Journal of Child Psychology and Psychiatry 57:1 (2016), pp 55–64
doi:10.1111/jcpp.12444
Default mode network maturation and psychopathology in children and adolescents ~ o Ricardo Sato,1,2,3,4 Giovanni Abraha ~ o Salum,4,5 Ary Gadelha,2,4 Nicolas Crossley,6,7 Joa 3,8 4,5 Zugman,2,4 Felipe Almeida Picon,4,5 Pedro Gilson Vieira, Gisele Gus Manfro, Andre 2,4 2,4,9 s,4,5 Luciana Monteiro Moura,2,4 Mario Pan, Marcelo Queiroz Hoexter, Mauricio Ane 2,4 Marco Antonio Gomes Del’Aquilla, Edson Amaro Jr,3 Philip McGuire,6 Acioly Luiz 2,4 Tavares Lacerda, Luis Augusto Rohde,4,5 Euripedes Constantino Miguel,4,9 Andrea Parolin Jackowski,2,4 and Rodrigo Affonseca Bressan2,4 1
Center of Mathematics Computation and Cognition, Universidade Federal do ABC, Santo Andre; 2Interdisciplinary Lab for Clinical Neurosciences (LiNC), Universidade Federal de Sao Paulo (UNIFESP), Sao Paulo; 3Department of Radiology, School of Medicine, University of Sao Paulo, Sao Paulo; 4National Institute of Developmental Psychiatry for Children and Adolescents, CNPq, Sao Paulo; 5Department of Psychiatry, Federal University of Rio Grande do Sul, Porto Alegre, Brazil; 6Institute of Psychiatry, King’s College London, London, United Kingdom; 7Institute for Biological and Medical Engineering, Faculties of Engineering, Medicine and Biological Sciences, P. Catholic University of Chile, Santiago, Chile; 8Bioinformatics Program, Institute of Mathematics and Statistics, University of Sao Paulo, Sao Paulo; 9Department of Psychiatry, School of Medicine, University of Sao Paulo, Sao Paulo, Brazil
Background: The human default mode (DMN) is involved in a wide array of mental disorders. Current knowledge suggests that mental health disorders may reflect deviant trajectories of brain maturation. Method: We studied 654 children using functional magnetic resonance imaging (fMRI) scans under a resting-state protocol. A machinelearning method was used to obtain age predictions of children based on the average coefficient of fractional amplitude of low frequency fluctuations (fALFFs) of the DMN, a measure of spontaneous local activity. The chronological ages of the children and fALFF measures from regions of this network, the response and predictor variables were considered respectively in a Gaussian Process Regression. Subsequently, we computed a network maturation status index for each subject (actual age minus predicted). We then evaluated the association between this maturation index and psychopathology scores on the Child Behavior Checklist (CBCL). Results: Our hypothesis was that the maturation status of the DMN would be negatively associated with psychopathology. Consistent with previous studies, fALFF significantly predicted the age of participants (p < .001). Furthermore, as expected, we found an association between the DMN maturation status (precocious vs. delayed) and general psychopathology scores (p = .011). Conclusions: Our findings suggest that child psychopathology seems to be associated with delayed maturation of the DMN. This delay in the neurodevelopmental trajectory may offer interesting insights into the pathophysiology of mental health disorders. Keywords: Neurodevelopment, default mode network, neuroimaging, psychopathology, MVPA.
Introduction Previous investigations have suggested that neuropsychiatric disorders are deviations from typical neurodevelopmental trajectories (Ducharme et al., 2012, 2014; Insel, 2010; Shaw et al., 2006, 2007; Thompson et al., 2001; Uddin, Supekar, & Menon, 2010). The results of such deviant trajectories can be observed already in childhood and adolescence, when the majority of mental disorders begin (KimCohen et al., 2003). Thus, investigation of the typical brain is important not only to achieve a better understanding of the neural substrates of mental disorders but also to develop early intervention strategies (Cannon et al., 2008). Functional magnetic resonance imaging (fMRI) has played a key role in understanding the brain circuitry changes that take place throughout neurodevelopment (Power, Fair, Schlaggar, & Petersen,
Conflict of interest statement: See Acknowledgements for disclosures.
2010; Sato et al., 2015; Smyser, Snyder, & Neil, 2011; Uddin et al., 2010; Vogel, Power, Petersen, & Schlaggar, 2010). Previous resting-state fMRI studies have found that certain functional networks are associated with developmental disorders (Castellanos et al., 2008; Sato, Hoexter, Castellanos, & Rohde, 2012). Recent studies have investigated the changes that occur to the brain networks during their maturation process in different stages of childhood and adolescence; they have examined the default mode network (DMN), in particular. The default mode network (DMN, Raichle et al., 2001; Greicius, Krasnow, Reiss, & Menon, 2003; Buckner, Andrews-Hanna, & Schacter, 2008) is usually associated with introspection and is less active in externally directed tasks (McGuire, Paulesu, Frackowiak, & Frith, 1996). Fair et al. (2008) found that the functional connectivity within the DMN was very sparse between 7–9 years of age but the intensity of connectivity among these regions increased with age (Sato et al., 2014). In addition, two multimodal studies investigating the functional
© 2015 Association for Child and Adolescent Mental Health. Published by John Wiley & Sons Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main St, Malden, MA 02148, USA
56
Jo~ ao Ricardo Sato et al.
and structural connectivity of DMN in children reported convergent results from both modalities. The findings suggest that the functional connectivity between the posterior cingulate and medial prefrontal cortex (PCC-mPFC) is the most immature connection in young children, and the strength of this connection was correlated with fractional anisotropy of the cingulum bundle (Gordon et al., 2011; Supekar et al., 2010). Interestingly, the DMN has also emerged as being associated with a myriad of psychiatric diseases (Broyd et al., 2009; Chang et al., 2014). Thus, it is important to investigate associations between psychopathological symptoms and the functional maturation of this network. One possible approach to investigating neurodevelopment is by using machine-learning methods. Resting-state fMRI studies based on machine-learning techniques have been able to predict subjects’ chronological ages (from 7- to 30-years old) (Dosenbach et al., 2010; Supekar, Musen, & Menon, 2009; Wang, Su, Shen, & Hu, 2012), even after controlling for head movement (Satterthwaite et al., 2013). The handling of motion artifacts is indeed a challenge in neuroimaging. One possible approach in the analysis of resting-state fMRI might be to use fractional amplitude of low frequency fluctuations (fALFF, Zou et al., 2008), found to be less affected by motion artefacts (Yan, Craddock, Zuo, Zang, & Milham, 2013; Yan, Cheung et al., 2013). In this hypothesis-driven study, we investigated whether the DMN maturation status could be associated with behavioural disturbances in a large nonreferred sample of 654 subjects ranging in age from 6- to 15-years old using a machine-learning method on resting-state fMRI data. This developmental period seems to be crucial in the formation of DMN (Fair et al., 2008; Sato et al., 2014). We first used machine-learning techniques to predict the chronological age of subjects based on the fALFF of regions-of-interests belonging to the DMN. We then defined a ‘maturational status’ for each subject, based on the difference between the predicted and actual age of the participants. We hypothesized that children and adolescents whose predicted age was much younger than their actual age, which we described as a delayed maturation pattern, would present higher general psychopathological manifestations when compared to those with a more precocious developmental status.
Methods Subjects Students from 22 state schools in the city of Porto Alegre and 35 in S~ ao Paulo were assessed using a screening procedure. For a detailed description of the Brazilian High Risk Cohort (HRC), see Salum et al. (2013, 2014). The ethics committee of the University of S~ ao Paulo approved the study (IORG0004884, 1138/08), and parents/guardian and the children provided written consent and verbal assent respectively.
J Child Psychol Psychiatr 2016; 57(1): 55–64 In the current investigation, 654 subjects (352 male; 53.8%) from the HRC were submitted to a resting-state fMRI protocol. The mean age of the sample was 10.71 years (SD = 1.89), and mean IQ was 102.58 (SD = 16.72). The vocabulary and block design subtests of the Wechsler Intelligence Scale for Children (Wechsler, 1991) were used to estimate IQ, using the method described in Tellegen and Briggs (1967). The sample averaged 4 years of education (SD = 1.67), and 558 subjects (85.3%) were right-handed. As declared by the mother), the sample was 60% Caucasian, 10.55% Afro-Brazilian, 28.29% mixed African background and 1.16% were classified as ‘other’. The Brazilian rating scale (ABIMEP) was used to determine subjects’ socio-economic status classifications, which were as follows: 4.28% were low to very low (E and D classes), 66.82% were medium (C and B classes) and 28.9% were comfortable (A class) groups. Subjects were also assessed for psychiatric disorders using the Development and Well-Being Assessment (DAWBA, Goodman, Ford, Richards, Gatward, & Meltzer, 2000). Of the 654 participants, 447 did not present any disorder. Because this study focuses on a dimensional approach to psychopathological manifestation, none of the subjects were excluded from the analyses. Finally, 356 participants reported a familiar history of psychiatric conditions based on the Family History Survey (FHS).
