Profiling intra- and inter-individual differences in brain development across early adolescence

2.1. Participants

Longitudinal data were collected as a part of the ongoing Adolescent Brain and Cognitive Development (ABCD) Study, and included in the annual 4.0 data release (https://doi.org/10.15154/1523041). The ABCD Study enrolled 11,880 children 9 to 10 years of age (mean age = 9.49; 48% female) in a 10-year longitudinal study. Participants were recruited at 21 study sites across the United States from elementary schools (private, public, and charter schools) in a sampling design that aimed to represent the nationwide sociodemographic diversity (Garavan et al., 2018). All experimental and consent procedures were approved by the institutional review board and human research protections programs at the University of California San Diego. ABCD Study inclusion criteria included age at enrollment (9.0 to 10.99 years); English fluency; lack of MRI contraindications; no history of traumatic brain injury or major neurological disorder; absence of any non-correctable sensory and/or motor impairments that would preclude the youth’s participation in study procedures; current intoxication at appointment; diagnosis of any DSM-I psychotic, autism spectrum, or substance use disorder; an intellectual disability reported by their caregiver; premature birth, very low birth weight, or perinatal complications; and caregiver knowledge at baseline of an impending move to an area beyond reasonable traveling distance to an ABCD Study site. Each participant provided written assent to participate in the study and their legal guardian provided written agreement to participate. For more information, see Garavan et al. (2018) and (Volkow et al., 2018). Here, we use a subset of data from the ABCD Study, including magnetic resonance imaging (MRI), in addition to variables regarding participants’ age at time of study enrollment, sex at birth and pubertal development. Data include assessments from two time points: baseline enrollment (i.e., at ages ~9–10 years) and year 2 follow-up (i.e., at ages ~11–13 years). Participants’ sex was taken from the baseline time point, while pubertal stage was considered at baseline, year 1 follow-up, and year 2 follow-up.

Participants were excluded from these analyses if they did not have imaging data collected at the 2-year follow-up visit. From this subset, data were further excluded based on image quality. Structural and diffusion-weighted data were included if a participant’s T1- and diffusion-weighted images, respectively, met the ABCD-recommended criteria for inclusion. For T1-weighted images, quality assessments were based on motion, intensity inhomogeneity, white matter and pial surface estimation by Freesurfer, and susceptibility artifacts. Exclusion was recommended if an image exhibited severe artifact in any of those categories. For diffusion-weighted images, quality assessments were based on residual B₀ distortion after processing, coregistration to the participant’s T1, image quality, and segmentation quality, with exclusion recommended if an image exhibited severe artifact in any of those categories. For resting-state scans, quality assessments were based on the number of frames remaining after high-motion (i.e., framewise displacement > 0.3 mm) frames were censored and periods with fewer than 5 contiguous frames were excluded, coregistration to T1 success, B₀ distortion maps, Freesurfer tissue type segmentation quality, and the number of usable time points per run (after censoring and exclusion, runs with fewer than 100 usable time points were excluded). We further excluded any resting-state fMRI data from participants with fewer than 750 usable (i.e., low-motion) frames, or 10 low-motion minutes, across all runs. This decision was based on estimations of scan lengths necessary for reliable resting-state functional connectivity estimates (Birn et al., 2013; Noble et al., 2017). The final sample characteristics for the current study are described in Table 1. Because each structural, diffusion-weighted, and resting-state functional data are included based on their image quality, there are different numbers of subjects included for analyses with different imaging modalities. Those values are in Supplemental Table 2.

2.2. Developmental variables and imaging measures

For clarity, “variables” will be used to refer to non-brain assessments of development and demographics (i.e., age, sex, pubertal status) (i.e., age, sex, pubertal status) and “measures” will be used to refer to brain imaging outcomes throughout.

2.3. Analyses

An analysis plan was registered for this project with the Open Science Framework (OSF; 10.17605/OSF.IO/ZCRW8). Prior to data analysis, missingness was assessed across imaging measures and demographic variables and for the purposes of assessing intra- and inter-individual variability only complete cases were used. As the number of participants with high-quality MRI data differed based on the scan (i.e., T1- and diffusion-weighted, resting-state functional), assessments of imaging measures comprised slightly different numbers of participants. A total of 7547 participants had data from the 2-year follow-up visit, 7115 (95%) of these had high-quality T1-weighted, structural data from both baseline and 2-year follow-up visits (i.e., assessing cortical thickness, volume, area, and gray/white matter contrast), 6248 (84%) of these had high-quality diffusion-weighted data (i.e., assessing fractional anisotropy; mean, longitudinal, and transverse diffusivity; directional and isotropic intracellular diffusion; white matter tract volume), and 4119 (55%) of these had high-quality resting-state functional data (i.e., assessing BOLD variance, within-/between-network connectivity, and subcortical-network connectivity). Further details can be found in Supplementary Table 2.

