Temporal stability of network centrality in control and default mode networks: Specific associations with externalizing psychopathology in children and adolescents

Abstract

Abnormal connectivity patterns have frequently been reported as involved in pathological mental states. However, most studies focus on “static,” stationary patterns of connectivity, which may miss crucial biological information. Recent methodological advances have allowed the investigation of dynamic functional connectivity patterns that describe non‐stationary properties of brain networks. Here, we introduce a novel graphical measure of dynamic connectivity, called time‐varying eigenvector centrality (tv‐EVC). In a sample 655 children and adolescents (7–15 years old) from the Brazilian “High Risk Cohort Study for Psychiatric Disorders” who were imaged using resting‐state fMRI, we used this measure to investigate age effects in the temporal in control and default‐mode networks (CN/DMN). Using support vector regression, we propose a network maturation index based on the temporal stability of tv‐EVC. Moreover, we investigated whether the network maturation is associated with the overall presence of behavioral and emotional problems with the Child Behavior Checklist. As hypothesized, we found that the tv‐EVC at each node of CN/DMN become more stable with increasing age (P < 0.001 for all nodes). In addition, the maturity index for this particular network is indeed associated with general psychopathology in children assessed by the total score of Child Behavior Checklist (P = 0.027). Moreover, immaturity of the network was mainly correlated with externalizing behavior dimensions. Taken together, these results suggest that changes in functional network dynamics during neurodevelopment may provide unique insights regarding pathophysiology. Hum Brain Mapp 36:4926–4937, 2015. © 2015 Wiley Periodicals, Inc.

INTRODUCTION

In the last few years, neuroimaging studies using network approaches have provided novel insights into the organization of brain structure [Bullmore and Sporns, 2009; Sporns et al., 2005] and function [Power et al., 2013]. It has been long known that the functional brain networks have non‐stationary properties [Sato et al., 2006a; Smith et al., 2012; Whitlow et al., 2011], which may be modified by cognitive states [Moussa et al., 2011], reflecting the reconfiguration of the network in preparation of current environmental demands. Abnormal dynamic connectivity patterns have been associated with pathological states such as absence seizures [Liao et al., 2013] and schizophrenia [Ma et al., 2014]. In this context, current approaches consider the average of brain interactions in extended periods (fMRI session). However, this focus on stationary signals may miss biological information. Recent methodological have allowed for the description of networks functioning in a time‐varying fashion [Cribben et al., 2013; Eavani et al., 2013; Kang et al., 2011; Sato et al., 2006a, 2006b; Sotero and Shmuel, 2012] which has attracted the interest of scientific community [Hutchison et al., 2013]. From a biological perspective, it has been suggested that the brain modifies its configuration across time as a way to minimize metabolic requirements, while maintaining the brain in a responsive state [Zalesky et al., 2014].

Human brain structural and functional networks exhibit a complex but consistent developmental trajectory [Dosenbach et al., 2010; Franke et al., 2012]. Studies have shown consistent developmental changes in static functional networks [Fair et al., 2007; Sato et al., 2014, in press; Satterthwaite et al., 2013b; Zhong et al., 2014], and their behavioral correlates [Velanova et al., 2008]. A compelling further hypothesis states that deviations of such a typical developmental trajectory may lead to psychiatric disorders, as evidenced by structural neuroimaging studies [Shaw et al., 2013].

The Default Mode (DMN) and Control (CN) networks have attracted particular interest, due to their central roles in internally‐ and externally‐directed cognition [Fornito et al., 2012; Satterthwaite et al., 2013a]. Notably, they are among the most commonly implicated networks in a myriad of mental symptoms and disorders [Broyd et al., 2009 ; Chabernaud et al., 2012; Cortese et al., 2012; Sripada et al., 2014]. In this context, we sought to include the time component into an analysis of functional networks, and investigate whether the manifestation of psychopathology in children is related to the stability of the dynamic functional connectivity between nodes of the Control and Default Mode Network (CN/DMN).

