Multimodal Integration: Constraining MEG Localization with EEG and fMRI

Abstract

I review recent methodological developments for multimodal integration of MEG, EEG and fMRI data within a Parametric Empirical Bayesian framework [1]. More specifically, I describe two ways to incorporate multimodal data during distributed MEG/EEG source reconstruction under linear Gaussian assumptions: 1) the simultaneous inversion of EEG and MEG data using a common generative model [2], and 2) the addition of spatial priors from fMRI data when inverting MEG or EEG data [3]. In the former, the addition of EEG data was shown to increase the conditional precision of source estimates relative to MEG alone; in the latter, the inclusion of each suprathreshold cluster in the fMRI data as a separate spatial prior was shown to increase the Bayesian model evidence for MEG and EEG reconstruction. The former is an example of “symmetric” integration, or “fusion”, in which a single generative model of all data modalities is inverted; the latter is an example of “asymmetric” integration, in which the data from one modality is used to inform inversion of another. I will conclude by considering whether symmetric fusion of MEG/EEG and fMRI data is worthwhile.