Participants
This study used data from a community-based sample of 175 children (91 children with ADHD and 84 non-ADHD controls) between the ages of 9 and 14. All participants were a part of the longitudinal neuroimaging cohort, Neuroimaging of the Children’s Attention Project (NICAP)(32), in Melbourne, Australia. Subjects underwent up to three waves of repeated MRI scans with approximately 18-month intervals. Screening for ADHD was undertaken using parent and teacher reports on Conners 3 ADHD Index (33), followed by diagnostic confirmation using face-to-face diagnostic interview with parents (NIMH Diagnostic Interview Schedule for Children IV [DISC-IV] (34)). Further information regarding participants and study design is detailed in (35). Diagnostic confirmation was initially conducted at recruitment (3 years before neuroimaging) and subsequently repeated at the first wave of neuroimaging. Children with a history of ADHD (i.e., met ADHD criteria at either wave) were included in the ADHD group. The control group had to screen negative to parent and teacher Conners 3 ADHD Index, and not meet criteria for ADHD in diagnostic interviews.
To ensure the quality of imaging data, functional scans with excessive head motion (mean frame-wise displacement greater than 0.5mm (36), n = 10), scans missing field maps (n = 25) and poor-quality DWI scans (hyperintense cerebellum, omission of white matter, problematic bias correction, problem with Freesurfer mask, the presence of excess non-brain voxels n = 45) were excluded from the final analysis. No significant differences were observed between included and excluded participants in the age distribution of control or ADHD groups in one, two or three waves (range p = 0.062–0.960). However, those children with ADHD who were excluded had more severe ADHD symptoms than included ADHD subjects (p < .05).
The final sample with both structural and functional data comprised 278 scans (139 Control, 139 ADHD) across the three assessment waves (see Fig. 1 and Table 1). At any given wave 7–21% of the ADHD group were taking medication related to their diagnoses, and of this subset, medications comprised methylphenidate: 90–100%, atomoxetine: 0–10%. In addition to one of the former, 33–50% were concurrently taking clonidine or fluoxetine.
Table 1
Demographic characteristics of participants
|
ADHD
|
Control
|
Difference
|
Participants wave 1 (% male)
Participants wave 2 (% male)
|
57 (70%)
53 (74%)
|
48 (56%)
56 (57%)
|
χ2 = 0.77
χ2 = 0.08
|
Participants wave 3 (% male)
|
27 (62%)
|
34 (56%)
|
χ2 = 0.80
|
Age wave 1, Mean (SD)
Age wave 2, Mean (SD)
Age wave 3, Mean (SD)
|
10.42 (0.51)
11.83 (0.62)
13.23 (0.74)
|
10.41 (0.42)
11.70 (0.44)
13.08 (0.51)
|
t = 0.71
t = 1.16
t = 0.73
|
dMRI mean head motion wave 1, Mean (SD)
dMRI mean head motion wave 2, Mean (SD)
dMRI mean head motion wave 3, Mean (SD)
rs-fMRI mean head motion wave 1, Mean (SD)
rs-fMRI mean head motion wave 2, Mean (SD)
rs-fMRI mean head motion wave 3, Mean (SD)
DSM inattentive symptom count, Mean (SD)
DSM hyperactive-impulsive symptom count, Mean (SD)
Baseline symptom severity count (Conner 3 ADHD index), Mean (SD)
|
0.44 (0.26)
0.39 (0.14)
0.33 (0.12)
0.19 (0.16)
0.15 (0.11)
0.11 (0.07)
5.59 (2.49)
6.71 (1.69)
13.17 (4.69)
|
0.36 (0.11)
0.38 (0.22)
0.29 (0.06)
0.14 (0.16)
0.14 (0.10)
0.09 (0.05)
1.02 (1.42)
0.62 (0.95)
1.12 (1.97)
|
t = 1.98
t = -0.03
t = 1.48
t = 1.12
t = 0.45
t = 1.02
t = -21.80*
t = -10.60*
t = -16.38*
|
Medicated wave 1 (%)
Medicated wave 2 (%)
Medicated wave 3 (%)
|
12 (21%)
10 (19%)
2 (7%)
|
-
-
-
|
-
-
-
|
NB: *p<0.0001, dMRI – diffusion MRI, rs-fMRI – resting-state functional MRI
[Approximate location of Fig. 1]
[Approximate location of Table 1]
MRI acquisition
All participants underwent a 30 min mock (practice) scanner session to get familiarized to the MRI environment. Subsequently, MRI scans were acquired using a 3-Tesla Siemens scanner at a single site. However, waves 1 and 2 were collected on a TIM Trio scanner and wave 3 was collected after an upgrade to a MAGNETOM Prisma scanner (note that scanner upgrade was accounted for within statistical modelling). Using a 32-channel head coil, functional images were acquired using multi-band accelerated EPI sequences (MB3), with the following parameters: repetition time (TR) = 1500ms, echo time (TE) = 33ms, field of view (FOV) = 255x255mm, flip angle (FA) = 85 deg, 60 axial slices, matrix size = 104x104, voxel size = 2.5mm3, and 250 volumes acquired covering the whole brain in a 6min 33s sequence. Participants were instructed to keep their eyes open and look at a fixation cross. High Angular Resolution Diffusion Imaging (HARDI) data were acquired using a multi-band factor of 3 with the following parameters: b = 2800 sec/mm2, 63 slices, matrix size = 110x100, voxel size = 2.4mm3, FOV = 260x260mm, bandwidth = 1748Hz, acquisition time = 3min 57s. T1 weighted images were acquired using a multi-echo magnetization prepared rapid gradient-echo (MEMPRAGE) sequence along with navigator based prospective motion correction with parameters: TR = 2530 ms, TE = 1.77, 3.51, 5.32 and 7.20ms, FOV = 230x230mm, FA = 7 deg, axial slices = 176, matrix size = 256x232, voxel size = 0.9mm3, acquisition time = 6min 52s (37).
