Independent Channel Attention and Cross-Subject Data Generation for EEG-Based Patient-Independent Epileptic Seizure Detection Using ConvLSTM.

Epilepsy is a common chronic neurological disease. As a non-invasive detection method, scalp EEG plays an important role in the diagnosis and continuous management of epilepsy. Currently, EEG interpretation is mainly done by experienced physicians through visual inspection, which is labor-intensive and time-consuming. Most of recent studies were patient-specific, but inter-patient variability has a great impact on the detection results. This paper proposes a multi-stage training strategy for cross-subject data generation and constructs a ConvLSTM network with an independent attention module, which greatly improves the detection performance on unseen subjects. The proposed model was evaluated on the public dataset CHB-MIT from two levels: segment-based and event-based epileptic seizure detection. The accuracy and F1-score of segment-based detection was 0.7957 and 0.7824 respectively, and the recall and false alarm rate of event-based detection was 0.9660 and 3.07/h respectively, which outperforms compared state-of-the-art deep learning models.