Convergent and Divergent Brain-Cognition Development

Abstract

How brain networks and cognition co-evolve during development remains poorly understood. Here, we use resting-state functional magnetic resonance imaging (rs-fMRI) and cognitive data at baseline and Year 2 of 2,949 individuals in the Adolescent Brain Cognitive Development (ABCD) Study to examine how stable and changing features of brain network organization predict cognitive development during early adolescence. We find that baseline resting-state functional connectivity (FC) more strongly predicts future cognitive ability than baseline cognitive ability. Models trained on baseline FC to predict baseline cognition generalize better to Year 2 FC and cognition, suggesting that brain-cognition relationships strengthen over time. Intriguingly, baseline FC outperforms longitudinal FC change in predicting future cognitive ability. One potential reason is the lower reliability of FC change compared to baseline FC: ICC = 0.24 vs. 0.56. However, reducing baseline FC's reliability by shortening scan duration only partially narrows the predictive gap, suggesting reliability alone cannot be the full explanation. Furthermore, neither baseline FC nor FC change meaningfully predicts longitudinal change in cognitive ability. We also identify converging and diverging predictive network features across cross-sectional and longitudinal models of brain-cognition relationships, revealing a multivariate twist on Simpson's paradox. Together, these findings suggest that during early adolescence, stable individual differences in brain functional network organization play a more critical role than dynamic changes in shaping future cognitive outcomes.

Figure 1.

Illustration of how cross-sectional and longitudinal estimates can converge or diverge. (a) Classical (Simpson’s paradox) illustration, where within-individual changes (arrows) diverge from between-individual differences (dashed line). (b) Convergent cross-sectional and longitudinal brain-cognition models. For example, greater salience network connectivity predicts better cognitive ability among children cross-sectionally (dashed line). Assuming a causal relationship, individuals with larger increases in salience network FC should enjoy greater cognitive gains longitudinally (arrows). (c) Divergent cross-sectional and longitudinal brain-cognition models. Similar to panel (b), greater salience network connectivity predicts better cognitive ability among children cross-sectionally (dashed line). However, individuals with larger reductions in salience network FC enjoy greater cognitive gains longitudinally (arrows). Figure S1 illustrates other possible divergences between cross-sectional and longitudinal estimates of brain–cognition relationship. We hypothesize that convergence and divergence vary across brain networks.

Figure 2.

Individual differences in longitudinal cognitive change during the transition from childhood to adolescence. (a) Spearman’s correlation between baseline and Year 2 cognitive measures. The positive correlations indicate that children with higher baseline cognition generally maintained their cognitive advantage over their peers at Year 2. (b) Longitudinal cognitive change at the group level estimated from a linear mixed effects model. (c) Individual variability in longitudinal cognition change. For each individual and each cognitive measure, cognition change was defined as the difference in scores between the two timepoints. Sex, baseline age and age interval (between baseline and Year 2) were regressed out. Standard deviation was then computed across individuals. (d) Longitudinal changes of the eight cognitive measures for three participants. RAVLT: Rey Auditory Verbal Learning Test (verbal memory); LMT: Little Man Task (spatial reasoning); PicVocab: Picture Vocabulary Task (vocabulary); Flanker: Flanker Task (executive function); Pattern: Pattern Comparison Processing Speed Test (processing speed); Picture: Picture Sequence Memory Test (episodic memory); Reading: Oral Reading Recognition Task (reading ability). PC1: the first principal component of the above seven cognitive measures. Asterisks (*) indicate statistical significance after false discovery rate (FDR) correction at q < 0.05 (Benjamini & Hochberg, 1995).

Figure 3.

Individual differences in longitudinal functional connectivity (FC) change. (a) Cortical parcellation of 400 regions (Yan et al., 2023), which is the homotopic variant of the Schaefer parcellation (Schaefer et al., 2018). Parcel colors are assigned corresponding to 17 large-scale networks (Kong et al., 2021). (b) 19 subcortical regions (Fischl et al., 2002). 419 × 419 FC matrices were computed based on the 419 cortical and subcortical regions. (c) Spearman’s correlation (stability) between baseline (FCY0) and Year 2 FC (FCY2) for each FC edge. Positive correlations indicate that children exhibiting stronger brain connectivity at baseline continued to exhibit stronger brain connectivity than their peers at Year 2. 99.995% of entries were significant after FDR correction with q < 0.05. (d) Visualization of FC stability at the regional level, by summing the rows of panel (c). (e) Longitudinal FC change at the group level based on a linear mixed effects model. 54.96% of entries were significant after FDR correction with q < 0.05. (f) Visualization of longitudinal FC change at the regional level, by summing the rows of panel (e). (g) Individual variability in longitudinal FC change. FC change (z value) was computed for each FC edge (Afyouni, Smith, & Nichols, 2019). Sex, baseline age, age interval (between baseline and Year 2) and head motion at two timepoints were regressed out. Standard deviation was then computed across individuals. (h) Visualization of individual variability in longitudinal FC change at the regional level, by summing the rows of panel (g). (i-k) Individual level FC change for three participants.

Figure 4.

