Classification of Gait Phases by Using SVM and ANN Based on EMG Signals

Advanced technology in rehabilitation aims to improve gait patterns through innovative mechanisms and powerful motors. Interestingly, a good performance of the control system of those devices can be achieved when it is paired with a functional gait phase detection algorithm using wearable sensors such as electromyography (EMG) signals. Since emerging machine learning in EMG signals has a significant impact on the development of exoskeletons, machine learning such as artificial neural networks (ANN) has been widely utilized, especially for gait patterns and gait phases. Although support vector machines (SVM) are seen as having great potential for interpreting EMG signals, few studies have been observed in gait phases, especially stance and swing. Therefore, this study proposes a classification gait phase by using SVM and compares the performance with ANN. Combinations of two and more of the five time domain (TD) features were extracted from EMG signals and fed into the SVM and ANN models. Then, the SVM and ANN models with different kernel functions and training algorithms were compared, respectively. As a result, combinations of all five TD features enhanced the classification accuracy more than two or fewer combinations of TD features. Besides, SVM with a radial basis function (RBF) achieved better performance than a linear function with 98% accuracy. This model also performed better than ANN, which only gained up to 95.8% of classification accuracy. Thus, this study demonstrates that SVM is not only able to discriminate between stance and swing phases but also improves the accuracy of gait phases. Therefore, SVM with an RBF kernel function should be considered for analyzing EMG signals in near future.