Abstract:
Brain connectivity in vivo is usually estimated either with generic statistical models or more biophysically oriented Neural Mass Models. We compare both approaches with simulated and real data (resting-state eyes-open fMRI). Towards this end, we leverage the Tigramite Suite of programs for information theory-based causal discovery in time series. The approach uses a PC-type algorithm to strictly search for casual links starting from different dependency measures (DM), discovering the directionality of connections with transfer entropy. As expected, simulated linear time series networks were best for the Partial Correlation DM. For generic nonlinear simulations, the best DM was CMIknn, a fully nonparametric and nonlinear test based on Conditional Mutual Information with the k nearest neighbors. Surprisingly CMIKnn performed poorly for specific neural mass models. We hypothesize that our negative results may be due to misspecification of the delays in the neural system modeled, especially regarding the need to have a distributed delay Connecotmic Tensor. A new method and toolbox to explore this type of hypothesis are presented
