Cognitive and psychiatric relevance of dynamic functional connectivity states in a large (N > 10,000) children population

The study investigates how dynamic functional connectivity (DFC) during resting-state fMRI relates to cognitive performance and mental health in children aged 9–11. Traditional functional connectivity (FC) is often treated as static, yet accumulating evidence indicates FC is time-varying and linked to cognition and psychopathology. Prior pediatric studies have associated resting-state FC with cognitive abilities and dimensions of psychopathology, and emerging work suggests DFC captures transient brain organization relevant to development. However, previous child DFC studies typically used small samples and did not jointly examine interactions among DFC, cognition, and mental health or establish test–retest reliability. The authors hypothesized that: (1) children’s resting-state FC exhibits reproducible DFC states reflecting distinct information transfer capacities; (2) time spent in distinct states relates to cognitive and psychiatric measures; and (3) associations between DFC and cognition are partly mediated by psychiatric symptoms (e.g., attention problems/ADHD). Using large-scale ABCD data, the study aims to establish robust pediatric DFC states and quantify their cognitive and psychiatric relevance.

The paper situates DFC within a broader literature indicating that FC relates to cognitive performance, sex differences, and neuropsychiatric disorders, but that FC is not stationary and varies over time. In children, resting-state FC is a reliable indicator of cognition and is altered in various psychopathologies; abnormalities have been observed in neurocognitive networks linked to internalizing and externalizing symptoms. Prior pediatric DFC studies (e.g., relating state occurrence to self-generated thoughts or social motivation) used small samples and did not address the interplay among mental health, cognition, and DFC. The authors highlight gaps: limited sample sizes, lack of replication/test–retest reliability, and absence of mediation frameworks integrating cognition and psychopathology with DFC.

Data: ABCD Study release 3.0 (NDA) including >11,800 children aged 9–11; after preprocessing and quality control (QC), 10,988 subjects with two good baseline scans and 3,622 with two good second-year scans were retained. Scanners: Siemens Prisma, Philips, GE 3T; common protocol TR/TE = 800/30 ms, voxel size 2.4 mm isotropic, 60 slices, FA 52°, FOV 216×216 mm, multiband 6. Preprocessing: FSL v6.0 and SPM12 pipelines; post-processing of component time courses included detrending (linear–cubic), regression of 6 motion parameters and derivatives, outlier removal, and band-pass filtering (0.01–0.15 Hz). QC included mask correlation thresholds; subjects required ≥2 good scans per session. Feature extraction (NeuroMark): Group ICA (GIFT) with spatially constrained ICA using NeuroMark templates to obtain comparable individual-level components; 53 components (ROIs) grouped into 7 networks: subcortical (5), auditory (2), sensorimotor (9), visual (9), cognitive-control (17), default-mode (7), cerebellar (4). DFC estimation: Sliding-window approach with window size 40 TRs (~32 s) using a tapered (rectangular convolved with Gaussian σ=3 TRs) window; robustness checked with alternative window sizes/σ. For each window, covariance estimated from regularized precision matrices via graphical LASSO (L1 penalty); lambda optimized per scan by held-out log-likelihood. Concatenated windowed FC matrices formed the DFC array. Clustering: K-means (Euclidean) applied to subsampled windows (exemplars selected by local maximum FC variance across pairs) with 500 random initializations. Clustering used 1,378 upper-triangular FC features (53×52/2). Optimal k determined by elbow criterion; k=5 chosen, with robustness checks for k=3–6. Fractional rate (percentage of windows per state) computed as state occurrence. Reliability analyses included state stability, null models (VAR and phase randomization) to reject stationarity, and replication across scans/sessions. Associations: Linear mixed-effects models (LMM) related fractional rates to cognitive (NIH Toolbox raw scores) and psychiatric (CBCL raw scores) measures. Fixed effects: each score, age, sex (birth sex), race/ethnicity, height, weight; random effects: family nested within site. Multiple comparisons controlled via FDR. Robustness checks included motion controls, scanner effects, parental education, medications. Additional validations used ABCD Pearson Scores and KSADS-5/PGBI measures. Test–retest/replication: Split-half by site sets; equalized subsamples per site; replication in baseline scan 2 and both second-year scans; states from replication matched to discovery centroids via centroid correlations. Mediation analysis: Standard three-variable mediation (Mediation Toolbox) tested whether attention and DSM-5 ADHD (CBCL) mediated relationships between state fractional rate (independent variable) and cognitive performance (primarily fluid composite; also crystal and total composite). Confounds regressed; significance via bias-corrected bootstrap (10,000 samples). Graph analysis: For each subject, averaged DFC within state to form mean state-based FC; constructed positive, negative, and absolute graphs to compute node strength, global efficiency, and local efficiency (Brain Connectivity Toolbox). Pairwise t-tests compared states (especially 1 vs 5) with FDR correction.

