Optimized ICA-based removal of ocular EEG artifacts from free viewing experiments

Abstract

Combining EEG with eye-tracking is a promising approach to study neural correlates of natural vision, but the resulting recordings are also heavily contaminated by activity of the eye balls, eye lids, and extraocular muscles. While Independent Component Analysis (ICA) is commonly used to suppress these ocular artifacts, its performance under free viewing conditions has not been systematically evaluated and many published findings display residual artifacts. Here I evaluated and optimized ICA-based correction for two tasks with unconstrained eye movements: visual search in images and sentence reading. In a first step, four parameters of the ICA pipeline were systematically varied: the (1) high-pass and (2) low-pass filter applied to the training data, (3) the proportion of training data containing myogenic saccadic spike potentials (SP), and (4) the threshold for eye tracker-based component rejection. In a second step, the eye-tracker was used to objectively quantify correction quality of each ICA solution, both in terms of undercorrection (residual artifacts) and overcorrection (removal of neurogenic activity). As a benchmark, results were compared to those obtained with an alternative spatial filter, Multiple Source Eye Correction (MSEC). With commonly used settings, Infomax ICA not only left artifacts in the data of both tasks, but also distorted neurogenic activity during eye movement-free intervals. However, correction could be drastically improved by training the ICA on optimally filtered data in which SPs were massively overweighted. With optimized procedures, ICA removed virtually all artifacts, including the SP and its associated spectral broadband artifact, with little distortion of neural activity. It also outperformed MSEC in terms of SP correction. Matlab code is provided.\n\nAuthor NoteI would like to acknowledge Maarten De Schuymer, who conducted early explorations of the effects of high-pass filtering that helped to initiate the current work. I am also grateful to Lisa Spiering for assisting with the MSEC correction and Werner Sommer for providing an excellent working environment. Collection of one of the datasets was supported by a grant from DFG (FG-868-A2). Comments or corrections are highly appreciated.