Conceived and designed the experiments: KS VM. Performed the experiments: KS VM. Analyzed the data: KS. Contributed reagents/materials/analysis tools: VM. Wrote the paper: KS VM.
The authors have declared that no competing interests exist.
Cognitive skills undergo protracted developmental changes resulting in proficiencies that are a hallmark of human cognition. One skill that develops over time is the ability to problem solve, which in turn relies on cognitive control and attention abilities. Here we use a novel multimodal neurocognitive network-based approach combining task-related fMRI, resting-state fMRI and diffusion tensor imaging (DTI) to investigate the maturation of control processes underlying problem solving skills in 7–9 year-old children. Our analysis focused on two key neurocognitive networks implicated in a wide range of cognitive tasks including control: the insula-cingulate salience network, anchored in anterior insula (AI), ventrolateral prefrontal cortex and anterior cingulate cortex, and the fronto-parietal central executive network, anchored in dorsolateral prefrontal cortex and posterior parietal cortex (PPC). We found that, by age 9, the AI node of the salience network is a major causal hub initiating control signals during problem solving. Critically, despite stronger AI activation, the strength of causal regulatory influences from AI to the PPC node of the central executive network was significantly weaker and contributed to lower levels of behavioral performance in children compared to adults. These results were validated using two different analytic methods for estimating causal interactions in fMRI data. In parallel, DTI-based tractography revealed weaker AI-PPC structural connectivity in children. Our findings point to a crucial role of AI connectivity, and its causal cross-network influences, in the maturation of dynamic top-down control signals underlying cognitive development. Overall, our study demonstrates how a unified neurocognitive network model when combined with multimodal imaging enhances our ability to generalize beyond individual task-activated foci and provides a common framework for elucidating key features of brain and cognitive development. The quantitative approach developed is likely to be useful in investigating neurodevelopmental disorders, in which control processes are impaired, such as autism and ADHD.
The human brain undergoes significant maturational changes between childhood and adulthood that are thought to enable increasingly sophisticated thoughts and behaviors. One of the skills that we develop over time is the ability to problem solve, which relies in turn on the ability to control our attention and successfully direct our cognitive efforts. Using a novel multi-pronged neuroimaging approach, we identify for the first time the dynamic brain systems underlying the maturation of problem solving abilities. We find that the anterior insula, part of a larger network of regions previously shown to be important for salience processing and generating influential control signals, shows weaker influences over other key brain regions important for task performance in children compared to adults. In addition, structural connections between the anterior insula and other key regions were found to be weaker in children compared to adults. Importantly, measures of causal influences between key regions could be used to predict individual differences in behavioral performance. Our study is the first to show that the anterior insula, by virtue of its dynamic influences on other key brain regions, shows developmental differences in both structural and functional connectivity, which may contribute to more mature cognitive abilities in adulthood compared to childhood.
The development of increasingly sophisticated cognitive skills relies on the maturation of control processes for orienting attention and allocating resources for task relevant information
Based on experimental studies across a wide range of cognitive domains, a number of cortical areas within the frontal lobe, including the anterior cingulate cortex (ACC), ventrolateral prefrontal cortex (VLPFC), dorsolateral prefrontal cortex (DLPFC) and the fronto-insular cortex (FIC) have emerged as putative sites for implementing different aspects of control
In this study we use a theoretically motivated approach to this problem based on neurocognitive network models derived from studies of intrinsic brain connectivity. Studies in adults have shown that the human brain is intrinsically organized into distinct functional networks
Over the past decade, several studies have examined developmental changes in the recruitment of brain areas belonging to these networks using cognitive tasks ranging from response inhibition, attention, and memory, to decision-making, reasoning and problem solving
A systematic network approach has the potential for providing insights into general development mechanisms mediating dynamic control processes during cognition. However, in both adults and children, the differential role and primacy of control signals has been difficult to disentangle, partly because these areas are typically coactivated during a wide range of cognitive tasks
Children (ages 7–9) and adults (19–22) did not differ on IQ (
The two main networks of interest – SN and CEN – were identified using ICA applied to resting-state fMRI data (
We first examined fMRI responses within the five SN and CEN ROIs during the arithmetic problem solving task. Task-related brain activation was identified using a general linear model with arithmetic problem solving task versus rest/null condition contrast. Only correct trials were included in the analysis. All five right-hemisphere nodes showed significant task-related activation in both children and adults (