3.1. Intra-individual changes in brain development

Intra-individual differences as measured by APΔ are presented for measures of brain macrostructure, microstructure, and function in Fig. 1 and Table 3. Consistent with extant literature on this age range, cortical thickness and gray matter volume decreased annually across brain regions, while white matter volume increased. Likewise, measures of white matter organization (i.e., fractional anisotropy) largely increased, while magnitudes of diffusion (i.e., longitudinal, transverse, and mean diffusivity) decreased. However, isotropic intracellular diffusion increased in both gray and white matter regions, while directional intracellular diffusion increased in white matter tracts but decreased in gray matter regions. Moreover, both isotropic and directional intracellular diffusion exhibited greater variability in gray matter than in white matter. Finally, while macro- and microstructural measures largely either decreased or increased across the brain (with the exception of cortical area), measures of brain function exhibited both positive and negative APΔ. Finally, standard deviations of functional measures were an order of magnitude larger, on average, than those of macrostructure and microstructure measures.

3.2. Inter-individual changes in brain development

Inter-individual differences of APΔ showed that brain changes were more similar between individuals in macro- and microstructural measures (Fig. 3A), compared to functional measures, which were more variable between individuals. (Fig. 3B).

Comparisons of variance in APΔ across age bins, sex at birth, and pubertal status revealed significant heteroscedasticity across neuroimaging modalities, with inter-individual variability in the rate of brain macrostructure, microstructure, and functional development (Table 5). Fig. 4 displays Fligner-Killeen (F-K) statistics, representing the amount of inhomogeneity between levels of each variable (i.e., age, sex, puberty) for each brain outcome examined per imaging measure. showing how unevenly individual differences in developmental change are distributed. Higher F-K statistics indicate greater differences in variance, or in other words greater inhomogeneity, in developmental change between age groups, sexes, and pubertal stages, respectively. For example, microstructure measures demonstrate less heterogeneity in inter-individual variance, or more equal distributions of individual differences, across participants of different ages at study enrollment, compared to macrostructure and functional measures. In contrast to the correlations between age and APΔ described above, heteroscedasticity tells us that there is greater variability between individuals of a certain age group at study enrollment as compared to others who enrolled in the study at an earlier or later age. Additional plots of variance seen for each neuroimaging measure by age, sex, and puberty categories are included in Supplemental Figures 2–4.

Overall, most brain measures displayed similar variance in APΔ, or evenly distributed individual differences, based on initial age at study enrollment (binned in equal-sized increments from 9.0–10.99 years) as shown by low F-K scores. Less than or equal to 1% of regions or tracts demonstrated significant differences in variance across children based on their age at study enrollment across macrostructural (Fig. 5A), microstructural (Fig. 6A), and functional (Fig. 7A) variance. With respect to sex, macrostructure changes demonstrated considerably more heteroscedasticity (~13% of regions; Fig. 5B), compared to microstructure (<1% of regions or tracts; Fig. 6B), or functional changes (<1% of regions or networks; Fig. 7B). Unequal distribution of variance in APΔ was more prominent between individuals classified as being in prepubertal, early puberty, and midpubertal stages at study enrollment. With respect to puberty, 24% of regions or tracts across macrostructural measures (Fig. 5C), 6% of GM and WM microstructural measures (Fig. 6C), and 4% of functional measures (Fig. 7C) demonstrated significant heteroscedasticity. While the majority of measures that were heteroscedastic across pubertal stages demonstrated the greatest variance in mid-puberty as compared to early and pre-pubertal development, white matter volume and longitudinal diffusivity demonstrated the greatest variance in pre-pubertal individuals as compared to early and mid-pubertal participants (Supplemental Figure 4).

