A Spatio-Temporal Multi-Subgraph Convolutional Network for Parkinson's Disease Detection Using Gait Data

Gait disturbance is a key motor symptom of Parkinson's disease (PD) and has been widely studied for diagnosis and severity assessment using non-invasive sensor data. However, most existing methods for PD detection struggle to adaptively model sensor positions and inter-signal correlations within a multi-sensor system. Additionally, these sensors form a complex system with multiple subsystems monitoring local regions, and their interrelations are often overlooked. In this work, we propose a spatio-temporal multi-subgraph convolutional network (ST-MSGCN), a hybrid framework using graph neural networks to capture gait differences between healthy individuals and PD patients from multi-sensor data. We first decouple the overall sensor system and encode inter-subsystem correlations into multiple subgraphs. The ST-MSGCN then employs a spatio-temporal graph convolutional network to analyze 3D-graph representations of foot force signals, exploring spatial and temporal characteristics within each subgraph. We also introduce channel-wise topology and temporal attention mechanisms into the graph convolutional layers to dynamically learn discriminative features. Finally, a decision-making layer using a voting strategy fuses observations and state estimations from multiple graphs. The proposed method is validated on the PDgait dataset and compared with several existing models. Additionally, it is evaluated on a normal skeleton-based dataset to explore the relationships among body parts.