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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning for electroencephalogram (EEG)-based seizure prediction is advancing.
• Existing methods struggle to integrate spatial and temporal EEG data effectively.
• This limitation hinders the accurate modeling of complex spatiotemporal patterns in seizures.

Purpose of the Study:

• To develop a novel deep learning network for improved EEG-based seizure prediction.
• To address the limitations of existing models in capturing spatiotemporal dynamics.
• To enhance the accuracy and reliability of seizure prediction systems.

Main Methods:

• Proposed a hierarchical self-attention-based dynamic spatiotemporal fusion network (HSA-DSTF Net).
• Utilized hierarchical self-attention to extract multi-scale spatial features from EEG time-frequency representations.
• Employed a Convolutional LSTM-based fusion network with residual connections to model dynamic spatiotemporal variations.

Main Results:

• The HSA-DSTF Net achieved high performance on two public datasets (CHB-MIT and Kaggle).
• On CHB-MIT: AUC of 0.949, sensitivity of 96.5%, false positive rate of 0.025/h.
• On Kaggle: AUC of 0.838, sensitivity of 90.0%, false positive rate of 0.018/h.

Conclusions:

• The proposed HSA-DSTF Net significantly outperforms mainstream approaches for EEG-based seizure prediction.
• The network effectively captures complex spatiotemporal evolutionary patterns.
• This advancement holds promise for more accurate and reliable seizure forecasting.