4.1. Moving away from the mean: widespread individual differences in brain changes from 9–13 years of age

The existing literature on human brain development relies largely on cross-sectional or sequential cohort designs, characterizes development across a much broader age range, and/or focuses on normative or group-level changes over time. Comparatively little focus has been on variability in change. Our findings highlight that even within the narrow developmental window of early adolescence, intra-individual change is vastly noticeable and varies considerably from person to person. On a whole-brain scale, we identified average annualized percent changes (APΔ) consistent with literature on normative brain development across this age range. These include decreases, on average, in gray matter volume (Bethlehem et al., 2022; Lenroot and Giedd, 2006), cortical thickness (Tamnes et al., 2017; Wierenga et al., 2014), gray-to-white matter contrast (Norbom et al., 2019; Paus et al., 2001), mean diffusivity (Schmithorst and Yuan, 2010), and transverse diffusivity (Asato et al., 2010). We additionally replicated increases in white matter volume, fractional anisotropy (Lebel and Deoni, 2018; Schmithorst and Yuan, 2010; Tamnes et al., 2018), white and gray matter isotropic intracellular diffusion, white matter directional intracellular diffusion (in the same dataset; Palmer et al., 2022), and within-network functional connectivity (Fair et al., 2007; Grayson and Fair, 2017; Satterthwaite et al., 2012). Also consistent with prior findings, we identified both increases and decreases in gray matter directional intracellular diffusion (Palmer et al., 2022), subcortical-network functional connectivity (Ji et al., 2019; Langen et al., 2018; van Duijvenvoorde et al., 2019), and BOLD variance (Nomi et al., 2017; Wang et al., 2021), depending on the brain region or network. Larger decreases in gray matter volume were identified in parietal regions than in frontal, temporal, or occipital regions, consistent with prior findings (Lenroot et al., 2007). Fine-grained comparisons with prior literature across microstructural measures are more difficult, unfortunately, as many studies of white matter microstructural development capture a much broader age range (e.g., 5 to 30 years), across which trajectories follow curvilinear trajectories, and studies of gray matter microstructural development are uncommon. However, there are a few notable similarities, including greater annualized percent change in fractional anisotropy of the cingulum than other tracts, along with virtually no change in the fornix (Lebel et al., 2008). Across functional measures, regional and network-wise comparisons with prior literature are difficult, too, given the narrow age range and overall lack of consensus across functional imaging studies of development (Oldham and Fornito, 2019), and differences in large-scale network definitions across the literature (Uddin et al., 2019).

On the other hand, the current estimates of developmental change conflicted with some prior findings. Compared with Sowell et al. (Sowell et al., 2004), this work only found increased cortical thickness in a much more restricted area, the bilateral entorhinal cortex, instead of extended temporal pole and orbitofrontal regions. However, our age range was a bit older than that of the sample in that study. In contrast to integration within networks, findings regarding segregation between large-scale functional brain networks were contrary to prior literature. We found that each network’s changes in connectivity to other networks were zero on average, indicating that changes in functional connectivity strength occur on a connection level, including weakened connectivity with some networks and strengthened connectivity with others, rather than as broad network-level segregation. Finally, there is not much literature on the development of intracellular diffusion across this age range. Here, gray matter directional intracellular diffusion showed opposite directions of change between cortical and some subcortical regions, potentially reflecting differential cytoarchitectonic processes underlying the development of cortical and subcortical regions. While this work moves toward a more comprehensive understanding of developmental change in brain structure and function in late childhood and early adolescence, further time points will be crucial to clarifying these different trajectories.

This work also extends recent findings demonstrating that individual rates of change (i.e., APΔ) are significantly associated with their starting point (here, values at ages 9–10 years) (Mills et al., 2021). Our current study not only extends prior findings into a larger and more generalizable sample of children, it provides evidence that this association generalizes to all brain measures included here; although to varying degrees across brain regions. That is, after controlling for age effects, changes in all brain regions were negatively associated with their initial values at ages 9–10 years, across all measures of brain macrostructure, microstructure, and function. It is worth noting that the “starting points” in this research refer only to the start of measurement in a particular study. A wide range of influences contribute to individual differences in brain development, including genetics and gene-by-environment interactions, as well as socioeconomic status and maternal experiences during pregnancy (Blokland et al., 2012; Brito and Noble, 2014; Farah, 2017; Gao et al., 2015; Gilmore et al., 2018; Harden et al., 2007; Turkheimer et al., 2003; Ursache and Noble, 2016). Such factors can also influence the pace of development (Tooley et al., 2021) and are, thus, likely reflected in the overall changes in brain development reported here. Brain development is further complicated by changes in the relative influence of genes versus the environment throughout childhood and adolescence (Lenroot and Giedd, 2006).

Our findings may reflect curvilinear intra-individual developmental trajectories from 9–12 years, such that higher values at ages 9–10 years are associated with less positive change in the next two years. This is consistent with much of the group-level effects reported in the literature on normative trajectories in structural brain development (Bethlehem et al., 2022; Mills et al., 2016; Wierenga et al., 2014), although there is some evidence of linear developmental trajectories, for example in gray-to-white matter contrast and in white matter area (Norbom et al., 2019; Paus et al., 2001; Wierenga et al., 2014). Finally, these associations may reflect equipotentiality as presented in developmental psychobiology, which implies that many paths in development can lead to the same outcomes or consequences (Bornstein, 2018). For example, while the rate may differ between individuals, the sequence of development is similar, so that individuals with lower brain metric values at ages 9–10 years had either more increasing or less decreasing to do towards neurodevelopmental milestones, depending on the neuroimaging measure and brain region.

