Improving User-in-the-Loop Myoelectric Control Using Context Informed Incremental Learning
DOI:
https://doi.org/10.57922/mec.2498Abstract
Screen or prosthesis guided training is typically used to train pattern recognition-based myoelectric control by providing controlled calibration samples with known labels. When these models are used with a user-in-the-loop, however, the observed patterns are much more variable, resulting in poor model extrapolation to these conditions and thus poor usability. Incremental and reinforcement learning approaches can continue learning from user-in-the-loop settings, but are limited in their reliability due to the lack of supervised labels and increased training times. In this work, we propose context informed incremental learning (CIIL), which adapts by drawing contextual information from the control task, to solve these issues. We test our claims across two conditions: a short training data scenario and a simulated electrode shift scenario. With only one second of initial training data per class, CIIL achieves similar throughput as SGT in a Fitts' law-style usability test after only two minutes of adaptation (the same amount of time taken for SGT). In the harder electrode shift scenario, CIIL significantly outperformed the pre-shifted SGT model after 5 minutes of adaptation, offering a promising direction for future clinical validation of user-in-the-loop training.
Downloads
Published
2024-08-15
How to Cite
[1]
E. Campbell, E. Eddy, U. Côté-Allard, and E. Scheme, “Improving User-in-the-Loop Myoelectric Control Using Context Informed Incremental Learning”, MEC Symposium, Aug. 2024.
Conference Proceedings Volume
Section
Myoelectric Control Algorithms
