Explainable Machine Learning for Early Detection of Parkinson's Disease in Aging Populations Using Vocal Biomarkers

Bright Egbo¹Zhanbota Nigmetolla¹Naveed Ahmad Khan²Prashant Kumar Jamwal^1*

• ¹Nazarbayev University, Astana, Kazakhstan
• ²University of Canberra, Canberra, Australia

Parkinson's Disease (PD) is a progressive neurodegenerative disorder that disproportionately affects the aging population and imposes a growing burden on global health systems. Early-stage detection remains clinically challenging due to the gradual and often ambiguous onset of symptoms. This study introduces an accurate and interpretable machine-learning framework for early PD identification using non-invasive biomedical voice biomarkers from the UCI Parkinson's dataset (195 sustained phonation recordings from 31 participants: 23 PD and 8 healthy controls, ages 46–85). The pipeline employs subject-level stratified splitting and normalization, applies BorderlineSMOTE to mitigate class imbalance in diagnostically critical regions, and uses an initial XGBoost to select the top 10 acoustic features for a Bayesian-optimized XGBoost classifier; the decision threshold is tuned by F1-maximization on validation data. On the held-out test set, the model attains 98.0% accuracy, 0.97 macro-F1, and 0.991 ROC-AUC, outperforming a strong neural baseline (DNN) by +4.0 percentage points (pp) in accuracy (94.0→98.0), +4.3 pp in macro-F1 (92.7→97.0), and +0.050 in AUC (0.941→0.991), and a classical SVM by +7.0 pp accuracy (91.0→98.0), +6.5 pp macro-F1 (90.5→97.0), and +0.089 AUC (0.902→0.991). Model decisions are explained with SHAP, providing global and patient-specific insights into influential voice features. These results demonstrate the feasibility of a scalable, non-invasive, and explainable voice-based tool for early PD screening, with strong potential for integration into mobile or telehealth diagnostic platforms.