Changes in the interaction of resting‐state neural networks from adolescence to adulthood

Abstract

This study examined how the mutual interactions of functionally integrated neural networks during resting‐state fMRI differed between adolescence and adulthood. Independent component analysis (ICA) was used to identify functionally connected neural networks in 100 healthy participants aged 12–30 years. Hemodynamic timecourses that represented integrated neural network activity were analyzed with tools that quantified system “causal density” estimates, which indexed the proportion of significant Granger causality relationships among system nodes. Mutual influences among networks decreased with age, likely reflecting stronger within‐network connectivity and more efficient between‐network influences with greater development. Supplemental tests showed that this normative age‐related reduction in causal density was accompanied by fewer significant connections to and from each network, regional increases in the strength of functional integration within networks, and age‐related reductions in the strength of numerous specific system interactions. The latter included paths between lateral prefrontal‐parietal circuits and “default mode” networks. These results contribute to an emerging understanding that activity in widely distributed networks thought to underlie complex cognition influences activity in other networks. Hum Brain Mapp 2009. © 2009 Wiley‐Liss, Inc.

INTRODUCTION

Neural constructivist models of development predict that the increasing sophistication of cognitive control ability that arises between childhood and adulthood is the likely result of more interconnected and efficient neural network function [Blakemore and Choudhury,2006; Quartz and Sejnowski,1997; Westermann et al.,2006]. Therefore, refinement of the degree of functional integration within and across neural networks is likely a defining characteristic of normal maturation. The term “functional integration” generically refers to brain regions whose activity is empirically associated. In recent years, such associations have been described as “functional connectivity” or “effective connectivity.” Although these two terms continue to be used variably by different researchers, their general meaning has become somewhat well established [see recent reviews by Friston,2002; Friston and Price,2001]. When distributed brain regions display strongly correlated patterns of neural activity change, it is taken as evidence that those regions are functionally connected, likely via reciprocal excitatory neurotransmission through long‐distance white matter pathways [Bressler and Kelso,2001; Fingelkurts and Kahkonen,2005; Tononi et al.,1998]. The presence of stronger within‐network functional connectivity compared to without is believed to be a basic empirical characteristic of neural network structure. “Effective connectivity” can be thought of as a special case of functional connectivity, in which inferences can be made about the cognitive relevance of regional brain function from the presence or strength of statistically determined “causal” influences among various regions [Friston,2002; Friston and Price,2001]. Simply put, functional connectivity analysis methods can identify functionally integrated networks, whereas effective connectivity can measure how brain regions within those networks interact in ways that might show how the network operates to mediate cognitive demands.

Functional connectivity is often measured as inter‐regional correlations among spontaneous fluctuations of hemodynamic activity during a “resting state” while participants lie passively in the MRI machine, but no active cognitive or behavioral demands are imposed [Raichle et al.,2001; Shulman et al.,1997]. The finding that there was significant intercorrelation among precuneus/posterior cingulate, ventral anterior cingulate, and ventromedial prefrontal cortex (i.e., regions comprising the “default mode” of brain activity) [Greicius et al.,2003] helped to spark interest in identifying functionally integrated neural networks during resting state. Since these initial studies, other examinations of resting‐state brain activity have found there are numerous functionally connected circuits during rest, including classically described motor, sensory, language, and visual networks [Cordes et al.,2000]. Studies applying independent component analysis (ICA) or similar data‐driven methods to resting‐state fMRI data have identified additional discrete neural circuits comprised of brain regions often engaged by higher‐order cognitive tasks, including fronto‐cerebellar, parietal‐cerebellar, fronto‐parietal, and cinguloopercular networks [Beckmann et al.,2005; Dosenbach et al.,2007; Fransson,2005; Seeley et al.,2007]. In sum, fMRI resting‐state research has found reproducible evidence for at least 11 distinct networks engaged during rest [Beckmann et al.,2005; Calhoun et al.,2008; Damoiseaux et al.,2006; De Luca et al.,2006].

Although there is evidence that some functional networks are formed by puberty [Fair et al.,2007; Stevens et al.,2007a,2007b], the structure and interconnectivity of these networks continue to change throughout adolescence [Fair et al.,2007,2008; Stevens et al.,2007a,2007b]. Fair et al. found greater correlation of activity in brain regions connected by long‐distance pathways with increasing age [Fair et al.,2007]. In addition, the amplitude of hemodynamic change in these distributed networks generally increases with age [Stevens et al.,2007a,2007b], though there is evidence that regional decreases also occur [Fair et al.,2007; Stevens et al.,2007b]. Implicit in the idea of functional integration among brain regions is that activity in these regions is somehow mediated by its effective connections with other brain regions. This concept is easily extended to systems of neural networks. Given the recognition that complex cognition arises not simply from the local processing in any given task‐engaged brain region, but from widely distributed groups of brain regions [Fuster,2006; Mesulam,1998], it is reasonable that activity in one entire network could have a demonstrable and reliable effect on another. This is not a new idea, as several theorists have proposed similar constructs for EEG or fMRI‐measured brain activity [e.g., “microstates,” or “cognits,” etc.; Fuster,2006; Koenig et al.,2002]. These constructs generally are defined as an integrated network of distributed brain regions showing the same activity profiles whose conjoint activity is presumed to influence activity in other brain regions or networks.

The profile and strength of such network‐to‐network influences (i.e., “functional network connectivity”) likely will shed important light on changes to brain organization underlying cognitive development. Independent components analysis (ICA) of fMRI is well suited for characterization of multiple functional networks, because by definition the brain regions in each component have the same profile hemodynamic signal change. Londei et al. were the first to suggest studying functional network connectivity by combining methods to identify networks with ICA and methods to calculate effective connectivity among their hemodynamic timecourses [Londei et al.,2006]. We recently examined functional network connectivity in healthy and schizophrenic adults during resting state, active working memory, and attention tasks [Demirci et al., submitted to Neuroimage; Jafri et al.,2008]. These studies not only found evidence for measurable, directed influences among large‐scale functional networks, but also found a diagnosis of schizophrenia resulted in widespread disruption, greater dependency, and greater variability of network inter‐relationships. Thus, healthy adult brain function is characterized by the presence of distinct, directed relationships among numerous distributed neural networks. Moreover, this study validates the notion of networks influencing networks and implies that normal cognitive development may result from changes to neural processes that transform a system of relatively less specialized functional circuits into a more greatly organized system where interactions among entire networks become more distinct, efficient, and influential.

This study examined changes across normal adolescent to adult development in interactions among numerous functional networks identified during resting state. Despite evidence that there are many functionally connected networks at rest, few studies have examined the characteristics of systems of these circuits. Our approach was to apply tools that quantified the connection of among neural networks [Seth,2005] within a system comprising network activity timecourses identified by ICA. ICA timecourses represent the conjoint pattern of hemodynamic activity across each distributed neural network, as in our previous work [Stevens et al.,2007a,2007b]. We hypothesized that there would be a linear relationship of decreasing system causal density with increasing age. If so, this would support the premise that maturation leads to a trimming of unneeded or unused interdependencies across networks. We hypothesized there would be linear effects of age. We also tested for possible curvilinear age‐related changes. These tests were prompted by the recognition that many neurobiological changes have nonlinear trajectories and resolve in mid‐ to late‐adolescence and because it is not yet clear whether any of these structural and functional changes would be associated with the causal density estimates examined in this work. A secondary aim was to characterize specific age‐related changes to effective connectivity among networks that might be related to developmental changes in the profile of network causal dynamics. Therefore, we conducted a series of supplemental analyses designed to better characterize what aspects of brain activity changed in association with system causal density changes. We examined the relationship of age and number of significant in‐degree and out‐degree causal connections for each network. We also examined the effects of age on the strength of specific network connections and the association of network spatial maps of functional connectivity with measures of causal density.

MATERIALS AND METHODS

Participants

Participants were 100 healthy volunteers (50% male) between the ages of 12 and 30; mean (SD) = 19.9 (6.76). Age median‐split subsamples were n = 48 participants ages 12–20 [mean (SD) = 16.6 (2.40)] and n = 52 participants ages 21–30 [mean (SD) = 23.8 (2.59)] (t ₉₈ = −14.32, P <.001). Participants were recruited via advertisements at the Olin Neuropsychiatry Research Center (Olin NRC), Hartford, CT and provided written informed consent in protocols approved by Hartford Hospital's institutional review board. For legal minors, parents provided written permission and minors provided assent. All research was conducted in adherence to ethical standards required for human subjects protection. Participants were drawn from numerous ongoing studies at the Olin NRC. As such, assessment of intellectual ability was not standardized. However, all participants were in high school at grade appropriate levels or had a high school or greater level of academic achievement. All had IQ estimates (evaluated by various screening instruments) at least in the average range or better.

