Dynamic Spatio-Temporal Fusion Network Via Hierarchical Self-Attention for Seizure Prediction

The use of deep learning for EEG-based seizure prediction has grown rapidly in recent years. However, existing studies fail to effectively encode spatial variations under temporal dynamics. This limitation impedes the modeling of complex spatiotemporal evolutionary patterns, consequently leading to suboptimal performance. To address this issue, we propose a hierarchical self-attention-based dynamic spatiotemporal fusion network (HSA-DSTF Net). The hierarchical self-attention backbone first extracts multi-scale spatial features from EEG time-frequency representations, which are then reconstructed into spatiotemporal sequences with multiple time steps and fed into a Convolutional LSTM-based fusion network to model dynamic spatial variations over long temporal horizons. Moreover, residual connections are introduced into the fusion process to enrich contextual information and reduce spatiotemporal feature loss. The efficacy of the proposed methodology is validated through comprehensive experiments conducted on two public datasets: CHB-MIT and Kaggle. The HSA-DSTF Net attained an AUC of 0.949, sensitivity of 96.5%, and false positive rate of 0.025/h on CHB-MIT, and an AUC of 0.838, sensitivity of 90.0%, and false positive rate of 0.018/h on Kaggle. These findings establish that the proposed methodology significantly outperforms mainstream approaches in EEG-based seizure prediction.