A Bayesian model for dynamic functional connectivity estimation in the human brain with structural priors

Studies of dynamic functional connectivity have demonstrated that anatomical linkage is related to persistent functional connectivity. Bayesian models can leverage this connection by regularizing estimates of functional connectivity according to the strength of the corresponding structural connectivity. We proposed and evaluated the ability of such a model to recover covariance matrices. The model performed well in a high dimensional, small sample simulated setting. In addition, it exhibited robustness to temporal transformations and an ability to recover simulated data generated according to both discrete and continuous temporal dynamics. Finally, it outperformed sliding window baselines and anatomically un-informed baselines on estimating instantaneous covariances according to out-of-sample log likelihood on two task datasets