Low-Rank + Sparse Decomposition (LR+SD) for EEG Artifact Removal

Abstract

Concurrent EEG-fMRI recordings are advantageous over serial recordings, as they offer the ability to explore the relationship between both signals without the compounded effects of nonstationarity in the brain. Nonetheless, analysis of simultaneous recordings is challenging given that a number of noise sources are introduced into the EEG signal even after MR gradient artifact removal with balistocardiogram artifact being highly prominent. Here, we present an algorithm for automatically removing residual noise sources from the EEG signal in a single process using low rank + sparse decomposition (LR+SD). We apply this method to both experimental and simulated EEG data, where in the latter case the true EEG signature is known. The experimental data consisted of EEG data collected concurrently with fMRI (EEG-fMRI) as well as alone outside the scanning environment while subjects viewed Gabor flashes, a perceptual task known to produce event related power diminutions in the alpha spectral band. On the simulated data, the LR+SD method was able to recover the pure EEG signal and separate it from artifact with up to three EEG sources. On the experimental data, LR+SD was able to recover the diminution in alpha spectral power that follows light flashes in concurrent EEG-fMRI data, which was not detectable prior to artifact removal. At the group level, we found that the signal-to-noise ratio was increased ~34\% following LR+SD cleaning, as compared independent component analysis (ICA) in concurrently collected EEG-fMRI data. We anticipate that this method will be highly useful for analyzing simultaneously collected EEG-fMRI data, and downstream for exploring the coupling between these two modalities.