Cross-domain self-supervised EEG feature learning network based on collaborative attention

This study addresses challenges in epileptic detection, including limited single-scale feature representation, scarcity of EEG labeled data, and poor generalization performance due to distribution differences across epileptic EEG datasets. We propose a collaborative attention-based cross-domain self-supervised EEG feature learning network (CD-SSLANet). The model first establishes a self-supervised pre-training module that designs tasks across temporal, channel, and spectral dimensions to model dynamic temporal dependencies, multi-electrode spatial correlations, and pathologically sensitive frequency band features. A parameter-sharing encoder learns unified features with strong generalization capabilities, effectively mitigating overfitting caused by insufficient labeled data. Building on this foundation, a multi-branch convolution structure extracts multi-scale local features to enhance the model's ability to capture information at different granularities. Additionally, a collaborative attention module composed of deformable attention and channel attention is introduced to achieve adaptive focus on key time points and recalibration of channel features, thereby optimizing feature discriminability through temporal and spatial coordination. Experimental results on three public epileptic EEG datasets (CHB-MIT, Seina, and TUSZ) demonstrate that CD-SSLANet significantly outperforms existing mainstream methods in accuracy, F1 score, sensitivity, precision, and AUC. This validates its strong generalization performance and stability across different acquisition scenarios, providing a reliable deep learning solution for precise epileptic seizure detection.