Figure 5

Canonical networks are significantly more temporally independent than random cortical subsystems. The prior analysis suggests that the constitution of brain networks may change over time as parts of those networks decohere or modularise. Given this information, the utility of the canonical network architecture in a dynamic framework is unclear. We tested this by asking the following statistical question: given the dynamic states among a selection of nodes, how well do these dynamic states explain dynamic variations in global (whole-brain) connectivity? If local changes in canonical brain networks (such as the DMN) largely recapitulate global changes across the entire connectome, then canonical networks are perhaps not as useful as whole-brain states as an organizational framework for dynamic connectivity. If, by contrast, the connectivity of canonical brain networks changes in a manner that is relatively independent of the rest of the connectome, then network-level metrics could provide an important complement to global metrics. To assess these possibilities, we examined how well each canonical brain network accounted for dynamic variations in the whole brain, by calculating the within-cluster sum-of-squares error (see methods for details), as illustrated in (A). This metric takes the whole-brain connectivity at each time point (grey windows at left), identifies the nearest whole-brain context among a network’s NC-states (depicted by the arrow pointing to the small square, the color of which signifies a different NC-state and the grey boundary its corresponding whole-brain context), and calculates a distance or error measure (indicated by the horizontal bar plot). Summing across all time points provides a measure of how well a canonical network – given its NC-states – explains variation in the whole brain, or conversely, how informationally independent the network is – given its NC-states – from the rest of the brain. A null distribution was built by applying the same approach to random pseudo-networks. (B) The WCSS approach demonstrated that canonical networks (colored circles) were more independent than random pseudo-networks (grey distribution plot, background) (p < 0.05 for all except SOM, p = 0.05 for SOM). (C) A separate metric, the mean Bayesian concordance, recapitulated the results obtained using the WCSS error. Here, greater concordance corresponds to more interdependence. Each canonical network (colored bars) was compared against a null distribution of pseudo-networks (black bars). With one exception, canonical brain networks were significantly more independent (p < 0.01).