Source-linked AI summary
The Dynamics of Functional Brain Networks: Integrated Network States during Cognitive Function
James M. Shine, Patrick G. Bissett, Peter T. Bell, Oluwasanmi Koyejo, Joshua H. Balsters, Krzysztof J. Gorgolewski, Craig A. Moodie, Russell A. Poldrack
TL;DR
It is unclear how brain network integration unfolds over time during cognition. The study maps time-varying network structure and finds that greater integration relates to faster and more effective cognitive performance.
Problem
It is unclear how fluctuations in the brain’s network structure unfold over time.
Method
The study uses a novel analysis technique to map spatiotemporal dynamics in human brain network structure.
Results
Greater brain integration during the N-back task correlated with faster drift rate and shorter non-decision time.
Takeaways & Limitations
The findings suggest that integration relates to fast and effective cognitive performance.
Takeaways & Limitations
The role of network topology in cognition requires further exploration.
Abstract
from arXiv · showhide
Higher brain function relies upon the ability to flexibly integrate information across specialized communities of brain regions, however it is unclear how this mechanism manifests over time. In this study, we use time-resolved network analysis of functional magnetic resonance imaging data to demonstrate that the human brain traverses between two functional states that maximize either segregation into tight-knit communities or integration across otherwise disparate neural regions. The integrated state enables faster and more accurate performance on a cognitive task, and is associated with dilations in pupil diameter, suggesting that ascending neuromodulatory systems may govern the transition between these alternative modes of brain function. Our data confirm a direct link between cognitive performance and the dynamic reorganization of the network structure of the brain.
Results · Fluctuations in Network Cartography
Resting-state brain networks fluctuated aperiodically between integrated and segregated states identified from whole-brain cartographic profiles. The integrated state dominated rest and showed greater inter-modular communication and global efficiency, whereas the segregated state showed greater modularity.
- Fluctuations in Network Cartography: A novel analysis classified whole-brain cartographic profiles into two temporal states without assigning regions to predefined cartographic classes.Subject-level k-means clustering used k = 2 and remained stable at higher k values.
- Fluctuations in Network Cartography: The resting brain fluctuated aperiodically between Integration and Segregation, spending 70.32 ± 1.4% of the rest session in the integrated state.The two states represented alternative whole-brain cartographic extremes characterized by integration or segregation.
- Fluctuations in Network Cartography: Most group-level fluctuations involved inter-modular connectivity, while individual parcels also showed window-to-window changes in intra-modular connectivity.Inter-modular connectivity transitioned between high and low states en masse, whereas intra-modular connectivity fluctuated within individual parcels.
- Fluctuations in Network Cartography: The integrated state produced a global increase in inter-modular communication across the brain, significant for all 375 individual parcels.This pattern was also reflected in network-wide measures of integration and segregation.
- Fluctuations in Network Cartography: Modularity was higher in the segregated state than the integrated state (QS = 0.55 ± 0.1 vs. QI = 0.42 ± 0.2; Cohen’s d = 0.9; p = 10-11).The reported modularity difference indicates stronger segregation in the segregated state.
- Fluctuations in Network Cartography: Global efficiency was higher in the integrated state than the segregated state (ES = 0.18 ± 0.03 vs. EI = 0.24 ± 0.05; Cohen’s d = 1.5; p = 10-8).The shift toward integration was most prominent in sensory and attentional networks.
- Fluctuations in Network Cartography: The segregated state showed relatively higher participation within default mode regions, whereas the integrated state emphasized sensory and attentional networks.These regional patterns suggested that cartographic profiles may track changing engagement of attention and cognition over time.
- Fluctuations in Network Cartography: Global topology fluctuations were independent of mean framewise displacement, nuisance signals, and the number of modules estimated per temporal window.The reported correlations were mean r = 0.01 ± 0.01 for displacement, mean r = -0.02 ± 0.01 for nuisance signals, and mean r = 0.03 ± 0.01 for module number.
Task-‐‑based Alterations in the Cartographic Profile
During the cognitively challenging N-back task, the brain shifted toward greater global integration relative to rest, with reconfiguration varying according to cognitive demands. This shift was strongest in frontoparietal, default mode, and subcortical regions and supported communication across specialist regions.
