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

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Medical Informatics

Background:

• Voice analysis shows potential for diagnosing Parkinson's disease (PD).
• Clinical translation is hindered by limited data and insufficient comparison of feature selection techniques.
• Overfitting due to data scarcity is a significant challenge in PD diagnosis models.

Purpose of the Study:

• To comparatively evaluate original and augmented datasets for PD diagnosis using voice analysis.
• To assess the performance of deterministic recursive feature elimination (RFE) and stochastic genetic algorithm (GA) feature selection.
• To investigate the impact of data augmentation on machine learning model performance for PD detection.

Main Methods:

• Evaluated eight machine learning classifiers on original and augmented datasets with signal-level transformations.
• Employed 10-fold stratified group k-fold cross-validation to quantify performance of selected feature subsets.
• Compared RFE and GA feature selection approaches for PD diagnosis via voice analysis.

Main Results:

• Models trained on the original dataset exhibited low performance (max F1-score 0.699).
• Data augmentation substantially enhanced performance across all evaluated pipelines.
• k-Nearest Neighbors with RFE achieved 98.64% accuracy and 0.987 F1-score; GA-CatBoost achieved 98.30% accuracy and 0.983 F1-score.

Conclusions:

• Data augmentation effectively addressed data scarcity and boosted model performance for PD diagnosis.
• Cepstral and spectral voice features demonstrated strong discriminative capabilities.
• Feature selection methods and classification techniques interact, influencing diagnostic performance.