Unmixing concurrent EEG-fMRI with parallel independent component analysis

Abstract

Concurrent event-related EEG-fMRI recordings pick up volume-conducted and hemodynamically convoluted signals from latent neural sources that are spatially and temporally mixed across the brain, i.e. the observed data in both modalities represent multiple, simultaneously active, regionally overlapping neuronal mass responses. This mixing process decreases the sensitivity of voxel-by-voxel prediction of hemodynamic activation by the EEG when multiple sources contribute to either the predictor and/or the response variables. In order to address this problem, we used independent component analysis (ICA) to recover maps from the fMRI and timecourses from the EEG, and matched these components across the modalities by correlating their trial-to-trial modulation. The analysis was implemented as a group-level ICA that extracts a single set of components from the data and directly allows for population inferences about consistently expressed function-relevant spatiotemporal responses. We illustrate the utility of this method by extracting a previously undetected but relevant EEG-fMRI component from a concurrent auditory target detection experiment.

Introduction

Processing of simple stimuli and tasks produces spatially and temporally extensive event-related neuronal responses in the brain. For example, auditory target detection induces hemodynamic activation in about fourty cortical, subcortical and cerebellar regions (Kiehl et al., 2005), complementing the results from intracranial recordings (Baudena et al., 1995, Halgren et al., 1995a, Halgren et al., 1995b). These neuronal mass responses can be observed across scales and modalities from single unit recordings, intracranial and scalp electrophysiology, as well as metabolic and hemodynamic signals, but no single technique provides a sufficient view of the full temporal, spatial and functional extent of these responses. Visibility can be improved with techniques that integrate data across different neuroimaging modalities (Debener et al., 2006, Hopfinger et al., 2005, Horwitz and Poeppel, 2002, Makeig et al., 2002). In the case of concurrent EEG-fMRI recordings, one can complement the temporal resolution provided by scalp potentials with the spatial precision of fMRI. This can be done for example by finding correlations between single-trial modulation at a selected latency in the event-related EEG and activation in the fMRI volume employing mass univariate voxel-by-voxel analysis (Bénar et al., 2007, Debener et al., 2005b, Eichele et al., 2005). Implicit in this approach is the critical assumption that the scalp EEG data from a selected channel and latency can predict the fMRI activation in single voxels (Friston et al., 1995, Friston, 2005b). This is imposed by the sampling properties of the recordings, and the way fMRI time-series data are commonly analyzed. While this assumption provides a workable solution to ‘integration-by-prediction’, it is not necessarily physiologically plausible for many of the samples from both modalities. The reason for this is that a salient event can induce multiple, simultaneously active, regionally overlapping, and functionally separable responses which add to existing neuronal background activity, in other words, event-related processes are spatially and temporally mixed across the brain. The scalp EEG samples a volume-conducted, spatially degraded version of the responses, where the potential at any location and latency can be considered a mixture of multiple independent timecourses that stem from large-scale synchronous field potentials (Makeig et al., 2004a, Onton et al., 2006). Similarly, the neurovascular transformation of the distributed neuronal activity into hemodynamic signals (Lauritzen and Gold, 2003, Logothetis, 2003) affords detection of blood oxygenation level dependent responses (BOLD, Ogawa et al., 1990) that are temporally degraded and spatially mixed across the fMRI volume (Calhoun and Adali, 2006, McKeown et al., 2003).

This physiological spatiotemporal mixing process creates situations in which prediction of fMRI activity by EEG features has to contend with the fact that neither the predictor, nor the response variables are any likely to represent a single source of variability. For example, the point-to-point correlation between the two data mixtures fails when the trial-to-trial modulation in the EEG receives different contributions from several function-relevant spatially separate sources such that no single regional fMRI response represents the predicted signal. Also, this applies to the case where the EEG feature captures a single source, but the fMRI activity at corresponding locations is buried in the spread of other, unrelated sources, leading to underestimation of the spatial extent of the response. Although denoising and inclusion of parametric modulations into the stimulus paradigm (Eichele et al., 2005), and temporal unmixing of the EEG (Debener et al., 2005b) solve parts of the problem and make way for refined spatiotemporal mapping, there is still need for improvement of the analysis tools for integration of concurrent recordings (cf. Debener et al., 2006). One such improvement is to unmix both modalities in parallel at the single-trial level, which follows naturally from the recent work (Calhoun et al., 2006b, Debener et al., 2005b, Eichele et al., 2005) and the reasoning laid out above.

