Continuous prediction of knee and ankle joint angles based on sEMG and hip joint angle fusion signals by NRBO-LSTM network

Surface electromyography (sEMG) serves as a crucial information source for capturing human motion. Joint angle prediction based on sEMG holds significant application value in the field of rehabilitation medicine. This paper investigates continuous prediction of knee and ankle joint angles under three walking modalities (Level walking, Ramp ascent, and Stair ascent). Experimental data comprising sEMG and inertial measurement unit (IMU) signals were collected from 10 healthy subjects, with time-domain features extracted from sEMG signals and joint angles calculated from IMU data. Model performance was evaluated using the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE). Three input modalities were compared: single 7-channel sEMG signals, single 4-channel sEMG signals, and fused 4-channel sEMG with hip joint angle signals. Comparative analyses were conducted between long short-term memory networks (LSTM), MHA-LSTM(Multi-Head Attention LSTM), and NRBO-LSTM(Newton-Raphson-Based Optimization LSTM) models. Results demonstrate that when using fused signal input with the NRBO-LSTM model, the RMSE values for knee and ankle joint prediction were 4.11 ± 0.74° and 3.48 ± 0.54° (Level walking), 4.33 ± 0.76° and 3.6 ± 0.78° (Ramp ascent), 4.91 ± 1.0° and 4.3 ± 0.7° (Stair ascent), respectively. The MAE values were 3.04 ± 0.6° and 2.66 ± 0.41° (Level walking), 3.37 ± 0.63° and 3.29 ± 0.61(Ramp ascent), 3.74 ± 0.74° and 3.29 ± 0.61° (Stair ascent), respectively. The R 2 values were 0.93 ± 0.03/0.89 ± 0.06 (Level walking), 0.9 ± 0.04/0.76 ± 0.08 (Ramp ascent), and 0.95 ± 0.02/0.64 ± 0.21 (Stair ascent), respectively. This research provides theoretical references for exoskeleton control.