Imaging Parameters and Processing

Imaging was done with the Siemens Allegra 3T system located at the Olin NRC. Each participant's head was firmly secured using a custom head holder. Localizer images were acquired for use in prescribing the functional image volumes. The echo planar image gradient‐echo pulse sequence (TR/TE 1500/28 ms, flip angle 65°, FOV 24 × 24 cm, 64 × 64 matrix, 3.4 by 3.4 mm in plane resolution, 5 mm effective slice thickness, 30 slices) effectively covered the entire brain (150 mm) in 1.5 s. Head motion was restricted using a custom built cushion inside the head coil. During scanning, participants were asked to fixate gaze on a small cross presented on the screen, remain alert and awake, and not to close their eyes or move their head. These instructions to keep eyes open and fixated on a point in the middle of a visual display helped to reduce head motion and served as an experimental control over visual input. All participants were judged to be awake and alert at the start and conclusion of this fMRI session. The stimulus run consisted of 210 time points (5 min 15 s). The initial six images during which T₁ effects stabilized were discarded. Remaining functional images were reconstructed offline and each run was separately realigned using INRIAlign [Freire et al.,2002] as implemented in statistical parametric mapping (SPM2). Each participants' translation and rotation corrections were examined to ensure there was no excessive head motion. A mean functional image volume was constructed for each session from the realigned image volumes. This mean image volume was then used to determine parameters for spatial normalization into Montreal Neurological Institute standardized space employed in SPM2 (http://www.fil.ion.ucl.ac.uk/spm/). The normalization parameters determined for the mean functional volume were then applied to the corresponding functional image volumes for each participant. These normalized images were corrected with a custom algorithm that used linear interpolation to remove variation in BOLD signal intensity due to slice acquisition temporal onset differences. Finally, the normalized functional images were smoothed with a 12‐mm full width at half‐maximum Gaussian filter.

Identification of Functionally Connected Neural Networks

These fMRI timeseries data were analyzed to identify spatially independent and temporally coherent networks using a group ICA algorithm [Calhoun et al.,2001] available in a Group ICA of FMRI Toolbox (GIFT v1.3b) implemented in Matlab (http://www.icatb.sourceforge.net). In this approach, a single ICA analysis was performed on the 100 participants, followed by a back reconstruction of single‐subject timecourses and spatial maps from the raw data. For computational feasibility, data were reduced through three data reduction stages (two PCA and one ICA reduction). At each stage, data were concatenated for further reduction. Twenty‐five components were estimated using the infomax approach [Bell and Sejnowski,1995]. The number of components was determined using the minimum length description criteria adjusted to account for correlated samples [Li et al.,2007]. Component timecourses and spatial maps were reconstructed for each individual participant, then calibrated (scaled) using the raw data to reflect percent fMRI signal strength [Calhoun et al.,2001] for between participant comparison. Participant spatial maps represented regional strength of functional connectivity, defined as the statistical correspondence of each voxel to the average network timecourse. Because all components were scaled to an average baseline across voxels and subjects, the zero‐point of the component maps did not necessarily represent true absence of signal change. To more accurately represent positive versus negative signal change, each set of component maps were converted to z scores using a transformation that added a bias term that centered the distribution's maximum point of the normal curve at zero.

Components with low correlation with spatial templates of grey matter or high correlation with maps of white matter or CSF were discarded. A secondary criterion was whether the component depicted connectivity in at least two distal brain regions. Therefore, 13 components were retained for further examination. To visualize the components we created statistical parametric maps that quantified the conservation of spatial structure over subject‐to‐subject variability. This was done by calculating a one‐sample SPM2 t test on each component set of individual participant's spatial maps. Brain regions were considered to be within each network if they met a P < 0.0001 threshold, corrected using false discovery rate (FDR) adjustments for multiple comparisons [Genovese et al.,2002]. Components depicted in Figures 1 and 3 were mapped to cortical surface renderings for ease of visualization [Van Essen et al.,2001]. For visual comparability, the statistical activation maps are displayed using a uniform critical t score of 4.90, which is the lowest critical t score for FDR P < 0.05 across the components.

Figure 1.

Cortical renderings [Van Essen et al.,2001] of the spatial structure of 13 ICA components identified during resting‐state fMRI assessment. These renderings depict the spatial structure of neural networks identified during resting‐state fMRI assessment. The data in each rendering was produced using SPM2 one‐sample t test to quantify how the spatial structure of these components is conserved over all n = 100 subjects.

Figure 3.

Age‐differences in effective connections among ICA components. These are results from a supplemental analysis that examined the correlation of age with estimates of the strength of “causal connectivity”, calculated as the log transform of the F statistic testing the significance of Granger Causality between any two network hemodynamic timecourses [Seth,2005]. These exploratory results are evaluated at a liberal statistical threshold of P < 0.05, uncorrected.

