A Novel Transformer-Based Method for Full Lower-Limb Joint Angles and Moments Prediction in Gait Using sEMG and IMU data

--This study presents a transformer-based deep learning framework for the long-horizon prediction of full lower-limb joint angles and joint moments using surface electromyography (sEMG) and inertial measurement unit (IMU) signals. Two separate Transformer Neural Networks (TNNs) were designed: one for kinematic prediction and one for kinetic prediction. The model was developed with real-time application in mind, using only wearable sensors suitable for outside-laboratory use. Two prediction horizons were considered to evaluate short-and long-term performance. The network achieved high accuracy in both tasks, with Spearman correlation coefficients exceeding ρ = 0.96 and R Notably, the model consistently outperformed a recent benchmark method in joint angle prediction, reducing RMSE errors by an order of magnitude. The results confirmed the complementary role of sEMG and IMU signals in capturing both kinematic and kinetic information. This work demonstrates the potential of transformer-based models for real-time, full-limb biomechanical prediction in wearable and robotic applications, with future directions including input minimization and modality-specific weighting strategies to enhance model efficiency and accuracy. CRUCIAL requirement in developing real-world systems--especially those that involve repetitive tasks--is optimization. Without an optimized system, we risk excessive energy consumption, increased physical or computational effort, and ultimately higher operational costs, all of which are undesirable. However, achieving such optimization requires a foundational step: analyzing the system's dynamics throughout task execution.