Abstract
Background: Patient-specific seizure prediction remains challenging because preictal EEG patterns and channel relevance vary substantially across individuals. Methods: We propose a CNN-GCN-Diffusion Network (CGD-Net) integrating channel-wise temporal encoding, montage-informed graph modeling, bag-level aggregation, and diffusion-prior-guided soft channel gating. The patient-specific channel prior was estimated exclusively from training data. CGD-Net was evaluated on 19 CHB-MIT recording cases using seizure-group-aware chronological training, validation, and test partitions. Robustness was assessed using random seeds 42, 52, and 62. Results: Across the three seeds, mean accuracy ranged from 74.84% to 75.87%, and mean AUC ranged from 81.82% to 82.01%. In the seed-42 component comparison, the full model achieved the highest mean accuracy, AUC, F1-score, and specificity. Recording-case-level Wilcoxon tests with Holm correction showed significant improvements in accuracy, AUC, and F1-score over the No bag, No diffusion, and No prior configurations, whereas differences from Random graph were not significant after correction. Channel-importance patterns were highly stable across seeds, with a mean pairwise Spearman correlation of 0.9819 and a mean Top-5 Jaccard index of 0.8730, while hub importance showed moderate patient-dependent stability. Conclusions: These results support the combined use of temporal aggregation, montage-informed graph modeling, and prior-guided channel modulation for retrospective preictal-versus-interictal classification. The explanations represent model-level attributions rather than clinically validated seizure-onset regions. Event-level sensitivity and false-alarm rates were not evaluated, and continuous-recording validation is required before clinical warning-system deployment.