P091 Dynamic causal modelling (DCM) for effective connectivity in MEG High-Gamma-Activity: a data-driven Markov-Chain Monte Carlo approach P. García-Rodríguez, M. Gilson, J.-D. Lemaréchal, A. Brovelli
CNRS UMR 7289 - Aix Marseille Université, Institut de Neurosciences de la Timone, Campus Santé Timone, Marseille, France
Email: pedro.garcia-rodriguez@univ-amu.fr
Introduction
Model inversion in DCM traditionally considers the application of Bayesian variational schemes,i.e.quadratic approximations in the vicinity of minima in the parameters space [1]. On the other hand, more general Markov Chain Monte Carlo methods (MCMC) opt for an intense use of random numbers to sampling posterior probability distributions. The successful application of either of them highly depend on the correct choice of prior distributions.
Methods
Here we propose an automated workflow combining MCMC with more conventional optimization Gradient-Descent (GD) techniques. Following the bi-linear model [2], a simpler DCM is considered with a matrixAforeffective connectivity and a matrixCfor sensory driving inputs.AlphaandGammafunctions for input profiles complete the modeling scenario.
Model’s parameters are estimated in three parts.Firstly, the matrixAis initialized from a Gaussian distribution with null-mean and variance given by observer-specific or group-level Granger Causality (GC) computed from the data. Next, GD algorithms implement a constraint bounded optimization to keep input parameters within plausible (positive) intervals. The adequacy of the parameters values found are further tested through a Levenberg-Marquard GD form. Finally, a MCMC Bayesian scheme incorporates the covariance of the observation noise in a Multivariate Gaussianlikelihood model. A Generative Model is so completed with parameter's prior distributions based on the GD optimizations mentioned above. Normal or Log-Normal distributions are alternatively used, the later to ensure positive values after sampling when needed.
Results
The approach is applied to High-Gamma Activity induced responses during visuomotor transformation tasks executed by 8 subjects, as reported in [3]. Methods were applied to hundred of trials for each subject, providing a handy data-driven DCM framework to evaluate the plausibility of various model configurations. Observation noise is empirically estimated from the pre-stimulus periods in original trials. Model inversion pipeline tends to support the most realistic model configuration tested, with an apparent relation between the estimated effective connectivityAandmatrixGC(Fig. 1).
Discussion Comparison of prior and posterior distributions can help distinguish informative from non-informative parameters.Initialization of matrixAwith structural connectivity instead was tested.
Figure 1. A DCM for high-gamma-activity (HGA). First column: brain regions and model configurations tested (top) and corresponding Granger-Causality (GC) matrix (bottom). Second column: model predictions compared to experimental HGA profiles (top) and relation between GC and estimated effective connectivty matrix A (bottom). Acknowledgements A.B. and P.G-R were supported by EU’s Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreements No. 101147319 (EBRAINS 2.0 Project). References
[1] Zeidman P, Friston K, Parr T. (2023) A primer on Variational Laplace (VL). Neuroimage 279:120310. doi: 10.1016/j.neuroimage.2023.120310.[2] Chen CC, Kiebel SJ, Friston KJ (2008). Dynamic causal modelling of induced responses. Neuroimage 41(4):1293-1312. DOI: 10.1016/j.neuroimage.2008.03.026. PMID: 18485744.
[3] Brovelli A., Chicharro D., Badier J-M., Wang H., Jirsa V. (2015). Characterization of Cortical Networks and Corticocortical Functional Connectivity Mediating Arbitrary Visuomotor Mapping. J. Neuroscience 35(37):12643-12658. doi: 10.1523/JNEUROSCI.4892-14.2015.
