(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Differences in the functional brain architecture of sustained attention and working memory in youth and adults [1]

['Omid Kardan', 'University Of Chicago', 'Chicago', 'Illinois', 'United States Of America', 'University Of Michigan', 'Ann Arbor', 'Michigan', 'Andrew J. Stier', 'Carlos Cardenas-Iniguez']

Date: 2023-01

Sustained attention (SA) and working memory (WM) are critical processes, but the brain networks supporting these abilities in development are unknown. We characterized the functional brain architecture of SA and WM in 9- to 11-year-old children and adults. First, we found that adult network predictors of SA generalized to predict individual differences and fluctuations in SA in youth. A WM model predicted WM performance both across and within children—and captured individual differences in later recognition memory—but underperformed in youth relative to adults. We next characterized functional connections differentially related to SA and WM in youth compared to adults. Results revealed 2 network configurations: a dominant architecture predicting performance in both age groups and a secondary architecture, more prominent for WM than SA, predicting performance in each age group differently. Thus, functional connectivity (FC) predicts SA and WM in youth, with networks predicting WM performance differing more between youths and adults than those predicting SA.

Funding: OK was supported by National Institute on Alcohol Abuse and Alcoholism T32 AA007477. This research was supported by the National Science Foundation (BCS-2043740 to MDR, S&CC-1952050 to MGB, DGE-1746045 to KES, BCS-1558497 to MMC), National Institutes of Health (MH 108591 to MMC), and the University of Chicago Micro-Metcalf Program to YD and LT. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data Availability: All ABCD curated data are available at https://nda.nih.gov/edit_collection.html?id=2573 . The sustained attention network masks are available at https://github.com/monicadrosenberg/Rosenberg_PNAS2020 . Specific data used in the analyses of the current study are shared in the National Institute of Mental Health Data Archive (NDA) study 1849 with doi: 10.15154/1528288 . Users with an NDA account and approved Data Use Agreement can download the shared data. HCP data are available at https://db.humanconnectome.org . Analysis scripts to generate results and figures in the manuscript are available at https://github.com/okardan/ABCD_SA-WM .

Copyright: © 2022 Kardan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

To achieve this, in Study 2, we combined the behavioral and fMRI data from the youth sample ABCD Study (9 to 11 years old) and a large adult sample from the Human Connectome Project (HCP; 21 to 36+ years old). We first investigated the developmental differences in SA within the constraints of adult networks by (a) benchmarking the adult models’ fit to novel adults compared to the preadolescents; and (b) computationally lesioning different components of the adult networks and comparing how much different regions of the networks were contributing to SA in the youths versus novel adults. Second, we used a multivariate method to find the set of connections that differentiated the youth and adult connectomes with regards to SA and WM performances.

In the second approach, we aimed to characterize the functional connections that were differentially related to SA and WM performance in children compared to adults. This was performed both within the constraints of the adult networks and in a whole-brain data-driven manner. This approach can complement both Study 1 as well as existing work that has revealed, for example, changes in the coupling of structural and functional connectivity (FC) profiles that may support improvements in working memory and executive abilities in adolescence [ 25 ].

To address these gaps, in Study 1, we utilized previously developed neuromarkers in the form of large-scale functional networks defined to predict SA [ 23 ] and WM [ 24 ] in adults. We applied these adult connectome-based models to data from the Adolescent Brain Cognitive Development (ABCD) Study to predict individual differences and block-to-block changes in SA and WM task performance in youth. In addition, to characterize relationships between sustained attention, working memory, and long-term memory, we asked whether these same models not only predicted ongoing task performance, but also predict subsequent recognition memory for task stimuli. Successful model generalization would suggest that the functional networks underlying SA and WM overlap between children and adults. Furthermore, a dissociation such that the neuromarker of sustained attention captures sustained attention performance, whereas the neuromarker of working memory captures working memory performance would provide evidence that separable networks support these processes in development.

Despite the popularity of connectome-based predictive modeling of behavior, cross-dataset and cross-population testing is rare. In other words, brain-based predictive models defined in 1 dataset are rarely validated in other samples, and even less so in other participant populations (e.g., different ages or diagnoses, see [ 20 ]). Hence, existing “publication preregistered” brain markers are currently underutilized and under scrutinized, which obscures both their potential and limitations [ 21 ]. Testing the generalizability of connectivity-based models across ages can inform the degree to which adults and children share common network predictors of cognition and delineate models’ predictive boundaries. Cross-age model validation may also provide insight into how networks underlying cognitive and attentional processes change with development. Additionally, validating models of different cognitive processes (e.g., sustained attention and working memory) to evaluate their unique contributions to predicting behavior can determine if distinctions between the models are behaviorally relevant (e.g. see [ 22 ]) and whether those distinctions generalize to different stages of development.

