Differences in EMG Feature Space between Able-Bodied and Amputee Subjects for Myoelectric Control

Difficulties accessing amputee populations has resulted in the widespread adoption of able-bodied subjects in virtual environments for the development of myoelectric prostheses. Factors such as scar tissue, different physiologies or surgical outcomes, and reduced visual and proprioceptive feedback, however, may contribute to differences in electromyogram (EMG) patterns between these groups. As such, studies have consistently found worse results when comparing the performance of amputee subjects to that of their able-bodied counterparts under the same conditions. To identify the source of this performance degradation, a topology-based data analysis method, called Mapper, was employed to visualize the "shape" of EMG feature spaces derived from amputee and able-bodied subjects. The information content of amputee EMG features was found to differ from those of non-amputee subject in three ways: 1) the loss of nonlinear complexity and frequency information, 2) the loss of time-series modeling information, and 3) the segmentation of unique information. The empirical effects of these differences were visualized by classifying motion classes using consistent and migratory features from functional feature groups. In summary, this work characterized inconsistencies in EMG features between amputee and able-bodied populations by theoretical means, highlighted the empirical effects when these are ignored, and proposed a solution for future studies with able-bodied subjects.