Following the above arguments, we develop an analysis framework for group data that employs Infomax independent component analysis (ICA, Bell and Sejnowski, 1995, Lee et al., 1999; for an overview see Stone, 2002) to recover a set of statistically independent maps from the fMRI (sICA), and independent time-courses from the EEG (tICA) separately, and match these components across modalities by correlating their trial-to-trial modulation. ICA was developed to address linear mixing problems similar to the ‘cocktail party problem’ in which many people are speaking at once and multiple microphones pick up different mixtures of the speakers' voices (Bell and Sejnowski, 1995). The algorithm used here attempts to separate mixed signals into maximally independent sources by maximization of information transfer between them. ICA has general applicability to normally distributed two-dimensional mixtures, and regarding psychophysiological data it has been used for decomposition of averaged ERPs (Makeig et al., 1997), single trial EEG (Makeig et al., 2004b, Onton et al., 2006), fMRI (Calhoun and Adali, 2006) and EEG-fMRI (Calhoun et al., 2006b, Debener et al., 2005b, Feige et al., 2005). ICA can be used for EEG-fMRI integration assuming that the different recording modalities faithfully sample features from the same set of sources, expressed in the covariation between single trials (Debener et al., 2005b) or subjects (Calhoun et al., 2006b).

Unlike univariate methods such as the general linear model, ICA is not naturally suited to generalize results from a group of subjects. There are two strategies to allow for matching of independent components across individuals: one is to combine individual ICs across subjects with clustering techniques (Esposito et al., 2005, Onton et al., 2006). Another approach is to create aggregate data containing observations from all subjects, estimate a single set of ICs and then back-reconstruct these in the individual data (Calhoun et al., 2001, Schmithorst and Holland, 2004). We adopted the latter strategy for the group EEG temporal ICA analysis, because it directly estimates components that are consistently expressed in the population, involves the least amount of user interaction and is straightforward to combine with the existing framework for group ICA of fMRI data (Calhoun et al., 2001).

In summary, possible ways for EEG-fMRI integration include predicting both modalities, a mass-univariate framework testing all voxel timeseries in the fMRI, as well as channels and timepoints in the EEG employing a pre-defined model function as is commonly done in fMRI timeseries analysis (however, to the best of our knowledge this has not yet been realized). Another option is to predict the fMRI data with the measured EEG single trial amplitudes, assuming that some EEG timepoints and channels represent functional processes in some voxels without much overlap, representing a point-to-point correlation between mixtures (Bénar et al., 2007, Eichele et al., 2005). A third solution is to unmix the EEG and predict the fMRI mixture with the modulation of a temporally independent component (Debener et al., 2005b, Feige et al., 2005). The method developed here un-mixes both modalities separately, and matches temporal ICs in the EEG with spatial ICs in the fMRI.

The utility of this method is demonstrated in previously published data that were collected in an auditory oddball with varying degrees of target predictability. The parametric modulation induced distinct EEG-correlated fMRI activation patterns at the latencies of the P2, N2, and P3 (Eichele et al., 2005; see also Jongsma et al., 2006). We have re-analyzed these data with the open search question whether systematic EEG-fMRI covariation was missed out in our previous analysis and if it could be recovered by parallel ICA. A likely candidate for such a miss is the auditory onset response and the subsequent low-level orienting/change detection processes. Although being expressed in the N1-ERP (Naatanen and Picton, 1987, Woods, 1995) and in bilateral temporal fMRI activation (Kiehl et al., 2005, Liebenthal et al., 2003, Linden et al., 1999) this process did not support a significant correlation between the modalities (cf. Eichele et al., 2005).