Rather than addressing temporal changes (at an fMRI session timescale) on individual connections, we focused on one network characteristic that integrates much of this information, namely the eigenvector centrality (EVC) of brain regions [Bonacich, 1987; Sato et al., in press]. This measure is extracted for each time‐point during a resting state fMRI scanning session, providing a time‐varying EVC (tv‐EVC) for each node of the network. We first hypothesized that the stability of the CN/DMN connections increased with age, and that the variance across time of the nodes’ centrality would predict age using multivariate statistical learning techniques. Furthermore, we hypothesized that deviations in the maturation of these particular functional networks, reflected in the time domain, would be associated with psychiatric symptoms.

MATERIAL AND METHODS

Sample

This study is part of a larger community sample (the Brazilian “High Risk Cohort Study for Psychiatric Disorders,” HRC) consisting of a total of 2.512 students aged between 7 and 15 years old, from 57 public schools in the cities of Porto Alegre and São Paulo, Brazil, The main aim of this large cohort study is to explore the neurodevelopmental trajectories subsiding psychopathological phenomena and mental disorders [Salum et al., in press; Sato et al., 2014, in press]. From the whole cohort of 2,512 participants, 741 subjects participated in MRI scanning. Some of these subjects (86) were discarded due to excessive motion during acquisition (as observed by the technician) or because they did not completed the MRI session (e.g., due to anxiety or discomfort). From the remaining participants, 655 completed the resting state fMRI acquisition with successful preprocessing steps.

The mean sample age was 10.70 years (s.d. = 1.89; median = 10.54), 344 male (52.52%) and 558 (85.32%) right handed. Mean IQ was 102.5 (s.d = 16.72; median = 100); estimated using vocabulary and block subtests of the Weschler Intelligence Scale for Children [Tellegen and Briggs, 1967; Wechsler, 1991]. Ethnicity was assessed as declared by the mother: 60% Caucasian/white, 11% African‐American, 28% mixed background, and 1% others (i.e., Asian, native‐American). Socioeconomic status was assessed with the Brazilian rating scale (ABIPEME): 4.28% were classified as very low/low (E and D classes), 66.82% medium (C and B classes), and 28.9% comfortable (A class) groups. Further information is shown in Figure 1 (see Table 1 for detailed summary statistics). Finally, we have not discarded any of the 655 subjects due to excessive motion after preprocessing. Actually, from these, only five subjects presented a median frame‐displacement greater than 0.5 mm being the maximum of 0.94 mm. Thus, we prefer to conduct the analysis using all subjects of our sample without further exclusions involving arbitrary choices.

Figure 1.

Processing flowchart. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Table 1.

Summary statistics for demographical and behavioral data

	Min.	1st Qu.	Median	Mean	Std. Dev	3rd Qu.	Max.	Missing
Age	7.0	9.3	10.5	10.7	1.9	12.0	15.0	0
IQ	52.0	91.0	100.0	102.5	16.7	114.0	152.0	3
Anxiety	0.0	3.0	6.0	7.0	5.0	10.0	24.0	3
Withdrawn	0.0	1.0	3.0	3.4	3.3	5.0	16.0	3
Somatic	0.0	1.0	3.0	3.6	3.3	5.0	19.0	3
Social	0.0	2.0	3.0	4.3	3.7	6.0	19.0	3
Thought	0.0	1.0	3.0	3.9	4.1	6.0	25.0	3
Attention	0.0	3.0	6.0	6.8	4.9	10.0	20.0	3
Rule Breaking	0.0	1.0	2.0	2.9	3.1	4.0	19.0	3
Aggression	0.0	4.0	8.0	10.1	7.5	15.0	35.0	3
Internalizing	0.0	7.0	11.0	14.1	9.9	20.0	55.0	3
Externalizing	0.0	5.0	10.0	13.0	10.1	18.0	51.0	3

Written consents were obtained from all the parents or legal guardians and all the children provided verbal assent. The University of São Paulo's Ethical Committee approved the experimentation protocol.

