Enhancing sEMG contaminant classification accuracy through statistical feature analysis and machine learning: Enabling improved sEMG signal quality

Abstract

Can the most prevalent contaminants in Electromyography (EMG) signals be accurately identified and classified during an EMG examination? EMG with applications are highly affected by signal contaminants, decreasing efficiency and accuracy. The thesis aims in particular to assist home users and non-technical hospital personnel in detecting EMG contaminations, and in the future, guide how to remove contaminants without filtering. This thesis presents two machine learning models for EMG contaminant classification: a single-label model and a multi-label model. A comprehensive feature evaluation was conducted to identify signal features that could differentiate signals of different contaminations and signal-to-noise ratios. Combining time and frequency domain features improved the ability to distinguish between different contaminants at various signal-to-noise ratios, enhancing overall classification accuracy. EMG signals from public databases were artificially contaminated with the three most common EMG contaminants: Electrocardiography (ECG) interference, Motion Artefact (MA) and White Gaussian Noise (WGN). Among several machine learning algorithms, a Random Forest model type achieved the highest accuracy. Two different models, one single-label model and one multi-label model, provide the possibility to either detect the most prominent or all present contaminants. The models utilise five-second segments to classify the EMG signal, allowing for quick feedback on possible contaminations. While offline performance was strong, online validation revealed challenges related to signal variability and generalisation. Still, the approach demonstrates promising potential for classification of EMG signal contaminants in clinical settings, providing the opportunity to improve the quality of EMG signals during examinations or when used with an assistive technology, improving the user’s quality of life.