Subject-independent Parkinson's Disease Detection through A Lightweight Spectrotemporal Hybrid Network

Vocal impairment constitutes a prominent prodromal biomarker for Parkinson's disease (PD). However, existing deep learning solutions often rely on computationally intensive architectures or biased validation protocols. This study proposes the lightweight spectrotemporal hybrid network (LST-Net), a resource-efficient framework for robust PD screening.

The model employs a dual-stream architecture that concurrently processes raw one-dimensional waveforms and two-dimensional Mel-spectrograms to capture complementary cues, specifically transient temporal jitters and spectral energy patterns. The proposed architecture was evaluated on the Italian Parkinson's Voice and Speech dataset using a strict subject-independent 5-fold cross-validation protocol to prevent data leakage.

The proposed LST-Net achieved a robust segment-level accuracy of 96.69%. By implementing a subject-level majority voting strategy, mimicking clinical diagnosis where patients are assessed over time, the model successfully classified all subjects in the held-out test sets (100% subject-level accuracy). Quantitative Gradient-weighted Class Activation Mapping analysis confirmed the model's focus on clinically relevant low-frequency tremors (mean energy: 61.56% ±27.90%).

This framework demonstrates that explicit spectrotemporal fusion yields superior diagnostic performance without heavy computational burdens, offering a verifiable and scalable solution for real-time deployment on edge devices.