Domain-Adaptive Fusion of Wavelet-Entropy and wPLI Graph Attention for Cross-Dataset EEG Seizure Detection

Cross-dataset EEG seizure detection remains challenging because models trained on one dataset often degrade under differences in patient populations, acquisition settings, and background EEG distributions. This paper proposes a domain-adaptive dual-stream attention-fusion framework for bidirectional seizure detection between the CHB-MIT and Siena Scalp EEG datasets. The local stream combines wavelet-entropy features with channel-attention CNN learning, while the global stream models inter-channel functional connectivity using weighted Phase Lag Index graphs and an edge-aware graph attention network. Their embeddings are integrated through sample-adaptive attention fusion, and a domain-adversarial network with a gradient reversal layer uses unlabeled target-adaptation data to reduce source-target mismatch. Under patient-disjoint evaluation, the framework achieves 92.84% accuracy, 90.75% F1-score, and 92.96% AUC for CHB-MIT→Siena, and 91.34%, 89.42%, and 91.18% for Siena→CHB-MIT. Domain adaptation improves AUC by 2.60 and 2.85 percentage points, while post-hoc domain-probe accuracy decreases from 94.1% to 55.3%, indicating reduced dataset separability while preserving seizure-class structure. These findings support complementary local-global representation learning and adversarial alignment for robust cross-dataset EEG seizure detection.