4.2. Inter-individual variability in brain maturation is not equally distributed across imaging measures

Recently, more research has focused on leveraging large datasets to study individual differences in development, but these studies still focus largely on differences in mean trajectories. Contextualizing mean trajectories by assessing developmental variance is necessary to comprehensively understand human brain development. Characterizing developmental variance illuminates the ability of normative trajectories to describe individual-level phenomena. Across this sample, functional changes were much more variable than were changes in macrostructure and microstructure. Within measures of macrostructural change, white and gray matter volume showed the greatest inter-individual variability, whereas heterogeneity in annualized change in cortical thickness was minimal. In general, measures of microstructural change showed more inter-individual variability than did macrostructural measures, with the greatest variability seen in directional intracellular diffusion in gray matter. Inter-individual variability in functional measures was an order of magnitude greater than in structural measures, with subcortical-cortical network connectivity showing the greatest variability and within-network connectivity, the least. With the current MRI data it is difficult to disentangle variability inherent to the MRI scanner and scanning sequence from variability due to underlying neurobiological differences between individuals, though this certainly warrants further study. Differences in variability between macrostructural, microstructural, and functional measures may be due to differences in neurobiological variability or they may be due to differences in image acquisition, processing, and inherent measurement error to a given modality (Birn et al., 2013; Botvinik-Nezer et al., 2020; Kirilina et al., 2016).

4.3. Inter-individual variability: brain changes vary between individuals in early adolescents

This work represents a novel application of a classic statistical concept, using heteroscedasticity to assess distributions of neurodevelopmental variability across age, sex, and puberty in children ages 9–13 years. Thus, we describe not only the variability in neurodevelopmental change, but its distribution across developmental factors.

We identified few age-related differences in inter-individual variability (i.e., heteroscedasticity) across measures, including changes in gray matter volume, white matter isotropic intracellular diffusion, and BOLD variance. Interestingly, variability in gray matter volume increased with increasing age, while variability in the other measures demonstrating significant heteroscedasticity across age showed the greatest variability in individuals ages 9.5–10 years at the beginning of data collection. We expect these nuanced differences in distributions of inter-individual variability, within both measures of brain change and specific age ranges, likely reflect the differential timing and neuroanatomy of aspects of brain development in this age range.

Sex differences in variability of brain macrostructure have been previously noted across the lifespan, dominated by greater variability in males compared to females (Wierenga et al., 2022), often referred to as the greater male variability hypothesis. However, this work has been largely cross-sectional, with less information concerning variability in change. Our findings address that gap and show heteroscedasticity in cortical volume, area, and intracellular diffusion; white matter volume; and functional network connectivity exhibit sex differences in variability as it pertains to developmental change. However, these findings only support the greater male variability hypothesis for a few measures of brain change. Male children exhibited greater variability in tracts with significantly heteroscedastic white matter volume changes, and in heteroscedastic network connectivity (both cortical and subcortical) changes. On the other hand, female children showed greater variability in regions of significantly heteroscedastic cortical area, volume, and intracellular diffusion. Changes in most brain regions across each of these measures displayed no sex differences in variability, nor did most brain measures overall. Together, this work extends Wierenga and colleagues’ findings to suggest greater male-than-female variability in changes in structural (i.e., white matter) and functional connections exist, in addition to established cross-sectional differences, along with greater female-than-male variability in cortical macro- and microstructure changes during early adolescence. Thus, we provide some neurobiological limitations to the larger variability hypothesis that posits greater male-than-female variability across a range of psychological and physical attributes (Hyde, 2014; Johnson et al., 2008; Lehre et al., 2009; Winsor, 1927).

