Robust and explainable machine learning for leakage aware subject level voice based Parkinson disease detection

Abstract

Parkinson's disease (PD) is a progressive neurodegenerative disorder in which early diagnosis is essential for effective treatment and symptom management. Voice impairment is among the earliest indicators of PD, making voice-based machine learning (ML) approaches a promising non-invasive diagnostic alternative. However, many existing studies suffer from subject-level data leakage, single-split evaluation artifacts, and limited interpretability of the models they actually deploy. This study proposes a comprehensive, leakage-aware ML framework for PD detection using 22 biomedical voice features, evaluating five models (Decision Tree, Random Forest, XGBoost, Logistic Regression, and Support Vector Machine) under both random sample-level and subject-level data splitting, Stratified K-Fold cross-validation, and RandomizedSearchCV tuning. Under a single fixed subject-level split, Random Forest's testing accuracy rose from 79 to 95%; a class-composition diagnostic showed this split's test set contained no healthy individuals, rendering it unsuitable for cross-model comparison and motivating three split-independent validation protocols instead: subject-aware group cross-validation, ten repeated subject-level splits, and Leave-One-Subject-Out (LOSO) validation. Under these protocols, SVM and Logistic Regression matched or exceeded the ensemble models on specificity and the Matthews Correlation Coefficient (MCC), supporting the selection of SVM as a stable, conservative deployment candidate rather than as the single highest-accuracy model. XGBoost achieved the highest ROC-AUC (0.98), followed by Random Forest (0.96). SHAP and permutation importance, applied to both the XGBoost model and the deployed SVM model, identified jitter- and nonlinearity-related features as consistent with known patterns of Parkinsonian speech impairment. An SVM-based predictive system was developed as a proof-of-concept for telemedicine-oriented screening; external validation on an independent dataset remains the principal precondition for any clinical-use claim.