An explainable method based on sinc filters to diagnose Parkinson’s disease severity by gait cycle analysis

This paper presents SincPD, an explainable deep learning model for diagnosing Parkinson’s disease (PD) and assessing its severity using vertical ground reaction force (vGRF) signals. Traditional machine-learning approaches for gait-based PD analysis often rely on handcrafted features, while many deep learning methods learn representations directly from raw signals but provide limited interpretability. In contrast, SincPD employs adaptive Sinc-based filters to directly capture informative frequency bands from raw gait data, improving the interpretability of the learned representations. The model achieved 98.77% segment-level accuracy for PD diagnosis and 97.22% segment-level accuracy for severity detection. Furthermore, pruning reduced the number of required filters from 800 to 30 with negligible performance loss, enhancing efficiency and transparency. Robustness tests under multiple simulated noise conditions showed stable performance under controlled perturbations. These results suggest that SincPD combines high predictive performance, computational efficiency, and clinically relevant interpretability, while further validation on larger and real-world cohorts remains necessary.