EEG Anomaly Detection using Generative Adversarial Networks (GANs)

Abstract

Abnormality is the unusual behavior or feature of a sample compared to the whole population. In medical applications, the task of anomaly detection is meritorious in early or in-time diagnosis of diseases by detecting disease markers in for example different neuroimaging data such as Electroencephalogram (EEG) etc. Anomaly detection approaches include supervised, unsupervised, and semi-supervised machine learning algorithms. However unsupervised and semi-supervised approaches are more preferred due to the arduous task of labeling the data and due to the intrinsic unknown definition of anomaly. The development of applications of generative adversarial networks in recent years has made them a competent tool for anomaly detection problems. They learn (normal) data distribution and detect anything different as an anomaly. The task of EEG anomaly detection is one proposed interdisciplinary problem that can have applications in early pathology detection as a starting point for later treatment plans, or in brain-computer interface studies. In this work, we used a generative adversarial model and an image representation of EEG data to detect abnormal EEG samples. The whole pipeline and the results indicate the proper standing of this approach among peers on Temple University Hospital's (TUH) abnormal EEG corpus data set.