Continuous estimation of joint angle from electromyography using multiple time-delayed features and random forests

To estimate the continuous human motion from surface electromyography (sEMG), it is required to extract hidden information from sEMG and generalize an estimation model. In this study, we proposed that the multiple time-delayed feature (MTDF) signals of sEMG improve the performance of elbow joint motion estimation. Among different learning algorithms, we found Random Forests (RF) to be the best in terms of execution time and accuracy of estimation. Features of sEMG that were best describing the joint motion included: mean absolute value, waveform length, zero crossing, slope signs changes, and difference absolute standard deviation value. The speed of joint movement ranged from 15°/s to 180°/s. The time-delay coefficient and the optimal time-delayed coefficient of RF using MTDF method were 2 and 317, respectively. Mean difference and the standard difference between the actual angle and the estimated angle, using the Bland-Altman analysis, were 0.08 and 5.01, respectively. The average root mean square difference value was 0.0543 ± 0.0071. In addition, the average execution time of motion estimation of a 17.57-s signal was 0.2642 seconds. We found the RF algorithm using MTDF features as the most robust method to estimate the elbow joint motion.