Impact of Feature Selection Techniques for EEG-Based Seizure Classification

Abstract

A neurological condition called epilepsy can result in a variety of seizures. Seizures differ from person to person. It is frequently diagnosed with fMRI, magnetic resonance imaging and electroencephalography (EEG). Visually evaluating the EEG activity requires a lot of time and effort, which is the usual way of analysis. As a result, an automated diagnosis approach based on machine learning was created. To effectively categorize epileptic seizure episodes using binary classification from brain-based EEG recordings, this study develops feature selection techniques using a machine learning (ML)-based random forest classification model. Ten (10) feature selection algorithms were utilized in this proposed work. The suggested method reduces the number of features by selecting only the relevant features needed to classify seizures. So to evaluate the effectiveness of the proposed model, random forest classifier is utilized. The Bonn Epilepsy dataset derived from UCI repository of Bonn University, Germany, the CHB-MIT dataset collected from the Children’s Hospital Boston and a real-time EEG dataset collected from EEG clinic Bangalore is accustomed to the proposed approach in order to determine the best feature selection method. In this case, the relief feature selection approach outperforms others, achieving the most remarkable accuracy of 90% for UCI data and 100% for both the CHB-MIT and real-time EEG datasets with a fast computing rate. According to the results, the reduction in the number of feature characteristics significantly impacts the classifier’s performance metrics, which helps to effectively categorize epileptic seizures from the brain-based EEG signals into binary classification.