Therefore, in the current study, we aimed to understand the development of sustained attention (SA) and working memory (WM) through the lens of network neuroscience. To do this, we used 2 approaches. First, we assessed the degree to which connectome-based predictive models of SA and WM defined in adults generalize to predict SA and WM in preadolescents. Second, we characterized the functional brain connections that are differentially related to SA and WM performance in preadolescents compared to adults, both within the constraints of the adult networks and in a whole-brain data-driven manner.

Network neuroscience proposes that cognitive and attentional processes are emergent properties of interactions between brain regions [ 13 ]. The success of recent work predicting behavior based on functional magnetic resonance imaging (fMRI) functional connectivity (i.e., the correlation between synchronous blood-oxygen-level-dependent (BOLD) activity among pairs of brain regions) supports the tenability of this position [ 14 – 18 ]. In other words, this work suggests that the degree to which activity is coordinated across large-scale brain networks may better characterize cognitive processes than the magnitude of activity in single regions in isolation [ 19 ].

Maintaining focus over time and information in working memory—related but separable functions [ 1 – 5 ]—are foundational cognitive processes critical for successfully performing everyday activities across the lifespan. In addition to being integral to everyday life, these cognitive processes vary greatly across individuals [ 6 – 8 ] and fluctuate over time within the same person [ 9 ]. These inter- and intra-individual differences are particularly important to study in development because of their consequences for life-long achievements. For example, research in children and adolescents has suggested that attention is more predictive of later academic achievement than more general problem behaviors (e.g., aggression and noncompliance) and interpersonal skills [ 10 – 12 ].

Results

Study 1 overview In Study 1, we tested the generalizability of network models previously defined to predict SA and WM in adults to children. In Study 1.1, we asked whether the degree to which children expressed FC markers of SA [23] and WM [24] previously defined in adult data during an in-scanner n-back task predicted their task performance (Fig 1). We hypothesized that the SA connectome-based predictive model would predict 0-back task performance because this low-working-memory-load task is essentially a target detection task similar to a continuous performance task (CPT) traditionally used to assess SA (e.g., [26]). The SA network may or may not predict 2-back task performance: Although working memory and attention fluctuate in tandem in adults [27], sustained attention is not sufficient for successful 2-back task performance. We hypothesized that the WM connectome-based predictive model, on the other hand, would predict 2-back performance, and that model predictions would be more closely related to 2-back than to 0-back performance because successful 2-back (but not 0-back) performance requires the continuous maintenance and updating of items in working memory. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Overview of Study 1. First, we constructed block-wise FC by correlating BOLD signal time series from all pairs of functional parcels (left). For each participant, we calculated whole-brain FC patterns from fMRI data collected during the eight 0-back and eight 2-back tasks blocks. That is, we calculated up to 16 FC matrices per individual: 1 using data from each 25-s (30–31 volumes) n-back block separately. Each of the 2 predefined predictive network masks were then applied to each of these matrices to generate block-specific WM or SA network strength scores (middle). Each child’s mean network strength during 0-back and 2-back blocks was compared to their mean accuracy in 0-back and 2-back blocks (Study 1.1) or their mean out-of-scanner recognition memory for n-back stimuli (Study 1.3). In Study 1.2, block-to-block changes in network strength were compared to corresponding block-to-block changes in 0-back and 2-back accuracy within-subjects. ABCD, Adolescent Brain Cognitive Development; BOLD, blood-oxygen-level-dependent; fMRI, functional magnetic resonance imaging; SA, sustained attention; WM, working memory. https://doi.org/10.1371/journal.pbio.3001938.g001 In Study 1.2, we asked whether these same adult-defined network models captured changes in SA and WM over time in children—i.e., whether block-to-block changes in network strength predicted block-to-block changes in n-back task accuracy (Fig 1, right panel). Again, we tested for specificity, asking whether the sustained attention and working memory networks better predicted sustained attention (0-back) and working memory (2-back) performance fluctuations, respectively. Finally, in Study 1.3, we asked: Does the degree to which an individual shows a FC signature of better sustained attention or working memory only affect concurrent task performance, or does it also impact later cognitive processes, such as long-term memory? To investigate this question, we evaluated the consequences of SA and WM network expression for long-term memory by testing whether network strength during the n-back task predicted post-scan recognition memory for task stimuli.

