One-Class SVM Using Domain-Based sEMG Features for Detection of Neuromuscular Disorders

Abstract

Neuromuscular Disorders (NMDs) are a group of conditions that affect nerves and muscles leading to muscle weakness, atrophy, and sometimes paralysis. The effects of these conditions cannot be reversed, causing permanent impairment if not diagnosed timely. Recent medical advancements have led to timely detection and successful prevention of many diseases with portable healthcare solutions backed by strong machine learning algorithms. A similar approach for NMDs can be designed using non-invasive Surface Electromyography (sEMG) signals. While machine learning approaches have been developed for detecting NMDs, generalized features and a lack of comprehensive datasets prevent the evaluation of the robustness of these methods. Thus, a systematic methodology that incorporates features representing NMD biomarkers and one-class support vector machine for unsupervised detection of neuromuscular disorders is proposed. The method is implemented on electromyography analysis of Human Activity - Database 1 (EMAHA DB-1) which is a comprehensive dataset for hand gesture recognition based on Indian population. For robust evaluation of the algorithm, functions are designed to simulate anomalies based on NMD bio-signals' knowledge by generating an augmented dataset that mimics the nature of NMDs. Multiple machine learning models are implemented and evaluated on the real healthy dataset as well as simulated NMDs dataset. One-class SVM emerged successful amongst other models with an accuracy of 81.45%, averaged over separately trained models for each limb action. © 2024 IEEE.