Data and Image Acquisition

Before scanning, children were engaged in recreational activities (including playing inside a fabric tunnel and training to keep still) in order to minimize anxiety. Imaging was performed using 1.5T MRI scanners (GE, Signa HDX and HD). High‐resolution T1‐weighted images up to 156 axial slices (TR = 10.916 ms, TE = 4.2 ms, thickness = 1.2 mm, flip angle = 15°, matrix size = 256 × 192, FOV = 245 mm, NEX = 1, bandwidth = 122.109) were acquired. BOLD fMRI data were acquired as a series of 180 echo planar imaging volumes (TR = 2,000 ms, TE = 30 ms, slice thickness = 4 mm, gap = 0.5 mm, flip angle = 80°, matrix size = 80 × 80, FOV = 240 mm, reconstruction matrix 128 × 128, 1.875 × 1.875 mm, NEX = 1, slices = 26) during a resting state protocol with eyes open at a fixation point (small cross). The duration of the resting state acquisition was 6 min.

Behavioral Assessment

On the same day of the MRI scanning, a parent or caregiver completed the Child Behavior Checklist [CBCL; Achenbach and Rescorla, 2001] translated to Portuguese. This tool provides a score for general psychopathology in children. In addition, by combining the different sub‐items it is possible to obtain internalizing (emotional) and externalizing (behavioral) symptom scores. Factor analysis in diverse samples and cultures [Heubeck, 2000] yielded an 8‐factor solution consisting in eight dimensions: anxiety/depression, withdrawn/depression, somatic complaints, social problems, thought problems, attention problems, rule‐breaking behavior, and aggressive behavior (see Fig. 1). The general psychopathology (total CBCL score) is the main focus of this study. The analyses of other domains were carried out only for exploratory purposes and in order to facilitate the interpretation of the findings relating to the total score.

From a household interview of the parents, child psychiatric diagnosis was obtained using the Development and Well‐Being Assessment [DAWBA, Goodman et al., 2000]. However, whereas the CBCL was obtained on the day of the imaging session, the average interval between the DAWBA and the scanning session was approximately 8.5 months. Due to this delay and the dimensional measures provided by the CBCL, in this article we used the CBCL as our primary behavioral measure. However, the diagnostic concordance between the CBCL and the DAWBA was excellent (AUC 0.832; CI 95% 0.814–0.850).

Image Preprocessing

Conventional preprocessing procedures were performed using AFNI v.2011 [Cox, 2012] and FSL v.5, [Jenkinson et al., 2012]. The following steps were conducted for BOLD time series: discarding the first four volumes; realignment for head‐motion correction; despiking; band‐pass filtering for 0.01 Hz and 0.1 Hz; detrending; spatial smoothing (Gaussian kernel, FWHM = 8 mm) and linear registration to the respective structural scan. Structural images were skull‐stripped and nonlinearly registered to the MNI152 standard template using FSL's FNIRT. Nuisance covariates (mean CSF, white matter and global signals and the six linear parameters of motion) were regressed out from the functional data. Although MNI152 is an adult template, Burgund et al. [2002] have demonstrated the feasibility of using this template from 7 years old, with minimal warping differences when compared to adults. BrainNetViewer (http://www.nitrc.org/projects/bnv/) was used for visualization purposes.

The analysis pipeline following the preprocessing steps comprised: (i) modeling time‐varying functional connectivity; (ii) calculating dynamic graph centrality; (iii) obtaining a network maturity index based on centrality stability; and (iv) testing for associations between the maturity index and behavioral/emotional problems dimensions. A flowchart depicting this pipeline is shown in Figure 2.

Figure 2.

Histograms of demographical and behavioral information of this sample.

Network Modeling

The regions‐of‐interest (ROIs, see Fig. 3) were defined using the coordinates from Dosenbach et al. [2006, Table III], which identifies a set of areas that consistently present a positive (control network, CN) or a negative (default‐mode network, DMN) BOLD response across several goal directed tasks. These included 15 spherical ROIs with 4 mm radius (8 mm diameter). The use of larger nodes would risk partial volume artifacts due to inclusion of white matter and non‐brain signal. Supporting Information Table S1 describes the coordinates of these regions. We refer to this set of ROIs as the CN/DMN network.

