TDNN-Based Ankle Angle Prediction from Lower-Limb Kinematics and Surface EMG

This study presents a Time Delay Neural Network (TDNN) framework for predicting sagittal-plane ankle angle from proximal lower-limb kinematics and surface electromyography (sEMG), with the long-term objective of supporting active ankle prosthesis control. Two TDNN configurations were evaluated: a kinematics-based model (TDNN-Kin) using hip and knee kinematics together with gait-phase information, and an extended model (TDNN-Kin-EMG) incorporating three additional sEMG channels. The models were evaluated using independently held-out treadmill-walking trials from nine able-bodied participants obtained from a publicly available biomechanics dataset. The TDNN-Kin model achieved a mean RMSE of 3.52 ± 0.55° with a mean Pearson correlation coefficient of 0.9338. Incorporating sEMG reduced the mean RMSE to 3.04 ± 0.44°, corresponding to a 13.8% improvement, while the mean correlation coefficient increased to 0.9419. RMSE improved in eight of the nine participants, indicating that muscle-activation information provides complementary information beyond proximal-joint kinematics alone. These findings demonstrate the feasibility of accurate offline ankle-angle prediction using wearable kinematic and physiological signals and provide a foundation for the future integration of data-driven trajectory generation into active ankle prostheses.