Dual-Stage Improvement with Domain Adaptation for Cross-Subject Epileptic Seizure Prediction

Electroencephalogram (EEG)-based epileptic seizure prediction is a critical adjunctive treatment strategy for drug-resistant epilepsy. To mitigate cross-subject EEG variability and improve generalization performance, existing methods use domain adaptation (DA) to minimize distribution discrepancy between source and target domains. However, these methods overlook the joint impact of data imbalance and noisy pseudo-labels. This results in biased attention toward imbalanced source samples and degraded understanding of the true target distribution, thereby impairing the model's generalization performance. To address these issues, a cross-subject seizure prediction method based on DA with dual-stage improvement is proposed to integrate data balancing and pseudo-label optimization. Specifically, a global context-aware generative network is introduced to generate preictal samples with global contextual consistency to rectify class imbalance between preictal and interictal samples in the source domain. A common spatial pattern clustering filter is then introduced, which uses a confidence-guided covariance weighting strategy to integrate classification confidence from source-domain labels and target-domain pseudo-labels, optimizing spatial filters to maximize inter-class discrimination. To mitigate the accumulation and propagation of noisy pseudo-labels during training, a dual-filtering mechanism iteratively eliminates them. Experimental evaluation on the CHB-MIT dataset demonstrates the effectiveness of the proposed method in mitigating interpatient domain discrepancies and achieving superior performance over recent leading approaches.