Machine Learning based Electromyography Signal Classification with Optimized Feature Selection for Foot Movements

Abstract

Electromyography (EMG) signals are bioelectric signals generated by the electrical activities of muscle fibers during contraction or relaxation. Detailed analysis and classification of the complex nature of this signal when related to movements is complicated but are particularly useful for controlling prosthesis and orthosis control systems. To determine optimal signal classification, in this paper the relevant set of features and the classifier that maps these features to carry out EMG signal classification for four different foot movements is proposed. These movements such as Plantar Flexion (PF), Dorsi Flexion (DF), Inversion (IV) and Eversion (EV) are chosen, since these are useful for rehabilitation of persons having a lower limb ankle joint injury which results in gait abnormality. In this study, EMG signals are acquired using BIOPAC System (MP 150). The features for EMG signals, in time and frequency domain has been extracted to find optimal features and are further classified using support vector machine (SVM), Neural Network (NN) and Logistic Regression (LR). From the results, it is depicted that the time domain features reflected better performance. The maximum classification accuracy achieved is 99.69% and average classification accuracy being 94.92 ± 3.03 % using linear SVM for RMS as optimal feature.