Figure 3.

Top: Regions of interest in anatomical space, with their size proportional to their contribution (measured by the correlation between predicted age and the standard deviation of tv‐EVC at each region) to the maturation index. Bottom: two illustrative examples of two subjects with high and low temporal stability of tv‐EVC at left posterior cingulate, aI/fO and dACC/msFC. Abbreviations: dACCmsFC: dorsal anterior cingulate/medial superior frontal cortex; aIfO: anterior insula/frontal operculum. Obs: The two subjects were matched by the amount of head motion. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Given regions‐of‐interest (ROI) X and Y from the CN/DMN network, dynamic connectivity analysis was carried out using a spline‐based smoothing estimator for time‐varying correlations. This estimator is given by:

$${\widehat{cor}}_{X,Y}\left(t\right)=\frac{{\widehat{cov}}_{X,Y}\left(t\right)}{{\widehat{sd}}_X\left(t\right){\widehat{sd}}_Y\left(t\right)},$$

where

$${\widehat{cov}}_{X,Y}\left(t\right)=\phi \left[\left(X_t‐\acute{X}\left(t\right)\right)‐\left(Y_t‐\acute{Y}\left(t\right)\right)\right],$$

$${\widehat{sd}}_X\left(t\right)=\sqrt{\phi \left[{\left(X_t‐\acute{X}\left(t\right)\right)}^2\right]},$$

$${\widehat{sd}}_Y\left(t\right)=\sqrt{\phi \left[{\left(Y_t‐\acute{Y}\left(t\right)\right)}^2\right]},$$

$$\acute{X}\left(t\right)=\phi \left[X_t\right],$$

$$\acute{Y}\left(t\right)=\phi \left[Y_t\right],$$

and where φ is the splines smoothing estimator with smoothing parameter (number of knots) tuned using generalized cross‐validation [Craven and Wahba, 1979; Golub et al., 1979]. The “smooth.spline” function from “splines package” of R platform [http://www.r-project.org/] was used to implement this procedure. By using this smoothing spline procedures (based on curve fitting), it was possible to estimate the local mean, standard deviation, covariances and correlations for each time‐point during the resting state session.

For each scan time‐point, we obtained an undirected weighted graph (without using any threshold or categorization). Each graph was then built considering the absolute values of ROIs’ pairwise time‐varying correlation (at the respective scan) as the weights. The eigenvector centrality of each node [EVC; Bonacich, 1987] was then calculated for each time‐point, the time‐varying EVC (tv‐EVC). The EVC of a node is a measure from graph theory that quantifies the relevance of a node in the context of the whole network, from a hierarchical perspective. Thus, The EVC of a node is proportional to the centrality of its neighbors. Therefore, this measure is suitable to investigate networks in which hierarchical relationships play a central role in the dynamics of information flow. A good property of EVC is that it accounts for the fact that connections to high‐scoring nodes are more important than connections to low‐scoring nodes. Moreover, the use of correlations as the connectivity strength for the weighted graph is a straightforward choice. Other centrality measures (e.g., closeness, betweenness, etc) require the definition of distances based on the correlations (e.g., 1‐r; 1/r; etc), which may be arbitrary.

Estimation of Network Maturity

Given our focus on the temporal stability of hierarchical organization in CN/DMN, we calculated the standard deviation (across time) EVC at each node for each subject. By using the time‐varying correlation at each time‐point, we have an EVC of each node and at each time‐point. The standard deviation (across scans) of tv‐EVC represents the node temporal stability regarding the hierarchical relationships of a functional network.

