UCLA

Biological processes often produce signals detectable by multiple means, but with entirely different information content. For both research and clinical applications concurrent multimodality data collection plays an important role in understanding the signal sources. Electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) are particularly interesting examples, in that each offers largely independent yet complementary information on neuronal activity. While we and others have made great strides in making concurrent EEG-fMRI recordings possible, the EEG data, in particular, still contain signal artifacts of cardiac origin (ballistocardiogram) that make EEG analysis difficult, or even impossible.

To date, no satisfying means to separate brain EEG signal and ballistocardiogram (BCG) exist especially for non-event-related-potential experiments and under 3-T MR scanner. The BCG presents high temporal non-stationarity due to variation in cardiac cycles, and its amplitude scales with magnetic field strength. This explains the considerable variation of success levels among studies, with more successful applications achieved at lower field strength. Previously published methods used one of blind source separation methods to remove the BCG. All such blind source separation approaches are limited to performing component extraction based on the contaminated data alone, agnostic of the structural difference between BCG and EEG. Another kind of approach is to utilize reference signals for the artifact itself. However, this requires purpose-built hardware and exploits no further denoising step besides a simple subtraction.

We have developed three algorithms to separate EEG signal and BCG artifacts. Firstly, we have designed a Direct Recording - Prior Encoding (DRPE) method to maximally incorporate prior knowledge of BCG/EEG subspaces described by bases learned from a modified recording configuration, and of the group sparsity characteristics in the signal. To further promote subspace separability, a Direct Recording Joint Incoherent Basis (DRJIB) method is proposed to learn a representative and sparse set of BCG and EEG bases by minimizing a cost function consisting of group sparsity penalties for automatic dimension selection and an energy term for encouraging incoherence. Reconstruction is subsequently obtained by fitting the contaminated data to a generative model using the learned bases subject to regularization. The third algorithm takes advantage of currently available high-density EEG cap, to reliably estimate the full-scalp BCG contribution from a near-optimal small subset (20 out of 256) of channels and a corresponding weight through our modified experimental setup using Orthogonal Matching Pursuit (OMP).

We show in carefully constructed simulations that the residual artifacts are reduced by several orders of magnitude to a tiny fraction of the true signal. In human studies we show that the methods work effectively, and validate our quantitative results. Beyond the application to the EEG-fMRI challenge, we expect that our algorithmic methods will have impact in many other domains where signal and contaminant have distinct information structures. Digital signal, imaging, or even financial data are also contaminated with noises so that studying and characterizing the information structures of desired and undesired signals would greatly improve the modeling power.