Automatic Seizure Classification Based on Domain-Invariant Deep Representation of EEG

Abstract

Accurate identification of the type of seizure is very important for the treatment plan and drug prescription of epileptic patients. Artificial intelligence has shown considerable potential in the fields of automated EEG analysis and seizure classification. However, the highly personalized representation of epileptic seizures in EEG has led to many research results that are not satisfactory in clinical applications. In order to improve the clinical adaptability of the algorithm, this paper proposes an adversarial learning-driven domain-invariant deep feature representation method, which enables the hybrid deep networks (HDN) to reliably identify seizure types. In the train phase, we first use the labeled multi-lead EEG short samples to train squeeze-and-excitation networks (SENet) to extract short-term features, and then use the compressed samples to train the long short-term memory networks (LSTM) to extract long-time features and construct a classifier. In the inference phase, we first adjust the feature mapping of LSTM through the adversarial learning between LSTM and clustering subnet so that the EEG of the target patient and the EEG in the database obey the same distribution in the deep feature space. Finally, we use the adjusted classifier to identify the type of seizure. Experiments were carried out based on the TUH EEG Seizure Corpus and CHB-MIT seizure database. The experimental results show that the proposed domain adaptive deep feature representation improves the classification accuracy of the hybrid deep model in the target set by 5%. It is of great significance for the clinical application of EEG automatic analysis equipment.