CLASSIFICATION OF EMG SIGNALS OF EYE MOVEMENT USING MACHINE LEARNING TECHNIQUES

Abstract

This approach aims to classify electromyography (EMG) signals from the extraocular muscles into six distinct eye movement classes: Blink, Normal Behavior, Left, Right, Downward, and Upward Movement and to apply in medical applications. The dataset used in this study consisted of two types of signal values: one obtained from horizontally connected electrodes and the other from vertically connected electrodes. We explored both signal types individually but found that the classification accuracy was lower when using the vertically connected electrodes, and determined by correlation between variables and scattering plot. To process the data, windowing technique was employed. This technique involves dividing the preprocessed data stream into smaller segments, or windows, to analyze and extract features. A total of 28 features were calculated from the preprocessed dataset, forming a feature matrix that also included the corresponding class labels. To ensure that the training process did not lead to overfitting, the rows of the feature matrix were randomized. We compared this approach to existing works in the literature and found that it outperformed previous methods in terms of accuracy. The evaluation of classification accuracies was performed using various classifier algorithms. Among them, the best accuracy achieved was 96.8% using the Cubic Support Vector Machine (SVM) algorithm.