Abstract
Myocontrol is control of a prosthetic device using data obtained from (residual) muscle activity. In most myocontrol prosthetic systems, such biological data also denote the subject's intent: reliably interpreting what the user wants to do, exactly and only when she wants, is paramount to avoid instability, which can potentially lead to accidents, humiliation and trauma. Indeed, instability manifests itself as a failure of the myocontrol in interpreting the subject's intent, and the automated detection of such failures can be a specific step to improve myocontrol of prostheses-e.g., enabling the possibility of self-adaptation of the system via on-demand model updates for incremental learning, i.e., the interactive myocontrol paradigm. In this work we engaged six expert myocontrol users (five able-bodied subjects and one trans-radial amputee) in a simple, clear grasp-carry-release task, in which the subject's intent was reasonably determined by the task itself. We then manually ascertained when the intent would not coincide with the behavior of the prosthetic device, i.e., we labeled the failures of the myocontrol system. Lastly, we trained and tested a classifier to automatically detect such failures. Our results show that a standard classifier is able to detect myocontrol failures with a mean balanced error rate of 18.86% over all subjects. If confirmed in the large, this approach could pave the way to self-detection and correction of myocontrol errors, a tighter man-machine co-adaptation, and in the end the improvement of the reliability of myocontrol.