Efficient Group Cosine Convolutional Neural Network for EEG-Based Seizure Identification

Epilepsy is a common neurological disease, and its diagnosis usually depends on labor-intensive visual inspection of electroencephalogram (EEG). Although various deep learning-based seizure detection models have been investigated, their outcomes generally lack spatial information related to EEG channels, limiting their accuracy and the capability of localizing seizure onset channels. In this study, we designed a group cosine convolutional neural network (group CosCNN) for end-to-end seizure identification based on a hardware-friendly and memory-efficient cosine convolutional operator containing only two learnable parameters. The multi-channel EEG recordings were fed into the group CosCNN model for identifying seizures, where the first module performs channel-wise convolutions and subsequent modules execute group convolutions to maintain EEG spatial information. Meanwhile, an algorithm for computing the normalized channel contribution scores was introduced for realizing real-time seizure onset channel localization. Comprehensive evaluations were conducted on the publicly available CHB-MIT database and our SH-SDU database collected in clinical settings, achieving sensitivities of 97.70% and 90.51%, and specificities of 97.54% and 95.48%, respectively. Our dynamic seizure onset channel localizing strategies were further validated on the CHB-MIT database with individual-level and event-level localization accuracies of 91.30% and 85.12%, respectively. These outstanding results demonstrated the superior efficacy of the proposed group CosCNN for seizure identification.