Abstract
Introduction: Parkinson's Disease (PD) is a progressive neurodegenerative disorder that significantly affects the aging population, creating a growing burden on global health systems. Early detection of PD is clinically challenging due to the gradual and ambiguous onset of symptoms. Methods: This study presents a machine-learning framework for the early identification of PD using non-invasive biomedical voice biomarkers from the UCI Parkinson's dataset. The dataset consists of 195 sustained phonation recordings from 31 participants (23 PD and 8 healthy controls, ages 46-85). The methodology includes subject-level stratified splitting and normalization, along with BorderlineSMOTE to address class imbalance. Initially, an XGBoost model is applied to select the top 10 acoustic features, followed by a Bayesian-optimized XGBoost classifier, with the decision threshold tuned via F1-maximization on validation data. Results: On the held-out test set, the model achieves 98.0% accuracy, 0.97 macro-F1, and 0.991 ROC-AUC. This performance exceeds that of a deep neural network baseline by 4.0 percentage points in accuracy (94.0% to 98.0%), 4.3 percentage points in macro-F1 (92.7% to 97.0%), and 0.050 in AUC (0.941 to 0.991). Compared to a classical SVM, it outperforms by 7.0 percentage points in accuracy (91.0% to 98.0%), 6.5 percentage points in macro-F1 (90.5% to 97.0%), and 0.089 in AUC (0.902 to 0.991). Discussion: Model decisions are elucidated using SHAP, offering global and patient-specific insights into the influential voice features. These findings indicate the feasibility of a non-invasive, scalable, and explainable voice-based tool for early PD screening, highlighting its potential integration into mobile or telehealth diagnostic platforms.