CBCL evaluation and description of the bi-factor model Parents/caregivers completed the Child Behavior Checklist (CBCL; Achenbach & Rescorla, 2001) on the day of MRI scanning. This is a well-established assessment tool containing short statements (e.g.: ‘he talks too much’ and ‘he cries too much’) with three different levels of agreement/frequency for each item. Raw internalizing and externalizing factors from the CBCL are usually correlated, which can be a problem regarding interpretation because it is not clear whether the findings are related to the general factor or to the specific internalizing/ externalizing construct. We therefore fitted a bifactor model to the responses using structural equation modelling. The model considers that the response to each CBCL item is driven by a single general factor accounting for the covariation among all symptoms (general factor), plus a specific factor (internalizing or externalizing, orthogonal to the general factor) and random error (individual variation). This approach model provides a framework to conceptualize both the commonality and the specificity of symptoms from distinct domains (Brunner, Nagy, & Wilhelm, 2012; Holzinger & Swineford, 1937; Tackett et al., 2013). The model was fitted using the structural equation module of the software MPLUS 7.0 (http://www.statmodel.com/). The bifactor model provided a good fit to our data. Briefly, the bifactor models were fitted to polychoric correlations among the CBCL items using the mean- and varianceadjusted weighted least squares (WLSMV) estimator (Muth en & Muth en, 2012). The goodness of fit was assessed through the following fit indices: chi-square, weighted root mean square residual (WRMR), comparative fit index (CFI), Tucker-Lewis Index (TLI) and the root mean square error of approximation (RMSEA). To demonstrate good fit to the data, it is suggested that an estimated model should have a WRMR near or below .9 (Muth en & Muth en, 2012), an RMSEA near or below .06, and both CFI and TLI near or above .95 (Hu & Bentler, 1999). The bifactor model for this sample provided acceptable fit to our data: X2 = 2688.5, df = 1650, p < .0001, number of free parameters = 239, RMSEA = 0.029 (95%CI: 0.027–0.031), CFI = 0.950, TLI = 0.947 and WRMR = 1.279.
Image acquisition On the neuroimaging day, children were engaged in recreational activities as a desensitization method. A mock-up of data acquisition procedures was carried out using a fabric play © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
tunnel and MRI noise, in addition to didactical explanations about the next steps and the importance of not moving during the fMRI procedure. Further information about the protocol can be found in Sato et al. (2014). Imaging was performed using two 1.5T MRI scanners (General Electric, Signa HDX and HD, in the cities of S~ ao Paulo and Porto Alegre, Brazil). To provide T2*-weighted BOLD contrast, fMRI data were acquired with a protocol of a series of 180 echo planar imaging volumes (TR = 2000 ms, TE = 30 ms, slice thickness = 4 mm, gap = 0.5 mm, flip angle = 80 degrees, matrix size = 80 9 80, FOV = 240 mm, reconstruction matrix size = 128 9 128, 1.875 9 1.875 mm, NEX = 1, slices = 26) during a resting state with eyes open and a fixation point (small cross) lasting 6 min. For spatial normalization and segmentation procedures, high-resolution T1-weighted images of the whole brain comprising up to 156 axial slices (TR = 10.916 ms, TE = 4.2 ms, thickness = 1.2 mm, flip angle = 15 degrees, matrix size = 256 9 192, FOV = 245 mm, NEX = 1, bandwidth = 122.109) were also used. To maximize participants’ cooperation, the MRI technician asked the child to look at a black dot painted inside the magnet and not to sleep. After scanning, contact was performed to ensure that the child was awake.
Image processing Analysis was performed using AFNI (version 2011_12_21_1014; Cox, 1996) and FMRIB’s Software library (FSL, version 5.0; Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). Functional MRI data preprocessing steps included: removal of the first four volumes, head movement correction, skull stripping and despiking, linear detrending, spatial smoothing (Gaussian kernel, FWHM = 6 mm) and grand-mean scaling (scripts from www.nitrc.org/projects/fcon_1000). In this instance, we used fractional amplitude of low frequency fluctuations (fALFF; Zou et al., 2008), which has been found to be robust against motion artefacts (Yan, Craddock et al., 2013; Yan, Cheung et al., 2013). This measure was calculated voxel-by-voxel in the frequency band from 0.01 to 0.1 Hz. Individual fALFF maps were scaled to z-scores across the whole brain and warped to standard space by using the Montreal Neurological Institute template (MNI152). Head movements were quantified by the mean frame-wise displacement (FD, equation 9, Yan, Cheung et al., 2013) and
Figure 1 Regions-of-interest from the default mode network © 2015 Association for Child and Adolescent Mental Health.
57
the mean temporal derivative of the RMS variance over the voxels (DVARS). The associations between these descriptors and the CBCL were tested using Pearson’s correlation.
Network maturation index The average coefficient of fractional amplitude of low frequency fluctuations (fALFF) of the preprocessed BOLD signal was extracted from each region-of-interest (see Figure 1), defined by 8-mm diameter spheres centred in the coordinates of the default mode network [DMN, 13 region of interest (ROIs), coordinates from Table 1 of Fair et al., 2008; transformed to MNI space using the method proposed by Lancaster et al., 2007]. We focused solely on the DMN to avoid problems with multiple comparisons and because it has been reported to be associated with many mental disorders (Broyd et al., 2009). The chronological age of the children and these fALFF variables were then considered as response and predictor variables (input) in a machine-learning method called Gaussian Process Regression with linear kernel (Rasmussen & Williams, 2006; Williams, 1998). We trained the algorithm leaving one subject out from the training set. We then computed the predicted age for the excluded subject and calculated the maturity index by subtracting this from his/her actual age. This procedure was then repeated for each subject in the sample. As noted above, we chose a leave-one-subject-out (LOSO) procedure instead of splitting the sample into two sets (training and test). We opted for the LOSO for the following reasons: (a) the data are very noisy, and thus, a large training set is required, (b) sample splitting and the k-fold procedure depend on the choice of the subjects in each sample, which would produce different results depending on the choice (note that this is not a problem for LOSO) and (c) the main focus of this study is the association between the neurodevelopment and psychopathological manifestations and not providing unbiased optimal accuracy estimates of the maturational index (thus, the most important is obtaining a maturation score for each subject which does not depend on the choice of training and test sets). Large positive values for the maturational index refer to delayed maturation (i.e. the actual age is much greater than the predicted age by using the fALFF at the nodes of each network). Conversely, large negative values refer to precocious
58
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
maturation. For a categorical approach, we split the sample into three status classes using the maturation index: precocious (less than the 25% percentile), typical (between 25% and 75% percentile) and delayed maturation groups (greater than the 75% percentile). These classes were defined a priori, based on the same criterion of box-plot construction (i.e. quartiles).
Statistical analysis The Pearson correlation coefficient between the actual and (leave-one-out) predicted age was used to quantify the accuracy of age predictions. Analyses based on the univariate general linear model (GLM) were conducted using the CBCL (general factor, internalizing and externalizing in three separate models) as the dependent variable, the maturation index as the main regressor, and age, site and gender as nuisance variables. The same analyses were repeated when considering the maturation status as a categorical variable (delayed, typical and precocious). The significance level of the tests was set to 5%, corrected for multiple comparisons using Bonferroni correction (considering three tests, one for each CBCL factor).
Results Age and psychopathological manifestations Histograms of age in addition to the general, internalizing and externalizing scores from the CBCL are
shown in Figure 2, highlighting the wide distribution of these variables. More details about the psychopathological manifestation across the CBCL subscales are provided in Table 1. Both the histograms and the table highlight a skewed distribution with a low prevalence of symptoms. We first looked at whether age was correlated with any of the psychopathological scores. None of the psychopathological scores was correlated with age (general factor p = .257, internalizing p = .238 and externalizing p = .947; Pearson correlation coefficient). In other words, no specific developmental stage was associated with an increased psychopathological manifestation.
