Resting-State Functional Magnetic Resonance Imaging (rs-fMRI) analysis

The fMRI data processing is described in detail in^101,102. Briefly, neuroimaging data minimally preprocessed by the Data Analysis, Informatics & Resource Center (DAIRC) of the ABCD study¹⁰⁵ were further processed using our group’s Next Generation Neural Data Analysis platform (NGNDA)¹⁰⁶. Each participant’s fMRI was coregistered to their structural MRI, slice-time and motion corrected, and normalized to MNI152 space. fMRI voxel time series were parcellated using 3 atlases (cortical, subcortical, and cerebellum) and individually denoised, resulting in 1088 signals with high signal-to-noise ratio¹⁰¹. Only runs with ≤ 10% frames censored for motion were considered for further analysis. Brain regions at rest are overall weakly synchronized, thus each participant’s run with the lowest overall median connectivity, typically coinciding with the run with the lowest number of censored frames, was included in the final analyses. These were conducted in the Harvard Medical School’s High-Performance Computing cluster, using the software MATLAB (Release 2021a; Mathworks, Inc).

Resting-state functional connectivity was estimated using signal cross-correlation and mutual information, to assess method dependence. Statistically similar correlation patterns were estimated by the two methods. Results based on peak cross-correlation between each pair of parcel BOLD signals are reported. Connectivity matrices were thresholded using the approaches described in¹⁰¹, using a relatively conservative population-based threshold, estimated as the upper 95% confidence interval of the moderately outlying correlation values (defined as median + 1.5*IQR), so that only moderate and high correlation values were included in the adjacency matrices used to calculate network properties.

Topological properties were estimated at three spatial scales: the entire brain, specific large-scale resting-state networks, and individual brain regions. Networks included those identified in¹⁰⁷ and the reward network, encompassing dorsal prefrontal, orbitofrontal and anterior cingulate cortex, ventral striatum and pallidum, amygdala, thalamus, and hippocampus¹⁰⁸. Algorithms from the Brain Connectivity Toolbox¹⁰⁹ and the NGNDA platform were used in these estimations. Global properties included median connectivity, community structure (modularity and global clustering), global efficiency, small-worldness, topological robustness—measured by natural connectivity¹¹⁰, and topological stability—using the largest eigenvalue of the adjacency matrix¹¹¹. Median connectivity was estimated both within each network and between nodes in and outside a network. Local properties included node eigenvalue centrality (regional topological importance), degree (number of node connections) and local clustering.