Multiscale Convolutional Attention Network for Epileptic Seizure Prediction

Timely and accurate prediction of epileptic seizures allows healthcare professionals to implement interventions promptly prior to a seizure, preventing secondary injuries. Many previous deep learning network models have achieved varying degrees of success in seizure prediction. However, most of the existing prediction models do not fully extract and integrate the temporal and spatial feature information in multichannel electroencephalogram (EEG). To address this issue, this study proposes a multiscale convolutional attention network for epileptic seizure prediction. Specifically, the network consists of a weighted attention mechanism and a multiscale convolutional module, which are designed respectively for highlighting the spatial information in important EEG channels and capturing both the local and the global temporal features. Tested on the CHB-MIT scalp EEG dataset, the seizure prediction method based on the multiscale convolutional attention network achieved an average sensitivity of 97.5%, an average false prediction rate of 0.30/h, and an average prediction time of 26.04 minutes. The results of multiple experiments further validate the feature extraction capability to multichannel EEG and the seizure prediction performance of the multiscale convolutional attention network.