Study 1.1. Predicting sustained attention and working memory across participants Do functional network models defined to predict SA and WM in adulthood generalize to a large, heterogeneous developmental sample to predict individual differences in these abilities? To test this possibility, we applied our adult connectome-based models of SA and WM to functional connectivity observed during 9- to 11-year-olds’ performance of 2 n-back task conditions.

Relationship between sustained attention and working memory networks We hypothesized that the predefined SA [23] and WM [24] network masks (Fig 2) capture related but distinct aspects of cognitive function (see Methods for descriptions of these networks). Prior to predicting behavioral performance in the ABCD Study sample, we assessed this hypothesis by (a) comparing the anatomy of the SA and WM network masks; and (b) comparing the strength of the SA and WM networks across participants in the ABCD Study sample. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Adult SA and WM networks and their differences. The circle plots (connectograms) show the SA and WM networks on the Shen-268 parcels grouped into 20 anatomical regions (10 per hemisphere). The networks positively related to the behavior are shown on the top row and the networks negatively related to the behavior are shown in the bottom row. The matrix plots show the percentage of edges belonging to each macroscale region in WM connectogram minus the percentage of edges belonging to each macroscale region in SA connectogram. Differences in networks predicting better attention and working memory are shown in the top matrix plot; differences in networks predicting worse attention and working memory are shown in the bottom matrix plot. The data for this figure are available at NDA study 1849 10.15154/1528288. Circular plots are made with Circos [28]. SA, sustained attention; WM, working memory. https://doi.org/10.1371/journal.pbio.3001938.g002 First, we found that although the SA and WM networks both include functional connections, or edges, representing coordinated activity across distributed brain regions, they show little overlap. Thirty-seven edges are common to the high-attention and high-working-memory networks (1.5% of combined edges in high-attention and high-working-memory networks, hypergeometric p = 0.351; see Methods), which predict better SA and WM performance, respectively. Thirty-three edges are common to the low-attention and low-working-memory networks (1.8% of combined edges, hypergeometric p = 0.005), which predict worse SA and WM performance, respectively. Most of these common edges involved prefrontal (32%), motor (21%), and temporal (16%) regions in the high-attention and high-working-memory networks; and cerebellar (45%), occipital (18%), and parietal (18%) regions in the low-attention and low-working-memory networks. There is no significant overlap between the high-attention and low-working-memory networks (19 edges, 0.9%, p = 0.89) or the low-attention and high-working-memory networks (12 edges, 0.5%, p = 0.99). At the macroscale region level, the SA networks are more dominated by cerebellar, temporal, and occipital connections, whereas the WM networks include more prefrontal connections (Fig 2). Anatomical differences between the SA and WM networks, however, do not guarantee that their strength does not covary together across participants. That is, the degree to which an individual expresses the networks may not be independent. As such, we correlated SA and WM network strength during the 0-back and 2-back tasks in the ABCD Study sample (see Methods). Briefly, in the 0-back task, children were instructed to detect a target image, shown in the beginning of the block, among a series of images by pressing index versus middle finger on the response box. In the 2-back task, children saw a series of images and determined if the image in each trial matched that of 2 trials prior to it or not, again by pressing middle versus index finger. In each block, images from 1 of 4 categories: faces with positive, negative, and neutral expressions and scenes were used in the task. Results revealed that SA and WM network strength values were positively correlated across children during the 0-back task (r = 0.16, p adj < 0.001), but negatively correlated during the 2-back task (r = –0.11, p adj < 0.001; black scatterplots in Fig 3). We found a similar pattern of results within participants, such that SA and WM network strength values were positively correlated across 0-back task blocks (mean r = 0.05, CI = [0.02, 0.07], p < 0.001) and negatively correlated across 2-back task blocks (mean r = −0.04, CI = [−0.02, −07], p = 0.001). Taken together, the anatomical overlap and network strength correlation analyses both across and within participants suggest that the SA and WM masks are separable functional networks in children and thus likely do not reflect a monolithic cognitive process. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Network strengths across the participants and tasks. Correlations between predictive networks strength values across the participants in the 2-back task and 0-back tasks. Task-congruent relationships are shown in black scatterplots. The data for this figure are available at NDA study 1849 10.15154/1528288. https://doi.org/10.1371/journal.pbio.3001938.g003

