Accurate real-time prediction of lower limb biomechanics is critical for enhancing assistive robotic control systems. While machine learning approaches have emerged as computationally efficient alternatives to traditional musculoskeletal multibody dynamics simulations, existing methods faced persistent challenges including limited prediction accuracy, inefficient multi-modal feature integration, and latency constraints inherent to current-time prediction frameworks. Therefore, this study aimed to propose a wearable sensors-driven deep learning model (KsFormer) for continuous multi-step ahead prediction of sagittal-plane joint moments (hip, knee, and ankle) and three-dimensional ground reaction forces (GRFs) across complete gait cycles. The encoder–decoder model KsFormer was specifically designed for cross-modal feature extraction and integration, with its encoder adopting a three-stage hierarchical processing pipeline. The preprocessed inertial measurement unit (IMU) kinematics and surface electromyography (sEMG) data recorded by wearable sensors served as the inputs of KsFormer. The prediction results were then validated by comparing them to those from gold-standard musculoskeletal simulations and force plate measurements. The results demonstrated exceptional predictive performance with mean Pearson correlation coefficients exceeding 0.9 across six walking speeds and three running speeds patterns, achieving low error rates ( RMSE ¯ = 0.092 N m/kg for joint moments; RMSE ¯ = 0.032 body weight for GRFs). Additionally, the proposed model enabled accurate and continuous biomechanical prediction 240-960 ms prior to motion initiation, significantly outperforming conventional current-time prediction approaches. This study provided a more practical method for real-time lower limb biomechanics feedback to the assistive robotic system in the real-world environment, enabling dynamic torque adjustment and pilot gait pattern recognition.
Zhou et al. (Thu,) studied this question.