Association of Intrinsic Brain Architecture With Changes in Attentional and Mood Symptoms During Development

Key Points Question Can brain imaging predict future psychiatric symptoms in children? Findings In this 4-year longitudinal cohort study, distinct patterns of resting-state functional connectivity in healthy children predicted changes in psychiatric symptoms. Weaker positive dorsolateral prefrontal connectivity with medial prefrontal cortex predicted a better developmental trajectory for attentional symptoms, whereas weaker positive dorsolateral prefrontal connectivity with subgenual anterior cingulate cortex predicted a worse trajectory for internalizing symptoms (eg, anxiety/depression). Meaning Brain imaging measures can contribute to early identification of children at risk for common psychiatric disorders and thus identify children in need of preventive treatments.


eMethods
Exclusion Criteria: Children were eligible for the LERD study if they met the following general exclusion criteria: No uncorrected vision or hearing problems, no mental retardation (IQ<70), no limited proficiency in English, no brain injury (e.g., history of head trauma, meningitis, epilepsy, etc.), no severe psychiatric disorders (major depression, Tourette's syndrome, obsessivecompulsive disorder), and no ferromagnetic material in their body (e.g., braces). However, because the LERD study related to reading disability (RD) and there is a comorbidity of RD with ADHD, children with ADHD (as well as other mild psychiatric conditions, e.g., oppositionaldefiant disorder, adjustment disorder, mild depression) were not excluded from participation. For the purposes of this paper, we excluded those children who were on medication and performed predictive analyses with and without the children who were diagnosed with ADHD.
ADHD Diagnosis: ADHD status was determined by DSM-IV criteria, which requires that symptoms be present in at least two settings. Therefore, two questionnaires were administered to each child's parent and teacher. For an ADHD diagnosis, participants had to meet the criterion of scoring above the 93rd percentile on at least one of two parent questionnaires/rating scales, and on at least one of the two teacher questionnaires/rating scales (ADHD Rating Scale-IV 1,2 ).
Participants classified as having ADHD also had to meet DSM-IV diagnostic criteria for ADHD based on Diagnostic Interview for Children and Adolescents-IV (DICA-IV 3 ) interview (past or present) conducted with the parent and signs/symptoms must have been present before age 7 and have persisted for longer than 6 months. Children were only considered free of ADHD if they did not meet criteria on the parent and teacher questionnaires/rating scales used to diagnose ADHD and on the DICA-IV. Seven patients who completed the study were diagnosed with ADHD and four of them were on medication. We statistically controlled for all of the ADHD subjects as well as for those four participants who were diagnosed with ADHD and were on medication. In addition, we also controlled for at-risk for reading difficulty as measured by the Woodcock Johnson scores. CBCL reliable change index: When calculating the changes of CBCL, we used a reliable change index (RCI). We assumed a reference Cronbach's alpha 0.90 for CBCL internalizing scores, and 0.86 for CBCL attentional problems 4,5,6 .
Resting state fMRI Analyses: Resting state fMRI data were analyzed in Conn (http://www.nitrc.org/projects/conn) 7 , which incorporates methods to both minimize the influence of head motion artifacts and allow for valid identification of correlated and anticorrelated networks 8 .

Denoising (e.g., Motion and Physiological Aliasing):
To address potential spurious correlations in resting state networks caused by head motion, we used a procedure to identify problematic time points during the scan, using the Artifact Detection Tools (ART, http://www.nitrc.org/projects/artifact_detect) which is implemented in Conn. Specifically, an image was defined as an outlier image if the head displacement in x, y, or z direction was greater than 1.0 mm from the previous frame, or if the global mean intensity in the image was greater than 3 standard deviations from the mean image intensity for the entire resting scan. The temporal timeseries characterizing the estimated subject motion (3 rotation and 3 translation parameters, plus another 6 parameters representing their first-order temporal derivatives) and artifactual covariates (one covariate per artifactual time point consisting of 0's everywhere and a "1" for the artifactual time point), were used as nuisance regressors in the first level General Linear Model (GLM). The anatomical image for each participant was segmented into white matter, grey matter, and cerebrospinal fluid (CSF) masks using SPM12. To minimize partial voluming, the white matter and CSF masks were eroded by one voxel, which resulted in substantially smaller masks than the original segmentations 9 . The eroded white matter and CSF masks were then used as noise regions of interest (ROI). Signals from the white matter and CSF noise ROIs were extracted from the unsmoothed functional volumes to avoid additional risk of contaminating white matter and CSF signals with grey matter signals. The BOLD timeseries within the subject-specific white matter mask (5 PCA parameters) and CSF mask (5 PCA parameters), were then used as temporal covariates and removed from the BOLD functional data using linear regression, and the resulting residual BOLD timeseries were band-pass filtered (0.01Hz < f < 0.10Hz).
Global signal regression, a widely used preprocessing method, was not used because it mathematically mandates negative correlations that prevent the interpretation of anticorrelations 10 and can contribute to spurious group differences in positive correlations 11 .
Instead, the anatomical CompCor (aCompCor) method of noise reduction 12 as implemented in Conn and described above, allows for interpretation of anticorrelations and yields higher specificity and sensitivity compared with global signal regression 9 .
Head Motion: The average number of outliers across all timepoints was 17 out of 160 timepoints. Excluding these timepoints preserved enough data to achieve a stable estimate of resting state networks 13 . Three subjects were dropped due to excessive head motion. Although rs-fMRI/behavior correlations have been called into question due to the fact that motion often correlates with the behavioral measure of interest 14

Seed Definitions:
The default mode network seed was defined as a 10mm sphere around the peak coordinates from literature (MPFC: (-1, 47, -4) 15 ). The selection of these coordinates was based on a number of papers illustrating that a) this MPFC seed region has significant anticorrelations with DLPFC, which correlates with executive function 16 19 . In order to define the sgACC seed to investigate the relationship between sgACC-DLPFC connectivity and the CBCL Internalization, we used Independent Component Analyses to define this component (see below). Previously, we reported that the MPFC is positively correlated with the right DLPFC in children (n=32, age 8) 17 . In the current sample, we applied the right DLPFC mask defined from the previous study to replicate that MPFC at Time 1 (n=94, age 7) similarly shows positive correlation with the a priori right DLPFC mask. The DLPFC mask was defined as a 10mm sphere around the peak coordinates from literature (46,46,6) 14. Specifically, we performed a onesample t-test of the MPFC-seed Fisher-transformed r-maps at Time 1 and then calculated the mean resting state correlations between the MPFC seed region and the a priori DLPFC mask 17 that was generated independently from the current sample ( Figure S1).

ICA analyses:
Because we did not have an a priori ROI for the sgACC, we derived the sgACC-DLPFC component by performing ICA purely on the subjects' functional data, without any reference to behavioral or psychiatric measures. Because this analysis was independent of the CBCL scores there is no potential for the introduction of artifactual biases towards those components that could be more strongly associated with CBCL scores, so we didn't need to perform LOOCV.

Longitudinal Analyses:
Importantly, there were no baseline brain differences between completers and non-completers at a liberal whole-brain threshold (p = .01 uncorrected for multiple comparisons) for all relevant ROIs (MPFC, DLPFC, sgACC).
Cross Validation: For each cross-validation iteration/fold, all subjects' data except one (the test subject) were used to perform an ANCOVA analysis looking for voxel-level associations between functional connectivity and CBCL changes. The resulting set of suprathreshold voxels (height level p<0.001) was then used as a mask to compute the average functional connectivity values for the test (out-of-fold) subject. This procedure was then repeated for every test subject.