A Bayesian framework for unifying data cleaning, source separation and imaging of electroencephalographic signals

Publication Type:
Journal Article
Citation:
2019
Issue Date:
2019-02-24
Full metadata record
Abstract Electroencephalographic (EEG) source imaging depends upon sophisticated signal processing algorithms for data cleaning, source separation, and localization. Typically, these problems are addressed separately using a variety of heuristics, making it difficult to systematize a methodology for extracting robust EEG source estimates on a wide range of experimental paradigms. In this paper, we propose a unifying Bayesian framework in which these apparently dissimilar problems can be understood and solved in a principled manner using a single algorithm. We explicitly model the effect of non-brain sources by augmenting the lead field matrix with a dictionary of stereotypical artifact scalp projections. We propose to populate the artifact dictionary with non-brain scalp projections obtained by running Independent Component Analysis (ICA) on an EEG database. Within a parametric empirical Bayes (PEB) framework, we use an anatomical brain atlas to parameterize a source prior distribution that encourages sparsity in the number of cortical regions. We show that, in our inversion algorithm, PEB+ (PEB with the addition of artifact modeling), the sparsity prior has the property of inducing the segregation of the cortical activity into a few maximally independent components with known anatomical support. Artifacts produced by electrooculographic and electromyographic activity as well as single-channel spikes are also segregated into their respective components. Of theoretical relevance, we use our framework to point out the connections between Infomax ICA and distributed source imaging. We use real data to demonstrate that PEB+ outperforms Infomax for source separation on short segments of data and, unlike the popular Artifact Subspace Removal algorithm, it can reduce artifacts without significantly distorting clean epochs. Finally, we analyze mobile brain/body imaging data to characterize the brain dynamics supporting heading computation during full-body rotations. In this example, we run PEB+ followed by the spectral analysis of the activity in the retrosplenial cortex, largely replicating the findings of previous experimental literature.
Please use this identifier to cite or link to this item: