Abstract
Epilepsy, characterized by recurrent seizures, is one of the most common neurological illnesses impacting people across all age groups. It is associated with atypical electrical activity in the brain.
This study proposes a novel approach for epileptic seizure detection from EEG signals using a statistical feature extraction method being derived from a cascaded Histogram of Oriented Gradients (HOG) and Gray Level Co-occurrence Matrix (GLCM) techniques for 117 normal and 117 abnormal diagnosed EEG signal images collected from Erbil Teaching Hospital. Four classification algorithms namely—Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Discriminator (DR)— with rigorous hyperparameter optimization using Bayesian techniques were utilized to improve classification: three feature extraction approaches: HOG, GLCM, and cascaded HOG-GLCM, based statistical features extraction, were calculated. The proposed comprehensive simulation results revealed that the cascaded HOG-GLCM approach significantly outperforms single-feature methods. The SVM and KNN classifiers achieved exceptional performance with the cascaded features, both approximately reaching 98.57% accuracy ensuring almost no epileptic events went undetected, which represents a substantial improvement over GLCM (best: 92.86% accuracy) and HOG approaches (best: 94.29% accuracy). The synergistic effect observed between gradient-based and texture-based features demonstrates how HOG captures directional patterns characteristic of seizure activity, while GLCM extracts spatial relationships within the signal. Neither feature type alone provides sufficient discriminative power, as evidenced by the 5-8% accuracy drop in single-feature approaches.