Causal Connectivity Analyses

Analyses of network effective interactions were facilitated by tools developed by Seth [Seth,2005] that characterize the overall profile of network causal connectivity. These tools combine Granger causality estimates [Granger,1969] and graph theory [Bollobas,1985; Ioannides,2007] to examine a directed graph of dynamical interactions among separate timecourses within a neural system, in which each graphed edge reflects a causal influence between two nodes (ICA components, or functional networks). Like numerous statistical measures that quantify directed relationships between two measurements, Granger causality can be used to infer that one hemodynamic profile “causes” or in some way influences the activity in another. This approach provides several useful statistics that describe network causal complexity that can be computed for samples, or for individual participants to permit random effects hypothesis testing. For instance, causal density reflects the fraction of significant mutual causal interactions present in a set of networks. In this analysis, network causal density reflected the fraction of interactions among nodes (components) that were causally significant below an a priori significance criterion. Because this procedure merely sets a threshold for counting the connections among nodes that likely do not occur by chance, the choice of criterion level for these analyses is arbitrary. Any P level is defensible as long as it balances specificity of effects against control over chance. In these analyses, we adopted the recommended P levels for individual (P < 0.05) and group (P < 0.01) effects [Seth,2005]. Low causal density (i.e., close to zero) reflects independent node dynamics. Causal density was defined as gc/(2N(N−1)), where gc was the total number of significant causal links observed, and N was the size of the network. The number of lagged correlations to use in the model was estimated using Bayesian Information Criterion estimates. Model order estimates for the group model were subsequently employed for all individual‐participant models.

Because Granger causality estimates are most valid when all system influences are modeled, we approached calculation of system dynamics including all valid “within‐brain” network timecourses determined by ICA. After confirming that the set of ICA hemodynamic timeseries were covariance stationary [Seth,2005], causal density of the resting‐state system of 13 components was calculated using component timecourse data concatenated across subjects (i.e., 13 ICA hemodynamic timecourses with 204,100 time points each). Causal density also was calculated separately for each individual participant to permit random effects hypothesis testing. Individual participant models generally find fewer statistically significant causal paths, consistent with the relative reduction in number of measurements and corresponding drop in statistical power (e.g., 204 time points versus 20,400 in the concatenated dataset). For illustrative purposes only, the causal density estimates of adolescent and adult subsamples of concatenated data and the average results from individual subject analyses also were calculated in separate models.

The primary study hypothesis was examined using multiple regression testing for linear and/or curvilinear effects of age on individual causal density estimates. This test employed a stepwise regression that first tested for a significant linear effect of age. Secondarily, we tested whether there was a significant improvement in regression model fit by adding either a logarithmic or a quadratic age term. A significant result (P < 0.05) would indicate that developmental changes to network weighted causal density would be best described as curvilinear. If the results indicated a nonlinear fit was best, we evaluated the F statistic for the combined model instead of the simple linear effect to determine the significance of the statistical relationship between age and causal density.

We also included a series of supplemental analyses—each with its own goals, statistical tests, and attention to the problem of multiple comparisons. Changes in system causal density could be related to changes in a subset of nodes that either exert stronger influence over others (i.e., more causal) or nodes that are less often the targets of causal influence (i.e., “sinks”). The tools provided by Seth permit empirical assessment of out‐degree and in‐degree causal sources and sinks and the overall balance of these paths [Seth,2005]. Supplemental analyses of the number of in‐ and out‐degree paths were conducted using multivariate regression. The use of the multivariate technique provided a single overall test of age on all available in‐degree pathway data. This permitted us to evaluate the overall significance using an uncorrected threshold of P < 0.05. Following a significant multivariate effect, we then documented which specific networks contributed to that effect by examining the results of univariate tests. A separate multivariate model was computed to determine the effect of age on the number of in‐degree and on the number of out‐degree pathways. Because these two multivariate tests are conceptually linked (i.e., performed on similar data addressing essentially the same question), we set the a priori significance level at P < 0.025 (i.e., P < 0.05, adjusted with Bonferroni‐type correction for two independent tests).

The results from the test of the primary hypothesis prompted an exploratory analysis of the relationship of age on the strength of individual connections among networks. The dependent measure was a weighted estimate of causal influence. This weighted estimate was calculated by scaling the contribution of each significant connection by its magnitude (i.e., defined as the log transform of the F statistic for a connection between any two nodes) [Seth,2005]. Thus, age‐related effects depicted in Figure 3 do not simply represent developmental effects on the presence or absence of connections, but rather changes to the strength of connectivity. Because these analyses were truly exploratory, all effects of P < 0.05 uncorrected are reported.

Finally, simple regression SPM2 models were used to examine the relationship between individual causal density estimates and ICA component spatial maps representing the regional strength of functional connectivity. Significant relationships between these measures would identify specific brain regions within each network where age‐related changes in functional connectivity corresponded to changes in causal density. Given the exploratory nature of these analyses and the large number of components examined, stringent correction for searching all available brain voxels were employed (P < 0.05 FDR) [Genovese et al.,2002]. Although a separate SPM2 t test was calculated for each component, the use of FDR corrections for searching the whole brain effectively provides adequate statistical control over possible false positive results.

