Deep learning and eye-tracking for accurate EOG rejection

Abstract

Electroencephalography (EEG) is a neuroimaging technique used to record the electrical activity generated by the brain. EEG recordings are often contaminated by various artifacts, notably those caused by eye movements and blinks (EOG artifacts). Independent component analysis (ICA) is commonly applied to isolate EOG artifacts and subtract the corresponding independent components from the EEG signals. However, ICA is an unsupervised technique that contains no knowl- edge of the eye movements during the task or the generative process by which these movements result in EOG artifacts. It is generally difficult to assess whether subtracting EOG components estimated through ICA removes some neurogenic activity. To address this limitation, we developed a deep learning model for EOG artifact removal that exploits information about eye movements available through eye-tracking. We leveraged the Large Grid task from the open-source EEGEyeNet dataset to develop and validate this approach. In this task, 30 participants looked at a series of dots appearing at 25 predetermined positions on the screen (810 trials/participant). EEG and eye-tracking were collected simultaneously. In this paper, we show that we can train a long short-term memory (LSTM) model to predict the component of EEG signals predictable from eye-tracking data. We further used this eye-tracking-informed evaluation of EOG artifacts to investigate the sensitivity and specificity of ICA, currently the dominant approach for EOG artifact correction. Our analysis indicates that although ICA is very sensitive to EOG, it has a comparatively low specificity.