The standard deviation of tv‐EVC of each node were then used as input to a linear support vector regression [SVR, Smola and Scholkopf, 1998; as coded in libSVM; http://www.csie.ntu.edu.tw/~cjlin/libsvm; Chang and Lin, 2011] to predict the age of the subjects. In other words, for each subject, the standard deviation of tv‐EVC of all included nodes (fifteen variables) was used as a regressor to predict the respective chronological age of the subject. By considering all subjects, we then defined a maturity index as the predicted age, from a leave‐one‐subject‐out procedure. In order to quantify the contribution of each ROI in age prediction, we calculated the correlation between predicted age and the standard deviation of tv‐EVC at each region. For comparison purposes with the “static” EVC, the same procedure was also carried out considering the standard EVC, that is, by calculating the pairwise Pearson correlations (of ROIs’ BOLD signal) of the whole run.

Statistical Analysis

The association between the chronological and predicted age (maturity index) was assessed using Pearson correlation. For each node, the association between the standard deviation of tv‐EVC and age was assessed using a General Linear Model (GLM) considering the former as the response variable, age as the main regressor and site, mean frame displacement [FD, Yan et al., 2013], Eq. (9), and standard deviation of FD as nuisance variables. Type I error was set to 5% assuming FDR correction for multiple comparisons [Benjamini and Hochberg, 1995]. The consideration of FD as a possible confounder was necessary, since head movement is one of the main sources of artifacts in functional connectivity measures [Power et al., 2012; Satterthwaite et al., 2012]. In addition, we expect the time‐varying correlations (and thus the EVC) to be also influenced by motion.

The association of CN/DMN maturity index and psychopathology was carried out using a GLM considering the total score of CBCL as the response variable, the maturity index as the main regressor and age, site, mean FD and standard deviation of FD as nuisance variables. Type I Error was set at 5% for this single comparison.

Additional exploratory analyses were carried out by repeating the previous GLM considering the externalizing, internalizing, anxiety, withdrawn, somatic, social, thought, attention, rule‐breaking, and aggression scores. Given the exploratory nature of these analyses, which were carried out solely to facilitate the interpretation of the main finding (total score CBCL), these results were not corrected for multiple comparisons.

Finally, in order to evaluate whether our findings with CBCL were specific to CN/DMN, we repeated all steps by using the sensory‐motor/visual ROIs described in Zhang et al. [2014], Supporting Information Table S1.

RESULTS

Age Effect in the Stability of the Network Role of Each ROI

We first examined how the variability of nodal tv‐EVC changed with age. The SD of tv‐EVC of all CN/DMN regions were found to be negatively correlated with age (P < 0.05, FDR corrected). Thus, the findings indicate that the network's hierarchical organization becomes more stable with age. Next, we explored whether the multivariate pattern of tv‐EVC variability provided important information related to maturation. A cross‐validated SVR revealed a significant positive correlation between actual age and predicted age (r = 0.109, P‐value = 0.005, see Supporting Information Fig. 1). Notably, neither sex nor socioeconomics status significantly impacted predicted age (P > 0.05). In addition, Figure 3 describes the contribution of each region‐of‐interest in the maturity index (see also Supporting Information Table S1). Note that with the exception of anterior fusiform region, all regions have a high contribution to the index. The anterior insula/operculum at both hemispheres were the regions with the highest contributions.

Association with Behavioral Data

As hypothesized, the CN/DMN maturity index was found to be negatively correlated with total CBCL score (P‐value = 0.027), indicating that poor maturation is associated with general behavioral and emotional problems. This result remained the same (P‐value = 0.029) if sex was included as covariate in GLM. Exploratory analyses across all CBCL dimensions revealed that most symptoms are related to externalizing factors (see Table 2). All estimated betas for CBCL dimensions were negative.

Table 2.

