EEG-based Seizure Detection Using Generative Model and Deep Learning

It is essential to automatically distinguish epileptic seizure forms of electroencephalography (EEG) signals in assisting in clinical diagnosis and monitoring of epilepsy. We found there is a special phenomenon that limits the ability to accurately identify epileptic forms. Some non-epileptic EEG signals show similar features to that of epileptic ones. We call such epileptic-like abnormalities boundary-type signals. Considering the existence of such type of signals, in this work a novel workflow employing deep learning is put forward for an automated seizure detection and epilepsy diagnosis. To distinguish the boundary type prior to the classification, a generative model, deep convolutional generative adversarial network, is employed, which is trained to find a representative time frequency diagram for the non-epileptic data, so as to distinguish more implicit boundary-type signals. Next, a structure-modified VGG16, a deep convolutional neural network is trained to detect normal and seizure classes. Tested on the Bonn University database, the proposed method achieved good performance.