RESULTS

Table I lists the networks identified by ICA. The brief descriptive labels are based on the major regions within each component. Illustrations and a full list of all brain regions within each network is available in Figure 1 and the Supplemental Table online, which includes detail on both positive and/or negative‐going signal changes.

Table I.

Age differences in number of in‐degree and out‐degree paths for networks identified during “resting state” by ICA

	Component	Number of in‐degree paths			Number of out‐degree paths
		Age 12–20, mean (SD)	Age 21–30, mean (SD)	Age p	Ages 12–20, mean (SD)	Ages 21–30, mean (SD)	Age p
1	Occipital‐parietal‐PFC	1.62 (1.59)	1.36 (1.37)	ns	1.68 (1.66)	1.40 (1.19)	0.022
3	Cerebellum‐frontopolar‐cingulate	2.10 (1.76)	2.09 (1.51)	ns	2.29 (1.91)	1.76 (1.4)	ns
5	Precuneus‐sensorimotor‐SMA	2.52 (2.06)	2.17 (1.47)	ns	1.97 (1.59)	1.82 (1.41)	ns
7	IPL‐precuneus‐PCC‐fusiform	2.16 (2.16)	1.78 (1.36)	0.013	1.95 (1.59)	1.53 (1.11)	ns
10	Right lateral PFC‐parietal	1.58 (1.93)	1.32 (1.46)	ns	1.33 (1.74)	1.07 (1.08)	0.014
11	Precuneus‐PCC‐vmPFC‐caudate	2.18 (1.9)	1.69 (1.5)	0.044	2.20 (1.52)	2.34 (1.39)	ns
14	Ventromedial striatal	1.72 (1.45)	1.30 (1.32)	ns	2.10 (1.49)	2.01 (1.85)	ns
15	Sensorimotor‐temporal	2.06 (1.89)	2.53 (1.67)	ns	1.81 (1.45)	1.71 (1.3)	ns
16	Left lateral PFC‐parietal	1.72 (1.72)	1.23 (1.79)	ns	1.33 (1.54)	0.98 (1.09)	<0.001
17	Cingulate‐insula‐temporal‐parietal	2.12 (1.89)	2.36 (1.6)	ns	2.29 (1.32)	2.55 (1.51)	ns
19	Temporal‐parietal‐PFC	2.29 (1.73)	2.25 (1.29)	0.030	2.06 (1.68)	2.05 (1.34)	ns
23	Cingulate‐insula‐striatum‐temporal	1.95 (1.71)	1.80 (1.44)	0.045	3.20 (2.03)	3.36 (1.95)	ns
25	Occipital‐PCC‐temporal	2.06 (1.56)	2.17 (1.71)	0.038	1.87 (1.68)	1.46 (1.25)	ns

The first column lists the brief summary label provided for each of the 13 networks examined in this study. The number to the left of each label denotes which of the 25 possible ICA components. The rightmost six columns compare the mean (SD) and significance level between groups of adolescents (ages 12–30) and adults (ages 21–30) examined in the analysis. The P levels represent the results of a test for logarithmic effect of age. These univariate effects were examined following a significant multivariate effect for each set of age group comparisons.

System Causal Density Analysis

We calculated a causal density estimate [Seth,2005] for the resting‐state system using the 13 ICA component timecourses concatenated across all 100 participants. Overall system causal density was 0.917, indicating that the system of distributed brain networks had very dense mutual causal interactions. To get a sense of possible differences between age groups, causal density estimates were calculated separately for each age subgroup. This was not a direct statistical comparison and was conducted merely to exemplify subgroup differences. Causal density estimates showed appreciably lower system mutual causality for adults (0.859) compared to adolescents (0.904). Again for illustration purposes, individual causal density statistical test results were consistent with those calculated using group‐concatenated data in each subsample [i.e., adolescent mean (SD) = 0.17 (0.07) versus adult 0.15 (0.04)].

The primary study hypothesis was addressed using multiple regression analysis on these individual causal density statistics. Consistent with the age subgroup illustrations above, there was a significant effect of age on causal density (t = −2.841, P < 0.005), which supported the primary study hypothesis that there would be a linear decrease in causal density with increasing age. Stepwise regression additionally tested whether the inclusion of either a logarithmic or a quadratic term to this model significantly improved model fit. The logarithmic term resulted in better overall model fit (F _1,97 = 4.118, P < 0.045; R ² changed from 0.076 to 0.099 by including the logarithmic term). There was no model improvement for the quadratic term (F _1,97 = 2.443, ns). Therefore, the age‐related differences in network causal density statistics were best characterized by a logarithmic profile of age‐related reduction in causal density. A scatterplot of these relationships is shown in Figure 2. It can be seen that although the logarithmic effect was a better statistical fit to the data, the differences between the linear and logarithmic fits were not strikingly different. Despite this, for consistency all supplemental tests tested for logarithmic effects of age.

