Enhancing Parkinson's Disease Detection by Combining SMOTE and Feature Selection for Improved Machine Learning Classification Using Voice Recordings.

Parkinson's disease (PD) presents significant diagnostic challenges, particularly in its early stages when timely intervention is most effective. Voice analysis offers a noninvasive approach for early detection but faces challenges of class imbalance and high-dimensional data that limit machine learning (ML) classification performance. This study investigates the combined effect of the synthetic minority over-sampling technique (SMOTE) for data balancing and feature selection (FS) methods, analysis of variance, Chi-squared (χ2), and mutual information to optimize PD detection from voice recordings. These preprocessing techniques were integrated with multiple classifiers, including extreme gradient boosting (XGBoost), random forest, logistic regression, and support vector machine, using the Parkinson's Disease Classification Dataset containing 754 features extracted from voice recordings. Results demonstrate that combining SMOTE with FS significantly improves model performance compared to using either technique independently. The highest performance was achieved by XGBoost with Chi-squared (χ2) FS (96.4% accuracy, 96.9% F1-score) using approximately 600 optimal features. These findings confirm that appropriate preprocessing techniques improve the effectiveness of voice-based PD detection and support the development of accurate, noninvasive diagnostic tools powered by ML.