Voice-Based Parkinson's Disease Diagnosis: A Data-Augmented Study of Deterministic and Stochastic Feature Selection Methods.

INTRODUCTION

Voice analysis is a promising method for diagnosing Parkinson's disease (PD); however, two major obstacles stand in the way of its clinical translation: a lack of clinical data that leads to overfitting and a lack of investigation into direct comparisons between feature selection techniques.

OBJECTIVE

Comparative evaluation of the performance of the original and augmented datasets using deterministic recursive feature elimination (RFE) and stochastic genetic algorithm (GA) feature selection approaches for the diagnosis of PD using voice analysis.

METHODOLOGY: The performance of the original and augmented datasets, which contained signal-level transformations, was evaluated using eight machine learning classifiers. These classifiers were used to quantify the performance of the selected feature subsets based on the results of 10-fold stratified group k-fold cross-validation.

RESULTS

Models trained on the original dataset showed low performance with a maximum F1-score of 0.699. Data augmentation significantly improved performance for all pipelines. The k-Nearest Neighbors classifier with RFE resulted in 98.64% accuracy and F1-score: 0.987. GA-CatBoost resulted in 98.30% accuracy and an F1-score of 0.983.

CONCLUSION

Data augmentation addressed data scarcity and improved model performance. The cepstral and spectral features showed good discriminative ability, and performance trends showed method-specific interaction effects of feature selection and classification.