(
We then examined differences in functional connectivity between children and adults. Functional connectivity here is measured as instantaneous correlations between pairs of ROIs after removal of drift and physiological noise. We found that rAI connectivity with ACC, rDLPFC, and rPPC, and between the rVLPFC and rDLPFC was significantly greater in adults, compared to children (
We examined differences in the onset latency of the event-related fMRI responses in the five right hemisphere ROIs. We extracted the mean time-course in each ROI, and used a linear basis function that is a combination of the SPM canonical hemodynamic response function and a temporal derivative to fit the event related BOLD response for each subject and event, and then averaged the fitted responses across events and subjects. Onset latencies were then computed as the time point at which the slope of the fitted response reached 10% of its maximum positive (or negative) slope in the initial ascending (or descending) segment
Onset latencies in the five key nodes of the SN (blue bars), and CEN (green bars) are shown in (
(
We used two different quantitative methods to examine causal interactions in fMRI data. Based on our previous studies, we first used multivariate Granger causal analysis (MGCA)
Briefly, MGCA detects causal interactions between brain regions by assessing the relative prediction of signal changes in one brain region based on the time-course of responses in another. We performed MGCA using a multivariate model on the time-courses extracted from each of the ROIs. We used bootstrap techniques to create null distributions of influence terms (F-values) and their differences. In children, MGCA revealed statistically significant direct causal influences from the rAI to the rVLPFC, ACC, rDLPFC, and rPPC (
An identical set of analyses were conducted using MDS methods which have the advantage of modeling causal interactions in the latent “neuronal” signals, rather than in the fMRI signal itself. Furthermore, MDS also takes into account inter-regional variations in hemodynamic response in an explicit manner
To quantify the causal interactions of each node of the network, we performed graph-based network analyses. Analysis of the causal network identified with MGCA revealed that the rAI had the highest number of causal outflow connections (out-degree), the lowest number of causal inflow connections (in-degree), and the shortest path length among all regions. The rAI also had a significantly higher net causal outflow (out-in degree) than all of the other regions (
We used DTI and quantitative tractography to investigate the anatomical correlates of developmental changes in causal interactions between the rAI and rPPC. The density of fibers along the superior longitudinal fasciculus linking the rAI and rPPC was significantly lower in children compared to adults (
(
We compared the relationship between functional and structural connectivity between the rAI to rPPC in children and adults. We found that functional connectivity, measured by instantaneous temporal correlations, and structural connectivity, measured by fiber density, between the rAI to rPPC was significantly correlated in adults (
Functional interactions is correlated with structural white matter connectivity in Adults (
We used multivariate sparse regression analysis, based on GLMnet
In children, strength of directed causal influences from rAI to rDLPFC, rAI to ACC, and rAI to rPPC cumulatively predicted reaction time. The remaining seven connections were non-significant (zero) i.e. they did not contribute to the prediction of reaction times. In adults, the strength of directed causal influences from rAI to rPPC and rAI to rVLPFC cumulatively predicted reaction time. The remaining eight connections were non-significant. Comparison of model fit revealed that causal network interactions better predicted reaction time in adults (
Reaction time | Accuracy | |||||||
Predictive causal connections | R2 | Mean square error |
|
Predictive causal connections | R2 | Mean square error |
|
|
|
|
0.50 | 0.75 | <0.01 | rAI→rACC | 0.43 | 0.83 | <0.01 |
|
rAI→rPPC | |||||||
|
rAI→rVLPFC | |||||||
|
|
0.66 | 0.54 | <0.01 | rAI→rACC | 0.47 | 0.76 | <0.01 |
|
rAI→rPPC |
Relationship between cumulative directed causal influences between nodes of the salience and central executive networks and behavior assessed using a multivariate GLMnet model. rAI = right Anterior Insula, ACC = Anterior Cingular Cortex, rVLPFC = right Ventrolateral Prefrontal Cortex, rDLPFC = right Dorsolateral Prefrontal Cortex, rPPC = right Posterior Parietal Cortex. Analysis of GLMnet model fits revealed that causal network interactions better predicted reaction times and accuracy in adults than in children.
In children, strength of directed causal influences from rAI to ACC, rAI to rPPC, and rAI to rVLPFC cumulatively predicted accuracy. The remaining seven connections were non-significant. In adults, the strength of directed causal influences from rAI to rACC, and rAI to rPPC cumulatively predicted reaction time. The remaining eight connections were non-significant. Comparison of model fit revealed that causal network interactions better predicted accuracy in adults (R2 = 0.47,p<0.01; mean square error = 0.76) than in children (R2 = 0.43, p<0.01; mean square error = 0.83) (
Similar results were observed when participant age was included as an additional predictor variable. More specifically, causal network interactions outcompeted age in the prediction of reaction time and accuracy in both children and adults.
