Seizure Types Classification Based on Multi-branch Hybrid Deep Learning Network

Identifying different types of seizure is essential for the treatment of epilepsy and reducing the pain and risk to patients. The task of classifying seizure types is made more challenging by the fact that EEG signals change more complexly in different types of epileptic seizures. In this paper, a multi-branch hybrid deep learning network is proposed to identify different types of seizure. Firstly, EEG signals are decomposed into multiple subcomponents by applying Hilbert Vibration Decomposition (HVD), which could provide subcomponents that retain phase information and choose subcomponents with high energy as inputs for the multi-branch hybrid deep learning network. Then, we designed a multi-branch hybrid deep learning network using convolutional neural network (CNN) and Mogrifier long short-term memory (Mogrifier LSTM) to extract spatial and temporal EEG features, followed cross-branch feature fusion based on attention mechanism. This study validates the effectiveness of the proposed method in identifying three types and four types of seizure using EEG data collected from the Temple University Hospital Epileptic Seizure Corpus (TUSZ). The highest accuracy achieved in identifying three types of seizure is 98%, with an F1-score of 0.98, while for four types of seizure, the highest accuracy reaches 97%, with an F1-score of 0.97. The results show that this multi-branch hybrid network is beneficial to the characterization of EEG signals and improves classification performance.