- N-back task: Cartographic fluctuations closely tracked N-back task blocks, and integration remained correlated with the task regressor after controlling for global signal and mean connectivity.The controlled correlations were mean r = 0.452 ± 0.21; p = 10^-10 and mean r = 0.393 ± 0.14; p = 10^-9.
- N-back task: All 375 regions showed a significant shift toward greater inter-modular connectivity (BT) during N-back compared with rest (FDR p < 0.05).The global shift toward integration was most pronounced in frontoparietal, default mode, and subcortical regions.
- N-back task: Task-related involvement of highly interconnected regions likely facilitated communication between otherwise isolated specialist regions, expanding potential responses to cognitive challenges.These regions were identified as frontoparietal, subcortical, and other highly interconnected hub regions.
- Task complexity: Other HCP tasks also increased global integration relative to rest, but less than N-back, particularly compared with the relatively simple Motor task.In N-back, 88.8% of parcels showed higher BT than in the Motor task (FDR p < 0.05).
- Task complexity: The extent of network reconfiguration during task performance varied as a function of cognitive demands.This was quantified by estimating each task’s affine transformation along the BT axis relative to rest.
Investigating the Relationship Between Cartography and Behavior · Network Cartography Fluctuates with Pupil Diameter
Greater global network integration was associated with faster and more effective cognitive processing, while fluctuations toward integration tracked pupil diameter, implicating ascending neuromodulatory input in dynamic network reorganization. These relationships were replicated across cohorts and localized most strongly to frontoparietal and subcortical regions.
- Investigating the Relationship Between Cartography and Behavior: The study fit an EZ-diffusion model to each subject’s response-time distributions and accuracy on cognitively challenging 2-back trials.The model estimated drift rate (v), non-decision time (t), and boundary separation (a).
- Investigating the Relationship Between Cartography and Behavior: Greater global network integration correlated positively with drift rate, inversely with non-decision time, and not at all with the boundary threshold.These patterns matched the predicted links between integration and information-processing speed while remaining independent of response caution.
- Investigating the Relationship Between Cartography and Behavior: The drift-rate and non-decision-time relationships replicated in a separate cohort of 92 subjects.Replication was observed in both the Discovery and Replication cohorts.
- Investigating the Relationship Between Cartography and Behavior: The associations between cognitive function and integration were most pronounced across frontoparietal and subcortical regions at FDR p < 0.05.Together, the findings support a globally efficient, integrated architecture for fast, effective computation throughout the cognitive processing stream.
- Network Cartography Fluctuates with Pupil Diameter: In a separate resting-state dataset of 14 individuals, pupil diameter was compared with the cartographic profile using a 10-TR window.The dataset had TR = 2s, 3.5mm3 voxels, and 204 volumes.
- Network Cartography Fluctuates with Pupil Diameter: Pupil diameter positively correlated with mean between-module connectivity, with group mean r = 0.241 +/-‐‑ 0.06; R2 = 0.06; p = 10-‐‑5.The result indicates that global fluctuations in network structure were related to dynamic ascending neuromodulatory input to cortex and subcortex.
- Network Cartography Fluctuates with Pupil Diameter: Across the 14-subject cohort, pupil diameter showed a positive relationship with network-level integration at FDR p = 0.05.This relationship linked pupillometry with the integrated end of the network cartographic profile.
Identifying Regions Related to Global Integration · Reproducibility · Discussion
The study identifies distributed right-lateralized regions associated with integrated network topology, replicates its time-resolved findings across sessions, cohorts, scanners, and protocols, and links global integration to cognitive performance and pupil-linked network dynamics. It also highlights limitations concerning neuromodulatory mechanisms, task specificity, and metric robustness.
- Identifying Regions Related to Global Integration: A right-lateralized network of frontal, parietal, thalamic, and striatal regions showed consistently elevated BT, whereas visual cortex and insula showed elevated WT.Together, these regions form a distributed network implicated in computational integration during effective cognitive processing.
- Discussion: The brain fluctuated between segregated and integrated network topologies, with integration increasing during more demanding cognitive tasks and correlating with faster drift rate and shorter non-decision time.Integration within the functional connectome also correlated with increased pupil diameter, a marker of arousal and behavioral engagement.