Neuromarkers differentially predict sustained attention and working memory abilities After confirming that the SA and WM networks are separable in children, we asked whether they generalize to specifically predict these abilities in children. To answer this question, we related adult SA and WM network strength values to task performance during 0-back and 2-back task blocks across the 9- to 11-year-old participants. Again, we predicted that the SA model would capture 0-back performance, whereas the WM model would capture 2-back performance. Supporting our hypothesis, we found that strength of the adult SA network predicted 0-back performance (r = 0.19, ρ = 0.15, p adj < 0.001) and strength of the adult WM network predicted 2-back performance (r = 0.13, ρ = 0.14, p adj < 0.001) in the preadolescent youth. This external validation demonstrates cross-dataset and cross-age generalizability of the SA and WM connectome-based predictive models (Fig 4). This result suggests that the FC features that predict individual differences in sustained attentional and working memory abilities in adults are present and predictive in 9- to 11-year-olds. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. Strength of adult SA and WM networks in preadolescents predict respective task performance. Correlations between SA (left, golden) and WM (right, blue) network strength and children’s 0-back (top) and 2-back (bottom) task performance. The data for this figure are available at NDA study 1849 10.15154/1528288. SA, sustained attention; WM, working memory. https://doi.org/10.1371/journal.pbio.3001938.g004 Finally, we adjusted for data exclusions on household income, age, maternal education, race/ethnicity, sex, household size, and mean Child Behavior Checklist scores using non-participation sensitivity analysis (see Methods). The corrected correlation between WM network strength and 2-back and 0-back accuracy (r corrected = 0.134 and r corrected = 0.115, respectively; ps < 0.001), as well as between SA network strength and 0-back accuracy (r corrected = 0.246, p < 0.001) were significant. Thus, these correlations are robust to the application of post-stratification weights (to account for differences between the full ABCD Study sample and nationally representative sociodemographics) and non-participation weights (to account for differences between the full ABCD Study sample and the 1,545 participants included our analyses with respect to representation of sociodemographics and psychopathology). To assess model specificity, we compared the predictive power of the SA and WM networks for 0-back and 2-back accuracy. SA network strength was not significantly related to 2-back accuracy (r = –0.04, ρ = −0.07, p = 0.11). This correlation was significantly weaker than the correlation between WM network strength and 2-back accuracy (William’s t [test of difference between 2 dependent correlations sharing 1 variable] = –4.65, p < 0.001). Thus, the WM network was a better predictor of performance on the high-working-memory load 2-back task than the SA network. We did not observe this dissociation for the 0-back task accuracy. Instead, WM network strength predicted 0-back accuracy (r = 0.15, ρ = 0.14, p adj < 0.001), and this correlation was numerically but not significantly lower than the correlation between SA network strength and 0-back accuracy (William’s t = –1.27, p = 0.20). SA network strength was more correlated with 0-back accuracy than it was with 2-back accuracy (William’s t = 10.93, p < 0.001), but the WM strength was not more predictive of 2-back than it was of 0-back accuracy (William’s t = −0.78, p = 0.44). Strength in the SA and WM networks was correlated across children (r = 0.16, p < 0.001 during 0-back; r = –0.11, p < 0.001 during 2-back; Fig 3), and performance in 0-back and 2-back tasks are typically correlated across individuals (r = 0.62, p < 0.001 in the current sample of 1,545 children). Thus, it is important to further assess the unique contributions of the SA and WM networks to 0-back and 2-back task performance. To this end, we included both SA and WM network strength in a regression model to predict either 0-back or 2-back accuracy (Table 1). The regression also included age, sex, and remaining head motion (after exclusion, see Methods) as covariates, as well as random intercepts for data collection sites. Echoing the correlation results, SA network strength predicted 0-back accuracy better than chance (β = 0.16, t = 6.35, p < 0.001) and better than it predicted 2-back performance (p < 0.001 based on bootstrapped distribution of the difference between β coefficients). In contrast, WM network strength predicted 0-back and 2-back accuracy above chance but equally well (β = 0.11, t = 4.19, and β = 0.10, t = 3.89, respectively; p values < 0.001). Therefore, we found partial support for the specificity of the models, such that the SA network predicts 0-back accuracy better than it predicts 2-back accuracy, whereas the WM network predicts both 0-back and 2-back accuracy. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 1. Strengths of both SA and WM networks as predictors of task performance. https://doi.org/10.1371/journal.pbio.3001938.t001 Two factors may contribute to the lack of specificity of the WM connecotme-based model. First, the 0-back task does require memory for the target image introduced at the start of each 0-back block, and thus is a low-load rather than a no-load task. Second, the adult WM model was originally defined to predict individual differences in 2-back task performance from adult connectomes comprised of both 0-back and 2-back fMRI data in the HCP sample [24], potentially increasing its sensitivity to n-back task performance overall. Future work assessing the SA and WM models’ generalizability to different datasets and behavioral measures will further inform their sensitivity and specificity. Finally, we compared the performance of the SA and WM models to that of 8 canonical functional networks (the medial-prefrontal, frontoparietal, default, subcortical-cerebellar, motor, visual I, visual II, and visual-association networks; [30]). To do so, we regressed n-back accuracy across participants on the average strength of all within-network connections in each network, with age, sex, and motion as covariates and site as random intercept (similar to Table 1). The SA and WM networks significantly outperformed all 8 canonical networks when predicting 0-back and 2-back accuracy, respectively (Z-test between β coefficients of SA network against each of the 8 canonical networks: Zs > 4.74, p values < 0.001 for 0-back accuracy and Zs > 2.39, p values < 0.017 for the WM network and 2-back accuracy). Furthermore, the SA and WM networks also significantly outperformed randomly selected size-matched sets of connections from outside these networks for predicting 0-back and 2-back accuracy, respectively (p values < 1/200 for both networks when compared to 200 random networks).