Movement analysis The average (mean across subjects) of the mean FD (across frames) was 0.16 (SD = 0.23), and the average DVARS was 25.82 (SD = 10.38). The average 95th percentile was 0.54 for FD (SD = 1.09) and 38 for DVARS (SD = 24.30). As expected, FD and DVAR were correlated with both actual age (both r = .16; p < .001; Pearson correlation coefficient) and predicted age (both r = .38; p < .001). However, neither FD (r = .05; p = .21) nor DVAR (r = .05; Histogram of general factor
0.4
Density
0.0
0.2
0.010 0.005 0.000
Density
0.6
0.015
Histogram of age
80
100
120
140
160
180
−1.0
−0.5
Age
0.0
0.5
1.0
1.5
Score
Histogram of externalizing factor
1.0
Density
0.0
0.5
1.0 0.0
0.5
Density
1.5
1.5
2.0
Histogram of internalizing factor
−0.4
−0.2
0.0
Score
0.2
0.4
−0.5
0.0
0.5
Score
Figure 2 Histograms (in density, i.e. the area of the bars sum to 1.0) of age (in months), general psychopathology, as well as internalizing and externalizing scores © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
59
Table 1 Descriptive statistics of CBCL scores across the subscales. The scores were normalized as a percentage of the maximum possible score at each domain Minimum (%)
1st Quart. (%)
Median (%)
Mean (%)
3rd Quart. (%)
Maximum (%)
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
11.5 6.3 4.5 9.1 3.3 15.0 2.9 11.1 5.9
23.1 18.8 13.6 13.6 10.0 30.0 5.9 22.2 14.7
27.0 21.4 16.5 19.6 12.9 34.1 8.5 28.2 16.2
38.5 31.3 22.7 27.3 20.0 50.0 11.8 41.7 23.5
92.3 100.0 86.4 86.4 83.3 100.0 55.9 97.2 55.9
Anxiety Depression Somatic Social Thought Attention Rule-breaker Aggression Other
p = .19) were correlated with the maturity index (actual age minus predicted age) used to define the maturation status of each participant. Moreover, head motion descriptors were not correlated with the general or specific factors (p > .05, Pearson correlation). Thus, we believe movement artefacts are not strongly influencing our results involving neither the maturation status nor the frequency of psychopathological manifestations.
t-test). In addition, we calculated the contribution of each brain region of DMN in the maturational index (Figure 4). The participation of each region was measured by the Pearson correlation coefficient between the index and the corresponding fALFF of each ROI. The most relevant regions composing the index were the anterior medial prefrontal cortex, the left and right lateral parietal cortex, the right parahippocampal cortex and the cerebellar tonsils.
Age prediction from the network activity
Maturational status and psychopathology
Following previous studies, we then used a machinelearning approach to predict the age of the children in our sample from their oscillatory resting-state activity. Our algorithm could significantly predict age from a resting-state fALFF from DMN. Figure 3 depicts the correlation between actual and predicted ages and the maturation groups for the DMN (r = .288; p < .001; Pearson correlation coefficient). There were no significant effects of site of acquisition on age, predicted age or maturational index (p > .05;
The GLM analyses, using the maturity index as a continuous variable, yielded no significant results with any of the CBCL factors. In contrast, GLM analysis considering the maturity index as categorical was associated with CBCL general (p = .011 for delayed vs. precocious). It is important to highlight that these analyses also considered age, gender and acquisition site as possible confounders. The site was the only significant (p < .001) nuisance variable. Further details are provided in Table 2. However, even in the categorical approach, no significant association was found with internalizing or externalizing symptoms. Figure 3 (bottom) depicts boxplots of the CBCL general factor and shows a graduation of symptom manifestations from the delayed (more symptoms) to the precocious group (less symptoms), with the typical group in between the two.
160 120
r = 0.288 p < 0.001
Precocious Typical Delayed
80
Actual age
Age prediction
100
120
140
160
1.0 0.0 1.0
CBCL general factor
Predicted age
Delayed
Typical
Precocious
Figure 3 Top: Scatter-plot comparing actual age (in months) with predicted age (default mode network maturity) among the three maturational groups (precocious, typical and delayed). Bottom: Box-plot of the CBCL general factor across the maturational groups © 2015 Association for Child and Adolescent Mental Health.
Discussion We have performed a large-scale study to evaluate children with fMRI scans to find deviant brain maturation. We have used a machine-learning method (Gaussian processes regression) to obtain age predictions of children based on a measure of brain spontaneous activity, the fALFF. Then, we compared the predicted age with the actual age, showing that DMN maturation delays are associated with more severe child psychopathology, assessed by the CBCL general psychopathology factor. As hypothesized, the maturational status of the default mode network was related to psychopathology. Our findings on age prediction can be compared with those obtained by Dosenbach et al. (2010) and
58
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
maturation. For a categorical approach, we split the sample into three status classes using the maturation index: precocious (less than the 25% percentile), typical (between 25% and 75% percentile) and delayed maturation groups (greater than the 75% percentile). These classes were defined a priori, based on the same criterion of box-plot construction (i.e. quartiles).
Statistical analysis The Pearson correlation coefficient between the actual and (leave-one-out) predicted age was used to quantify the accuracy of age predictions. Analyses based on the univariate general linear model (GLM) were conducted using the CBCL (general factor, internalizing and externalizing in three separate models) as the dependent variable, the maturation index as the main regressor, and age, site and gender as nuisance variables. The same analyses were repeated when considering the maturation status as a categorical variable (delayed, typical and precocious). The significance level of the tests was set to 5%, corrected for multiple comparisons using Bonferroni correction (considering three tests, one for each CBCL factor).
Results Age and psychopathological manifestations Histograms of age in addition to the general, internalizing and externalizing scores from the CBCL are
shown in Figure 2, highlighting the wide distribution of these variables. More details about the psychopathological manifestation across the CBCL subscales are provided in Table 1. Both the histograms and the table highlight a skewed distribution with a low prevalence of symptoms. We first looked at whether age was correlated with any of the psychopathological scores. None of the psychopathological scores was correlated with age (general factor p = .257, internalizing p = .238 and externalizing p = .947; Pearson correlation coefficient). In other words, no specific developmental stage was associated with an increased psychopathological manifestation.
Movement analysis The average (mean across subjects) of the mean FD (across frames) was 0.16 (SD = 0.23), and the average DVARS was 25.82 (SD = 10.38). The average 95th percentile was 0.54 for FD (SD = 1.09) and 38 for DVARS (SD = 24.30). As expected, FD and DVAR were correlated with both actual age (both r = .16; p < .001; Pearson correlation coefficient) and predicted age (both r = .38; p < .001). However, neither FD (r = .05; p = .21) nor DVAR (r = .05; Histogram of general factor
0.4
Density
0.0
0.2
0.010 0.005 0.000
Density
0.6
0.015
Histogram of age
80
100
120
140
160
180
−1.0
−0.5
Age
0.0
0.5
1.0
1.5
Score
Histogram of externalizing factor
1.0
Density
0.0
0.5
1.0 0.0
0.5
Density
1.5
1.5
2.0
Histogram of internalizing factor
−0.4
−0.2
0.0
Score
0.2
0.4
−0.5
0.0
0.5
Score
Figure 2 Histograms (in density, i.e. the area of the bars sum to 1.0) of age (in months), general psychopathology, as well as internalizing and externalizing scores © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
to predict maturation status are still not well understood. The neural mechanisms generating the signal fluctuations in low frequencies are still under investigation. Usually, fALFF is interpreted as a measure of spontaneous local neural activity during resting states. On the other hand, although the neural basis of fALFF is unclear, many studies have reported the clinical value of this measure (Han et al., 2011; Hoptman et al., 2010; Xuan et al., 2012). Another point is our choice of using fractional ALFF instead of ALFF. Because we are using the data from different scanners (although with the same parameters), the variance of the MRI signal may depend on the system of acquisition. Thus, proportional measures such as fALFF are more suitable because its denominator provides a ‘normalization factor’ for site effects. Moreover, Zuo et al. (2010) have shown that fALFF is more robust against physiological artefacts (cardiac and respiratory) than ALFF. consistent with recent studies (Chabernaud et al., 2012), we neither discriminated nor split the sample according to psychiatric diagnostic criteria, such as DSM-IV or DSM-5. Although significant results were found only when grouping the subjects by network maturity (precocious, typical and delayed), the whole spectrum of general psychopathology was considered. We believe that this approach is adequate for studying brain behaviour relationships without being limited by arbitrary concepts such as the current definition of specific psychiatric disorders (Cuthbert & Insel, 2013; Insel et al., 2010), particularly for children. The dimensional (as opposed to diagnostic) approach for relating brain and behaviour in children may provide novel insights into how variations in neural circuits influence behavioural expression. Our findings suggest that machine-learning multivoxel pattern analyses offer insights into our understanding of maturation and childhood psychopathology. Most functional neuroimaging studies showing potential applications of pattern recognition and classification analyses focus on discriminating patients from healthy subjects (Kloppel et al., 2012; Mourao-Miranda et al., 2012; Sato, de Araujo Filho et al., 2012). Although the modelling of causal relations is a relevant aspect of scientific research, one of the main aims of machine-learning methods is to obtain predictions. The ability to have predictions plays a crucial role to test model validity. In other words, although machine-learning methods are not used directly to explain the disorder or discriminate patients, they can be useful to obtain predictions of associated neurobiological features, such as brain maturity, as illustrated in this study. The main limitations of this study are the following: (a) the data acquisition was performed at two different sites with identical protocols used, allowing joint analyses as well, (b) this is a cross-sectional study that does not allow any inference about © 2015 Association for Child and Adolescent Mental Health.
