Abstract
The electroencephalogram (EEG) provides a direct measure of brain electrical activity but is typically contaminated by artifacts, most notably those arising from eye movements. Such artifacts are often removed using blind source separation techniques such as independent component analysis (ICA). However, it remains challenging to determine whether subtracting eye-movement-related components identified by techniques like ICA inadvertently removes neurogenic activity. Addressing this issue is critical to prevent bias in EEG analyses. To this end, we developed complementary deep learning and biophysical modeling approaches for the deconfounding of eye-movement artifacts, leveraging eye-tracking (ET) information. Using a multimodal, open-access dataset, we trained a deep learning model within-recording to predict the component of the EEG explainable from ET. In parallel, we employed a realistic head model to simulate artifacts generated by eye movements. Using this framework, we can distinguish neural and nonneural correlates of eye movements from neural activity not associated with eye movements. Further, this ET-informed framework enabled us to evaluate the sensitivity and specificity of techniques used to isolate and remove eye-movement artifacts. These advances provide a solid framework for future work aiming to further disentangle neural components using tailored experimental paradigms, for example, by dissociating eye movements and visual feedback through virtual reality.