Figure 2.

Results of stepwise regression showing significant linear and curvilinear inverse effects of age on systemwide causal density statistics. These results show a general reduction in distributed network interactions with increasing age, with the most significant changes occurring during adolescence.

Age Effects on Network Causal Dynamics

Causal flow can be defined as the net difference of significant “in‐degree” relative to “out‐degree” causal paths. Tests for whether the number of in‐degree or out‐degree connections decreased with age was evaluated with multivariate regression on individual subject statistics. These two multivariate regression models tested for a significant logarithmic effect. For the number of in‐degree pathways to each system node, there was a logarithmic (F _13,85 = 2.123, P = 0.020) effect of age that surpassed our P < 0.025 Bonferroni‐corrected criterion for significance. Although there was a logarithmic effect of age (F _13,85 = 1.939, P = 0.037) for average number of out‐degree pathways, that finding should be interpreted cautiously because it did not meet our a priori α adjustment for multiple comparisons (P < 0.025). However, age‐related changes in number of significant connections typically decreased or remained stable with increasing age, which supported the study's overall predictions. Univariate tests identified several specific networks that appeared to become less the target of influence by other networks with increasing age (Components 7, 11, 19, 23, and 25; Table I). There was another group of components for which greater age was associated with exerting fewer causal influences over other networks (Components 1, 10, and 16; Table I).

The finding that there was a general reduction in average number of paths prompted supplemental analysis of the relationship of age to the strength of each network connection within each participant. Because this analysis was exploratory, liberal statistical thresholds (P < 0.05) were used. Therefore, the results should be interpreted cautiously. Figure 3 depicts age‐related decreases in the average strength of specific pathways among networks in the system. Note, only age effects at P < 0.05 are displayed, as the dense mutual interactions prevented showing all possible significant causal paths among components. All significant age‐related changes were negative, such that increasing age was associated with decreased connection strength between specific networks. One notable finding was that there were numerous age‐related changes of influences over default mode networks (Components 5, 7, and 11). Specifically, the left lateral PFC‐parietal network significantly decreased its causal input to Components 5 and 11. We also observed less influence of IPL‐precuneus‐PCC‐fusiform default mode network (Component 7) over the right lateral‐PFC‐parietal network. There also was decreased integration between the right lateral PFC‐parietal network (Component 10) and cingulate‐insula‐striatum‐temporal (Component 23) networks. Although other age‐effects shown in Figure 3 have similar P levels, these particular findings were for networks for which there also were significant univariate effects of change in number of in‐degree or out‐degree paths (see above), which supports their validity.

Association of Causal Density With Within‐Network Functional Connectivity

We also were interested in exploring whether changes in the degree of regional functional connectivity within any given brain structure might have influenced the system as a whole. The Pearson correlation of individual casual density statistics with individual ICA spatial maps identified how the relative strengths of functional connectivity within specific brain regions comprising a given network were associated with changes to the overall system of network causality. Five networks (Components 1, 3, 7, 19, and 23) had brain regions that were significantly associated with overall causal density statistics (P < 0.05 FDR corrected for searching the whole brain; [Genovese et al.,2002]). Greater causal density was associated with greater functional integration of left inferior parietal lobule (x, y, z = −51, −51, 51, t = 5.58) into the occipital‐parietal‐PFC network (Component 1). The cerebellum‐parietal‐PFC circuit (Component 3) had less integration of left inferior/middle frontal gyri (x, y, z = −36, 39, −9, t = 5.55) with greater causal density. One of the default mode components (IPL‐precuneus‐PCC‐fusiform; Component 7) had greater integration of bilateral middle frontal gyri in dorsolateral prefrontal cortex (x, y, z = −42, 30, 33, t = 4.84; x, y, z = 45, 36, 36, t = 5.31), mid‐cingulate gyrus (x, y, z = −3, −12, 30, t = 4.22, and occipital cortex (x, y, z = 6, −81, −6, t = 4.54) with greater causal density. Causal density also was associated with greater functional connectivity of rostral anterior cingulate (x, y, z = −9, 33, 30, t = 4.67) and bilateral fusiform/parahippocampal gyri (x, y, z = −42, −21, −27, t = 4.98; x, y, z = 42, −18, −30, t = 5.19) in Component 19 (Temporal‐parietal‐PFC). Finally, Component 23, the cingulate‐insula‐striatum‐temporal network identified in several previous studies [Fair et al.,2007; Seeley et al.,2007], showed greater integration left frontopolar cortex (x, y, z = −12, 54, 39, t = 4.75) with greater causal density.

