Epileptic Seizure Detection Based on Semi-supervised Generative Adversarial Network

Abstract

Since the manual diagnosis of electroencephalograph (EEG) recordings requires a lot of labor and material costs for clinical seizure detection, the annotation for seizure data is of great challenge for seizure detection. To tackle the issue of small samples of epilepsy-labeled data, we propose a semi-supervised generative adversarial network-based seizure detection method. To begin with, a Butterworth filter is used to preprocess the EEG, and the filtered EEG signal is input into the SGAN model. Finally, the output of the SGAN model is subjected to post-processing operations including averaging filtering and threshold comparison, and the discriminative result of whether the tested EEG is a seizure is output. The method achieved an average sensitivity of 90.36%, an average specificity of 93.72%, and an average accuracy of 93.72% in the CHB-MIT EEG dataset. Experiments demonstrate that the semi-supervised generative adversarial network has more accurate classification performance in epilepsy detection.