An Approach for Detection of Epileptic Seizures from EEGs Considering Ictal-like Non-Ictal Signals

In the clinical diagnosis and treatment of epilepsy, it is a necessity to detect epileptic seizures automatically and accurately from the electroencephalography (EEG) data instead of manual detection. In most studies on this issue, the problem is simply regarded as a binary classification problem to distinguish ictal and non-ictal signals. However, EEG signals actually show characteristics of multimodal distribution, which should not be ignored in the detection. With consideration of the existence of such signals, a new workflow employing a k nearest neighbor algorithm and a generative adversarial network (GAN) is presented. The latter is developed to discern the ictal-like non-ictal EEGs before training the classifier. The GAN is used to find more implicit boundary-type samples. Consider the current task, a 16-layer CNN, which is the type most used to identify patterns in images and video, is transferred here to solve the problems of insufficient training data, long training time, and insufficient accuracy. Evaluated by a public EEG dataset, it proved effective to distinguish boundary signals prior to the classification of ictal and non-ictal signals.