Optimizing Speech-Based Parkinson’s Diagnosis with Hybrid Feature Selection and Machine Learning

Parkinson’s disease (PD) is a common neurological disorder that causes cognitive and motor impairments, including speech and gait abnormalities. Medication is often used to treat symptoms, but a cure is still elusive. Effective disease management depends on early detection. This study employs ML approaches, including K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), and Naive Bayes (NB), to distinguish between individuals with PD and healthy controls based on voice signal characteristics. To enhance classification performance, multiple feature selection techniques were applied, including filter-based methods (correlation analysis, Analysis of Variance [ANOVA]), wrapper-based approaches such as Recursive Feature Elimination (RFE), embedded models including Random Forest (RF) and XGBoost importance, and a Genetic Algorithm (GA). Furthermore, a novel hybrid method was proposed by combining GA and XGBoost feature importance to identify the most relevant features. The models were trained and tested using standard preprocessing techniques such as feature scaling and RandomOverSampler to address class imbalance. Experimental results demonstrate that the hybrid feature selection method significantly improved classification accuracy. Using rigorous 10-Fold Stratified Cross-Validation, the proposed method achieved a robust mean accuracy of 92.74% (XGBoost) and a geometric mean of 91.13% (KNN). These findings suggest that integrating evolutionary and model-driven feature selection can significantly enhance the diagnosis of PD, providing a promising approach for voice-based medical decision support systems.