61
disease causation, although this feature did not impact the main finding of the article, (c) it is not possible to infer the deviations from individual trajectories, which is the most relevant that could only be obtained by comparing the pattern with the whole sample and (d) regression and fALFF, two Gaussian processes, were arbitrary choices (i.e. the prediction error depends on the model and regressors considered, and thus, a large variety of methods and predictor variables could also be tested). Finally, despite the brain-behaviour analyses carried out using GLM (to remove the effects of site, gender and age as confounders), two different approaches based on brain maturity were attempted. The first considered the DMN maturity index as a continuous variable, and no significant association was found with the CBCL general factor. The second approach considered the maturity status as a categorical factor (delayed, typical and precocious), and it was found that the mean CBCL general factor was significantly different between delayed and precocious groups, with more typical scores in between (see Figure 3). It is important to emphasize that the definition of maturation classes were defined a priori based on the quartiles. We believe that these differences in statistical significance between the continuous and categorical approach were due to the high variability in CBCL (filled by the parents or guardian), particularly because symptoms in children are very unstable. The categorical approach in GLM focuses on mean differences, which are more robust against noise/variability than are correlations (i.e. the continuous approach). Thus, although our sample was large, we believe it may not have been sufficiently powered to detect correlations of a continuous nature. In summary, our investigation of the neurodevelopmental trajectories of a large cohort of children and adolescents provided results suggesting that the maturational status of the default mode network might be related to psychopathology in children. This finding should be further investigated in longitudinal studies and validated with additional clinical, neuropsychological, and bio-signatures, such as molecules related to neuroplasticity regulation as putative biomarkers of risk for psychiatric disorders. The relationship between brain networks development and psychopathology is a novel and exciting field of study. Novel computational approaches applied to neuroimaging data may help clarify how neurodevelopment relates to mental health and disease.
Acknowledgements L.A.R. has been on the speakers’ bureau/advisory board and/or acted as a consultant for Eli-Lilly, Janssen-Cilag, Novartis and Shire in the last three years. The ADHD and Juvenile Bipolar Disorder Outpatient Programs that he chaired received unrestricted
62
Jo~ ao Ricardo Sato et al.
educational and research support from the following pharmaceutical companies in the last three years: EliLilly, Janssen-Cilag, Novartis and Shire. He receives authorship royalties from Oxford Press and ArtMed. R.A.B. 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, ~ o E.J. Safra and Fundac ~ o ABAHDS and is a Fundacßa ßa shareholder in Biomolecular Technology Ltda. E.A. Jr. has received research grants from FAPESP, CNPq, ~ o E.J. Safra and Fundacßa ~ o ABAHDS. CAPES, Fundac ßa P.P. has received payment for the development of educational material for Janssen-Cilag and Astra-Zeneca. The opinions, hypotheses, conclusions and recommendations expounded upon in this study are those of the authors and do not necessarily represent the opinions of the funding agencies. The authors are
J Child Psychol Psychiatr 2016; 57(1): 55–64
grateful to the Sao Paulo Research Foundation FAPESP (J.R.S grants 2013/10498-6 and 2013/ 00506-1, A.J. grant 2013/08531-5), the National Institute of Developmental Psychiatry for Children and Adolescents, a science and technology institute funded by Conselho Nacional de Desenvolvimento Cientıfico e Tecnol ogico (CNPq; National Council for Scientific and Technological Development) and FAPESP (grant number 573974/2008-0). N.C. is supported by the Wellcome Trust.
Correspondence Jo~ ao Ricardo Sato, Av. dos Estados, 5001. Bairro Bangu. Santo Andr e - SP - Brazil. CEP 09210-580; Email: [email protected]
Key points
• • •
A community sample of 654 children from a developing country was scanned using functional magnetic resonance imaging (fMRI) under a resting-state protocol. We built a default mode network maturation (DMN) index based on resting-state fMRI, fractional amplitude of low frequency fluctuations and machine-learning techniques. We found an association between delay in DMN maturation and psychopathological manifestations.
References Achenbach, T.M., & Rescorla, L.A. (2001). Manual for the ASEBA school-age forms and profiles. Burlington, VT: University of Vermont, Research Center for Children, Youth, and Families. Broyd, S.J., Demanuele, C., Debener, S., Helps, S.K., James, C.J., & Sonuga-Barke, E.J. (2009). Default-mode brain dysfunction in mental disorders: A systematic review. Neuroscience and Biobehavioral Reviews, 33, 279–296. Brunner, M., Nagy, G., & Wilhelm, O. (2012). A tutorial on hierarchically structured constructs. Journal of Personality, 80, 796–846. Buckner, R.L., Andrews-Hanna, J.R., & Schacter, D.L. (2008). The brain’s default network: Anatomy, function, and relevance to disease. Annals of the New York Academy of Sciences, 1124, 1–38 Review. Cannon, T.D., Cadenhead, K., Cornblatt, B., Woods, S.W., Addington, J., Walker, E., . . . & Heinssen, R. (2008). Prediction of psychosis in youth at high clinical risk: A multisite longitudinal study in North America. Archives of General Psychiatry, 65, 28–37. Castellanos, F.X., Margulies, D.S., Kelly, C., Uddin, L.Q., Ghaffari, M., Kirsch, A., . . . & Milham, M.P. (2008). Cingulate-precuneus interactions: A new locus of dysfunction in adult attention-deficit/hyperactivity disorder. Biological Psychiatry, 63, 332–337. Chabernaud, C., Mennes, M., Kelly, C., Nooner, K., Di Martino, A., Castellanos, F.X., & Milham, M.P. (2012). Dimensional brain-behavior relationships in children with attentiondeficit/hyperactivity disorder. Biological Psychiatry, 71, 434–442. Chang, X., Shen, H., Wang, L., Liu, Z., Xin, W., Hu, D., & Miao, D. (2014). Altered default mode and fronto-parietal network subsystems in patients with schizophrenia and their unaffected siblings. Brain Research, 1562, 87–99.
Cox, R.W. (1996). AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Computers and Biomedical Research, 29, 162–173. Cuthbert, B.N., & Insel, T.R. (2013). Toward the future of psychiatric diagnosis: The seven pillars of RDoC. BMC Medicine, 11, 126. Dosenbach, N.U., Nardos, B., Cohen, A.L., Fair, D.A., Power, J.D., Church, J.A., . . . & Schlaggar, B.L. (2010). Prediction of individual brain maturity using fMRI. Science, 329, 1358– 1361. Ducharme, S., Albaugh, M.D., Hudziak, J.J., Botteron, K.N., Nguyen, T.V., Truong, C., . . . & Brain Development Cooperative (2014). Anxious/depressed symptoms are linked to right ventromedial prefrontal cortical thickness maturation in healthy children and young adults. Cerebral Cortex, 24, 2941–2950. Ducharme, S., Hudziak, J.J., Botteron, K.N., Albaugh, M.D., Nguyen, T.V., Karama, S., . . . & Brain Development Cooperative (2012). Decreased regional cortical thickness and thinning rate are associated with inattention symptoms in healthy children. Journal of the American Academy of Child and Adolescent Psychiatry, 51, 18–27 e12. Fair, D.A., Cohen, A.L., Dosenbach, N.U., Church, J.A., Miezin, F.M., Barch, D.M., . . . & Schlaggar, B.L. (2008). The maturing architecture of the brain’s default network. Proceedings of the National Academy of Sciences of the United States of America, 105, 4028–4032. Garrity, A.G., Pearlson, G.D., McKiernan, K., Lloyd, D., Kiehl, K.A., & Calhoun, V.D. (2007). Aberrant “default mode” functional connectivity in schizophrenia. American Journal of Psychiatry, 164, 450–457. Gentili, C., Ricciardi, E., Gobbini, M.I., Santarelli, M.F., Haxby, J.V., Pietrini, P., & Guazzelli, M. (2009). Beyond amygdala: Default mode network activity differs between patients with social phobia and healthy controls. Brain Research Bulletin, 79, 409–413. © 2015 Association for Child and Adolescent Mental Health.
