An Anatomy-Informed Cross-Attention Framework for sEMG-Driven Knee and Ankle Moment Prediction During Sit-to-Walk Transitions

Sit-to-walk (STW) is a short-duration, high-load, multijoint transition requiring rapid lower-limb neuromuscular coordination across seat-off, load transfer, and gait initiation. Surface electromyography (sEMG)-based prediction of knee and ankle joint moments may support motor function evaluation and inform future assistive-control applications, but existing models remain limited in modeling cross-muscle sEMG feature interactions and mitigating phase-dependent prediction errors. This study developed an anatomy-informed framework for sEMG-driven moment prediction during STW. The model encoded sEMG channels from thigh and shank muscles into separate anatomical branches. Cross-Attention was used to model task-relevant intersegmental interactions, and BiLSTM was applied to capture short-term temporal dependencies. Eighteen healthy participants performed STW trials while sEMG, three-dimensional kinematics, and ground reaction forces were synchronously collected. Knee and ankle moments were calculated using inverse dynamics and used as reference targets. Among six models, the Cross-Attention model achieved the lowest test-set overall error, with an Overall nRMSE Fixed of 4.51%; the knee peak error in the P3 unloading phase was 16.17%. Ablation experiments indicated that Cross-Attention, BiLSTM temporal modeling, anatomical branch separation, and joint-specific output mapping contributed to prediction performance. This framework provides an interpretable approach for sEMG-driven multijoint moment prediction in complex non-stationary movements.