EMG-based multi-joint kinematics decoding for robot-aided rehabilitation therapies

In recent years, a significant effort has been invested in the development of kinematics-decoding models from electromyographic (EMG) signals to achieve more natural control interfaces for rehabilitation therapies. However, the development of a dexterous EMG-based control interface including multiple degrees of freedom (DOFs) of the upper limb still remains a challenge. Another persistent issue in surface myoelectric control is the non-stationarity of EMG signals across sessions. In this work, the decoding of 7 distal and proximal DOFs' kinematics during coordinated upper-arm, fore-arm and hand movements was performed. The influence of the EMG non-stationarity was tested by training a continuous EMG decoder in three different scenarios. Moreover, the generalization characteristics of two algorithms (ridge regression and Kalman filter) were compared in the aforementioned scenarios. Eight healthy participants underwent EMG and kinematics recordings while performing three functional tasks. We demonstrated that ridge regression significantly outperformed the Kalman filter, indicating a superior generalization ability. Furthermore, we proved that the performance drop caused by the session-to-session non-stationarities could be significantly mitigated by including a short re-calibration phase. Although further tests should be performed, these preliminary findings constitute a step forward towards the non-invasive control of the next generation of upper limb rehabilitation robotics.