Multi-Source Motion Inputs and FA-TA-BiLSTM for Lower-Limb Joint Angle Prediction

Accurate lower-limb joint angle prediction can support motion-state perception and rehabilitation-oriented analysis. This offline feasibility study evaluated an FA-TA-BiLSTM model using surface electromyography (sEMG) features, together with historical hip and knee joint angles and angular velocities. Data were collected from five healthy adult male participants during level walking and sit-to-stand transitions, and the prediction horizon was 100 ms. Support vector regression (SVR), BiLSTM, and FA-TA-BiLSTM were compared under the same combined-input condition and evaluation protocol. The FA-TA-BiLSTM model achieved RMSE, MAE, and R2 values of 2.0684∘, 1.5920∘, and 0.9726 for hip prediction and 2.9604∘, 2.5142∘, and 0.9660 for knee prediction, respectively. These values were obtained using a mixed-participant chronological split and should be interpreted as preliminary within-cohort estimates rather than evidence of participant-independent generalization. Under this restricted protocol, FA-TA-BiLSTM produced lower errors than SVR and standard BiLSTM. The participant-level model ranking was consistent across the five participants, but exact pairwise comparisons did not reach significance after Holm adjustment. The current comparison does not isolate the incremental contribution of sEMG or individual attention modules; larger and more diverse cohorts, participant-independent validation, modality and module ablation, and causal online evaluation remain necessary.