Multimodal dynamic-static combined graph convolutional network for detecting freezing of gait

• Published in: Biomedical signal processing and control Vol. 112; p. 108368
• Main Authors: Xia, Yi, Cao, Aisheng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2026
• Subjects: Freezing of gait detection; Graph convolution neural network; Multimodal; Parkinson’s disease; Multimodal; Graph convolution neural network; Freezing of gait detection; Parkinson’s disease
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.108368

Freezing of gait (FoG) is a common motor disorder in Parkinson’s disease (PD) patients characterized by sudden, temporary walking cessation. Early detection and intervention of FoG are crucial for improving the quality of life for PD patients. This study proposes a FoG detection method based on a multimodal graph convolutional network. Specifically, human skeletal sequences are first extracted from the monitoring videos, and then wearable inertial sensor data are added into the human skeletal graph as extra nodes to form a unified gait graph structure. To extract gait features from the above graph data, this study proposes combining dynamically learnable graph structure with the static human physical skeleton structure to form a dynamic and static combined adjacency matrix. Then, feature aggregation is performed in different channels based of the aforementioned graph structure. After the joint spatial features are obtained, multi-scale temporal convolution is applied to extract multi-scale temporal features at different time scales. Comprehensive experimental results show that our model achieved very promising performance for FoG detection on the public multimodal FoG dataset, with an overall accuracy of 91.03% and a F1 score of 0.8629, when evaluated with a subject-independent cross-validation scheme. The code for MDS-STGCN is publicly available at: https://github.com/caoaisheng/MDS-STGCN. [Display omitted] •A multimodal approach is adopted to enhance the accuracy of FoG detection.•A dynamic-static combined adjacency matrix is proposed to boost graph modeling.•A non-shared topology is proposed to facilitate the FoG feature learning.