Voice-based Parkinson's disease detection: a clinically validated lightweight framework via multi-task knowledge distillation

Dysarthria in Parkinson’s disease (PD) is a crucial early biomarker. While advanced deep learning models achieve high diagnostic accuracy, their large size and ‘black-box’ nature hinder clinical adoption on portable devices. This study proposes a novel interpretable lightweight framework for PD diagnosis via multi-task knowledge distillation. We first train a high capacity ‘teacher’ model (a deep FCNN) and transfer its knowledge to compact ‘student’ models using a composite loss function that distills output logits, feature representations, and attention. Experiments on a public dataset of 674 voice samples with 41 acoustic features rigorously compare our distilled models against nine traditional ML models and modern deep architectures. Results show: 1) The teacher model attains 90.6% accuracy, competitive with top traditional classifiers (e.g. KNN: 91.8%) and a state of the art deep model (93.0%). 2) The 3 layer student model achieves 13.2x parameter compression while preserving 97% of teacher accuracy (88.7% vs. 90.6%). 3) The 2 layer student reaches 70x compression, ideal for edge deployment. 4) Noise robustness tests confirm graceful performance degradation. 5) SHAP interpretability analysis shows the student retains the teacher’s feature importance ranking (Kendall’s τ = 0.82), validating clinical reasoning transfer. This work provides a practical framework for building efficient, accurate, and trustworthy AI tools for early PD screening.