Physical Action Categorization Pertaining to Certain Neurological Disorders Using Machine Learning-Based Signal Analysis

Abstract

The daily life of thousands of individuals around the globe suffers due to physical or neurological disorders that can hinder their limb movement to varying degrees in addition to other effects. Assistive applications can alleviate the effects of the disease leading to improvement of quality of life for such individuals. In such computer-aided application, classification of physical actions is required. Surface electromyography (sEMG) presents a noninvasive mechanism which helps in translating physical movement to signals which can be used in applications for different neurological disorders such as tonic-clonic seizure and epilepsy. In this article, a framework for classification of four physical actions is proposed. Successful classification of these actions can help build systems that can passively monitor epileptic patients and others and provide useful data for improving the quality of life for such patients or providing better treatment. The framework makes use of various features from various signal signatures with contribution from time domain, frequency domain, and inter-channel statistics. Next, we conducted a comparative analysis of SVM, 3-NN, and ensemble learning with accuracy of 97.405%, 95.7%, and 96.5%, respectively. Finally, we have reported the effect which combinations of different signatures have on the classification of physical activities. It was observed that we can achieve an accuracy of 97.26% using SVM classifier via a subset of 48 features. These findings can help design an algorithm for a constrained environment such as real-time processing.

Keywords Physical action classification, sEMG signals, Signal processing and analysis, Feature engineering, Machine learning