Distribution of resting-state networks across reading areas
Reading-related regions were selected on the basis of the Neurosynth meta-analytical database [
22], comprising 11,406 studies as of October 31, 2017. NeuroSynth aggregates brain activation data from thousands of studies to return activation likelihood maps based on search terms. The database includes terms such as “word recognition” (74 studies), “visual word” (98 studies), and “language comprehension” (76 studies). We chose to use the term “reading” (427 studies) because it was inclusive of most of the studies returned by narrower results, and it also showed a high degree of consistency with the other terms. NeuroSynth provides two meta-analytic activation maps, one using forward-inference and the other using reverse-inference. The forward-inference map creates a map of brain areas associated with reading-related papers; the reverse-inference map returns the set of brain areas most likely to be active only in reading-related papers. The reverse-inference map thus de-weights domain-general functional areas and is more representative of reading-specific areas than the forward-inference map. Because domain-general processes are fundamental to skilled reading (especially comprehension), our primary interest was in the forward-inference map, but both maps were examined.
The 7-network cortical parcellation from Yeo and colleagues (2011) was used to represent canonical RSNs [
23]. On the basis of resting-state fMRI data from 1000 subjects, this atlas identifies the following RSNs: visual, somatomotor/auditory, limbic, ventral attention, dorsal attention, default mode, and fronto-parietal. The “Liberally Masked” volumetric data was downloaded in MNI152 space and co-registered to the NeuroSynth data. For each inference map, the percentage of activation falling into each RSN (cm
3) was calculated to provide a distribution across each RSN.
Relationship between brain modularity and reading skill
Original data for the modularity analyses were drawn from the third and fourth waves of a larger longitudinal study (NICHD R01 HD067254, 140 children at first wave). Participants completed out-of-scanner cognitive tests and an in-scanner language task. The fMRI task also included an extensive resting-state baseline in each run, which was the primary target of analysis here. Further information about the aims of this grant can also be found elsewhere [
24,
25].
Participants Participants were scanned in the summer and fall following completion of third or fourth grade (ages 8–11). All participants met the following criteria: native English speakers; normal hearing; normal or corrected-to-normal vision; no history of major psychiatric illness or traumatic brain injury/epilepsy; and no contraindication to MRI. Participants and their parents gave written consent to participate at the beginning of the study, with procedures carried out in accordance with Vanderbilt University’s Institutional Review Board (IRB).
Participants completed cognitive tests, including the Wechsler Abbreviated Scale of Intelligence (WASI) [
26] and the Test of Word Reading Efficiency (TOWRE) [
27]. Demographics and test data are summarized in Table
1.
Table 1
Demographics for study participants
Participants | 50 | 45 (15 new) |
Scan runs | 152 | 162 |
Gender | 24 F, 26 M | 23 F, 22 M |
Age at scan (SD) | 9.45 (0.3) | 10.5 (0.3) |
WASI Full-Scale IQ (SD) | 113.0 (15.5) | 111.0 (15.9) |
TOWRE - Total Word Efficiency (SD) | 109.9 (14.8) | 104.6 (17.4) |
Functional MRI data In the MRI scanner, participants performed up to four runs of a language comprehension task, which was crossed on two conditions: the modality of presentation (listening or reading) and the passage genre (expository or narrative). Each fMRI run had two baseline conditions: a modality-specific baseline task and a resting-state block with a fixation cross. The order and duration for each block varied slightly across runs but was approximately: paragraph 1 (70 s), baseline 1 (70 s), paragraph 2 (70 s), baseline 2 (70 s), and resting-state (270 s). Total scan time was 550 s for all runs, and the average amount of resting-state baseline was 272 s (4 m, 32 s) per run.
A scan run was included in the analysis only if a participant had both listening and reading scans in the same genre (e.g., auditory-expository and reading-expository). Therefore, for each year, a participant had data from either 2 or 4 scan runs (about 9 or 18 min of resting-state scan time, respectively). A scan session was excluded based on the following parameters: high-motion volumes exceeding 20%; poor task performance; and absence of a paired modality scan. In total, resting-state data from 50 children in the third wave (152 scans) and 45 children in the fourth wave (162 scans) met inclusion criteria.
Imaging acquisition and preprocessing All fMRI scans were acquired at Vanderbilt University Institute of Imaging Sciences on one of two Philips Achieva 3T MR scanners with a 32-channel head coil. Functional images were acquired using a gradient echo planar imaging sequence with 40 slices acquired parallel to the anterior-posterior commissure plane. Additional imaging parameters for functional images were 250 dynamics; TR = 2200 ms; TE = 30 ms; FOV = 240×240×120 mm; flip angle = 75∘ ; voxel size = 3×3×3.2 mm3.
