Comparative analysis of EMG-based classifiers for recognizing hand/wrist gestures and forces

Hand gesture recognition through surface electromyographic (sEMG) signals represents a well-consolidated approach for a natural and intuitive prosthesis control. Several studies showed how the classification performance can differ between healthy subjects and amputees, due to the high complexity and variability of amputations. Nonetheless, the evaluation of the goodness of Machine Learning classifiers in terms of adaptability to the subjects has never been studied deeply, and only few studies involved a sufficient number of amputees in their analysis. In this paper, the performance of four state-of-the art classifiers in decoding both motor intention and forces were evaluated, using a hierarchical classification approach. Algorithms have been tested on 29 healthy subjects and 15 trans-radial amputees. Results showed high performance for both the Non Linear Regression Classifier and the Random Forest with bootstrap aggregation, for both the gestures and the forces classification problem. No significant differences were found between the two classifiers, suggesting that they may be used interchangeably.