Machine learning algorithmic classification of surface electromyography (sEMG) of biceps & triceps agonist/antagonist activity

Surface electromyography (sEMG) signals are a powerful non – invasive tool for analyzing and assessing muscle activity, contributing in movement intention detection and muscle recruitment order. Muscle state classification is a complex procedure, especially for biceps and triceps which have an agonist – antagonist relationship, due to crosstalk, noise, artifacts and variability but also due to the limitations that accompany the traditional classification methods. This thesis aims to test and evaluate machine learning models in classifying sEMG signals of biceps & triceps working as agonist/antagonist pair. The sEMG data include recordings from 100 participants. The data went through a preprocessing stage which included dc artifacts removal, bandpass filtering, notch filter application, rectification, smoothing and normalization. From the feature extraction process, features retrieved from both Time – Domain (TDF) and Frequency Domain (FDF). The machine learning models of k-Nearest Neighbors (kNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) were used for the classification task. The XGBoost model achieved the highest accuracy of 97.68%, while RF and kNN achieved 97.60% and 92.85% respectively. Key features in this process were the Shape Factor (SF) and the Root Mean Square (RMS). The results prove that machine learning offers a valid classification of sEMG signals with applications in rehabilitation, exoskeleton control and prosthetics. Future improvements will include the real – time application of the classifiers as long as applications in wearables for Human – Computer Interfaces (HCI).