60
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
Figure 4 Contribution of the fALFF from each region-of-interest to the DMN maturational index. The participation of each region of interest (ROI) in this index was measured by the correlation between the index and the fALFF of the corresponding ROI. Abbreviations: aMPFC = anterior medial prefrontal cortex; vMPFC = ventral medial prefrontal cortex
Table 2 Results of GLM for the CBCL general factor as response variable. Note that the psychopathological manifestations are significantly smaller in the precocious when compared to the delayed group Regressor
Beta
SD
Corrected p
Group-delayed (ref) Group-typical Group-precocious Site Gender Age
– .144 .338 .277 .046 .004
– .075 .116 .042 .042 .002
– .167 .011
doi:10.1111/jcpp.12444
Default mode network maturation and psychopathology in children and adolescents ~ o Ricardo Sato,1,2,3,4 Giovanni Abraha ~ o Salum,4,5 Ary Gadelha,2,4 Nicolas Crossley,6,7 Joa 3,8 4,5 Zugman,2,4 Felipe Almeida Picon,4,5 Pedro Gilson Vieira, Gisele Gus Manfro, Andre 2,4 2,4,9 s,4,5 Luciana Monteiro Moura,2,4 Mario Pan, Marcelo Queiroz Hoexter, Mauricio Ane 2,4 Marco Antonio Gomes Del’Aquilla, Edson Amaro Jr,3 Philip McGuire,6 Acioly Luiz 2,4 Tavares Lacerda, Luis Augusto Rohde,4,5 Euripedes Constantino Miguel,4,9 Andrea Parolin Jackowski,2,4 and Rodrigo Affonseca Bressan2,4 1
Center of Mathematics Computation and Cognition, Universidade Federal do ABC, Santo Andre; 2Interdisciplinary Lab for Clinical Neurosciences (LiNC), Universidade Federal de Sao Paulo (UNIFESP), Sao Paulo; 3Department of Radiology, School of Medicine, University of Sao Paulo, Sao Paulo; 4National Institute of Developmental Psychiatry for Children and Adolescents, CNPq, Sao Paulo; 5Department of Psychiatry, Federal University of Rio Grande do Sul, Porto Alegre, Brazil; 6Institute of Psychiatry, King’s College London, London, United Kingdom; 7Institute for Biological and Medical Engineering, Faculties of Engineering, Medicine and Biological Sciences, P. Catholic University of Chile, Santiago, Chile; 8Bioinformatics Program, Institute of Mathematics and Statistics, University of Sao Paulo, Sao Paulo; 9Department of Psychiatry, School of Medicine, University of Sao Paulo, Sao Paulo, Brazil
Background: The human default mode (DMN) is involved in a wide array of mental disorders. Current knowledge suggests that mental health disorders may reflect deviant trajectories of brain maturation. Method: We studied 654 children using functional magnetic resonance imaging (fMRI) scans under a resting-state protocol. A machinelearning method was used to obtain age predictions of children based on the average coefficient of fractional amplitude of low frequency fluctuations (fALFFs) of the DMN, a measure of spontaneous local activity. The chronological ages of the children and fALFF measures from regions of this network, the response and predictor variables were considered respectively in a Gaussian Process Regression. Subsequently, we computed a network maturation status index for each subject (actual age minus predicted). We then evaluated the association between this maturation index and psychopathology scores on the Child Behavior Checklist (CBCL). Results: Our hypothesis was that the maturation status of the DMN would be negatively associated with psychopathology. Consistent with previous studies, fALFF significantly predicted the age of participants (p < .001). Furthermore, as expected, we found an association between the DMN maturation status (precocious vs. delayed) and general psychopathology scores (p = .011). Conclusions: Our findings suggest that child psychopathology seems to be associated with delayed maturation of the DMN. This delay in the neurodevelopmental trajectory may offer interesting insights into the pathophysiology of mental health disorders. Keywords: Neurodevelopment, default mode network, neuroimaging, psychopathology, MVPA.
Introduction Previous investigations have suggested that neuropsychiatric disorders are deviations from typical neurodevelopmental trajectories (Ducharme et al., 2012, 2014; Insel, 2010; Shaw et al., 2006, 2007; Thompson et al., 2001; Uddin, Supekar, & Menon, 2010). The results of such deviant trajectories can be observed already in childhood and adolescence, when the majority of mental disorders begin (KimCohen et al., 2003). Thus, investigation of the typical brain is important not only to achieve a better understanding of the neural substrates of mental disorders but also to develop early intervention strategies (Cannon et al., 2008). Functional magnetic resonance imaging (fMRI) has played a key role in understanding the brain circuitry changes that take place throughout neurodevelopment (Power, Fair, Schlaggar, & Petersen,
Conflict of interest statement: See Acknowledgements for disclosures.
2010; Sato et al., 2015; Smyser, Snyder, & Neil, 2011; Uddin et al., 2010; Vogel, Power, Petersen, & Schlaggar, 2010). Previous resting-state fMRI studies have found that certain functional networks are associated with developmental disorders (Castellanos et al., 2008; Sato, Hoexter, Castellanos, & Rohde, 2012). Recent studies have investigated the changes that occur to the brain networks during their maturation process in different stages of childhood and adolescence; they have examined the default mode network (DMN), in particular. The default mode network (DMN, Raichle et al., 2001; Greicius, Krasnow, Reiss, & Menon, 2003; Buckner, Andrews-Hanna, & Schacter, 2008) is usually associated with introspection and is less active in externally directed tasks (McGuire, Paulesu, Frackowiak, & Frith, 1996). Fair et al. (2008) found that the functional connectivity within the DMN was very sparse between 7–9 years of age but the intensity of connectivity among these regions increased with age (Sato et al., 2014). In addition, two multimodal studies investigating the functional
© 2015 Association for Child and Adolescent Mental Health. Published by John Wiley & Sons Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main St, Malden, MA 02148, USA
56
Jo~ ao Ricardo Sato et al.
and structural connectivity of DMN in children reported convergent results from both modalities. The findings suggest that the functional connectivity between the posterior cingulate and medial prefrontal cortex (PCC-mPFC) is the most immature connection in young children, and the strength of this connection was correlated with fractional anisotropy of the cingulum bundle (Gordon et al., 2011; Supekar et al., 2010). Interestingly, the DMN has also emerged as being associated with a myriad of psychiatric diseases (Broyd et al., 2009; Chang et al., 2014). Thus, it is important to investigate associations between psychopathological symptoms and the functional maturation of this network. One possible approach to investigating neurodevelopment is by using machine-learning methods. Resting-state fMRI studies based on machine-learning techniques have been able to predict subjects’ chronological ages (from 7- to 30-years old) (Dosenbach et al., 2010; Supekar, Musen, & Menon, 2009; Wang, Su, Shen, & Hu, 2012), even after controlling for head movement (Satterthwaite et al., 2013). The handling of motion artifacts is indeed a challenge in neuroimaging. One possible approach in the analysis of resting-state fMRI might be to use fractional amplitude of low frequency fluctuations (fALFF, Zou et al., 2008), found to be less affected by motion artefacts (Yan, Craddock, Zuo, Zang, & Milham, 2013; Yan, Cheung et al., 2013). In this hypothesis-driven study, we investigated whether the DMN maturation status could be associated with behavioural disturbances in a large nonreferred sample of 654 subjects ranging in age from 6- to 15-years old using a machine-learning method on resting-state fMRI data. This developmental period seems to be crucial in the formation of DMN (Fair et al., 2008; Sato et al., 2014). We first used machine-learning techniques to predict the chronological age of subjects based on the fALFF of regions-of-interests belonging to the DMN. We then defined a ‘maturational status’ for each subject, based on the difference between the predicted and actual age of the participants. We hypothesized that children and adolescents whose predicted age was much younger than their actual age, which we described as a delayed maturation pattern, would present higher general psychopathological manifestations when compared to those with a more precocious developmental status.
Methods Subjects Students from 22 state schools in the city of Porto Alegre and 35 in S~ ao Paulo were assessed using a screening procedure. For a detailed description of the Brazilian High Risk Cohort (HRC), see Salum et al. (2013, 2014). The ethics committee of the University of S~ ao Paulo approved the study (IORG0004884, 1138/08), and parents/guardian and the children provided written consent and verbal assent respectively.
J Child Psychol Psychiatr 2016; 57(1): 55–64 In the current investigation, 654 subjects (352 male; 53.8%) from the HRC were submitted to a resting-state fMRI protocol. The mean age of the sample was 10.71 years (SD = 1.89), and mean IQ was 102.58 (SD = 16.72). The vocabulary and block design subtests of the Wechsler Intelligence Scale for Children (Wechsler, 1991) were used to estimate IQ, using the method described in Tellegen and Briggs (1967). The sample averaged 4 years of education (SD = 1.67), and 558 subjects (85.3%) were right-handed. As declared by the mother), the sample was 60% Caucasian, 10.55% Afro-Brazilian, 28.29% mixed African background and 1.16% were classified as ‘other’. The Brazilian rating scale (ABIMEP) was used to determine subjects’ socio-economic status classifications, which were as follows: 4.28% were low to very low (E and D classes), 66.82% were medium (C and B classes) and 28.9% were comfortable (A class) groups. Subjects were also assessed for psychiatric disorders using the Development and Well-Being Assessment (DAWBA, Goodman, Ford, Richards, Gatward, & Meltzer, 2000). Of the 654 participants, 447 did not present any disorder. Because this study focuses on a dimensional approach to psychopathological manifestation, none of the subjects were excluded from the analyses. Finally, 356 participants reported a familiar history of psychiatric conditions based on the Family History Survey (FHS).