P‐values (and unstandardized betas) of the general linear model analyses with all CBCL factors

	Brain maturity P‐value (beta)		Age	Site	Mean FD	s.d. FD
Total	0.027	(−5.02)	0.926	<0.001	0.636	0.306
Internalizing	0.151	(−1.06)	0.705	<0.001	0.637	0.491
Externalizing	0.039	(−1.56)	0.473	<0.001	0.847	0.339
Anxious/depressed	0.171	(−0.51)	0.132	<0.001	0.216	0.126
Withdrawn/depressed	0.268	(−0.28)	0.002	<0.001	0.968	0.936
Somatic Complaints	0.279	(−0.27)	0.765	<0.001	0.622	0.867
Social Problems	0.093	(−0.47)	0.267	<0.001	0.346	0.245
Thought Problems	0.053	(−0.59)	0.187	<0.001	0.765	0.854
Attention Problems	0.027	(−0.82)	0.335	<0.001	0.537	0.380
Rule‐breaking behav.	0.014	(−0.58)	0.107	<0.001	0.756	0.339
Aggressive behav.	0.082	(−0.99)	0.773	<0.001	0.897	0.380

The highlighted cells depict the domains with significant associations (P < 0.05) with the brain maturity. Note that most domains with statistically significant effects refer to externalizing symptoms.

In contrast, the static EVC did not yielded significant results for the total CBCL score (P = 0.641) or any CBCL dimensions (all P > 0.1), reinforcing the importance of considering dynamic features such as the network temporal stability.

By repeating the same approach, we have not found significant associations between the maturation of sensory‐motor/visual and any CBCL score/dimension.

Motion Analysis

There was a significant association (P < 0.001 for all ROIs) between the temporal standard‐deviation of tv‐EVC and FD (both mean and standard deviation). As a consequence, both head motion measures are also associated with the maturity index. However, Table 2 demonstrates that the CBCL scores (which are the main outcome) were not correlated with the motion parameters. Moreover, this analysis highlights that although the maturity index is influenced by motion, it is associated with CBCL even when controlling for FD (both mean and standard deviation).

DISCUSSION

Previous studies have investigated age effects in typically developing children [Fair et al., 2007, 2008; Sato et al., 2014, in press; Satterthwaite et al., 2013b] showing the relevance of this period in neurodevelopment and networks formation. Here, we investigated whether the stability of the interactions between CN/DMN nodes changed across childhood and adolescence. Moreover, we also explored whether abnormal developmental trajectories of these measures were related to general psychopathology in a large sample of children. As hypothesized, the network role of the CN/DMN regions became more temporally stable with increasing age. Furthermore, the maturity index for these networks correlated with general psychopathology in children, in particular externalizing symptoms. To the best of our knowledge, this is the first report of abnormal development of CN/DMN dynamic connectivity predicting psychopathology in a community‐based sample.

The finding of increasing network stability with age is in line with a growing body of evidence describing the pattern of developmental trajectory of neural functional networks [Power et al., 2011; Vogel et al., 2010]. However, it is important to emphasize that most studies of functional connectivity across early development assume static, stationary connectivity. Here we explored time‐varying properties of these networks, which may be an essential aspect of functional brain development. Our results of increased stability with age echo a previous study [Hellyer et al., 2014] that used a complex dynamic system approach to explore the relationship between changes in whole‐brain meta‐stability and shifts between a focused or attentional state and an unfocused exploratory state. This study showed that increasing activity in the CN reduced the brains’ global meta‐stability (i.e., produced more stable neural dynamics), while increasing activity of DMN had the opposite effect. In other words, activities of CN and DMN nodes seem to be critical to actively shift between two global neural dynamics: a “resting” state, in which information capacity is maximized and stability is compromised, and an active attentional state in which stability is increased and information processing is more focused. We believe that our finding of increasing stability of CN/DMN network during development might reflect the acquisition of the capacity to actively change between an exploratory resting state and a focused task‐set state.