- Discussion: Right-lateralized frontoparietal and subcortical regions mediated the effects of integration on cognitive function, while pupil-linked fluctuations suggested a possible role for ascending neuromodulatory systems.The locus coeruleus may modulate these fluctuations through its influence on pupil diameter, but this mechanism remains inferential.
- Discussion: System-wide alterations in network topology were associated with more effective behavioral performance, extending metastability accounts that balance specialized processing with global coordination.The study demonstrates fluctuations in network topology consistent with metastability that relate to effective behavioral performance.
- Discussion: The relationship between global integration and cognition remains incompletely characterized because the N-back task may not capture dissociable components such as updating, set-shifting, and response inhibition.Future studies should test trial-by-trial performance and interactions between endogenous and exogenous attentional systems.
- Discussion: Direct links between ascending neuromodulatory input and network topology require confirmation, and the robustness of topology fluctuations across time-sensitive connectivity metrics remains unresolved.The authors suggest electrophysiological measures and further comparisons across connectivity-estimation techniques.
- Discussion: Global integration was closely related to N-back cognitive function by catalyzing communication between otherwise segregated specialist regions.The authors propose global integration as a candidate mechanism for complex brain networks and adaptive behavior.
Materials and Methods · Data acquisition · Data pre-‐‑processing
The study used minimally preprocessed resting-state fMRI data from HCP discovery and replication cohorts plus an out-of-sample NKI Rockland cohort. Preprocessing included motion and artifact control, nuisance regression, and temporal filtering, with exclusions for insufficient coverage or excessive motion-related artifacts.
- Data acquisition: 100 unrelated HCP participants provided the primary discovery resting-state fMRI dataset.Participants had a mean age of 29.5 years, and 55% were female.
- Data acquisition: Each participant underwent 14 minutes 30 seconds of resting-state acquisition using multiband gradient-echo echoplanar imaging.Acquisition used TR = 720 ms, echo time = 33.1 ms, multiband factor = 8, and 2x2x2 isotropic voxels.
- Data pre-‐‑processing: HCP data underwent bias-field and motion correction, and the first 100 time points were discarded because of scanner noise-related auditory signals.Motion correction used 12 linear degrees of freedom with FSL’s FLIRT.
- Data pre-‐‑processing: NKI scans were realigned for head motion and registered to each participant’s T1-weighted image and the MNI152 atlas.Registration used boundary-based registration and Advanced Normalization Tools, followed by manual inspection.
- Data pre-‐‑processing: 11 [6.3%] of 173 original participants were discarded after inspection because of insufficient anatomical coverage.Affected regions included orbitofrontal cortex and temporopolar cortex.
Brain parcellation · Multiplication of temporal derivatives · Time-‐‑resolved functional connectivity
The study represented whole-brain activity with 375 regions and estimated time-resolved functional connectivity using Multiplication of Temporal Derivatives (MTD). Connectivity was computed in 14-point windows, while the parcellation and covariance estimates carry stated methodological limitations.
- Brain parcellation: The analysis extracted mean time series from 375 predefined regions of interest to provide whole-brain coverage.These comprised 333 cortical, 14 subcortical, and 28 cerebellar regions.
- Brain parcellation: The parcellation was selected to interrogate temporal fluctuations in network architecture but does not necessarily reflect a ground-truth organization.Functional divisions may differ across subjects, and subcortical subdivisions may vary in network involvement over time.
- Multiplication of temporal derivatives: MTD estimated time-resolved connectivity by multiplying pairwise temporal derivatives and averaging the products over a moving temporal window.The method was used to improve temporal resolution relative to conventional sliding-window Pearson correlation while reducing high-frequency noise.
- Multiplication of temporal derivatives: The MTD calculation produced connectivity estimates between pairs of regions across the course of each time series.The computation used first temporal derivatives, their standard deviations, and a simple moving-average window length.
- Time-‐‑resolved functional connectivity: Time-resolved functional connectivity was calculated between all 375 brain regions using MTD within a 14-time-point sliding window.The window corresponded to 10.1 seconds for HCP data and 16 time points, approximately 10.4 seconds, for NKI data.
