A multimodal spatio-temporal graph neural network framework for fatigue detection in tennis serving

Background The tennis serve is a complex, high-velocity motion dependent on the efficient transfer of energy through the kinetic chain, from the lower extremities to the racquet. Fatigue-induced alterations in this chain are primary precursors to overuse injuries, such as rotator cuff tendinopathy and elbow medial collateral ligament stress. However, current wearable monitoring solutions predominantly rely on unimodal kinematic data, failing to capture the neuromuscular compensatory mechanisms that precede mechanical performance degradation. Methods This study introduces a novel multimodal framework integrating inertial measurement units (IMUs) and surface electromyography (sEMG) with a Spatio-Temporal Graph Convolutional Network (ST-GCN). We recruited 15 high-performance tennis players to perform a standardized specific endurance protocol (serving to exhaustion). A graph topology representing the human skeletal structure was constructed to model spatial dependencies, while sEMG signals were fused as node attributes to capture neural drive intensity. Statistical significance was assessed using paired t -tests with Bonferroni correction, and model performance differences were evaluated by one-way repeated measures ANOVA with post-hoc pairwise comparisons. Results The proposed ST-GCN framework achieved a fatigue state classification accuracy of 95.2% (F1-score: 0.94), significantly outperforming traditional Long Short-Term Memory (LSTM) (88.1%) and Convolutional Neural Network (CNN) architectures (85.6%), as well as a Spatial-Temporal Transformer network (ST-TR) (93.1%) ( p < 0.001 for all pairwise comparisons). Biomechanical analysis revealed a significant “neuromechanical lag” (p < 0.01) and a compensatory pattern characterized by reduced knee flexion (−12.4°) during the fatigued state, accompanied by a non-significant upward trend in shoulder internal rotation velocity (+8.5%, p = 0.012). Conclusion By decoding the hidden dependencies within the kinetic chain, this system provides a computationally efficient fatigue detection framework with a model inference latency of 8.4 ms, constituting a technically feasible step toward on-court deployment and offering a potential paradigm for smart sports engineering and precision rehabilitation.