We used a novel neurocognitive network approach with multimodal imaging to investigate the maturation of control processes underlying problem solving skills in 7–9 year-old children. Task-independent rsfMRI was first used to identify prominent nodes of the insula-cingulate salience and fronto-parietal central executive networks, two neurocognitive networks implicated in fundamental control processes. Functional connectivity and dynamic causal interactions between the major nodes of these networks were examined during arithmetic problem solving using multiple analytic procedures. The strength of both directed and undirected connections between five key nodes of the SN and CEN was compared between children and adults. Both instantaneous and causal functional connectivity analysis identified the rAI as a locus of immature control signals in children. Furthermore, our study provides novel converging evidence from structural connectivity analysis for weaker white matter pathways underlying fronto-parietal (rAI→rPPC) control signals in children. Remarkably, weaker control signals were associated with lower-levels of task performance in children. Below we describe findings from our quantitative analysis of dynamic functional interactions between key brain areas of insula-cingulate salience and fronto-parietal central executive networks and discuss their implications for understanding the maturation of fundamental control processes in the developing brain
The rAI showed strong causal influences on the ACC node of the SN and the DLPFC and PPC regions of the CEN in children and adults, suggesting that the role of the rAI as a primary node that drives the CEN is established early in development. Two additional analyses were performed to confirm these findings. First, we used a novel state-space MDS model, which estimates causal interactions in latent neuronal signals, rather than the recorded fMRI signals, after taking into account inter-regional variations in hemodynamic response
Our findings further suggest a novel pattern of temporal hierarchy among prefrontal and parietal regions implicated in control processes
An important novel finding of our study is that the strength of causal influence from rAI to rPPC was significantly weaker in children, compared to adults. Notably, this group difference was observed using both MGCA and MDS, two different and complementary methods for estimating causal interactions in fMRI data. In addition to differences in the causal interaction between the rAI and rPPC, MDS analysis revealed that the strength of causal influence from rAI to ACC was also significantly weaker in children. Here, we focus on the convergent findings from the two analyses on developmental differences in the causal link from the rAI to the rPPC. The only previous study to have examined developmental changes in causal interactions during cognition did not examine rAI connectivity and no connectivity differences were reported between the extended FIC or any regions of the inferior frontal gyrus with the PPC
Multimodal analysis of fMRI and DTI data revealed that functional connectivity differences between the rAI and rPPC were associated with weak structural links between these areas. Quantitative DTI-based tractography showed that the density of white-matter fiber tracts connecting the rAI and rPPC was significantly lower in children, compared to adults. This result is consistent with previous studies showing slow maturation of long-distance white matter tracts
As noted above, children were significantly slower and less accurate than adults. We examined whether this behavioral difference was the result of weak network interactions. We found that the strength of causal network interactions collectively were strongly predictive of reaction times; in contrast, the rAI→rPPC link by itself was only weakly correlated with response latency in children and adults. Using multivariate sparse regression analysis, we found that network interactions better predicted reaction time in both children and adults. In children, the strength of rAI→rPPC along with rAI→rVLPFC collectively predicted reaction times, while in adults the strength of rAI→rPPC along with rAI→ACC and rAI→rDLPFC collectively predicted reaction times. It is noteworthy that even though a different set of links predicted reaction times in both groups, the rAI→rPPC link was common to both. We also found that reaction times were better predicted in adults, compared to children. These results suggest that it is the multiple network interactions as a whole, rather than individual links by themselves, that moderate performance. Critically, similar results were observed when accuracy instead of reaction time was used as the performance measure. Thus, casual interactions between the rAI and rPPC are an important factor for mediating performance improvements in higher-order cognition with development.
It is noteworthy that the rAI showed the strongest causal signals, even though the rVLPFC has been most commonly implicated in control
Efficient control requires the concerted coordination between multiple brain regions and there is growing evidence to suggest that this is implemented via dedicated neurocognitive networks
A unified network approach – wherein we first specify intrinsic brain networks using rsfMRI data and then analyze interactions among anatomically discrete regions within these networks during cognitive information processing – enhances our ability to generalize beyond individual task activated foci and also provides a common framework for comparing brain response and connectivity in children and adults. Our findings are likely to have important implications for understanding the development of control mechanisms subserved by dynamic interactions between neurocognitive networks. Further studies are needed to examine whether similar control mechanisms underlie the functional maturation of specific cognitive processes involving inhibition, memory and decision-making. The quantitative approach developed here is likely to be useful in the investigation of neurodevelopmental disorders, such as autism and attention deficit/hyperactivity disorder, in which control processes are impaired.
