HIERARCHICAL ATTENTION SPATIOTEMPORAL GRAPH NEURAL NETWORK WITH DYNAMIC MODALITY WEIGHTING FOR STAGE-SPECIFIC PARKINSON’S DISEASE DETECTION

Abstract

Parkinson’s disease (PD), a progressive neurodegenerative disorder, affects over 10 million people globally, necessitating early detection to enable timely interventions that enhance quality of life. Current diagnostic methods, such as the MDS-UPDRS, rely on subjective clinical assessments, often missing subtle early-stage symptoms like minor gait changes or hand tremors. This paper proposes a Hierarchical Attention Spatiotemporal Graph Neural Network with Dynamic Modality Weighting (HAST-GNN-DMW) for stage-specific PD detection using multimodal computer vision. Integrating gait, hand movement, and speech data from the PPMI, PD-Posture-Gait, and a synthetic PD-MultiStage dataset (300 patients, Hoehn-Yahr labeled), our framework employs hierarchical attention to model intra- and inter-modality dependencies and dynamically weights modalities based on patient-specific symptom severity. Explainable AI (XAI) via Integrated Gradients identifies key biomarkers, such as stride length and tremor frequency, enhancing clinical interpretability. Evaluated on PD-MultiStage, HAST-GNN-DMW achieves 93.8% accuracy in early PD detection and 90.5% in stage classification, outperforming state-of-the-art methods like ST-GCN and DenseNet. Ethical protocols ensure fairness through balanced datasets and GDPR-compliant anonymization. Limitations include dataset size and real-world noise sensitivity, with future work targeting larger cohorts and edge-based telemedicine deployment. This framework offers a scalable, interpretable solution for early PD diagnosis, advancing clinical adoption and improving patient outcomes.