Optimal EMG Feature Selection for Classification of Normal and Post-Stroke from Hand-Reaching Movement

Abstract

Electromyography (EMG) signals have been used extensively in research related to muscle functioning rehabilitation of post-stroke patients. EMG signal is non-linear, non-stationary, and similar in domains of time and frequency hence needs an automated system to classify the signal acquired from normal and post-stroke patient. An accurate EMG signal classification system relies on the number of features extracted from datasets. More features involved mean a better classification performance but causing a heavier computational needed. In this paper, we selected the most optimal EMG signal feature to minimize the feature used. The criteria of selected optimal feature is should have the highest accuracy compared to the other features. In order to weigh the accuracy, we employed the tree-based classifier. The optimal features were selected among 13 time-domain (TD) of normal and post-stroke patient EMG hand-reaching movement. This study contributes to minimize the number of EMG features used in the classification system for a faster time processing yet still outcome the same accuracy than using all features, thus the computational resource kept efficient.