Study 1.2. Tracking changes in sustained attention and working memory over time Do the adult connectome-based models of sustained attention and working memory also vary with children’s performance fluctuations? To test this, we examined relationships between block-to-block fluctuations in network strength and block-to-block fluctuations in task performance within-participants. We also investigated whether changes in network strength and behavior were driven by stimulus types or were more spontaneous. Mixed-effects block-level regressions with random intercepts for participants (Table 2) showed that block-by-block changes in SA network strength tracked block-to-block fluctuations in 0-back accuracy (β = 0.07, t = 7.37, p < 0.001) and block-to-block fluctuations in WM network strength tracked block-by-block 2-back accuracy (β = 0.04, t = 3.83, p < 0.001). These results were consistent with our predictions in Study 1.1. Again, demonstrating partial specificity, adult SA network strength did not significantly track 2-back accuracy (β = –0.01, t = −1.46 p = 0.015), whereas adult WM network strength did track 0-back accuracy (β = 0.05, t = 5.56, p < 0.001) in youth. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 2. Block-by-block networks strength and task performance. https://doi.org/10.1371/journal.pbio.3001938.t002 The observed relationships between functional network strength and task accuracy are above and beyond the variance in block-by-block n-back accuracy explained by potential practice effects (i.e., run 2 versus run 1) or stimulus type (i.e., positive versus neutral faces, negative versus neutral faces, and places versus neutral faces; see Tables A and B in S1 Text) because these potential sources of variance are included as covariates in the regression model. Despite the numerically small effect sizes, it is noteworthy that the strength of SA and WM networks—developed in completely independent datasets to predict individual differences in adults—track within-person fluctuations in 0-back and 2-back accuracy in children. Similar to the across-participant results, we tested the SA and WM networks against the 8 canonical networks and 200 random size-matched networks in regressions with block motion, stimulus type, and run as covariates and subjects as random intercept (similar to Table 2). SA network strength was significantly more predictive of block-wise 0-back performance than any of the 8 canonical networks (comparison of beta coefficients: Zs > 6.70, ps < 0.001), and WM network strength was significantly more predictive of block-wise 2-back performance than any of the 8 canonical networks (Zs > 3.33, ps < 0.001). Both networks also outperformed 200 random size-match networks (ps < 1/200).

