Interpretable SincNet-Based Spatiotemporal Neural Network for Seizure Prediction.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Spatiotemporal convolutional neural networks (CNNs) are used for seizure prediction (SP) from electroencephalogram (EEG) signals.
• Existing CNNs face challenges in clinical application due to poor interpretability and numerous parameters.

Purpose of the Study:

• To develop an interpretable SincNet-based architecture for spatiotemporal CNNs to enhance SP performance.
• To enable direct visualization of temporal filter ranges and learned spatiotemporal features.

Main Methods:

• Proposed an interpretable SincNet-based architecture for spatiotemporal CNNs (EEGNet-8,2, ShallowConvNet, DeepConvNet, EEGWaveNet).
• Implemented a visualization analysis method to identify crucial spatiotemporal features.
• Evaluated performance on the CHB-MIT EEG dataset.

Main Results:

• ShallowConvNet and EEGWaveNet demonstrated significantly improved performance with fewer parameters.
• ShallowConvNet achieved 87.2% accuracy, 88.3% sensitivity, 87.1% F1-score, and 92.7% AUC for 21 epilepsy patients.
• Visualization confirmed the model's ability to extract significant spatiospectral energy differences in high-frequency EEG bands for SP.

Conclusions:

• The proposed interpretable SincNet-based architecture enhances SP performance and interpretability.
• The visualization method aids in understanding the learned features for seizure prediction.
• This approach offers a promising direction for developing clinically applicable AI tools for epilepsy management.