Directed functional connectivity using dynamic graphical models

Abstract

There are a growing number of neuroimaging methods that model spatio-temporal patterns of brain activity to allow more meaningful characterizations of brain networks. This paper proposes dynamic graphical models (DGMs) for dynamic, directed functional connectivity. DGMs are a multivariate graphical model with time-varying coefficients that describe instantaneous directed relationships between nodes. A further benefit of DGMs is that networks may contain loops and that large networks can be estimated. We use network simulations and human resting-state fMRI (N = 500) to investigate the validity and reliability of the estimated networks. We simulate systematic lags of the hemodynamic response at different brain regions to investigate how these lags potentially bias directionality estimates. In the presence of such lag confounds (0.4–0.8 s offset between connected nodes), our method has a sensitivity of 72%–77% to detect the true direction. Stronger lag confounds have reduced sensitivity, but do not increase false positives (i.e., directionality estimates of the opposite direction). In human resting-state fMRI, the default mode network has consistent influence on the cerebellar, the limbic and the auditory/temporal networks. We also show a consistent reciprocal relationship between the visual medial and visual lateral network. Finally, we apply the method in a small mouse fMRI sample and discover a highly plausible relationship between areas in the hippocampus feeding into the cingulate cortex. We provide a computationally efficient implementation of DGM as a free software package for R.

Differences in directed functional brain connectivity related to age, sex and mental health

“…Dual regression was applied to estimate specific spatial maps and corresponding time series from the group ICA for each subject (Beckmann & Smith, 2004; Filippini et al, 2009). As Schwab et al (2018) showed consistency in dFC between rsfMRI sessions, we included data from all 4 rsfMRI sessions in our analysis and as such dual regression was applied on the subjects that had completed all four rsfMRI sessions.…”

“…From the HCP data release, four subjects were excluded due to missing information about mean relative motion and 15 individuals were excluded due to missing information in cognitive or mental health data, yielding data from a total of 984 individuals aged 22-37 years (mean: 28.7 years, sd: 3.71 years, 52.8% females) for the analysis on all HCP subjects. Out of those, data from 495 individuals were not included by Schwab et al (2018) and were included for an additional replication analysis (mean: 28.6 years, sd: 3.72 years, 49.5% females).…”

“…For both HCP and UKB sample, we accessed the time series of decompositions performed with 25 independent components. In each sample, we chose ten resting-state networks (RSNs) that had the highest spatial correlation with the ten RSNs reported by Smith et al (2009), and in line with the procedure used in Schwab et al (2018). These RSNs comprised default mode (DMN), cerebellar (Cer), visual occipital (VO), visual medial (VM), visual lateral (VL), right frontoparietal (FPR), left frontoparietal (FPL), sensorimotor (SM), auditory (Au), and executive control (Ex) networks.…”

“…Dynamic graphical models (DGM) is a form of Dynamic Bayesian Networks, which describes the instantaneous directed relationships between nodes (Bilmes, 2010; Schwab et al, 2018). From this, one can study the spatiotemporal arrangement of links in the network, defined here as the directionality between a node pair.…”

“…This statistical method can give a meaningful characterization of the dynamic connectivity between network nodes. Initial implementation and validation of the approach in resting-state functional magnetic resonance (rsfMRI) data from mice (N=16) and humans (N=500) (Schwab et al, 2018) suggested consistent default mode network (DMN) influence on cerebellar, limbic and auditory/temporal networks, in addition to a stable mutual relationship between visual medial (VM) and visual lateral (VL) networks in human rsfMRI. Here, we aimed to replicate these findings using independent data from the Human Connectome Project (HCP, (Van Essen et al, 2013)) and the UK Biobank (Sudlow et al, 2015).…”