Fast Automatic Artifact Annotator for EEG Signals Using Deep Learning

Abstract

Electroencephalogram (EEG) is a widely used non-invasive brain signal acquisition technique that measures voltage fluctuations from neuron activities of the brain. EEGs are typically used to diagnose and monitor disorders such as epilepsy, sleep disorders, and brain death and also to help the advancement of various fields of science such as cognitive science, and psychology. EEG signals usually suffer from a variety of artifacts caused by eye movements, chewing, muscle movements, and electrode pops, which disrupts the diagnosis and hinders precise representation of brain activities. This paper proposes a deep learning based model to detect the presence of the artifacts and to classify the kind of the artifact to help clinicians resolve problems regarding artifacts immediately during the signal collection process. The model is optimized to map the 1-second segments of raw EEG signals to detect 4 different kinds of artifacts and the real signal. The model achieves a 5-class classification accuracy of 67.59%, and a true positive rate of 80% with a 25.82% false alarm for binary artifact classification with time-lapse. The model is lightweight and could potentially be deployed in portable machines.