Physiological classification of Parkinson’s disease severity using multimodal speech biomarkers with a hybrid CNN-Mamba framework

Abstract

Introduction – Hypokinetic dysarthria in Parkinson’s disease provides an accessible non-invasive biomarker, but multi-class severity grading remains difficult because of overlapping acoustic patterns and limited long-range temporal modeling in existing approaches. Methods – We developed a hybrid CNN-Mamba framework using multimodal speech features transformed into 2D representations. The model was trained and validated on speaker-disjoint PC-GITA Spanish data and tested on an independent Mandarin clinical cohort, with additional external evaluation on a public Parkinsonian speech corpus. Speaker-level results were obtained by aggregating segment predictions within each subject. Results – Segment-level accuracy reached 97.8% on PC-GITA and 95.4% on the Mandarin cohort. Speaker-level accuracy reached 94.0% and 91.2% using majority voting, improving to 94.8% and 91.9% with mean-probability aggregation. SHAP analysis supported physiological interpretability, and ablation studies showed advantages over CNN-BiLSTM, Transformer, and SVM baselines. Discussion – The proposed CNN-Mamba framework provides an interpretable, computationally efficient, and non-invasive approach for Parkinson’s disease severity assessment and remote monitoring, with promising cross-lingual transfer under structured clinical speech tasks.