An Interpretable and Auditable Pipeline for Robust Multi-Cohort Detection of Parkinson’s Disease from Voice Data with Data Leakage Prevention

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Abstract Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by early vocal changes thatserve as crucial prodromal biomarkers. This work presents an interpretable and auditable machine learningpipeline for robust PD detection through voice analysis, integrating three heterogeneous public datasets into aunified cohort of 996 recordings from 332 independent subjects. To strictly mitigate subject-level data leakage—astructural flaw endemic to the PD voice literature—the pipeline applies a Leave-One-Subject-Out (LOSO)validation protocol and confines SMOTE-based class balancing exclusively to training folds. A soft-votingensemble of Logistic Regression, Support Vector Machine, and Random Forest achieves a subject-level Area Underthe ROC Curve (AUC) of 0.8836, with a sensitivity of 88.60% and specificity of 70.19%. A label permutationtest, in which the AUC collapses to 0.556, mathematically confirms the absence of trivial target leakage. Featureimportance analysis reveals that Mel-Frequency Cepstral Coefficients (MFCCs) and the Harmonics-to-NoiseRatio (HNR) are the dominant predictors, consistent with the known pathophysiology of hypokinetic dysarthria.This study establishes a reproducible, realistic benchmark and a rigorously audited methodological foundationfor voice-based AI triage systems in telemedicine.

Related topics