Automatic estimation of continuous elbow flexion–extension movement based on electromyographic and electroencephalographic signals

Abstract

The aim of this study was to estimate the elbow joint angle based on EMG and EEG signals using signal processing and machine learning techniques. 21 subjects (ten females, eleven males) performed synchronous flexion–extension movements while EMG, EEG, and elbow kinematic signals were recorded. The EMG and EEG signals were used to estimate the elbow angle employing a long short-term memory neural network. The best results were obtained by training one network per subject (intra subject). The lowest error was reached using the EMG signal, RMSE = 8.59°±2.17° and R2=0.95 with a 95% CI (0.93–0.96). Employing EEG signals generated an RMSE = 9.27°±1.85° and R2=0.95 with a 95% CI (0.94–0.95). When both signals, EMG/EEG, were used, the results were RMSE = 9.53°±2.13° and R2=0.95 with a 95% CI (0.94–0.95). Statistically, for intra-subject data, there is no significant difference in RMSE on using a particular type of signal. In the case of inter-subject data, we obtained the lowest RMSE values considering the combination of EMG/EEG signals, for both, women and men, RMSE = 10.96°±5.28° and RMSE = 9.92°±4.62°, respectively. On the other hand, using subject-wise cross validation, errors increased as expected; however, men’s EMG/EEG signals proved to be robust increasing the RMSE only in 3.47°. A new methodology is proposed for estimating elbow angles based on EMG and EEG biosignals. This can be useful for generating control signals for prostheses and/or exoskeletons designed to provide the support that people with motor disabilities require.