Study 1.3. Working memory network strength predicts subsequent memory Do youth with FC signatures of stronger sustained attention and/or working memory function during memory encoding show better later visual recognition memory? To investigate this question, we measured the relationship between recognition memory performance for n-back task stimuli to individual differences in networks strength values during the n-back task. Recognition memory was assessed after scanning sessions. The n-back recognition memory test included 48 “old” stimuli (which has been presented during the n-back task) and 48 “new” stimuli (which has not been presented), with 12 images each of happy, fearful, and neutral faces as well as places. Participants were asked to rate each picture as either “old” or “new.” Memory performance was measured as the discrimination index (d’) based on all stimuli. Recognition memory d’ was related to strength in the SA and WM networks averaged across all blocks (i.e., both 0-back and 2-back blocks). Results revealed that strength in the WM (r = 0.12, ρ = 0.13, p adj < 0.001), but not the SA (r = 0.01, ρ = −0.01, p = 0.55), network predicted subsequent recognition memory (Fig C in S1 Text). The strength of the WM network was also a significantly better predictor of subsequent memory than any of the 8 canonical functional networks (comparison of beta coefficients: Zs > 5.37, ps < 0.001) or random networks (ps < 1/200). Unsurprisingly, in-scanner n-back performance was correlated with subsequent recognition memory performance across participants (r = 0.31, p < 0.001) [the correlation of recognition memory d’ with 0-back and 2-back accuracy separately is r = 0.26 and r = 0.30, respectively]. Nevertheless, the relationship between WM network strength and subsequent recognition memory remained significant even when in-scanner n-back performance accuracy was included in the regression model as a predictor (β = 0.04, t = 2.24, p = 0.025; Table 3) along with the age, sex, and residual head motion. Thus, the variance in recognition memory performance captured by the WM network is not fully accounted for by in-scanner task performance. This result highlights the unique contribution of the connectivity-based measures to long-term memory predictions. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 3. WM network strength during in-scanner n-back task performance is related to subsequent recognition memory for n-back task stimuli after adjusting for nuisance variables and even n-back performance itself. https://doi.org/10.1371/journal.pbio.3001938.t003

Study 2 overview In Study 2, we directly compared the adult and preadolescent brain networks that support SA and WM. In Study 2.1, we benchmarked the performance of the predefined adult network models in 2 ways to assess the effects of cross-dataset, cross-task, and cross-age generalization on models’ predictive power. In Study 2.2, we asked how networks predicting SA and WM are differently configured in children and adults.

Study 2.1. Benchmarking the predictive power of the adult sustained attention and working memory network models Although the adult SA and WM network models successfully generalized to predict inter- and intra-individual differences in n-back task performance in the ABCD Study sample, effect sizes were modest. In Study 2.1, we benchmarked these effect sizes in 2 ways. First, we asked how close the predictive power of the adult SA and WM models came to a model of general cognitive ability trained in the ABCD Study sample itself. (We did not train separate ABCD-specific SA and WM network models because the ABCD Study task battery does not include an out-of-scanner SA measure.) Second, we asked how close the predictive power of the SA model in Study 1 came to a theoretical maximum for the 0-back and 2-back tasks by applying the same model to data from the high-quality adult HCP sample. (We could not fairly perform this analysis with the WM model because it was defined using HCP data.) Finally, in a post hoc analysis, we trained a new network predictor of SA in adults using 0-back accuracy and tested its generalizability to ABCD sample’s 0-back performance (to maintain same SA task in adults and children).

Building a development-specific connectome-based predictive model To ask how close the predictive power of the adult SA and WM models came to that of a “youth-specific” network predictor of general cognitive ability trained in the ABCD sample itself, we defined a new connectome-based model—the cognitive composite network model—using leave-one-site-out cross-validation in the ABCD Study dataset (see Methods). The cognitive composite network model was defined to predict children’s average performance on 5 out-of-scanner NIH Toolbox tasks (i.e., their “cognitive composite” score; see Methods) because NIH Toolbox data were collected outside the scanner and the cognitive composite score was similarly correlated with 0-back accuracy (r = 0.31, ρ = 0.32, p < 0.001) and 2-back accuracy (r = 0.33, ρ = 0.38, p < 0.001). Thus, it was fair to use the cognitive composite network model to benchmark the predictive power of both the SA networks and WM network models. (In other words, a model built to predict cognitive composite scores would not be biased at the outset to better predict 0-back or 2-back accuracy.) Demonstrating its utility for this analysis, the cognitive composite model successfully predicted cognitive composite scores in left-out ABCD Study sites (r = 0.295, ρ = 0.27, p < 0.001 across all sites; see Figs D and E in S1 Text). The youth cognitive composite network (averaged over all the site-wise models and binarized at a threshold of 0.5) included edges spanning widespread cortical and subcortical-cerebellar areas (Fig 5). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 5. Left: The youth-defined cognitive composite network averaged over all the ABCD site iterations (binarized at a 0.5 threshold). Right: Cognitive composite network strength in 0-back and 2-back task blocks predict 0-back accuracy and 2-back accuracy across the ABCD sample, respectively. The data for this figure are available at NDA study 1849 10.15154/1528288. ABCD, Adolescent Brain Cognitive Development. https://doi.org/10.1371/journal.pbio.3001938.g005 Cognitive composite network strength during 0-back task performance predicted 0-back accuracy (r = 0.23, ρ = 0.23, p < 0.001) and strength during 2-back task performance predicted 2-back accuracy (r = 0.32, ρ = 0.33, p < 0.001) in children from left-out sites (Fig 5; Table A in S1 Text). This youth cognitive composite model significantly outperformed the adult WM model for predicting 2-back accuracy (β = 0.27, t = 10.76 versus β = 0.10, t = 3.89, p < 0.001 from bootstrapping). Surprisingly, however, the adult SA model’s prediction of 0-back accuracy in youth was comparable with that of this ABCD-specific cognitive composite model (β = 0.16, t = 6.35 versus β = 0.19, t = 7.57, p = 0.242, N.S.). Furthermore, including both adult SA and youth cognitive composite network strength of youths in a regression model to predict their 0-back accuracy results in comparable β coefficients for each (β = 0.19, t = 8.20 and β = 0.21, t = 7.59, respectively; Table B in S1 Text). Finally, intra-individual differences analyses revealed that block-to-block changes in the strength of the youth cognitive composite network tracked block-by-block changes in both 0-back and 2-back accuracy. Echoing the block-by-block results observed with the adult network models in Study 1.2 (Table 2; SA network tracking 0-back accuracy: β = 0.07, t = 7.37; WM network tracking 2-back accuracy: β = 0.04, t = 3.83), the effects were significant but subtle (Table C in S1 Text, cognitive composite network tracking 0-back accuracy: β = 0.05, t = 5.37; cognitive composite network tracking 2-back accuracy: β = 0.08, t = 8.33).

