Multimodal Spatiotemporal Biomechanical Modeling for Musculoskeletal Injury Risk Screening of Body Movements Based on Cross-Modal Attention Network

Abstract

Ethnic dance involves large-amplitude and repetitive movements that may impose substantial and uneven mechanical demands on the musculoskeletal system, potentially increasing the risk of joint and soft-tissue injuries. Conventional injury assessment relies largely on manual observation and clinical evaluation, which limits the ability to continuously and quantitatively characterize movement-related risk. To address this limitation, this study proposed a Cross-Modal Attention-based Injury Risk Perception Network (CMA-IRPN) for non-contact movement analysis, movement-pattern risk screening, and injury risk assessment in ethnic dance. The framework jointly modeled Red, Green, and Blue (RGB) video sequences and human skeletal keypoints. A dual-branch encoder combining a three-dimensional convolutional neural network (3D-CNN) and a graph convolutional network (GCN) was used to extract spatiotemporal visual and skeletal features, while spatiotemporal cross-modal attention modules were introduced to model interactions between movement appearance and joint-topological information. The model was evaluated for movement recognition using the publicly available MPII Human Pose and SportsPose datasets and for movement-pattern risk grading using a self-constructed ethnic dance dataset. CMA-IRPN achieved accuracies of 96.03% and 94.86% on the MPII and SportsPose datasets, respectively, with statistically significant improvements over the comparison models (p < 0.05). On the ethnic dance dataset, the model achieved a Macro-F1 score of 92.25% for movement-pattern risk grading and a sensitivity of 93.75% for the high-risk category. These results suggest that CMA-IRPN can identify risk-related movement patterns from non-contact multimodal data and provide quantitative support for movement screening and injury risk assessment in ethnic dance training. The predicted risk levels represent visual-kinematic movement patterns rather than direct measurements of joint torque, tissue stress, or other mechanical loads and therefore should not be interpreted as operational real-time injury early warnings or as substitutes for clinical biomechanical assessment or medical diagnosis.