Multi-Position Training Improves Robustness of Pattern Recognition and Reduces Limb-Position Effect in Prosthetic Control

摘要

Introduction: Electromyogram (EMG)-based pattern recognition control of prosthetic limbs is the current state of the art. However, these systems commonly fail when the user attempts to use the limb in a different position from which it was trained, resulting in significantly reduced functionality. Robust models for decoding EMG signals, accounting for specific changes that occur with positional variation, are needed to reduce this negative effect.Methods: Ten able-bodied participants and two participants with transradial amputation were included in the study. Participants were fitted with surface EMG electrodes as well as a network of inertial measurement units (IMUs) to monitor limb position during tasks. Positional covariates including elbow angle, hand height, and forearm angle were analyzed for impact on EMG signal features to drive the generation of unique LDA classifier algorithms. Offline analysis of classification error for each control scheme was then completed.Results: Elbow angle demonstrated the strongest impact on the EMG signal. Hand height also demonstrated a consistent increase in EMG signal with increasing height. Incorporating these specific covariates into classifier algorithms improved performance compared with classifiers trained in the conventional fashion (single-position EMG). However, able-bodied participants demonstrated lowest classification error when data from random-training positions were incorporated (10.3% vs. 17.2% single position, P < 0.001). These results were even more dramatic in participants with amputation (with five training repetitions: 7.14% vs. 32.08%, P < 0.001). Performance differences between single-position and random-position training for individuals with amputations were significantly larger when the user was wearing his/her prosthesis than otherwise.Conclusions: Incorporating position-specific covariates into myoelectric classification algorithms can dramatically improve robustness and classification accuracy when using the prosthesis in the user's entire workspace. In single-position training paradigms, classification error rates were 39.22% and 32.18%, respectively, for two participants with amputation and resulted in unusable classifiers. Conversely, classification errors were at 10% for able-bodied and near 7% for participants with amputation when at least five training repetitions were used to train either a random position or position-specific classifier. As position-tracking hardware becomes smaller and can be implemented into socket designs, incorporating this information into classifier algorithms can dramatically reduce the limb-position effect. Current users can experience reduction of the limb-position effect through training in multiple random positions.