Source space effective connectivity analysis of MEG/EEG data using Kalman filter based time-varying Granger causality

    In recent years, Granger causality analysis (GCA) has emerged as a primary tool for studying effective connectivity in neural systems. In this work we describe a processing strategy that addresses three major limitations of mainstream GCA techniques: (1) reliance on sensor space analyses that fail to fully exploit the localization literature, (2) the use of unrealistically small sets of factors that make analyses vulnerable to spurious correlation, and (3) failure to address the stochastic nature of neural processes.

Our approach relies on the spatiotemporal resolution provided by source space reconstructions of cortical activity based on MRI constrained MEG/EEG. These data provide sufficient spatial resolution (roughly on the order of a Brodmann area) to allow the functional interpretation of local activation based on the pathology and BOLD imaging literatures, and sufficient temporal resolution to provide enough time-points to do meaningful statistical time-series modeling and prediction. This approach offers less spatial resolution than more invasive measures such as intracortial EEG, but they allow inclusion of a more exhaustive set of ROIs – a fundamental requirement of Granger analysis that is required to avoid artifacts related to spurious correlation (Gow et al., 2009a).

We identify ROIs based on vertex by vertex measures of gamma phase locking to a reference point identified based on early and sustained activation during the period of interest.  Gamma phase locking is measured using the sdSPM wavelet technique developed by Lin et al. (2004).  Gamma phase locking provides both a measure of functional connectivity, and evidence of a mechanism for linking activation in disparate brain regions.

Using a liberal inclusion criterion to avoid eliminating potentially causal ROIs, this approach typically identifies a large set (>25 ) of ROIs in relatively simple language tasks. This large number of ROIs is a challenge because it typically makes windowing techniques used to achieve local stationarity unviable due to concerns about overfitting. We address this challenge by deriving time varying vector autogression coefficients using a Kalman filter technique developed by Milde et al. (2010). Unlike standard VAR techniques, Kalman filter modeling explicitly models brain activation as a stochastic process.

The viability of this approach is illustrated through discussion of previous work, which explores top-down articulatory and lexical processes in speech perception (Gow et al. 2008; 2009b). 

REFERENCES

Gow, D.W., Keller, C.J., Eskandar, E., Meng, N.,& Cash, S.S. (2009a). Parallel versus serial processing dependencies in the perisylvian speech network: A                         Granger analysis of intracranial EEG data.  Brain and Language, 110, 43-48.

Gow, D.W. & Segawa, A. (2009b).  Articulatory mediation of speech perception: A causal analysis of multi-modal imaging data. Cognition, 110, 222-236.

Gow, D.W., Segawa, J.A., Alfhors, S., & Lin, F-H. (2008). Lexical influences on speech perception: A Granger causality  analysis of MEG and EEG source estimates.          NeuroImage, 43,614-623.

Lin, F.-H., Witzel, T., Hämäläinen, M.S., Dale, A.M., Belliveau, J.W., & Stufflerbeam, S.M. (2004). Spectral spatiotemporal imaging of cortical oscillations and                       interactions in the human brain. NeuroImage, 23, 582-595.

Milde, T., Leistratz, L., Astolfi, L., Miltner, W.H..R., Weiss, T., Babiloni, F., & Witte, H. (2010). A new Kalman filter approach for the estimation of high-dimensional                 time-variant multivariate AR models and its application in the analysis of laser-evoked brain potentials NeuroImage, 50, 3, 960-969.