EMG-based decoding of motor tasks in amputees using a minimal set of EMG electrodes.

Lower-limb powered prostheses have been widely studied and optimized over the years to help amputees improving their life quality and their walking performances during different motor tasks. Prostheses control can rely on various kinds of sensors, like inertial sensors and accelerometers. Among these, electromyographic (EMG) control of powered prostheses seems to be a promising approach. In this study, we recorded EMG signals from four muscles (i.e. Adductor, Tensor Fasciae Latae, Rectus Femoris and Biceps Femoris) of the stump of thirteen transfemoral amputee patients while they were performing five motor tasks: ground level walking, stairs ascending, stairs descending, ramp ascending and ramp descending. Recorded signals were processed and the mean absolute value (MAV) was extracted over a 20ms moving window. Post-processed signals were then used as input of a Support Vector Machine-based algorithm to classify the motor tasks. The algorithm followed two steps of decoding of motor tasks, classifying ground level walking tasks first, and trying to identify stairs ascending, stairs descending, ramp ascending and ramp descending afterwards. Results showed that it was possible to identify motor tasks with one single muscle, with an accuracy ranging between 70% and 83%. Increasing the number of muscles used for the classification led to higher performances: combining EMG from three or four muscles resulted in performances above 90%. These results proved that it is possible to reliably decode motor tasks with a EMG-based algorithm relying on as little as three recording sites, ensuring a minimum discomfort of the system for the users.