Abstract

Previously, we introduced the use of individual cortical location and orientation constraints in the spatiotemporal Bayesian dipole analysis setting proposed by Jun et al. ([2005]; Neuroimage 28:84-98). However, the model's performance was limited by slow convergence and multimodality of the numerically estimated posterior distribution. In this paper, we present an intuitive way to exploit functional magnetic resonance imaging (fMRI) data in the Markov chain Monte Carlo sampling -based inverse estimation of magnetoencephalographic (MEG) data. We used simulated MEG and fMRI data to show that the convergence and localization accuracy of the method is significantly improved with the help of fMRI-guided proposal distributions. We further demonstrate, using an identical visual stimulation paradigm in both fMRI and MEG, the usefulness of this type of automated approach when investigating activation patterns with several spatially close and temporally overlapping sources. Theoretically, the MEG inverse estimates are not biased and should yield the same results even without fMRI information, however, in practice the multimodality of the posterior distribution causes problems due to the limited mixing properties of the sampler. On this account, the algorithm acts perhaps more as a stochastic optimizer than enables a full Bayesian posterior analysis.