Does an interpretable framework using entropy and Gram Matrix features in Beta-Gamma bands improve automated epileptic seizure detection accuracy in EEG datasets?
A lightweight, interpretable feature extraction framework using entropy and Gram Matrix modeling achieves up to 100% accuracy in automated epileptic seizure detection without relying on deep learning.
Epileptic seizure detection using EEG is vital for clinical diagnosis. This paper proposes a novel, interpretable framework for automated seizure classification based on entropy features from Beta (12–30 Hz) and Gamma (30–60 Hz) bands. The Discrete Legendre Transform (DLT) decomposes EEG signals, enabling extraction of four entropy measures: Approximate, Permutation, Sample, and Fuzzy Entropy. To capture inter-feature dependencies, Gram Matrix modeling is applied, yielding four global descriptors (trace, Frobenius norm, max eigenvalue, determinant), fused into a 12-dimensional feature vector. Evaluated with SVM and ProCRC classifiers on two datasets, Bonn (interictal D vs. ictal E) and independent Hauz Khas (HK, interictal vs. ictal), the method achieves 100% accuracy on Bonn and 97.5% on HK under 60–40 stratified split, and 99.5% (Bonn) and 97% (HK) under 10-fold cross-validation. Performance surpasses multiple state-of-the-art methods, including deep and hybrid models, while remaining fully interpretable, lightweight, and free of deep learning, image conversion, or GPU dependency. The framework’s transparency, computational frugality, and cross-dataset robustness make it highly suitable for portable, real-world clinical deployment in resource-limited settings. This work confirms that high diagnostic accuracy and interpretability are not mutually exclusive in seizure detection.
Salari et al. (Wed,) studied this question.