DISCUSSION

The primary goal of this study was to determine if typical adolescent maturation decreased the mutual influences of complex distributed networks integrated during rest. Consistent with our hypotheses, measurements of system causal density between networks identified by ICA decreased across development. Although these age‐related changes were found to best fit a logarithmic profile, it can be seen in Figure 2 that the overall changes closely matched a simple linear effect in this age range. Supplemental tests also found that increasing age was associated with a decrease in the number of both in‐degree and out‐degree effective connections each circuit had with one another. These developmental changes likely reflect greater independence of these networks from each other, each coactivating regions within the network to a greater degree than those outside of it. This presumably reflects the ongoing refinement of well‐described organization principals of neural segregation [Friston,2002] and hierarchical organization [Friston,2002; Ioannides,2007] believed to underlie network functional specialization and integration. These findings for developmental changes in between‐network effective connectivity are novel, but join previous evidence that established‐yet‐immature neural networks undergo changes to their structure and within‐network regional integration [Fair et al.,2007,2008; Stevens et al.,2007a,2007b] as a typical feature of adolescent development. These changes might reflect development of grey and white matter volume [Sowell et al.,2004], increasing white matter myelination [Mukherjee and McKinstry,2006], or experience‐dependent pruning of intercellular connections [Blakemore and Choudhury,2006]. The net result of such development likely is attainment of mature functional network profiles that not only are optimally capable of supporting cognitive or behavioral demands, but also are optimally efficient and flexible in adjusting local neuronal processing to mediate changing situational demands.

Because the resting‐state fMRI paradigm has been frequently used to study brain regions thought to comprise the “default mode,” we specifically examined the connections among ICA components that depicted the default mode brain regions. A key finding is that we found evidence for several different default mode networks. There were three networks (Components 5, 7, and 11) showing functional connectivity among brain regions closely resembling the default mode [Greicius et al.,2003; Raichle et al.,2001; Shulman et al.,1997]. Although all three networks engaged precuneus, they differed in their functional integration with other brain regions. This was similar to other reports using ICA to examine resting‐state fMRI data [Beckmann et al.,2005; De Luca et al.,2006; van de Ven et al.,2004]. The fact that these components had distinct timecourses raises the possibility that the default mode may comprise several unique circuits acting in close concert, likely with close mutual causal interactions. Our results also indicate that regions comprising the default mode system can engage in characteristically different ways during rest, raising the possibility that different profiles of activity could serve different neural processing roles.

Unlike several previous reports [Fair et al.,2008; Greicius et al.,2003; Margulies et al.,2007], we did not observe rostral anterior cingulate engagement into any of these three default mode circuits. Instead, rostral cingulate was integrated into Component 19. This comprised a network where activity in rostral cingulate, orbitofrontal cortex, and posterior cingulate were functionally integrated, but whose hemodynamic response was opposite to that seen in the precuneus. Because signal changes in key regions were in opposite directions, we did not label that circuit as default mode. In one study that also failed to find anterior cingulate integration into the default mode when using ICA [van de Ven et al.,2004], the authors suggested it might be due to modulatory effects on cingulate activity by other brain regions. It is also possible that the nature of the statistical method used to determine functional connectivity might explain the discrepancies among previous reports. ICA could be more sensitive to subtly differing spatiotemporal patterns in hemodynamic data than methods that quantify linear correlations with seed voxels. Alternatively, there has been recent evidence that default mode activation can be confounded with changes in BOLD signal in the brain resulting from respiration and changes in CO2 [Birn et al.,2008]. Because Birn et al. found ICA to have some success in separating respiratory artifact from functional connectivity measurements during resting state, it may be that ICA ultimately will prove to be more accurate in characterizing complex patterns of functional integration among default mode brain regions compared to correlation methods used in earlier work.

We conducted several sets of supplemental tests to characterize specific changes in brain activity that might contribute to altered system connectivity. One noteworthy finding was that there were developmental reductions in the strength of directed influences among prefrontal‐parietal networks and default mode circuits. Prefrontal‐parietal networks have been found to subserve verbal working memory and other forms of executive control [Collette et al.,2006; Dosenbach et al.,2007; Stevens et al.,2007b; Woodward et al.,2006]. We identified several networks that corresponded to “executive” networks previously described in studies of cognitive control and network maturation. The right lateral PFC‐parietal network close resembled that reported by Dosenbach and colleagues and others [Dosenbach et al.,2007; Fair et al.,2007,2008; Seeley et al.,2007]. Components 17 and 23 resembled the cingulo‐opercular network discussed in some of those studies. Interestingly, the current analysis did not find a network with a strong correlation to the parietal‐premotor executive network found in the Stevens et al. study of response inhibition [Stevens et al.,2007b], suggesting that some networks are integrated specifically to mediate task demands and do not appear during passive rest. The age‐related reduction in the connectivity strength between these executive and default mode types of networks suggests that they might be less tightly bound together with greater maturation, perhaps affording a greater degree of processing flexibility in unstructured situations. Fransson has suggested that passive rest might be characterized by a cycling between awareness and processing of external environmental cues (i.e., typically marked by engagement of prefrontal‐parietal brain regions) and attention to internal states, thoughts, and feelings (i.e., engagement of default mode regions) [Fransson,2005]. The current results are consistent with that proposal and imply that segregated lateral prefrontal cortex may characterize mature network system interactions.

