Automated Parkinson's Disease Detection from Sustained Phonation Using Acoustic Speech Features and Machine Learning

Abstract

Parkinson's disease (PD) is a progressive brain disorder that affects over ten million people worldwide. Dysphonia, which means a decline in vocal quality, is an important early symptom. This paper introduces a non-invasive machine-learning-based framework for detecting PD through speech analysis. It comprises 756 recordings from 252 subjects, including 564 recordings from subjects with PD and 192 recordings from healthy subjects. An initial baseline set of 752 acoustic features, mined from the intended speech recordings, is used. These attributes cover a range of speech-based measures, including jitter, shimmer, Mel-Frequency Cepstral Coefficients (MFCCs), Discrete Wavelet Transform (DWT) coefficients, and Tunable Q-Factor Wavelet Transform (TQWT) coefficients. The most important variables were identified using Mutual Information (MI) and Recursive Feature Elimination (RFE). To tackle class imbalance without risking data leakage, the Synthetic Minority Over-sampling Technique (SMOTE) was applied only within each cross-validation fold. Cost-sensitive weighting was also incorporated. We tested seven machine learning classifiers across three feature sets: Full, MI, and RFE. It was done using stratified five-fold cross-validation. The Support Vector Machine with a Radial Basis Function kernel (SVM-RBF), combined with RFE based 50 feature selection, significantly outperformed the other models. The optimized SVM-RBF setup achieved an accuracy of 94.84%, precision of 96.31%, recall of 96.80%, F1-score of 96.54%, ROC-AUC of 98.36%, and Cohen's Kappa of 86.37%. Importantly, this high-accuracy model is very lightweight, requiring only 0.24 seconds to train. These results show that the devised framework provides a highly accurate, non-invasive, and efficient option for early screening of Parkinson's disease.