and precise detection for effective management.
• Vocal characteristics offer a non-invasive and accessible method for PD assessment.
• Current diagnostic tools can be invasive or lack accessibility, highlighting the need for alternative methods.

Purpose of the Study:

• To evaluate machine learning algorithms for classifying Parkinson's Disease based on vocal features.
• To develop an optimized, non-invasive diagnostic tool for PD using acoustic analysis.
• To improve disease control and patient outcomes through early and accurate PD detection.

Main Methods:

• Utilized a public dataset of vocal samples from 188 PD patients and 64 controls.
• Extracted acoustic features including MFCCs and wavelet transform metrics.
• Employed Chi-Square for feature selection and optimized six ML classifiers (SVM, k-NN, DT, NN, Ensemble, Stacking) using Bayesian Optimization, Grid Search, and Random Search.

Main Results:

• Stacking models, optimized via Grid Search, achieved the highest performance with 92.07% accuracy and a 0.95 F1-score.
• Chi-Square feature selection significantly enhanced classification accuracy and computational efficiency.
• Ensemble models demonstrated robust diagnostic performance on complex datasets.

Conclusions:

• Combining advanced feature selection and hyperparameter optimization enhances ML-based PD diagnosis from vocal data.
• Ensemble models show significant potential for accurate and scalable clinical PD diagnosis.
• Future research should explore deep learning and temporal features for further diagnostic improvements.