A Pilot Study of EMG Pattern Based Classification of Arm Functional Movements

Abstract

Most previous studies of electromyography (EMG) pattern recognition with both able-bodied subjects and amputees for control of multifunctional prostheses had verified high performance in identifying different movements. While these movements mostly refer to single joint, it remains unclear whether the functional tasks involved in arm and hand could be discriminated by using EMG pattern based methods. In this pilot study, we investigated the performance of EMG pattern recognition in classifying eight functional movements plus a “no movement” task. Four kinds of EMG feature sets, time-domain (TD) features, auto-regression (AR) model features, combination of TD and AR features, and wavelet packet coefficients, were used to represent the EMG patterns, respectively. Using a linear discriminant analysis classifier, the TD features outperformed other three feature sets. The average classification accuracy of the TD features across four able-bodied subjects was greater than 94%. And the feasibility of EMG channels reduction was estimated with straightforward exhaustive search algorithm in terms of classification accuracy. The average classification accuracy of all 8-channel EMG combinations could achieve above 90%. This result was encouraging and suggested that it is feasible to use EMG pattern recognition for the classification of functional movements.