Quantification of Feature Space Changes With Experience During Electromyogram Pattern Recognition Control

Pattern recognition of the electromyogram (EMG) has been demonstrated in the laboratory to be a successful alternative to conventional control methods for myoelectric prostheses. Pattern recognition control is dependent upon both machine and user learning; the user learns to generate distinct classes of muscle activity while the machine learns to interpret them. With experience, users may learn to generate distinct classes by reducing intraclass variability or by increasing interclass distance. The goal of this study was to identify which of these strategies best explained differences in EMG patterns between subjects with and without experience using pattern recognition control. We compared classification errors of novice nonamputee subjects with experienced nonamputee subjects. We found that after brief exposure to the control method, classification error in novices was reduced, although not to the level of experienced subjects. While the level of intraclass variability in novices was similar to that of the experienced subjects, they did not achieve the same level of interclass distance. These differences can be used to guide the development of much needed rehabilitation methods to train subjects to use pattern recognition devices. In particular we recommend training protocols that emphasize increasing the interclass distance.