CBCL evaluation and description of the bi-factor model Parents/caregivers completed the Child Behavior Checklist (CBCL; Achenbach & Rescorla, 2001) on the day of MRI scanning. This is a well-established assessment tool containing short statements (e.g.: ‘he talks too much’ and ‘he cries too much’) with three different levels of agreement/frequency for each item. Raw internalizing and externalizing factors from the CBCL are usually correlated, which can be a problem regarding interpretation because it is not clear whether the findings are related to the general factor or to the specific internalizing/ externalizing construct. We therefore fitted a bifactor model to the responses using structural equation modelling. The model considers that the response to each CBCL item is driven by a single general factor accounting for the covariation among all symptoms (general factor), plus a specific factor (internalizing or externalizing, orthogonal to the general factor) and random error (individual variation). This approach model provides a framework to conceptualize both the commonality and the specificity of symptoms from distinct domains (Brunner, Nagy, & Wilhelm, 2012; Holzinger & Swineford, 1937; Tackett et al., 2013). The model was fitted using the structural equation module of the software MPLUS 7.0 (http://www.statmodel.com/). The bifactor model provided a good fit to our data. Briefly, the bifactor models were fitted to polychoric correlations among the CBCL items using the mean- and varianceadjusted weighted least squares (WLSMV) estimator (Muth en & Muth en, 2012). The goodness of fit was assessed through the following fit indices: chi-square, weighted root mean square residual (WRMR), comparative fit index (CFI), Tucker-Lewis Index (TLI) and the root mean square error of approximation (RMSEA). To demonstrate good fit to the data, it is suggested that an estimated model should have a WRMR near or below .9 (Muth en & Muth en, 2012), an RMSEA near or below .06, and both CFI and TLI near or above .95 (Hu & Bentler, 1999). The bifactor model for this sample provided acceptable fit to our data: X2 = 2688.5, df = 1650, p < .0001, number of free parameters = 239, RMSEA = 0.029 (95%CI: 0.027–0.031), CFI = 0.950, TLI = 0.947 and WRMR = 1.279.
Image acquisition On the neuroimaging day, children were engaged in recreational activities as a desensitization method. A mock-up of data acquisition procedures was carried out using a fabric play © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
tunnel and MRI noise, in addition to didactical explanations about the next steps and the importance of not moving during the fMRI procedure. Further information about the protocol can be found in Sato et al. (2014). Imaging was performed using two 1.5T MRI scanners (General Electric, Signa HDX and HD, in the cities of S~ ao Paulo and Porto Alegre, Brazil). To provide T2*-weighted BOLD contrast, fMRI data were acquired with a protocol of a series of 180 echo planar imaging volumes (TR = 2000 ms, TE = 30 ms, slice thickness = 4 mm, gap = 0.5 mm, flip angle = 80 degrees, matrix size = 80 9 80, FOV = 240 mm, reconstruction matrix size = 128 9 128, 1.875 9 1.875 mm, NEX = 1, slices = 26) during a resting state with eyes open and a fixation point (small cross) lasting 6 min. For spatial normalization and segmentation procedures, high-resolution T1-weighted images of the whole brain comprising up to 156 axial slices (TR = 10.916 ms, TE = 4.2 ms, thickness = 1.2 mm, flip angle = 15 degrees, matrix size = 256 9 192, FOV = 245 mm, NEX = 1, bandwidth = 122.109) were also used. To maximize participants’ cooperation, the MRI technician asked the child to look at a black dot painted inside the magnet and not to sleep. After scanning, contact was performed to ensure that the child was awake.
Image processing Analysis was performed using AFNI (version 2011_12_21_1014; Cox, 1996) and FMRIB’s Software library (FSL, version 5.0; Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). Functional MRI data preprocessing steps included: removal of the first four volumes, head movement correction, skull stripping and despiking, linear detrending, spatial smoothing (Gaussian kernel, FWHM = 6 mm) and grand-mean scaling (scripts from www.nitrc.org/projects/fcon_1000). In this instance, we used fractional amplitude of low frequency fluctuations (fALFF; Zou et al., 2008), which has been found to be robust against motion artefacts (Yan, Craddock et al., 2013; Yan, Cheung et al., 2013). This measure was calculated voxel-by-voxel in the frequency band from 0.01 to 0.1 Hz. Individual fALFF maps were scaled to z-scores across the whole brain and warped to standard space by using the Montreal Neurological Institute template (MNI152). Head movements were quantified by the mean frame-wise displacement (FD, equation 9, Yan, Cheung et al., 2013) and
Figure 1 Regions-of-interest from the default mode network © 2015 Association for Child and Adolescent Mental Health.
57
the mean temporal derivative of the RMS variance over the voxels (DVARS). The associations between these descriptors and the CBCL were tested using Pearson’s correlation.
Network maturation index The average coefficient of fractional amplitude of low frequency fluctuations (fALFF) of the preprocessed BOLD signal was extracted from each region-of-interest (see Figure 1), defined by 8-mm diameter spheres centred in the coordinates of the default mode network [DMN, 13 region of interest (ROIs), coordinates from Table 1 of Fair et al., 2008; transformed to MNI space using the method proposed by Lancaster et al., 2007]. We focused solely on the DMN to avoid problems with multiple comparisons and because it has been reported to be associated with many mental disorders (Broyd et al., 2009). The chronological age of the children and these fALFF variables were then considered as response and predictor variables (input) in a machine-learning method called Gaussian Process Regression with linear kernel (Rasmussen & Williams, 2006; Williams, 1998). We trained the algorithm leaving one subject out from the training set. We then computed the predicted age for the excluded subject and calculated the maturity index by subtracting this from his/her actual age. This procedure was then repeated for each subject in the sample. As noted above, we chose a leave-one-subject-out (LOSO) procedure instead of splitting the sample into two sets (training and test). We opted for the LOSO for the following reasons: (a) the data are very noisy, and thus, a large training set is required, (b) sample splitting and the k-fold procedure depend on the choice of the subjects in each sample, which would produce different results depending on the choice (note that this is not a problem for LOSO) and (c) the main focus of this study is the association between the neurodevelopment and psychopathological manifestations and not providing unbiased optimal accuracy estimates of the maturational index (thus, the most important is obtaining a maturation score for each subject which does not depend on the choice of training and test sets). Large positive values for the maturational index refer to delayed maturation (i.e. the actual age is much greater than the predicted age by using the fALFF at the nodes of each network). Conversely, large negative values refer to precocious
58
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
maturation. For a categorical approach, we split the sample into three status classes using the maturation index: precocious (less than the 25% percentile), typical (between 25% and 75% percentile) and delayed maturation groups (greater than the 75% percentile). These classes were defined a priori, based on the same criterion of box-plot construction (i.e. quartiles).
Statistical analysis The Pearson correlation coefficient between the actual and (leave-one-out) predicted age was used to quantify the accuracy of age predictions. Analyses based on the univariate general linear model (GLM) were conducted using the CBCL (general factor, internalizing and externalizing in three separate models) as the dependent variable, the maturation index as the main regressor, and age, site and gender as nuisance variables. The same analyses were repeated when considering the maturation status as a categorical variable (delayed, typical and precocious). The significance level of the tests was set to 5%, corrected for multiple comparisons using Bonferroni correction (considering three tests, one for each CBCL factor).
Results Age and psychopathological manifestations Histograms of age in addition to the general, internalizing and externalizing scores from the CBCL are
shown in Figure 2, highlighting the wide distribution of these variables. More details about the psychopathological manifestation across the CBCL subscales are provided in Table 1. Both the histograms and the table highlight a skewed distribution with a low prevalence of symptoms. We first looked at whether age was correlated with any of the psychopathological scores. None of the psychopathological scores was correlated with age (general factor p = .257, internalizing p = .238 and externalizing p = .947; Pearson correlation coefficient). In other words, no specific developmental stage was associated with an increased psychopathological manifestation.
