Cross-Frequency Coupling-Enhanced Graph-Transformer Network for Epileptic EEG Classification

This is the authors' abstract. We don't add key points for this paper.

This study presents a cross-frequency coupling (CFC)-enhanced Graph-Transformer Network for multichannel EEG seizure/non-seizure classification. The architecture combines PAC/PPC descriptors, a residual convolutional encoder, learnable graph attention over post-CNN temporal nodes, and a lightweight Transformer encoder. Evaluation on the CHB-MIT scalp EEG database comprises a random window split and case-wise five-fold cross-validation. The case-wise protocol yielded 67.16 ± 5.58% accuracy, 0.617 ± 0.214 recall, 0.630 ± 0.132 F1-score, and 0.748 ± 0.067 AUROC. Its mean AUROC was numerically higher than those of protocol-matched CNN-BiLSTM and EEGNet-like baselines (0.731 and 0.693, respectively). The random 70%/15%/15% split yielded 99.38% accuracy and 99.93% AUROC; an overlap diagnostic found shared raw samples in 91.0% of test windows, defining these scores as within-protocol reference results. Random-window ablation rankings are supplementary observations; the held-out-case comparisons provide the principal comparative results. CFC sensitivity analysis found an EMG-proxy-adjusted theta-to-low-gamma PAC association with class labels, while the circular-shift surrogate-corrected contrast was not significant. Continuous-event evaluation yielded 1.000 sensitivity and 22.39 false alarms per hour, and zero-shot Siena evaluation yielded 47.73% balanced accuracy and 0.427 AUROC. The study provides an integrated EEG representation framework and a protocol-specific performance profile. The five-fold results concern held-out case identifiers, with chb01 and chb21 retained in separate folds; they are distinguished from subject-grouped evaluation.