The CN was originally conceived as a network of brain regions critical for implementing maintenance of a task sets [Dosenbach et al., 2006]; further analysis demonstrated that this function was primarily implemented cingulo‐opercular regions [Dosenbach et al., 2007]. An alternative perspective for the functional role of anterior Insula considers this region as a hub that mediates dynamic interactions between large‐scale functional networks [Menon and Uddin, 2010]. In this view, aI/fO and dACC are nodes of a salience network that respond to the most relevant external or internal stimuli [Menon and Uddin, 2010; Seeley et al., 2007]. Remarkably, activation of right aI/fO was found to cause deactivation of the DMN and activation of fronto‐parietal CN [Sridharan et al., 2008], suggesting that aI/fO is critical for switching between large‐scale functional networks. Our finding of correlation between externalizing problems and impaired development of stability in the connections between aI/fO, dACC, and DMN nodes is also in agreement with this perspective. Sripada et al. [2014] suggest a developmental lag of the salience network may be associated with attentional problems. For instance, impaired development of aI/fO may produce distractibility by asserting excessive salience for irrelevant environmental stimuli. Such impairment could also explain attentional lapses, if the maturation of aI/fO leads to the ability to deactivate DMN [Christoff et al., 2009].

Our results seem accord with the current developmental model for the etiology of Attention‐Deficit Hyperactivity Disorder [ADHD, Fair et al., 2012] and other psychiatric disorders. In fact, a significant and specific maturational lag in CN/DMN connections in ADHD subjects when compared to typically developing controls was recently reported [Sripada et al, 2014]. However, as we performed a cross‐sectional assessment of a non‐clinical sample, predictive longitudinal analyses of psychopathology were not possible. Further longitudinal analysis of the same cohort may clarify the impact of atypical development of dynamic networks.

An important point to be mentioned is that the subjects participating in this study are from an under‐investigated population from a Latin American country. Brazilian population presents high genetic heterogeneity and environmental and socioeconomic conditions that are not typical in European or North American populations. This is relevant since mental disorders may be more prevalent in children from developing countries [Kieling et al., 2009].

One important point to be mentioned is that the correlation between chronological and predicted age (maturity index) was not high (r = 0.109), although significant. We believe that two main factors were driving this low correlation. The first is the inherent high heterogeneity of neurodevelopment during childhood and adolescence. In addition, this sample also includes subjects with psychopathology, which increases the heterogeneity. The second factor is that we are solely using nodes from CN/DMN networks and not the whole brain. However, these ROIs were in order to ensure the specificity of findings and interpretability of results.

Some limitations of this work need to be addressed. Motion artifacts impact in functional connectivity measures [Power et al., 2012; Satterthwaite et al., 2012]. This issue was at least partially assessed here by considering motion parameters as covariates in the statistical analysis [Yan et al., 2013]. However, since time‐varying correlations are based on a smoothing procedure involving all observed time‐points, it was not possible to implement scrubbing in this case.

CONCLUSIONS

Taken together, we conclude that the temporal stability of CN/DMN hierarchical organization increases with age in a large sample of children and adolescents. In addition, the delayed maturation of stability within these is associated with overall psychopathology, specially externalizing symptoms.

DISCLOSURES

João R. Sato, Euripedes C. Miguel, and Andrea P. Jackowsky are supported by grants of São Paulo Research Foundation (FAPESP).

Dr. Luis Augusto Rohde is supported by grants of CNPq and he has been on the speakers’ bureau/advisory board and/or acted as consultant for Eli‐Lilly, Janssen‐Cilag, Novartis and Shire in the last 3 years. The ADHD and Juvenile Bipolar Disorder Outpatient Programs chaired by him received unrestricted educational and research support from the following pharmaceutical companies in the last 3 years: Eli‐Lilly, Janssen‐Cilag, Novartis, and Shire. He receives authorship royalties from Oxford Press and ArtMed.

Dr. Rodrigo A. Bressan has been on the speakers’ bureau/advisory board of AstraZeneca, Bristol, Janssen and Lundbeck; he has received research grants from Janssen, Eli Lilly, Lundbeck, Novartis, Roche, FAPESP, CNPq, CAPES, Fundação E.J. Safra, and Fundação ABAHDS. He is a shareholder of Biomolecular Technology Ltda.

Dr. Edson Amaro Jr. has received research grants from FAPESP, CNPq, CAPES, Fundação E.J. Safra, and Fundação ABAHDS.

Theodore D. Satterthwaite was supported by NIMH K23MH098130.

All other authors reported no biomedical financial interests or potential conflicts of interest.

Supporting information

Supporting Information

Supporting Information