- Time-‐‑resolved functional connectivity: Each temporal window generated an unthresholded, signed, weighted three-dimensional functional connectivity matrix.The 14-point window was chosen to track slow cortical fluctuations and estimate signals at approximately 0.1 Hz.
- Time-‐‑resolved functional connectivity: The effects of small-sample noise on MTD covariance estimates remain unclear, although analyses were reliable and replicable across multiple datasets.MTD is more sensitive to covariance than connectivity, and covariance has been reported as a more reliable BOLD coupling marker.
Time-‐‑ resolved community structure · Time-‐‑ resolved hub structure
Time-resolved community structure was estimated from multilayer functional connectivity using Louvain modularity, while participation coefficients characterized each region’s cross-module hub connectivity. These measures quantified modular organization, within-module strength, and the distribution of connections across modules over time.
- Time-‐‑ resolved community structure: Louvain modularity was combined with the MTD to estimate both time-averaged and time-resolved community structure.The algorithm iteratively maximized Q to identify community assignments.
- Time-‐‑ resolved community structure: Modularity quantified the extent to which networks separated into communities with stronger within-module than between-module connectivity.Q measured the network’s subdivision into communities.
- Time-‐‑ resolved community structure: Graph measures used weighted and signed connectivity matrices, avoiding arbitrary thresholding.The modularity formulation incorporated positive and negative connections and defined community membership through δ_MiMj.
- Time-‐‑ resolved community structure: For each temporal window, community assignments were assessed 500 times and summarized with a consensus partition.This procedure produced time-resolved modularity (Q_T) and cluster assignments (C_iT) for each participant.
- Time-‐‑ resolved community structure: Within-module connectivity was estimated for each region using the time-resolved module-degree Z-score, W_iT.W_iT was based on a region’s connectivity to other regions in its module relative to the module’s mean and standard deviation.
- Time-‐‑ resolved hub structure: The participation coefficient, B_T, quantified the extent to which each region connected across all modules.B_T was calculated within each temporal window using Equation 4 and characterized hubs within brain networks.
- Time-‐‑ resolved hub structure: The participation coefficient used the strengths of positive connections from each region to regions in each module.κ_isT represented connections to module s, while κ_iT represented the sum of all positive connections at time T.
- Time-‐‑ resolved hub structure: A participation coefficient near 1 indicated uniformly distributed connections across modules, whereas 0 indicated links confined to the region’s own module.Thus, the coefficient distinguished broadly distributed cross-module connectivity from within-module connectivity.
Cartographic profiling
A novel cartographic-profile analysis tracked time-varying brain topology without predefined cartographic labels and identified two robust states. These states reflected significant fluctuations in topology, with distinct regional between-module connectivity patterns that were not explained by graph-density changes.
- Cartographic profiling: A novel cartographic profile tracked temporal fluctuations using joint histograms of within- and between-module connectivity without predefined cartographic class labels.Temporal windows were clustered from these histograms using k-means with k = 2.
- Cartographic profiling: 0.400 ± 0.02 mean mutual information showed that partitions across k = 2–20 were strongly similar to the k = 2 solution.Principal component analysis further associated the first two components with the integrated and segregated states.
- Cartographic profiling: 20.2 ± 1.4% variance was associated with the integrated state, whereas 4.9 ± 2.3% of variance was associated with the segregated state.Together, these results supported k = 2 as a relatively natural clustering pattern in the data.
- Cartographic profiling: 16.1 ± 1.1% of discovery-cohort temporal windows exceeded the 95th percentile of the VAR null model, indicating significant dynamic fluctuations in topology.Significant fluctuations along the BT axis remained after correcting for ongoing changes in the number of modules.
- Cartographic profiling: All 375 parcels demonstrated higher BT in the integrated state, whereas none showed significantly different WT between states at FDR p < 0.05.The two states were matched on graph density, indicating that BT fluctuations did not simply reflect changes in network sparsity.
Task-‐‑based alterations in the cartographic profile
Task-based functional connectivity was assessed across seven fMRI paradigms in 92 subjects using time-resolved cartographic profiling of 375 brain regions. The mean profile shifted rightward during 2-back relative to 0-back, and task-related mean BT associations persisted after controlling for global signal and mean MTD.
