Interaction-Driven Approaches for Efficient and Autonomous Calibration of Myoelectric Controllers

Abstract

Dexterous myoelectric prostheses of the hand and wrist typically use machine learning to translate muscle activations into motor commands. Their performance relies on the quality of the training data, posing challenges to the effectiveness of the calibration process. The lack of immediate quality assessment during standard data acquisition renders the need for additional data only apparent after the acquisition is complete. Additionally, accurate labeling of the training data for supervised learning relies on the user's ability to deliver precise muscle contractions, which necessitates professional supervision during preprosthetic signal assessment and training. These challenges may be addressed through novel calibration protocols that leverage continuous interaction between the user and the myocontrol system. The research initially focused on the inefficiency of existing multi-arm-position data acquisition protocols, which do not allow an instantaneous evaluation of model quality in different arm configurations. An interactive protocol was elaborated that combined real-time incremental model building with a feedback mechanism to direct users to acquire data in underperforming arm configurations. In subsequent work, the need for labeled training data was circumvented altogether by devising a novel unsupervised calibration paradigm, driven by an interaction protocol where the user and system synergistically identify muscle contractions suitable as myocontrol inputs. This was achieved by having the user practice an abstract motor mapping between adaptively extracted muscle synergies and arbitrarily associated prosthetic functions. An initial version of this method focused on the simultaneous learning of multiple functions, whereas a successive version enabled users to learn the prosthetic functions progressively. The studies highlighted that interactive procedures for myoelectric data acquisition and labeling increase the efficacy of supervised model calibration, holding practical relevance for prosthetic control and warranting further investigation in a broader range of myocontrol applications. Additionally, the potential of unsupervised calibration methods for myocontrol was demonstrated, especially in enabling users with varied residual motor abilities to engage quickly and autonomously with myocontrol systems, and to concurrently explore or even expand their muscle capabilities. Ultimately, this work presented a shift in perspective toward greater user involvement in myocontrol model calibration, contributing to more personalized and accessible myoelectric control.