Predicting n-back accuracy in adults Compared to the original studies in which these networks were identified, both the SA [23] and WM [24] models show lower predictive power in the current study than they did in adults. This could arise for many reasons, including those related to developmental change (i.e., differences between adults and children) and unrelated to development (e.g., differences in scan sites and parameters and differences in the to-be-predicted behavioral task). We used the HCP dataset to assess the degree to which differences unrelated to developmental change—scan site and parameters and task differences—impacted the predictive power of the SA model. To do so, we replicated the analyses in Studies 1.1 and 1.2 with n-back task HCP data and compared model performance to that achieved in the ABCD dataset. A result that the model predicted adults’ 0-back accuracy better than it predicted children’s would suggest that adult models do not capture well the functional networks underlying SA performance at age 9 to 11 and/or that predictive power was lower in the ABCD sample because of data quality. On the other hand, a result that the model did not predict adults’ 0-back accuracy better than it predicted children’s would suggest that adult models do capture the functional networks underlying sustained attention at age 9 to 11. In this case, predictive power may be lower in the ABCD Study sample than in adult datasets (e.g., [17]) because of site- or scanner-related differences or differences in the to-be-predicted behavioral measure of sustained attention (gradCPT d’ in [23,17] versus 0-back accuracy in the ABCD and HCP samples). HCP analyses included behavioral and fMRI data from 754 adults (405 female, 22 to 25 years old: 174, 26 to 30 years old: 321, 31 to 35 years old: 249, and 36+ years old: 10; see Methods). We applied the adult SA network mask to FC patterns of novel adults from HCP calculated during 0-back and 2-back blocks of the n-back task and related network strength to task performance both across and within subjects. (Again, we did not apply the adult WM network mask to HCP data because it was previously defined in this sample; [24]). Demonstrating cross-dataset validity—and replicating the pattern of results observed in the ABCD sample—the adult SA network predicted individual differences in novel adults in 0-back accuracy (r = 0.17, ρ = 0.12, p < 0.001) but not 2-back accuracy (r = 0.07, ρ = 0.07, p = 0.07; Fig F in S1 Text), with the former correlation being significantly larger than the latter (Steiger’s Z [test for the difference between 2 dependent correlations with different variables] = 2.29, p = 0.02). Results were consistent after adjusting for age, sex, and remaining head motion covariates (see Table D in S1 Text), and the β coefficient was significantly larger for 0-back than 2-back accuracy (β = 0.16, t = 4.51 versus β = 0.07, t = 1.95; p = 0.034 from a bootstrap test). Mixed-effects regressions showed that, within-subject, block-by-block changes in SA network strength tracked block-by-block changes in 0-back accuracy (β = 0.08, t = 6.36, p < 0.001; Table E in S1 Text). Thus, the adult SA network generalized to a novel sample of adults to predict 0-back, but not 2-back, task performance, which is similar to what we had also observed in the ABCD sample (Study 1). The predictive power of the adult SA network was numerically similar for children’s and novel adults’ 0-back task performance (between-subjects: ABCD r = 0.19 golden line in Fig 6, HCP r = 0.17 purple line in Fig 6). This was also true for tracking changes in performance within subjects (ABCD β = 0.07, t = 7.37; HCP β = 0.08, t = 6.36). This suggests that the SA network model captures children’s 0-back (i.e., sustained attention) performance just as well as it captures adults’. