Evaluation of Hand Gesture Classification Based on Motor Units Activation for Prosthetic Application

Recent advancements in upper-limb prosthetics have enabled users to control multi-degree-of-freedom prostheses using pattern recognition algorithms applied to electromyography (EMG). These methods are built on the extraction of EMG features that treat EMG signals as colored noise, neglecting their biological origin. In contrast, recent works have explored the use of features based on the decomposition of EMG signal into the activity of the motor units. These features hold great potential for control applications as they reflect the actual neural drive to the muscles. Here, we evaluate the feasibility of using a motor unit based feature to classify hand gestures. To this purpose, high-density EMG signals were recorded from 5 able-bodied participants performing 14 different hand gestures, including 10 fundamental movements and 4 grasps. The recorded EMG signals were offline decomposed with a blind source separation algorithm. Then, motor unit features were computed and a classification paradigm inspired by previous literature was applied. The accuracy of the motor unit based classification was compared to a machine learning approach based on root-mean-square (RMS). Different classification models were tested for both approaches. The accuracy of the motor unit based classification was lower and exhibited a greater inter-subject variability with respect to RMS based classification. However, accuracy values ranged between 80.6% and 90.3% and in some cases were comparable or even higher than those obtained from RMS classification. Our study suggests that motor units have the potential to be used for prosthetic control, though additional research is required to effectively translate these findings into practical applications.Clinical relevance- This study builds on previous literature to demonstrate the feasibility of using motor unit derived features to decode multi-degree-of-freedom hand gestures. The study also highlights technological and methodological limitations hindering real-life application of such approach.