SEViT: STFT Enhanced Vision Transformer for Robust Epileptic Seizure Monitoring

This is the authors' abstract. We don't add key points for this paper.

Monitoring seizures is a critical challenge for caregivers and individuals in managing refractory epilepsy. Several methods have proposed the use of Electroencephalograph (EEG) signals to detect or predict epileptic seizures to aid in seizure monitoring and personalizing . In this study, a novel neural network architecture inspired from Vision Transformers (ViT) is outlined that achieves state-of-art performance on cross-patient three-stage (i.e. ictal, preictal and inter-ictal) EEG classification problem. The proposed method uses carefully-ablated STFT (Short-term Fourier Transform) based embeddings to extract rich spatial-temporal representations from raw EEG segments. The proposed method outperforms existing benchmark works on single randomized test as well as 10-fold cross-validation tests. Several performance metrics indicating overall accuracy, class-wise balance, clinical relevance and inference latency have been reported on the CHB-MIT dataset to enable comparative evaluation. The proposed deep neural architecture can be used for the development of robust seizure detection and seizure prediction systems.