User adaptation in Myoelectric Man-Machine Interfaces

Abstract

State of the art clinical hand prostheses are controlled in a simple and limited way that allows the activation of one function at a time. More advanced laboratory approaches, based on machine learning, offer a significant increase in functionality, but their clinical impact is limited, mainly due to lack of reliability. In this study, we analyse two conceptually different machine learning approaches, focusing on their robustness and performance in a closed loop application. A classification (finite number of classes) and a regression (continuous mapping) based projection of EMG into external commands were applied while artificially introducing non-stationarities in the EMG signals. When tested on ten able-bodied individuals and one transradial amputee, the two methods were similarly influenced by non-stationarities when tested offline. However, in online tests, where the user could adapt his muscle activation patterns to the changed conditions, the regression-based approach was significantly less influenced by the changes in signal features than the classification approach. This observation demonstrates, on the one hand, the importance of online tests with users in the loop for assessing the performance of myocontrol approaches. On the other hand, it also demonstrates that regression allows for a better user correction of control commands than classification.