Enhanced FC–cognition relationships during development. (a) Correlation between the actual cognitive principal component (PC1) score and the predicted PC1 score. Model 1 predicted baseline cognition using baseline FC (FCY0 → CogY0). Model 2 predicted Year 2 cognition using baseline FC (FCY0 → CogY2). Model 3 predicted Year 2 cognition using Year 2 FC (FCY2 → CogY2). Insets show null distributions based on 1000 permutations with red dash line corresponding to actual prediction accuracy. (b) Comparison of prediction accuracies across the three models. Each value in the violin plot represents the accuracy (r) for a single cross-validation fold. Asterisks (*) denote above chance prediction after multiple comparisons correction (FDR q < 0.05). Carets (^) denote statistically significant differences between models based on the corrected resampled t-test (Nadeau & Bengio, 2003) with FDR q < 0.05. (c) Predictive network feature (PNF) matrix for each model. PNF was computed using the Haufe transformation (Haufe et al., 2014). Positive values indicate that higher FC were associated with higher predicted cognitive scores, while negative values indicate higher FC were associated with lower predicted cognitive scores. For visualization purposes, each predictive network feature matrix was normalized by dividing all values by the standard deviation of the entire matrix. (d) Models trained on baseline FC to predict baseline cognition improve in accuracy when applied to Year 2 FC and Year 2 cognition. Model1 was used to predict Y2 cognition from Year 2 FC, which we refer to as “model transfer”. Figures S2 and S3 repeat panels (b) and (d) for the other seven cognitive measures, respectively. FC: functional connectivity; PC1: the first principal component of the seven cognitive measures.

Figure 5.

Cross-sectional baseline FC outperforms longitudinal FC change in predicting cognition at Year 2 even accounting for reliability differences. (a) Comparison of prediction accuracies for Pattern Comparison Processing Speed Test at Year 2 across three models. “FCY0” refers to the model trained on baseline FC to predict Year 2 cognition (FCY0 → CogY2). “FCY0 (4min)” uses baseline FC computed from the first 4 minutes of fMRI data to predict Year 2 cognition (FCY0 (4min) → CogY2). “ΔFC” employs the FC change (between Year 2 and baseline) to predict cognition at Year 2 (ΔFC→ CogY2). (b) Same as panel (a) but for PC1. Figure S5 repeats panels (a) and (b) for other 6 cognitive measures. Asterisks (*) denote above chance prediction after multiple comparisons correction (FDR q < 0.05). Carets (^) denote statistically significant difference between models based on the corrected resampled t-test (FDR q < 0.05). (c) Estimated intra-class correlation (ICC) for baseline FC based on 20 minutes of fMRI data. (d) Estimated ICC for FC change based on 20 minutes of fMRI data. (e) Estimated ICC for baseline FC based on 4 minutes of fMRI data. (f) Comparable ICCs were observed for 4-minute baseline FC and 20-minute FC change. Each dot represents the ICC for an FC edge, computed from baseline (x-axis) and FC change (y-axis); dot color indicates point density, with warmer colors reflecting a higher concentration of edges. ICCs were strongly positively correlated (r = 0.68 without subcortical regions; r = 0.67 with subcortical regions). Statistical significance was determined using a spatial permutation (“spin”) test (see Methods) (Alexander-Bloch et al., 2018; Vasa et al., 2018). The yellow dashed line denotes the identity line (y = x) for reference.

Figure 6.

Limited prediction of cognitive change from baseline FC and longitudinal FC change. Each panel corresponds to the prediction accuracy of a different cognitive measure. Within each panel, baseline FC and longitudinal FC change (ΔFC) were used to predict cognitive change (ΔCog). Predictions of baseline cognition from baseline FC (FCY0 → CogY0) were also shown for reference. Asterisks (*) denote above chance prediction after multiple comparisons correction (FDR q < 0.05). RAVLT: Rey Auditory Verbal Learning Test (verbal memory); LMT: Little Man Task (spatial reasoning); PicVocab: Picture Vocabulary Task (vocabulary); Flanker: Flanker Task (executive function); Pattern: Pattern Comparison Processing Speed Test (processing speed); Picture: Picture Sequence Memory Test (episodic memory); Reading: Oral Reading Recognition Task (reading ability). PC1: the first principal component of the above seven cognitive measures.

Figure 7.

Convergent and divergent predictive network features (PNFs) between cross-sectional and longitudinal estimates of FC–cognition relationship. (a) PNFs from cross-sectional model using baseline FC (FCY0) to predict baseline Little Man Task (LMT) score (CogY0). (b) PNFs from longitudinal model using changes in FC (ΔFC) to predict changes in LMT performance (ΔCog) across the two timepoints. (c) Network blocks with consistent PNFs in the cross-sectional model. (d) Network blocks with consistent PNFs in the longitudinal model. PNFs were considered consistent if the average feature value for a network block had the same sign in both cross-sectional and longitudinal models. (e) Network blocks with inconsistent PNFs in the cross-sectional model. (f) Network blocks with inconsistent PNFs in the longitudinal model. PNFs were considered inconsistent if the average feature value had opposite signs across models. (g) Same as (c), visualized using a chord diagram. (h) Same as (e), visualized using a chord diagram. (i) Same as (f), visualized using a chord diagram. For visualization purposes, each predictive network feature matrix was normalized by dividing all values by the standard deviation of the entire matrix.