Electromyogram based muscle stress estimation of Gastrocnemius medialis using Machine learning algorithms

Surface electromyography (sEMG) is the primary technique for recording muscle activity by capturing the electrical signal from the muscles. The present study has been undertaken to analyze lower limb muscle stress of gastrocnemius medialis without and with a workout (exercise) using machine learning algorithms. Ten healthy subjects (Seven male & three female) have been chosen on the basis of their age, height and mass. The experiment has been conducted without a workout and with a workout having a set of 10 kg resistance band loads for the gastrocnemius medialis muscle of the right leg. Fourth-order Butterworth filters and full wave rectifiers have been employed in data preprocessing. Many features have been extracted, including mean absolute value (MAV), average, skewness, kurtosis, zero crossing (ZC), slope sign change (SSC), waveform length (WL), mean frequency (MNF), median frequency (MDF), power spectral density (PSD), short-time Fourier transform (STFT) and wavelet transform (WT). Feature selection has been used to reduce the dimensionality of the data after feature extraction, which lowers computing costs and the risk of overfitting. This increases the accuracy of the classifiers. Principal component analysis (PCA) and independent component analysis (ICA) are two dimensionality reduction approaches used for EMG signal classification. Moreover, in the present study, decision trees (DT), random forests (RF) and support vector machines (SVM) have been used as classifiers. After the data has been preprocessed, the RMS value is obtained as the best-performing feature among all the other features. Then, the stress (RMS value) is analyzed for workout (warm-up), i.e., exercise and without workout cases. The stress value for a workout (warm-up) is found to be less than that without a workout case. Moreover, individuals taking up warming-up exercises proved to be giving more accuracy (79%) with respect to their contemporaries (70%). So, in order to improve the accuracy of the workout (exercise), dimensionality reduction techniques (PCA and ICA) have been applied. The dataset is divided into training, testing and validation at a ratio of 70:20:10 respectively. Better accuracy (99.2%) is obtained when PCA is used as a dimensionality reduction technique, and SVM is used as a classifier. The proposed technique attains 99.2% accuracy with RMS as feature extraction, PCA as a dimensionality reduction and SVM as a classifier.