Voice-Based Parkinson Disease Detection Using SVM & XGBoost

Parkinson's disease (PD) diagnosis remains largely subjective, relying on clinical assessment despite the availability of objective biomarkers. Early, accurate detection is crucial for effective management, motivating non-invasive machine-learning approaches. In this study, we evaluate two classical classifiers-Support Vector Machine (SVM) and XGBoost on the UCI Parkinson's voice dataset (195 recordings, 22 acoustic features). We apply SMOTE oversampling (ratio=0.8) on the training split, remove highly correlated features (Pearson $r > 0.9$), and optimize hyperparameters via 5-fold stratified RandomizedSearchCV. To ensure transparency, we adopt a strict 60 / 20 / 20 train/validation/test split on the held-out test set, both SVM and XGBoost achieve 92.31% accuracy, with AUC-ROC above 0.94, specificity of 0.90, and inference latency under 50 ms on a midrange smartphone. For comparison, 5-fold cross-validation on the 80% train+validation pool yields 99.37% accuracy, illustrating the gap between optimistic and conservative estimates. Permutation and gain-based importance analyses identify key voice markers (e.g., HNR, ShimmerAPQ3). Our contributions include a reproducible feature-selection pipeline, dual evaluation protocol, and clinicallatency assessment. Future work will extend validation to larger, multi-modal cohorts and explore real-world telehealth deployment.