Training Movement Velocity Significantly Affects the Performance of Myoelectric Control

Abstract

Our native hands are uniquely capable of operating across a wide range of speeds and forces. In contrast, most commercial myoelectric prostheses typically provide limited speed and force output. One approach to endow myoelectric prostheses with variable speed and/or force output is to use continuous kinematic positions of the prosthesis based on electromyography (EMG). Within the field of machine learning, it is well established that homogeneous training data can lead to bias that negatively impacts the run-time performance of the algorithm. Yet, most continuous decoders are trained on a homogeneous dataset involving only a single kinematic speed. To this end, we systematically investigated how different training speeds influence myoelectric control with two common continuous decoders on multiple performance metrics. We compared a Kalman filter (KF) and Convolutional Long Short-Term Memory (C-LSTM) neural network trained on slow, medium, fast, and mixed-speed datasets, evaluating their performance in offline analyses and in two real-time online tasks with the user actively in the loop. We found that training speed significantly affected algorithm performance, but effects were often algorithm dependent. Linear algorithms, like the KF, are likely to exhibit lower unintended movement errors and smoother control when trained on slow-speed data but will also struggle to generalize to higher movement speeds. In contrast, nonlinear algorithms like the C-LSTM can likely provide greater adaptability, with mixed-speed training leading to improved accuracy and task success rates across conditions. Although an often-overlooked implicit parameter, these findings explicitly demonstrate that a lack of diverse training speeds in existing myoelectric control training paradigms leads to worse decoder performance. By incorporating a range of movement speeds into training protocols or decoder design, myoelectric continuous decoders could achieve more dexterous and robust control, potentially improving prosthetic usability and retention.