Abstract
Epileptic seizure detection from electroencephalography (EEG) remains challenging due to the non-stationary, multi-channel, and patient-specific nature of ictal activity. Handcrafted spectral features and fixed-basis wavelet transforms fail to fully capture the irregular, multi-scale oscillatory patterns that characterize seizure onset, while wavelet scattering transforms, although stable to time-warping deformations, are typically applied directly to the raw multi-channel signal, where seizure-related activity remains spectrally entangled with background brain activity and inter-channel redundancy is not explicitly modeled. This paper introduces an attention-guided Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)–wavelet scattering framework that first decomposes each EEG channel into intrinsic mode functions (IMFs), applies a wavelet scattering network independently to each retained mode, and aggregates the resulting representations through a dual-level learned attention mechanism operating jointly over intrinsic mode functions and EEG channels. Our approach achieves an overall accuracy of 99.68% and an F1-score of 98.91% on the CHB-MIT dataset, with a cross-dataset generalization accuracy of 99.10% when training on CHB-MIT and testing on the Bonn University dataset. Robustness experiments further show that performance remains above 96% accuracy under a 50% channel reduction and above 95% accuracy under moderate additive noise.