Abstract <p>Ictal–Interictal–Injury Continuum (IIIC) patterns are crucial indicators of epilepsy-related disorders, where early detection is essential for diagnosis and treatment. We propose a deep-learning approach for IIIC classification from electroencephalogram (EEG). For efficient feature extraction, we compress 20 EEG leads into 4 using the Banana montage and convert signals to Mel-spectrograms to capture time–frequency–amplitude characteristics. Our architecture comprises a Symmetry-Aware Intelligent IIIC Classifier Network (SAI3CNet) with a Spectrogram-Attention (SpAtt) block and a Lead Relationship Encoder (LRE) that models inter-lead symmetry and correlations; we further adopt an Expert-Distilled EEG learning strategy to reflect expert uncertainty in ambiguous brain-wave patterns. In clinically realistic class-imbalanced settings, the method outperformed strong baselines, reducing over-prediction of the “Other” class and improving minority-pattern recognition such as LRDA while maintaining well-calibrated probabilities. These properties highlight the system’s technical robustness and potential utility as a supportive screening tool. By pre-screening continuous EEG to prioritize high-risk segments and highlighting symmetry-aware saliencies, the proposed method aims to assist specialists in rapid triage and reduce review workload, thereby potentially facilitating more timely assessments in ICU and epilepsy-monitoring workflows.</p> Graphical abstract <p></p>