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 6. The strength of the adult SA network predicts 0-back accuracy in youth and novel adults. Even though the discriminability of individuals’ task performance is not significantly different within each dataset, i.e., there is no significant difference between the correlations (ABCD = gold, r = 0.19, ρ = 0.15, ps < 0.001; HCP = violet, r = 0.17, ρ = 0.12, ps < 0.001; difference between rs is not significant Z = 0.46, p = 0.456), the mean performance and mean network strength are both significantly larger in adults. Overall, adults show stronger SA networks and better 0-back performance than youth (red + signs show the mean of SA strength and 0-back accuracy for the scatterplots in each dataset). The data for this figure are available at NDA study 1849 10.15154/1528288. ABCD, Adolescent Brain Cognitive Development; HCP, Human Connectome Project; SA, sustained attention. https://doi.org/10.1371/journal.pbio.3001938.g006 Notably, the behavioral performance (mean 0-back accuracy) for adults is higher than youth by 6.2% ± 0.8% (Fig 6; Welch 2 sample t(2,130.6) = 14.3, p < 0.001), suggesting that SA ability is on average stronger in adults than youth. There are multiple possibilities that could explain this. One is that the preadolescent SA network is a “pre-mature” version of the adult SA network and predicts individual differences in youths similar to adults but is on average expressed less strongly at 9 to 11 years old. For example, in Fig 6, the SA network expressed in the functional connectome of the adults is stronger than the 9- to 11-year-olds (Welch 2 sample t(1,448.9) = 35.8, p < 0.001). FC matrices were z-scored within each participant when used to quantify the SA network strength values for this analysis, but scanner differences between HCP and ABCD studies (e.g., spatially non-homogenous differences between HCP and ABCD scans unrelated to development) could still bias the group-average network strength values. Therefore, the current analyses cannot verify this explanation until longitudinal data from the same youths are processed, allowing mediation tests. It is also possible that different subcomponents of the adult SA network predict task performance in adults and youth. To investigate this possibility, we “computationally lesioned” edges with at least 1 node in each of 10 macroscale brain regions from the SA model. We compared the effects of computational lesioning on the prediction of 0-back accuracy in the HCP and ABCD samples by comparing the ΔR2 in lesioned versus the full SA network strength models. We found that lesioning the prefrontal and temporal lobes decreased prediction power more in adults than it did in children (p = 0.015 for the prefrontal and p = 0.007 for the temporal lobe based on bootstrap distribution of ΔR2 values). Lesioning the subcortex, on the other hand, decreased prediction power more in children than it did in adults (p = 0.009). Therefore, the full SA network generalized equally well to adults and youth, though features within this network may contribute differentially to prediction at different ages. Together, Studies 2.1.1 and 2.1.2 demonstrate that the adult SA model captures youth’s individual differences and fluctuations in attention just as well as it captures novel adults’—and it is no worse at predicting attention in youth than is a youth-specific model of cognition defined in the ABCD dataset itself. Furthermore, the adult WM model captures general aspects of attention and memory in youth and is outperformed by a youth model of general cognition, potentially suggesting less consistency in the functional architecture of WM versus SA from age 9 to 11 to young adulthood.

Alternative adult-defined connectivity-based predictor of sustained attention To further assess the finding that there are more differences in the functional architecture of WM versus SA from preadolescence to adulthood, we trained a new connectivity-based predictive model for 0-back performance in adults from the HCP dataset’s n-back fMRI data in a post hoc analysis. We then applied this HCP-based network predictor of SA instead of the Rosenberg and colleagues [23] SA model (which is trained on adults performing a gradual-onset CPT) to ABCD Study data. Consistent with our findings using the Rosenberg and colleagues [23] SA model, the HCP-based network predictor of 0-back in adults predicts 0-back accuracy in the youth sample (r adj = 0.25, p < 0.001) but does not predict 2-back performance (r adj = 0.05, p = 0.056). Importantly, we found that this alternative adult SA model fits the 0-back performance of the youths significantly better than the adult WM model predicts 2-back accuracy in youth (r adj = 0.11; p < 1/1,000 when comparing fits based on bootstrapped distribution of r adj between the 2 models). This analysis further supports our finding that networks supporting SA are more consistent between youths and adults compared to those supporting WM.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3001938

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.