Pre-processing of functional data
Pre-processing of resting state fMRI images was done using FSL 5.0.9 (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki). Standard pre-processing steps such as discarding of four initial volumes to account for initial signal inhomogeneity, motion correction using MCFLIRT (FMRIB’s Linear Registration Tool), B0 unwarping, spatial smoothing using 5mm FWHM, spatial normalization to the MNI template using a 12-parameter affine transformation and registration of fMRI images to MNI space via high resolution T1 images using FSL FLIRT and FNIRT were undertaken (37–39). Further, each preprocessed dataset was decomposed using Multivariate Exploratory Linear Decomposition into Independent Components (MELODIC) in FSL. Following MELODIC, the resulting components from 20 subjects were manually classified as signal or noise based on previously mentioned criteria (40, 41). FIX (FMRIB’s ICA-based Xnoisefier)(42) classifier was trained using these classifications. FIX was then run on all single-session MELODIC output to auto-classify Independent Component Analysis (ICA) components into good vs bad components and denoise the data (37).
Pre-processing of structural data
Diffusion weighted imaging (DWI) data was pre-processed using MRtrix3tissue (https://3tissue.github.io), a fork of the MRtrix software (43). Various commands in MRtrix that work with the help of external software programs such as FSL (44) and ANTS (45) were used to pre-process the raw diffusion images. Pre-processing steps such as denoising (46), Gibbs unringing (47), correction for eddy current, motion (48) and bias field (48), and brain mask estimation was performed on all the subjects. Mean frame wise displacement (49) calculated in each subject’s diffusion space was used for further analysis to reduce motion confounds in diffusion images. After pre-processing the structural data, response functions (50) for white matter, gray matter, cerebero-spinal fluid and the orientation of fibers in each voxel were estimated (Fiber Orientation Distribution [FOD]) (51). Further, global intensity differences among the data were corrected using intensity normalization.
Functional and structural connectome
For each subject, at each wave, functional and structural connectivity matrices were defined using the multi-modal parcellation of human cerebral cortex (HCP-MMP) atlas (360 distinct regions) (52). The volumetric version of the HCP-MMP atlas available in AFNI (53) was used for the analysis, with the atlas converted and mapped into each subject’s surface space using Freesurfer (54–56). For the functional connectivity (FC) matrix, Pearson correlation coefficient between each pair of ROIs was calculated using CONN (Functional Connectivity toolbox, CONN20b), resulting in a connectivity matrix of size 360x360. Structural connectivity (SC) matrix for each subject at each wave was created by following the steps for estimating the whole brain tractogram outlined in Basic and Advanced Tractography (BATMAN) (57). Streamlines were created using anatomically constrained tractography (58), and spherical-deconvolution informed filtering of tracks (SIFT) (59). Further, the SC for each subject at each wave was created using the HCP-MMP atlas by scaling contribution of each streamline to the connectome edge by the inverse of the two node volumes (52). The options “-symmetric” and “zero_diagonal” in MRtrix were used for generating symmetric SC with diagonals set to zero. Further, structural connectivity matrices were thresholded using consistency-based thresholding at the 75th percentile for edge weight coefficient of variation to reduce the influence of false positives and false negatives, and nodes with just zero values were excluded, as suggested in prior research (6).
Structure-function coupling
The structural and functional connectome of 278 datasets (139 Control, 139 ADHD) were used to examine structure-function coupling between late childhood and mid-adolescence (Fig. 2). To calculate structure-function coupling for each region, Spearman’s rank correlation was performed between the non-zero elements of structural and functional connectivity of each region to the average of every other region of the brain (24).
Developmental trajectories of structure-function coupling
Developmental changes of structure-function coupling were examined using generalized additive mixed models (GAMM), to account for longitudinal data and the possibility of linear and nonlinear trajectories. All models included mean frame wise displacement of structural and functional connectivity of each subject, scanner effect (pre vs post upgrade), sex and medication as covariates. GAMMs were implemented in R 4.0.3, with the package ‘mgcv’ (60).
First, we examined age-related changes in structure-function coupling in typically developing children. We compared i) a null model to ii) a main effect of age, in predicting structure-function coupling. Next, we included children with ADHD. To estimate the differential developmental trajectories in children with ADHD relative to typically developing children four different models were used: i) a null model, ii) main effect of age iii) main effect of group, and iv) the interaction of group and age. For all models, the basis dimension for the smooth term was set to 4 (maximum degrees of freedom for smooth term) as recommended by Wood (61). Each model was examined using maximum likelihood function. Models were compared with likelihood ratio tests (LRT) and Akaike Information Criterion (AIC) to identify the best-fitting model. More complex models were chosen over lower models based on significant LRT (p < 0.05) and AIC units < 2 (62). Further, False Discovery Rate (FDR) (p < 0.05) was used to test the statistical significance of coefficients. All the whole brain maps and trajectory plots were created using Pysurfer v0.10.10 (https://pysurfer.github.io/) and RStudio (60) respectively.