Prediction of lower limb joint angles from surface electromyography using XGBoost

Surface electromyography (sEMG) is an important source of information for capturing human movement. Based on sEMG, joint angles can be predicted, which has important application value and significance in the fields of rehabilitation medicine, sports science, and bionic prosthetics. In this paper, eXtreme gradient boosting (XGBoost) regression model is used to predict the joint angle of lower limbs based on surface sEMG signal. In the experiment, the sEMG signals and joint angles were collected from 18 healthy subjects (11 men, 7 women) at different walking speeds and muscle synergy was employed to reduce the 10-channel sEMG signals to 3-channel. In the process of selecting input features, a comparative analysis of time-domain (TD), frequency-domain (FD), and time-frequency domain (TFD) features was conducted. In the end, the trained regression models were evaluated using metrics R-square (R2), root mean square error (RMSE), and correlation coefficient (gamma). The prediction results of extreme gradient boosting (XGBoost) were compared with those of multi-layer perception (MLP), recurrent neural networks (RNN), attention long short-term memory (ATN-LSTM), and dynamic convolution neural network (DCNN). The results clearly demonstrate the excellent performance of XGBoost regression model for joint Angle prediction. It can meet the needs of joint Angle prediction and provide reference for exoskeleton control. For the hip and knee joints the predicted RMSE and gamma values were 3.233 f 1.959 degrees, 3.811 f 1.904 degrees and 0.984 f 0.018, 0.984 f 0.021.