Movement analysis The average (mean across subjects) of the mean FD (across frames) was 0.16 (SD = 0.23), and the average DVARS was 25.82 (SD = 10.38). The average 95th percentile was 0.54 for FD (SD = 1.09) and 38 for DVARS (SD = 24.30). As expected, FD and DVAR were correlated with both actual age (both r = .16; p < .001; Pearson correlation coefficient) and predicted age (both r = .38; p < .001). However, neither FD (r = .05; p = .21) nor DVAR (r = .05; Histogram of general factor
0.4
Density
0.0
0.2
0.010 0.005 0.000
Density
0.6
0.015
Histogram of age
80
100
120
140
160
180
−1.0
−0.5
Age
0.0
0.5
1.0
1.5
Score
Histogram of externalizing factor
1.0
Density
0.0
0.5
1.0 0.0
0.5
Density
1.5
1.5
2.0
Histogram of internalizing factor
−0.4
−0.2
0.0
Score
0.2
0.4
−0.5
0.0
0.5
Score
Figure 2 Histograms (in density, i.e. the area of the bars sum to 1.0) of age (in months), general psychopathology, as well as internalizing and externalizing scores © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
59
Table 1 Descriptive statistics of CBCL scores across the subscales. The scores were normalized as a percentage of the maximum possible score at each domain Minimum (%)
1st Quart. (%)
Median (%)
Mean (%)
3rd Quart. (%)
Maximum (%)
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
11.5 6.3 4.5 9.1 3.3 15.0 2.9 11.1 5.9
23.1 18.8 13.6 13.6 10.0 30.0 5.9 22.2 14.7
27.0 21.4 16.5 19.6 12.9 34.1 8.5 28.2 16.2
38.5 31.3 22.7 27.3 20.0 50.0 11.8 41.7 23.5
92.3 100.0 86.4 86.4 83.3 100.0 55.9 97.2 55.9
Anxiety Depression Somatic Social Thought Attention Rule-breaker Aggression Other
p = .19) were correlated with the maturity index (actual age minus predicted age) used to define the maturation status of each participant. Moreover, head motion descriptors were not correlated with the general or specific factors (p > .05, Pearson correlation). Thus, we believe movement artefacts are not strongly influencing our results involving neither the maturation status nor the frequency of psychopathological manifestations.
t-test). In addition, we calculated the contribution of each brain region of DMN in the maturational index (Figure 4). The participation of each region was measured by the Pearson correlation coefficient between the index and the corresponding fALFF of each ROI. The most relevant regions composing the index were the anterior medial prefrontal cortex, the left and right lateral parietal cortex, the right parahippocampal cortex and the cerebellar tonsils.
Age prediction from the network activity
Maturational status and psychopathology
Following previous studies, we then used a machinelearning approach to predict the age of the children in our sample from their oscillatory resting-state activity. Our algorithm could significantly predict age from a resting-state fALFF from DMN. Figure 3 depicts the correlation between actual and predicted ages and the maturation groups for the DMN (r = .288; p < .001; Pearson correlation coefficient). There were no significant effects of site of acquisition on age, predicted age or maturational index (p > .05;
The GLM analyses, using the maturity index as a continuous variable, yielded no significant results with any of the CBCL factors. In contrast, GLM analysis considering the maturity index as categorical was associated with CBCL general (p = .011 for delayed vs. precocious). It is important to highlight that these analyses also considered age, gender and acquisition site as possible confounders. The site was the only significant (p < .001) nuisance variable. Further details are provided in Table 2. However, even in the categorical approach, no significant association was found with internalizing or externalizing symptoms. Figure 3 (bottom) depicts boxplots of the CBCL general factor and shows a graduation of symptom manifestations from the delayed (more symptoms) to the precocious group (less symptoms), with the typical group in between the two.
160 120
r = 0.288 p < 0.001
Precocious Typical Delayed
80
Actual age
Age prediction
100
120
140
160
1.0 0.0 1.0
CBCL general factor
Predicted age
Delayed
Typical
Precocious
Figure 3 Top: Scatter-plot comparing actual age (in months) with predicted age (default mode network maturity) among the three maturational groups (precocious, typical and delayed). Bottom: Box-plot of the CBCL general factor across the maturational groups © 2015 Association for Child and Adolescent Mental Health.
Discussion We have performed a large-scale study to evaluate children with fMRI scans to find deviant brain maturation. We have used a machine-learning method (Gaussian processes regression) to obtain age predictions of children based on a measure of brain spontaneous activity, the fALFF. Then, we compared the predicted age with the actual age, showing that DMN maturation delays are associated with more severe child psychopathology, assessed by the CBCL general psychopathology factor. As hypothesized, the maturational status of the default mode network was related to psychopathology. Our findings on age prediction can be compared with those obtained by Dosenbach et al. (2010) and
58
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
maturation. For a categorical approach, we split the sample into three status classes using the maturation index: precocious (less than the 25% percentile), typical (between 25% and 75% percentile) and delayed maturation groups (greater than the 75% percentile). These classes were defined a priori, based on the same criterion of box-plot construction (i.e. quartiles).
Statistical analysis The Pearson correlation coefficient between the actual and (leave-one-out) predicted age was used to quantify the accuracy of age predictions. Analyses based on the univariate general linear model (GLM) were conducted using the CBCL (general factor, internalizing and externalizing in three separate models) as the dependent variable, the maturation index as the main regressor, and age, site and gender as nuisance variables. The same analyses were repeated when considering the maturation status as a categorical variable (delayed, typical and precocious). The significance level of the tests was set to 5%, corrected for multiple comparisons using Bonferroni correction (considering three tests, one for each CBCL factor).
Results Age and psychopathological manifestations Histograms of age in addition to the general, internalizing and externalizing scores from the CBCL are
shown in Figure 2, highlighting the wide distribution of these variables. More details about the psychopathological manifestation across the CBCL subscales are provided in Table 1. Both the histograms and the table highlight a skewed distribution with a low prevalence of symptoms. We first looked at whether age was correlated with any of the psychopathological scores. None of the psychopathological scores was correlated with age (general factor p = .257, internalizing p = .238 and externalizing p = .947; Pearson correlation coefficient). In other words, no specific developmental stage was associated with an increased psychopathological manifestation.
Movement analysis The average (mean across subjects) of the mean FD (across frames) was 0.16 (SD = 0.23), and the average DVARS was 25.82 (SD = 10.38). The average 95th percentile was 0.54 for FD (SD = 1.09) and 38 for DVARS (SD = 24.30). As expected, FD and DVAR were correlated with both actual age (both r = .16; p < .001; Pearson correlation coefficient) and predicted age (both r = .38; p < .001). However, neither FD (r = .05; p = .21) nor DVAR (r = .05; Histogram of general factor
0.4
Density
0.0
0.2
0.010 0.005 0.000
Density
0.6
0.015
Histogram of age
80
100
120
140
160
180
−1.0
−0.5
Age
0.0
0.5
1.0
1.5
Score
Histogram of externalizing factor
1.0
Density
0.0
0.5
1.0 0.0
0.5
Density
1.5
1.5
2.0
Histogram of internalizing factor
−0.4
−0.2
0.0
Score
0.2
0.4
−0.5
0.0
0.5
Score
Figure 2 Histograms (in density, i.e. the area of the bars sum to 1.0) of age (in months), general psychopathology, as well as internalizing and externalizing scores © 2015 Association for Child and Adolescent Mental Health.
doi:10.1111/jcpp.12444
Default mode network and psychopathology in children and adolescents
to predict maturation status are still not well understood. The neural mechanisms generating the signal fluctuations in low frequencies are still under investigation. Usually, fALFF is interpreted as a measure of spontaneous local neural activity during resting states. On the other hand, although the neural basis of fALFF is unclear, many studies have reported the clinical value of this measure (Han et al., 2011; Hoptman et al., 2010; Xuan et al., 2012). Another point is our choice of using fractional ALFF instead of ALFF. Because we are using the data from different scanners (although with the same parameters), the variance of the MRI signal may depend on the system of acquisition. Thus, proportional measures such as fALFF are more suitable because its denominator provides a ‘normalization factor’ for site effects. Moreover, Zuo et al. (2010) have shown that fALFF is more robust against physiological artefacts (cardiac and respiratory) than ALFF. consistent with recent studies (Chabernaud et al., 2012), we neither discriminated nor split the sample according to psychiatric diagnostic criteria, such as DSM-IV or DSM-5. Although significant results were found only when grouping the subjects by network maturity (precocious, typical and delayed), the whole spectrum of general psychopathology was considered. We believe that this approach is adequate for studying brain behaviour relationships without being limited by arbitrary concepts such as the current definition of specific psychiatric disorders (Cuthbert & Insel, 2013; Insel et al., 2010), particularly for children. The dimensional (as opposed to diagnostic) approach for relating brain and behaviour in children may provide novel insights into how variations in neural circuits influence behavioural expression. Our findings suggest that machine-learning multivoxel pattern analyses offer insights into our understanding of maturation and childhood psychopathology. Most functional neuroimaging studies showing potential applications of pattern recognition and classification analyses focus on discriminating patients from healthy subjects (Kloppel et al., 2012; Mourao-Miranda et al., 2012; Sato, de Araujo Filho et al., 2012). Although the modelling of causal relations is a relevant aspect of scientific research, one of the main aims of machine-learning methods is to obtain predictions. The ability to have predictions plays a crucial role to test model validity. In other words, although machine-learning methods are not used directly to explain the disorder or discriminate patients, they can be useful to obtain predictions of associated neurobiological features, such as brain maturity, as illustrated in this study. The main limitations of this study are the following: (a) the data acquisition was performed at two different sites with identical protocols used, allowing joint analyses as well, (b) this is a cross-sectional study that does not allow any inference about © 2015 Association for Child and Adolescent Mental Health.