All scans were first preprocessed using a standard pipeline in FSL (version 5.0.9) [
28], and connectivity analysis was performed in the CONN toolbox [
29]. fMRI data were high-pass filtered at 0.008 Hz, motion-corrected, co-registered to a structural image, normalized to MNI space and smoothed by a 5-mm FWHM spherical kernel. Outlier volumes were identified as individual fMRI volumes in which the RMS framewise-displacement exceeded 0.7. fMRI timeseries were corrected using anatCompCorr methods, which uses signal from white matter tissue and cerebrospinal fluid areas to reduce noise not related to brain activity [
30]. Other covariates of no interest included six rigid motion parameters, six derivative motion parameters, and outlier volumes. Finally, we used a weighted general linear model (GLM) to model the resting block and averaged within-subject across scan sessions to get a de-noised resting-state timeseries.
Network definition To build off of previous work, we created networks using 264 nodes originally published by Power and colleagues (2011) [
31]. The node set covers the entire brain, including subcortical areas, and has been extensively used in graph theory analyses since its publication (e.g., [
32‐
34]). Suggested RSN assignments for each node, totaling 13 unique networks, are also available and were used to partition the network into different RSNs. fMRI timeseries correlations were calculated between each of the the 264 nodes, resulting in a single connectivity array for each subject at each time point. Matrices were then thresholded into binary maps at
r = 0.15. (To confirm that this particular threshold did not unduly influence results, we also tested thresholds at
r = 0.05, 0.10 and 0.20. No significant effect on the results was found.)
Network analysisGlobal modularity (
Q) quantifies how well the whole-brain network segregates into component RSNs. High modularity indicates that the network has much higher connectivity within RSNs compared to between RSNs; low modularity suggests that nodes do not segregate into RSNs well. For undirected and binary networks, each connection in the array is given a positive value if it links nodes in the same RSN, and it is then weighted based on the degree of each node and the total network. (For a detailed treatment, see [
35].) Connection-level values were aggregated to get node-by-node and global measures of modularity for each individual, which were then mean-centered and scaled to unit variance.
A GLM was used to determine the relationship between global modularity and individual performance on the TOWRE (total word efficiency, standard score). If subject data was available at multiple timepoints (n = 30), it was averaged together to produce a single value. Multiple supplementary analyses were also completed to ensure that effects were not driven by cohort or motion confounds: models were also analyzed for grades 3 and 4 separately, and also when including a measure of subject motion (mean global signal change). We also examined whether there were differences in the modularity relationship between TOWRE subtests (sight word efficiency (SWE) and phonemic decoding efficiency (PDE)).
To test whether there was an RSN-level trend in the modularity-to-reading relationships, node modularity values were also investigated. For the set of nodes comprising each RSN, a one-sample t test was performed to see whether the RSN average was significantly greater than the global node average.
Mapping dyslexia abnormalities onto hub areas
Two decades of neuroimaging research have allowed a relative consensus to form as to which brain regions are commonly dysfunctional in dyslexia. To determine whether there was any pattern related to network architecture in these areas, we gathered all activations from three meta-analyses comparing fMRI responses for individuals with dyslexia to typical readers [
7‐
9]. The meta-analyses encompassed a total of 68 studies comparing dyslexic and typically developing individuals. Sample populations included children and adults under different experimental conditions. All brain areas that showed atypical activation in dyslexia (either greater or less activity) were included. When an activation spanned a large area, all reported local maxima were included.
To get measures of hubness across the brain, we used data from a connectomics study by Power and colleagues [
36]. That study reports the
participation coefficient for each of the 264 nodes previously described. The participation coefficient reflects the diversity of a node’s connectivity to different RSNs, where a higher value indicates that the node is correlated with many different RSNs. Activations from the dyslexia meta-analyses were then mapped to the geometrically closest node from this dataset, resulting in a small set of dyslexia-related nodes and a larger, unaffected set. The nodes and descriptions, along with their suggested system and atlas label, are available as an Additional file
1.
The distribution of participation coefficients across the 264 nodes was non-normal, with a large group of areas having low participation coefficients (i.e., affiliated with few RSNs) and a smaller hub-like group. Therefore, a Wilcoxon rank-sum test was performed on the participation coefficients for the two groups, which tests for the equivalence of two distributions in a non-parametric fashion.