Investigating Causality Between Interacting Brain Areas with Multivariate Autoregressive Models of MEG Sensor Data

In this work, we investigate the feasibility to estimating causal interactions between brain regions based on multivariate autoregressive models (MAR models) fitted to magnetoencephalographic (MEG) sensor measurements. We first demonstrate the theoretical feasibility of estimating source level causal interactions after projection of the sensor-level model coefficients onto the locations of the neural sources. Next, we show with simulated MEG data that causality, as measured by partial directed coherence (PDC), can be correctly reconstructed if the locations of the interacting brain areas are known. We further demonstrate, if a very large number of brain voxels is considered as potential activation sources, that PDC as a measure to reconstruct causal interactions is less accurate. In such case the MAR model coefficients alone contain meaningful causality information. The proposed method overcomes the problems of model nonrobustness and large computation times encountered during causality analysis by existing methods. These methods first project MEG sensor time-series onto a large number of brain locations after which the MAR model is built on this large number of source-level time-series. Instead, through this work, we demonstrate that by building the MAR model on the sensor-level and then projecting only the MAR coefficients in source space, the true casual pathways are recovered even when a very large number of locations are considered as sources. The main contribution of this work is that by this methodology entire brain causality maps can be efficiently derived without any a priori selection of regions of interest.