• Brain parcellation: 53 ICA-derived ROIs grouped into 7 functional networks.
• Reproducible DFC states: Five states identified (k=5). Four of five states replicated across scans/sessions with centroid correlations r>0.8. Example occurrences (discovery baseline scan 1): State 1 ~8.07%, State 2 ~39.05%, State 3 ~17.53%, State 4 ~15.60%, State 5 ~19.74%; replication scans showed similar distributions (e.g., State 2 ~47–48%).
• State characteristics: • State 1: Strong within-network synchrony; strong anticorrelations between networks, notably SM–VS and cerebellum with other networks; lowest occurrence (~8–13%). • State 2: Sparsely connected; highest occurrence (~39–48%). • State 5: Positive FC between SM and VS; positive CC–CB and negative VS–CB; reflects sensory integration with DM–SM antagonism and weaker CB segregation.
• Cognition associations (NIH Toolbox): • Cognitive scores negatively correlated with occurrence of strongly connected states, especially State 1. For State 1, 10/10 cognitive scores significant with r from −0.0446 to −0.0919 (FDR q<0.05). Strongest: cardsort r=−0.0687 (p=7.97×10⁻¹³), fluid composite r=−0.0906 (p=4.52×10⁻²¹), total composite r=−0.0919 (p=1.30×10⁻²¹). • Cognitive scores positively correlated with occurrence of sparsely/weakly connected states, especially State 5. For State 5, 10/10 scores significant with r from 0.0208 to 0.0567. Strongest: cardsort r=0.0567 (p=3.44×10⁻¹⁰), fluid composite r=0.0553 (p=3.13×10⁻⁸), total composite r=0.0490 (p=3.72×10⁻⁷). Results replicated across additional scans/sessions.
• Psychiatric associations (CBCL): • 18/20 psychiatric scores positively correlated with State 1 occurrence (r=0.0310 to 0.0778) and negatively with State 5 (r=−0.0232 to −0.0818), FDR q<0.05. • Strongest effects: attention problems and DSM-5 ADHD with State 1 (attention r=0.0778, p=3.36×10⁻¹⁶; ADHD r=0.0742, p=7.36×10⁻¹⁵) and inverse with State 5 (attention r=−0.0818, p=9.58×10⁻¹⁸; ADHD r=−0.0736, p=1.16×10⁻¹⁴). Replicated across scans.
• Mediation analysis: Mental health mediates DFC–cognition relationships (fluid composite as outcome): • Attention mediates State 1 → cognition: indirect beta=−0.7979, p=2.49×10⁻⁴ (14.08% total effect). • Attention mediates State 5 → cognition: indirect beta=0.8732, p=1.70×10⁻⁴ (23.79%). • ADHD mediates State 1 → cognition: indirect beta=−0.6982, p=2.79×10⁻⁴ (12.29%). • ADHD mediates State 5 → cognition: indirect beta=0.7121, p=1.68×10⁻⁴ (19.35%). • Additional CBCL mediators: For State 1, 18/20 mediated fluid composite (0.84–14.08%); for State 5, 16/20 (2.03–23.79%). Similar mediation patterns for crystal and total composite scores.
• Graph topology differences (State 1 vs 5): • Global efficiency higher in State 1 (p<1×10⁻¹⁰⁰). • Node strength: Most ROIs higher in State 1 for positive and absolute graphs (except VS); VS and CB higher negative strength in State 1; SPL and PCC higher negative/absolute strength in State 5. • Local efficiency: Generally higher in State 1 across ROIs (especially CB) for positive/absolute graphs; SPL and PCC higher in State 5. Negative graph showed VS/CB higher efficiency in State 5, others higher in State 1.
• Additional observation: State 3 showed lowest occurrence and characteristics consistent with reduced vigilance/drowsiness across runs, aligning with prior literature on vigilance fluctuations.

The findings demonstrate robust, replicable DFC states in a large pediatric cohort, linking transient brain network topologies to cognitive performance and psychiatric symptoms. State 1 reflects high within-network synchrony and strong inter-network antagonism (greater segregation), including pronounced cerebellar anticorrelations, and is associated with poorer cognition and higher psychopathology. In contrast, State 5 exhibits greater sensory integration and DM–SM antagonism with weaker cerebellar segregation, and is associated with better cognition and lower psychopathology. These patterns align with theoretical frameworks of functional segregation versus integration: greater segregation relates to attentional deficits and psychopathology, whereas integration supports cognitive performance. The mediation results indicate that attention problems and ADHD partially explain DFC–cognition links, suggesting that atypical allocation of time to segregated versus integrated states may lead to attentional issues that impact cognitive outcomes. Differences in DM and cerebellar connectivity across states may reconcile mixed literature on DM–cognition associations and underscore cerebellar contributions to cognition. The reproducibility across scans and sessions and consistency under parameter variations support the robustness of these DFC patterns in children.

This study identifies five reproducible resting-state DFC states in over 10,000 children and shows that time spent in these states relates to cognitive performance and psychiatric symptoms. A strongly segregated state (State 1) is linked to lower cognitive scores and higher psychopathology, while a more integrated sensory state (State 5) shows the opposite. Attention problems and ADHD partially mediate the DFC–cognition relationships. Graph analyses reveal distinct topologies underpinning these states, with higher global and local efficiencies in the segregated state except for specific nodes (e.g., PCC, SPL). These results suggest that tracking transient DFC states may help characterize cognitive and mental health problems in childhood and inform early interventions. Future work should model longitudinal DFC changes as more ABCD waves accrue and consider integrating univariate and multivariate approaches to improve predictive power while maintaining interpretability.

• Longitudinal changes were only preliminarily explored; comprehensive longitudinal modeling awaits additional ABCD waves.
• Effect sizes are small (|r|≈0.02–0.10), consistent with large-scale brain–behavior studies, but may limit individual-level predictive utility.
• Primary analyses were univariate (fractional rate per state), which enhances interpretability but may understate multivariate predictive effects.
• Although extensive confound controls were applied (motion, scanner, demographics, parental education, medications), residual confounding and head motion effects cannot be entirely excluded.
• State 3 showed lower reproducibility, possibly reflecting vigilance fluctuations, indicating some state heterogeneity across sessions.