Hilbert-domain sub-band feature framework for EEG-based seizure detection.

Epilepsy is a chronic neurological disease in which the brain's activity deviates from normal. The classification and analysis of EEG signals is the first step in diagnosing epilepsy. Various machine learning algorithms have been used in the past to classify epileptic EEG recordings. The proposed approach for EEG seizure detection is based on processing signals in the Hilbert domain. In this study, the EEG signal is divided into short time frames. Each frame is decomposed into sub-bands using a Butterworth Bandpass Filter, and the Hilbert Transform is applied to each sub-band of a frame. Next, three categories of features-entropy-based, spike-related, and statistical features-are extracted from each sub-band. A high-dimensional feature vector is produced by concatenating the features obtained from each sub-band. A filter-based technique, Minimum Redundancy Maximum Relevance (mRMR) feature selection method, is used to select a discriminative subset of features. The mRMR method ranks the features based on calculated weights. The proposed framework was evaluated using the publicly available University of Bonn EEG dataset, which contains five EEG classes (A-E), and the CHB-MIT scalp EEG dataset, which consists of long-term pediatric EEG recordings. Experiments conducted on the University of Bonn and Boston Children's Hospital datasets show that the proposed method achieves classification accuracies of 99.27% and 99.04%, surpassing several state-of-the-art methods with performance improvements of up to 1.43%, demonstrating its effectiveness for epileptic seizure detection. Performance evaluation was conducted using repeated 15-fold cross-validation with multiple machine learning classifiers, including Linear SVM, RBF-SVM, Random Forest (RF), Ensemble Tree (ET), and Linear Discriminant Analysis (LDA). Furthermore, Friedman ANOVA and Tukey-Kramer post hoc statistical analyses confirmed the robustness and reliability of the proposed framework.