We also tested to see whether changes in the degree of regional functional connectivity within any given brain structure might have influenced the system as a whole. Collectively, these findings indicated that there was no systematic influence of within‐network integration of particular brain regions over system causal density. Instead, as different regions specific to each network were more greatly integrated into the functional circuit, causal density was higher. Therefore, it is possible that some of these regions may lessen their engagement with these circuits throughout adolescent development, playing a small role in influencing the overall system dynamics. Fair et al. recently observed that default mode regions are increasingly integrated into a cohesive, interconnected network across adolescent development [Fair et al.,2008]. The current results add to that finding by observing that changes specifically in regional integration of bilateral dorsolateral prefrontal cortex into the default mode (Component 7) was associated with reductions in overall system causal density.

The methods we employed to address our primary questions imposed several limitations to the current study. Foremost, although the metrics we used successfully characterized the overall systemic interactions among functionally integrated networks, it is important to recognize that statistical evidence for influence of one network over another does not necessarily prove causality. Granger causality estimates do provide evidence for a directed relationship, indicating the likely influence of activity in one network over another. Moreover, consistent changes in the strength of these relationships were found among networks identified in both adults and teens, which strengthens the argument that the age‐related findings are specific to interaction strength. However, the relationship of two hemodynamic signals theoretically could be influenced by physiological factors (e.g., age differences in blood supply, etc.), despite the encouraging results of ICA at disentangling these factors [Birn et al.,2008]. Second, the analyses are based on the assumption that the conjoint profile of signal change seen across brain regions within a given network implies that those regions act in unison to mediate task demands. Given the correspondence of activity in functionally specialized brain regions to task demands in dozens of previous ICA studies (e.g., visual network engagement during visual aspects of tasks, etc.), we believe this is a reasonable assumption for hemodynamic data. Third, we did not employ methods that allowed us to describe if or how these networks may form a neural hierarchy. Hierarchical arrangement functional networks recently has been proposed as the likely organizational principle that underlies complex system interaction [Ioannides,2007]. Fourth, because networks were measured using ICA, the distinction between hemodynamic increases versus decreases is not meaningful between components. Because this was a passive paradigm, there was no way to infer the true directionality of BOLD signal change by reference to paradigm demands. This limitation prevents us from making strong conclusions regarding the relationship of activity within specific brain regions in one component with those in another. Therefore, we have limited our interpretation of the data to discussing age‐related changes in strength of network effective connections. Fifth, because we focused on exploring developmental effects on numerous indices of network function, we did not explore other possible individual differences (e.g., gender, intellectual ability, etc.). Instead, we balanced gender in our age‐groups and ensured through screening tests that patients had no lower than average IQ estimates. These other factors are interesting and could influence brain activity measurements. However, detailed exploration of these factors must wait for future research focused on fully exploring these questions. Finally, it is worth noting that the exact neural mechanisms that promote synchronized activity within networks and the substrates through which causal influences operate are the topics of active research. It will require additional studies of age‐related changes to functional network connectivity using a variety of neuroscience measurements before firm conclusions can be made about the mechanisms of age‐related changes in system function.

Supporting information

Additional Supporting Information may be found in the online version of this article.

Table 2. List of brain regions in each component identified during ‘resting state’ fMRI. All regions surpass corrections for multiple comparisons (p < .000001 FDR) in SPM2 one‐sample t test.

Fig. 1. This is a sample causal connectivity graph for n=100 participants created using tools published by Seth (2005) to examine Granger Causality among the 13 ‘resting state’ fMRI neural network time courses we examined in this study. Each node is labeled using the component number generated from the ICA analysis. Red paths denote significant (p < .001) bidirectional paths while green paths indicate significant unidirectional connections. The thickness of each path represents the relative strength of each causal connection. As noted above in the response to Rev #1, Comment 10, this type of figure cannot usefully convey information about adolescent and adult differences in system causal relationships. Not only are there too many nodes to make the figure graphically appealing and clearly interpretable, but also the type of information represented does not lend itself to conveying specific age‐related findings. We favor Manuscript Figure 3 to present these findings.