61
disease causation, although this feature did not impact the main finding of the article, (c) it is not possible to infer the deviations from individual trajectories, which is the most relevant that could only be obtained by comparing the pattern with the whole sample and (d) regression and fALFF, two Gaussian processes, were arbitrary choices (i.e. the prediction error depends on the model and regressors considered, and thus, a large variety of methods and predictor variables could also be tested). Finally, despite the brain-behaviour analyses carried out using GLM (to remove the effects of site, gender and age as confounders), two different approaches based on brain maturity were attempted. The first considered the DMN maturity index as a continuous variable, and no significant association was found with the CBCL general factor. The second approach considered the maturity status as a categorical factor (delayed, typical and precocious), and it was found that the mean CBCL general factor was significantly different between delayed and precocious groups, with more typical scores in between (see Figure 3). It is important to emphasize that the definition of maturation classes were defined a priori based on the quartiles. We believe that these differences in statistical significance between the continuous and categorical approach were due to the high variability in CBCL (filled by the parents or guardian), particularly because symptoms in children are very unstable. The categorical approach in GLM focuses on mean differences, which are more robust against noise/variability than are correlations (i.e. the continuous approach). Thus, although our sample was large, we believe it may not have been sufficiently powered to detect correlations of a continuous nature. In summary, our investigation of the neurodevelopmental trajectories of a large cohort of children and adolescents provided results suggesting that the maturational status of the default mode network might be related to psychopathology in children. This finding should be further investigated in longitudinal studies and validated with additional clinical, neuropsychological, and bio-signatures, such as molecules related to neuroplasticity regulation as putative biomarkers of risk for psychiatric disorders. The relationship between brain networks development and psychopathology is a novel and exciting field of study. Novel computational approaches applied to neuroimaging data may help clarify how neurodevelopment relates to mental health and disease.
Acknowledgements L.A.R. has been on the speakers’ bureau/advisory board and/or acted as a consultant for Eli-Lilly, Janssen-Cilag, Novartis and Shire in the last three years. The ADHD and Juvenile Bipolar Disorder Outpatient Programs that he chaired received unrestricted
62
Jo~ ao Ricardo Sato et al.
educational and research support from the following pharmaceutical companies in the last three years: EliLilly, Janssen-Cilag, Novartis and Shire. He receives authorship royalties from Oxford Press and ArtMed. R.A.B. 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, ~ o E.J. Safra and Fundac ~ o ABAHDS and is a Fundacßa ßa shareholder in Biomolecular Technology Ltda. E.A. Jr. has received research grants from FAPESP, CNPq, ~ o E.J. Safra and Fundacßa ~ o ABAHDS. CAPES, Fundac ßa P.P. has received payment for the development of educational material for Janssen-Cilag and Astra-Zeneca. The opinions, hypotheses, conclusions and recommendations expounded upon in this study are those of the authors and do not necessarily represent the opinions of the funding agencies. The authors are
J Child Psychol Psychiatr 2016; 57(1): 55–64
grateful to the Sao Paulo Research Foundation FAPESP (J.R.S grants 2013/10498-6 and 2013/ 00506-1, A.J. grant 2013/08531-5), the National Institute of Developmental Psychiatry for Children and Adolescents, a science and technology institute funded by Conselho Nacional de Desenvolvimento Cientıfico e Tecnol ogico (CNPq; National Council for Scientific and Technological Development) and FAPESP (grant number 573974/2008-0). N.C. is supported by the Wellcome Trust.
Correspondence Jo~ ao Ricardo Sato, Av. dos Estados, 5001. Bairro Bangu. Santo Andr e - SP - Brazil. CEP 09210-580; Email: [email protected]
Key points
• • •
A community sample of 654 children from a developing country was scanned using functional magnetic resonance imaging (fMRI) under a resting-state protocol. We built a default mode network maturation (DMN) index based on resting-state fMRI, fractional amplitude of low frequency fluctuations and machine-learning techniques. We found an association between delay in DMN maturation and psychopathological manifestations.
References Achenbach, T.M., & Rescorla, L.A. (2001). Manual for the ASEBA school-age forms and profiles. Burlington, VT: University of Vermont, Research Center for Children, Youth, and Families. Broyd, S.J., Demanuele, C., Debener, S., Helps, S.K., James, C.J., & Sonuga-Barke, E.J. (2009). Default-mode brain dysfunction in mental disorders: A systematic review. Neuroscience and Biobehavioral Reviews, 33, 279–296. Brunner, M., Nagy, G., & Wilhelm, O. (2012). A tutorial on hierarchically structured constructs. Journal of Personality, 80, 796–846. Buckner, R.L., Andrews-Hanna, J.R., & Schacter, D.L. (2008). The brain’s default network: Anatomy, function, and relevance to disease. Annals of the New York Academy of Sciences, 1124, 1–38 Review. Cannon, T.D., Cadenhead, K., Cornblatt, B., Woods, S.W., Addington, J., Walker, E., . . . & Heinssen, R. (2008). Prediction of psychosis in youth at high clinical risk: A multisite longitudinal study in North America. Archives of General Psychiatry, 65, 28–37. Castellanos, F.X., Margulies, D.S., Kelly, C., Uddin, L.Q., Ghaffari, M., Kirsch, A., . . . & Milham, M.P. (2008). Cingulate-precuneus interactions: A new locus of dysfunction in adult attention-deficit/hyperactivity disorder. Biological Psychiatry, 63, 332–337. Chabernaud, C., Mennes, M., Kelly, C., Nooner, K., Di Martino, A., Castellanos, F.X., & Milham, M.P. (2012). Dimensional brain-behavior relationships in children with attentiondeficit/hyperactivity disorder. Biological Psychiatry, 71, 434–442. Chang, X., Shen, H., Wang, L., Liu, Z., Xin, W., Hu, D., & Miao, D. (2014). Altered default mode and fronto-parietal network subsystems in patients with schizophrenia and their unaffected siblings. Brain Research, 1562, 87–99.
Cox, R.W. (1996). AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Computers and Biomedical Research, 29, 162–173. Cuthbert, B.N., & Insel, T.R. (2013). Toward the future of psychiatric diagnosis: The seven pillars of RDoC. BMC Medicine, 11, 126. Dosenbach, N.U., Nardos, B., Cohen, A.L., Fair, D.A., Power, J.D., Church, J.A., . . . & Schlaggar, B.L. (2010). Prediction of individual brain maturity using fMRI. Science, 329, 1358– 1361. Ducharme, S., Albaugh, M.D., Hudziak, J.J., Botteron, K.N., Nguyen, T.V., Truong, C., . . . & Brain Development Cooperative (2014). Anxious/depressed symptoms are linked to right ventromedial prefrontal cortical thickness maturation in healthy children and young adults. Cerebral Cortex, 24, 2941–2950. Ducharme, S., Hudziak, J.J., Botteron, K.N., Albaugh, M.D., Nguyen, T.V., Karama, S., . . . & Brain Development Cooperative (2012). Decreased regional cortical thickness and thinning rate are associated with inattention symptoms in healthy children. Journal of the American Academy of Child and Adolescent Psychiatry, 51, 18–27 e12. Fair, D.A., Cohen, A.L., Dosenbach, N.U., Church, J.A., Miezin, F.M., Barch, D.M., . . . & Schlaggar, B.L. (2008). The maturing architecture of the brain’s default network. Proceedings of the National Academy of Sciences of the United States of America, 105, 4028–4032. Garrity, A.G., Pearlson, G.D., McKiernan, K., Lloyd, D., Kiehl, K.A., & Calhoun, V.D. (2007). Aberrant “default mode” functional connectivity in schizophrenia. American Journal of Psychiatry, 164, 450–457. Gentili, C., Ricciardi, E., Gobbini, M.I., Santarelli, M.F., Haxby, J.V., Pietrini, P., & Guazzelli, M. (2009). Beyond amygdala: Default mode network activity differs between patients with social phobia and healthy controls. Brain Research Bulletin, 79, 409–413. © 2015 Association for Child and Adolescent Mental Health.
60
Jo~ ao Ricardo Sato et al.
J Child Psychol Psychiatr 2016; 57(1): 55–64
Figure 4 Contribution of the fALFF from each region-of-interest to the DMN maturational index. The participation of each region of interest (ROI) in this index was measured by the correlation between the index and the fALFF of the corresponding ROI. Abbreviations: aMPFC = anterior medial prefrontal cortex; vMPFC = ventral medial prefrontal cortex
Table 2 Results of GLM for the CBCL general factor as response variable. Note that the psychopathological manifestations are significantly smaller in the precocious when compared to the delayed group Regressor
Beta
SD
Corrected p
Group-delayed (ref) Group-typical Group-precocious Site Gender Age
– .144 .338 .277 .046 .004
– .075 .116 .042 .042 .002
– .167 .011
