An investigation of transition-informed classifier adaptation for myoelectric control

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


Myoelectric prostheses use pattern recognition of surface electromyography (SEMG) to interpret a user’s intent. Over time, changes in the SEMG worsen the usability of these prostheses, requiring cumbersome retraining. Adaptive learning, although able to update the classifier, suffers from mislabelling errors during unsupervised use. This study aimed to overcome this by investigating the impact of transitions between classes, often associated with elevated misclassification, on the adaptation process. Several adaptation techniques, some based on explicitly avoiding transitions and others based on leveraging awareness of transitions to improve decision stream labelling, were explored. Finally, these transition-informed adaptation techniques were tested on two datasets that included sequences of transitions between known classes. Results suggest that an awareness of transience in the SEMG can inform the data selection process and improve the labelling of unsupervised data for adaptation. A resulting LC-SSL technique yielded significant (p¡0.05) improvement to several offline classifier performance metrics.