Interpretable Multi-Branch Architecture for Spatiotemporal Neural Networks and Its Application in Seizure Prediction

Currently, spatiotemporal convolutional neural networks (CNNs) for electroencephalogram (EEG) signals have emerged as promising tools for seizure prediction (SP), which explore the spatiotemporal biomarkers in an epileptic brain. Generally, these CNNs capture spatiotemporal features at single spectral resolution. However, epileptiform EEG signals contain irregular neural oscillations of different frequencies in different brain regions. Therefore, it may be underperforming and uninterpretable for the CNNs without capturing complex spectral properties sufficiently. This study proposed a novel interpretable multi-branch architecture for spatiotemporal CNNs, namely MultiSincNet. On the one hand, the MultiSincNet could directly show the frequency boundaries using the interpretable sinc-convolution layers. On the other hand, it could extract and integrate multiple spatiotemporal features across varying spectral resolutions using parallel branches. Moreover, we also constructed a post-hoc explanation technique for multi-branch CNNs, using the first- order Taylor expansion and chain rule based on the multivariate composite function, which demonstrates the crucial spatiotemporal features learned by the proposed multi-branch spatiotemporal CNN. When combined with the optimal MultiSincNet, ShallowConvNet, DeepConvNet, and EEGWaveNet had significantly improved the subject-specific performance on most metrics. Specifically, the optimal MultiSincNet significantly increased the average accuracy, sensitivity, specificity, binary F1-score, weighted F1-score, and AUC of EEGWaveNet by about 7%, 8%, 7%, 8%, 7%, and 7%, respectively. Besides, the visualization results showed that the optimal model mainly extracts the spectral energy difference from the high gamma band focalized to specific spatial areas as the dominant spatiotemporal EEG feature.