Abstract
Parkinson’s disease (PD) is a progressive neurodegenerative disorder for which timely diagnosis remains challenging. Voice-based machine learning (ML) diagnostic tools have shown promise for PD detection, but many studies do not jointly address class imbalance, subject-level data leakage, and sex-specific acoustic biomarkers. To address these limitations, we benchmarked state-of-the-art ML approaches for voice-based PD telediagnosis and examined whether sex-specific modeling improved detection and biomarker identification. Using voice recordings from 252 participants (188 PD, 64 controls), we evaluated four dataset configurations: sex-agnostic, sex-aware, male-only, and female-only. Adaptive Synthetic Sampling (ADASYN) addressed class imbalance within the training folds, and subject-level nested cross-validation was used to prevent data leakage. Two feature selection methods, Boruta and minimum Redundancy Maximum Relevance (mRMR), were compared across multiple feature counts, and six classifiers were evaluated (Random Forest, Neural Network, Naive Bayes, Logistic Regression, AdaBoost, SVM-RBF). The best configurations included a sex-agnostic neural network (Boruta-100 features; Matthews Correlation Coefficient (MCC) = 0.587), a sex-aware Random Forest (Boruta-80; MCC = 0.579), a male-only AdaBoost model (mRMR-20; MCC = 0.466), and a female-only Random Forest (mRMR-10; MCC = 0.674). Under subject-level validation, accuracies ranged from 0.827 to 0.868, providing realistic generalization estimates.SHapley Additive exPlanations (SHAP) supported interpretability and revealed sex-dependent vocal signatures: female classification was driven primarily by Mel-Frequency Cepstral Coefficient (MFCC)-derived features; male classification relied more on Tunable Q-factor Wavelet Transform (TQWT)-derived and baseline features, with minimal overlap among top predictors. This robust ML framework supports more reliable voice-based PD detection and highlights the importance of sex-specific modeling and addressing class imbalances for effective telediagnosis and early screening.