- Analysis: Time-resolved connectivity was computed from 375 regional time series after regressing task-evoked activity, using MTD windows of 14 TRs, approximately 10 seconds.The residual data were then analyzed with cartographic profiling, paralleling the resting-state analysis.
- Task-related profile shifts: The mean cartographic profile deviated rightward during 2-back compared with 0-back blocks.Task and rest were compared using two-sided one-sample t-tests across subjects with FDR p < 0.05.
- Robustness analyses: The relationship between the task regressor and mean BT survived separate regression of global signal and mean MTD.For global-signal regression, the reported mean correlation was r = 0.452 ± 0.21.
Investigating the Relationship Between Cartography and Behavior
The study used EZ-diffusion modeling of 2-back performance to examine relationships between the brain’s cartographic profile and behavioral outcomes. Integration showed a similar relationship with median reaction time and accuracy across both cohorts.
- Modeling cartography and behavior: EZ-diffusion modeling used mean reaction time, reaction-time variance, and accuracy to estimate drift rate, boundary separation, and non-decision time.These are the three main diffusion-model parameters.
- Modeling cartography and behavior: The analysis correlated each task’s mean joint-histogram bins with drift rate, non-decision time, and boundary threshold.No boundary-threshold bins survived multiple-comparisons correction.
- Modeling cartography and behavior: The diffusion model was fit to 2-back blocks because many participants made no errors in the 0-back condition, preventing parameter estimation.The 0-back data therefore could not support fitting the drift-diffusion parameters for those participants.
- Behavioral relationships: Integration showed a similar relationship with median reaction time and accuracy in both cohorts.The study also compared the cartographic profile with median reaction time and accuracy.
- Modeling cartography and behavior: The researchers found no evidence of contaminant trials, including extremely short responses below 300 msec, and proceeded with EZ-diffusion modeling.The few extremely short responses that existed were statistically above chance.
Network Cartography Fluctuates with Pupil Diameter · Identifying Regions Related to Global Integration · Replication analysis
Pupil-linked fluctuations in network cartography were quantified in a separate resting-state dataset, while conjunction analysis identified parcels jointly related to network measures and behavioral or pupil outcomes. The findings replicated strongly across sessions, cohorts, and datasets, including behavioral relationships with drift rate and non-decision time but not diffusion boundary.
- Network Cartography Fluctuates with Pupil Diameter: Pupil regressors were derived by preprocessing and down-sampling pupillometric data to 0.5 Hz, then convolving each run’s pupil vector with an informed basis set.This yielded three pupil regressors of interest per participant.
- Network Cartography Fluctuates with Pupil Diameter: The mean pupil regressors correlated positively with cartographic profiles across temporal windows in 14 subjects, with mean correlation r = 0.241 ± 0.06.One-sample tests assessed whether cartographic-profile bins and the relationship between mean BT and pupil diameter differed from zero.
- Identifying Regions Related to Global Integration: Parcel-wise conjunction analysis identified regions where BT and WT were significantly related to drift rate, non-decision time, and pupil diameter above chance.The analyses used FDR p < 0.05 and were conducted separately for WT and BT before surface projection.
- Replication analysis: Across sessions, graph measures and mean cartographic profiles showed strong positive correlations, with all statistical tests yielding p < 0.001.Session-level correlations were graph measures – 𝑟!! = 0.982, 𝑟!! = 0.957; mean cartographic profiles 𝑟!"#$ = 0.982.
- Replication analysis: In 92 unrelated HCP participants, graph measures and mean cartographic profiles replicated positively, with 𝑟!! = 0.971, 𝑟!! = 0.967, and 𝑟!"#$ = 0.973.The same fluctuations observed in HCP data were also present in the NKI dataset.
- Replication analysis: Behavioral relationships replicated across discovery and replication datasets for drift rate and non-decision time, but not diffusion boundary.Drift rate showed r = 0.613; R2 = 0.37; p = 10^-11, and non-decision time showed r = 0.681; R2 = 0.46; p = 10^-15; diffusion boundary had p > 0.500.