BiXCA-DualFormer: A time-frequency fusion model for epileptic seizure prediction

Epileptic seizure prediction based on electroencephalography (EEG) is important for developing early-warning technologies and improving patient safety. However, reliable prediction remains challenging because preictal EEG changes are subtle and non-stationary, with discriminative information distributed across both temporal and spectral domains. The purpose of this study is to develop a seizure prediction framework that can jointly model complementary temporal and spectral EEG information and capture their interactions. To this end, we propose BiXCA-DualFormer, a dual-branch time-frequency fusion model. Raw EEG signals and short-time Fourier transform representations are processed through separate temporal and spectral branches to extract complementary waveform and frequency-domain features. The main methodological contribution of the proposed model is the introduction of bidirectional cross-attention to explicitly exchange information between the two branches, followed by a Transformer encoder and learnable query pooling for global dependency modeling and feature aggregation. Experiments on the CHB-MIT dataset showed that the proposed model achieved 95.29% accuracy, 96.40% sensitivity, 95.05% specificity, and a false positive rate of 0.0495. Ablation analysis indicated that time-frequency interaction and global contextual modeling improved sensitivity and discriminative performance. These results suggest that BiXCA-DualFormer provides an effective end-to-end biomedical signal processing framework for EEG-based seizure prediction.