Identifying cybersickness features from EEG data using deep learning

Abstract

Virtual Reality (VR) is emerging as a useful technology in a wide range of areas, such as entertainment, education, and rehabilitation. However, one major challenge hindering its rapid and widespread adoption is cybersickness, a condition characterized by symptoms similar to motion sickness, including nausea, loss of balance, and headaches. Numerous studies have been conducted to determine the cause of cybersickness using Electroencephalography (EEG), but so far, only classification of cybersickness time segments in EEG data has been attempted. None of these studies explored the underlying psychophysiological patterns that are correlated with cybersickness beyond the classification step. This thesis trained Convolutional Neural Networks (CNNs) to classify cybersick and non-cybersick EEG samples, which were intended to extract learned features from the model. The features that were extracted showed an increase in the alpha frequency power in high cybersick samples and also Event Related Potential (ERP) component patterns related to the N1, P2 and P3 components. Although the models that were trained in this thesis did not yield high classification accuracies, the best model having an accuracy of 59.97%, the work shows that CNN can be used as a tool to uncover interesting biosignal patterns hidden in the brain.