Entropy-Based Graph Learning Framework for Cross-Subject Detection of Electrical Status Epilepticus During Sleep...

Electrical Status Epilepticus during Sleep (ESES) is a severe childhood epilepsy syndrome marked by sleep-induced and near-continuous epileptiform discharges, often leading to cognitive and behavioral impairments. Early and reliable detection is clinically critical but remains challenging due to the complex nonlinear EEG dynamics, strong inter-channel dependencies, and the limitations of conventional feature representations and fixed connectivity assumptions. To address these challenges, we propose GCN-AGLCA, a data-driven graph learning framework for cross-subject ESES detection, which jointly models multiband entropy-based complexity features, adaptively learns subject-specific functional connectivity graphs without predefined assumptions, and incorporates a channel attention mechanism to prioritize clinically relevant EEG channels during graph convolution. Experimental results show that combined entropy features outperform individual entropy measures and conventional time-frequency representations, achieving a classification accuracy of 84.00% and yielding more stable training and improved generalization. Furthermore, evaluated via leave-one-subject-out cross-validation on EEG recordings from 58 pediatric patients, the proposed framework attains a superior average accuracy of 77.82%, along with a lower average error rate of 6.39%, and enhanced training stability compared to the reproduced state-of-the-art baseline model. These results demonstrate that entropy-informed adaptive graph learning offers a robust and clinically interpretable approach for automated ESES identification.