Twenty-three children and twenty-two IQ-matched adults participated in this study after providing written informed consent. For those subjects who were unable to give informed consent, written, informed consent was obtained from their legal guardian. The study protocol was approved by the Stanford University Institutional Review Board. Children (10 males, 13 females) ranged in age from 7 to 9 (mean age 7.95) with an IQ range of 88 to 137 (mean IQ: 112). Adults (11 males, 11 females) ranged in age from 19 to 22 (mean age 20.4) with an IQ range of 97 to 137 (mean IQ: 112).
The fMRI experiment consisted of 52 arithmetic problems presented in a jittered event-related design along with “rest” or “null” trials in which participants passively viewed a cross on the screen. In the arithmetic trials, participants were presented with an equation involving two addends and a resultant and were asked to indicate via a button box whether the resultant was correct or incorrect. Half the addition trials consisted of problems with addends different from ‘1’ (e.g. 3+4 = 7). One operand ranged from 2 to 9, the other from 2 to 5 (tie problems such as ‘5+5 = 10’, were excluded), and resultants were correct in 50% of the trials. Incorrect answers deviated by ±1 or ±2 from the correct sum. The other half of the addition trials had the same format but one addend was ‘1’ (e.g. 5+1 = 7). Stimuli were displayed for 5 seconds with an inter-trial interval of 500 msec followed by a blank screen for 500 msec and an inter-trial jitter that varied between 0 to 3500 msec with an average duration of 1814 msec. Each subject underwent a math task scan and 8-min resting-state scan.
fMRI data acquisition, preprocessing, analysis of task data with General Linear Model (GLM), Independent component analysis (ICA) of resting data, and functional connectivity analysis procedure is described in detail in Supplementary Information (
We defined regions of interest (ROIs) in five key nodes of the SN, right CEN, and left CEN based on the peaks of the ICA clusters. ROIs were selected from respective combined-group ICA clusters: in the rAI, VLPFC and ACC (on the SN ICA map); in the rDLPFC and rPPC (on the right CEN ICA map); in the lDLPFC and lPPC (on the left CEN ICA map). After visually selecting a voxel with the highest
MGCA was performed in accordance with the methods of Seth et al.
Multivariate Dynamical Systems (MDS) is a novel state-space Multivariate Dynamical Systems (MDS) model to estimate intrinsic and experimentally-induced modulatory causal interactions between multiple brain regions
We model the fMRI time series in region
In Equation (3),
Estimating causal interactions between
The statistical significance of the causal connections was assessed by using non-parametric analysis. Empirical null distribution of the parameters
To describe the interactions between brain regions in the causal network generated by MGCA, we examined the following graph metrics: (1) Out-degree: Number of causal outflow connections from a node in the network to any other node. (2) In-degree: Number of causal in-flow connections to a node in the network from any other node. (3) (Out – In) degree: Difference between out-degree and in-degree is a measure of the net causal outflow from a node. (4) Path length: Shortest path from a node to every other node in the network (normalized by the number of nodes minus one). Shorter path lengths indicate a more strongly interconnected or “hub-like” node. A two-sample
To investigate whether causal network interactions predict behavior differently in children and adults, we examined causal connectivity patterns in the two groups. The causal functional connectivity patterns – strength of causal connectivity of 10 pairs of anatomical regions – along with behavioral measures were used as the input to a sparse regression algorithm. The sparse regression algorithm identifies causal network connections that predict behavior by modeling the relationship between the dependent variable (RT) and the independent variables (strength of pair-wise causal connectivity). An added advantage of using a sparse regression algorithm, as opposed to traditional regression, is that it performs feature selection wherein the coefficients of independent variables that do not contribute to the prediction of the dependent variable are set to zero. In our case, this entails that the regression analysis would identify the causal network connections that predict behavior while the non-contributing connections would be set to zero. Such sparse methods are particularly elegant when the number of possible predictor variables is large. In sum, we used sparse regression analysis instead of the more conventional regression analysis so that we could not only investigate whether causal connectivity predicts behavioral measure(s) but also identify in a purely data-driven manner which subset of causal connections, if any, predicts behavior. GLMNet
DTI data was obtained from 18 of the 23 children subjects and 15 of 22 adults. Acquired images underwent the following preprocessing steps: eddy-current correction, alignment with T1-weighted anatomical images, resampling, and tensor computation. Fiber tracts between the rAI and rPPC were computed as previously described
(TIF)
(TIF)
(TIF)
(TIF)
(TIF)
(DOC)
(DOC)
(DOC)
We thank Drs. Lucina Uddin, Srikanth Ryali, Miriam Rosenberg-Lee, Sarit Ashkenazi-Zoro, and Arron Metcalfe for valuable feedback.