Global classification of intentional movement across upper limb myoelectric pattern recognition-controlled prosthesis users

One idealized vision of advanced upper limb prosthetic control is a plug-and-play design that any new user can don and instantly control attached devices as intended. Traditional body-powered prostheses are likely the closest available option but are limited in their range of motion and can have negative long-term impacts. Basic dual-site myoelectric-controlled prostheses require only minor adjustments prior to users being able to control their prosthetic device, but still requires training to learn despite only having a limited number of motions available. This contrasts with state-of-the-art myoelectric pattern recognition-controlled prostheses where a machine learning algorithm also learns the individual user; specifically, their unique patterns of muscle activity corresponding to prosthesis motions. However, the wide variation in muscle activity patterns both within and between users, mainly due to physiological differences, has been the primary reason why it is difficult to develop a true off-the-shelf prosthesis component for myoelectric pattern recognition control. In this paper, we take a small step towards this vision by investigating statistical and machine learning methods that classify prosthesis motion for any pattern recognition user. Specifically, we use a large dataset of EMG training data collected from 191 users over a six-month period to develop, as a first step, a binary classifier built to simply identify intended motion or no motion for all users. Our results could have an immediate impact on prosthesis performance for current users and justify further development of a potential global classification model which can be used by any persons with upper limb difference who wish to use a myoelectric-controlled prosthesis.