Puberty plays a role in neurodevelopmental trajectories (Dai and Scherf, 2019; Goddings et al., 2014; Herting et al., 2018; Vijayakumar et al., 2018), such that individual differences in hormones (Herting et al., 2014; Vijayakumar et al., 2021), pubertal tempo (Vijayakumar et al., 2018; Vijayakumar et al., 2021, 2021), and pubertal staging are associated with brain development (Goddings et al., 2019; Herting and Sowell, 2017). Here, we have identified differences in the variability of annualized brain changes across Tanner stages of pubertal development during the narrow age-range of ages 9–10 years. Interestingly, these changes are more evenly distributed (i.e., less heteroscedastic) across pubertal stages at ages 10–11 years and even more so at ages 11–12 years. Moreover, we show the patterns of variability in neurodevelopment seen across Tanner stages at ages 9–10 years are distinct from and more widespread than those seen across ages and between sexes. Globally, prepubertal individuals exhibit greater variability in heteroscedastic white matter changes (i.e., tract volume and mean diffusivity), while mid-pubertal individuals exhibit greater variability across cortical and subcortical heteroscedastic brain changes (cortical thickness, area, volume, and intracellular diffusion; BOLD variance, and functional connectivity). Taken together, the diversity of inter-individual variability in brain development suggests only some aspects of neurodevelopment display globally monotonic increasing or decreasing variability as a function of puberty at ages 9–10 years. Further research is needed to better understand this diversity in neurodevelopmental trajectories with respect to puberty and whether these patterns reflect differential susceptibility of neurobiological mechanisms to pubertal influencess, between brain regions.

4.4. The importance of heteroscedasticity to developmental cognitive neuroscience

The current study provides novel insight into intra-individual and inter-individual differences in brain development, with notable heterogeneity in annualized patterns of change across the narrow age range of 9–13 years-old. Understanding variability in brain changes and factors contributing to this variability across common developmental categories is essential. Large-scale, univariate studies have highlighted the importance of “unmodeled noise”, or the often overlooked sources of intra- and inter-individual variability that can confuse or obscure brain-phenotype associations (Bandettini et al., 2022; Dubois and Adolphs, 2016). Thus, teasing apart this unmodeled noise is crucial for robust study of brain-phenotype associations in development. Furthermore, profiling variability and sources thereof (e.g., via heteroscedasticity) is crucial for studying normative development and deviations therefrom, which can help illuminate the etiology of neurodevelopmental disorders and the emergence of psychopathology during this crucial period. Well-characterized normative models require robust descriptions of both central tendency and spread of the data. While developmental cognitive neuroscience has made recent, substantial advancements in describing average trajectories, the variation in these trajectories as highlighted here has received comparably little attention. Studying how deviations from normative development underlie neurodevelopmental disorders, the emergence of psychopathology, and social determinants of health depends on, first, understanding the scope and magnitude of normative variations in development. Profiling variation associated with age, sex, and puberty both across various brain regions and brain measures can provide information about which brain regions, tracts, and networks may be more susceptible to risk factors during adolescent development, and the neurobiological processes underlying this susceptibility. Finally, identifying developmental periods and groups of individuals that demonstrate greater inter-individual differences provides clear targets for further study. Thus, the context provided by this work has sizable utility for studying individual differences in brain development and the broad survey of imaging measures maximizes this utility for researchers with a range of interests and methodologies.

4.5. Limitations & future directions

Given that developmental trajectories can be better characterized by three or more timepoints, this line of investigation, assessing variability in brain trajectories, will continue to be important to investigate as additional ABCD Study data is released in future years.

Noise introduced by participant head motion is a nontrivial confound in resting state fMRI analyses and there are significant age, clinical, cognitive, and sociodemographic differences between participants exhibiting greater and less head motion in the ABCD cohort (Cosgrove et al., 2022). These differences limit the generalizability of resting state fMRI, confound studies that use these data, and may also contribute to some of the heteroscedasticity in functional measures identified here. Further, another factor that may confound estimates of variability between MRI modalities (i.e. macrostructure volumes, white matter tracts, and functional networks) is the size of brain regions, tracts, and networks. That is, larger regions represent average values for a greater number of data points (voxels) and, thus, may be more robust to both small, meaningful differences and random noise, resulting in less variability. On the other hand, within-region, -tract, or -network heterogeneity may increase variability of any given measure. Relatedly, it is impossible to remove the impact of differences in test-retest reliability between MRI sequences used to estimate macrostructural (i.e., T1-weighted), microstructural (i.e., diffusion-weighted), and functional (i.e., BOLD fMRI) scans on between-measure differences in variability. For example: although not included here, the task-based fMRI data from the ABCD Study demonstrate notably low reliability even within a session (Kennedy et al., 2022). On the other hand, between-individual differences have been shown to swamp scanner-induced variability across measures of macrostructure, microstructure, and function in an adult sample (Hawco et al., 2018).

Moving to population-level study brings new challenges to neuroimaging research, which has historically depended on small, homogeneous samples, as opposed to the large, diverse samples required for generalizable research. In this transition, researchers should consider the role that various sources of bias play in their work, such as those outlined in bias assessment tools commonly used in environmental epidemiology (Eick et al., 2020), and assess how the structure of variance is associated with sample characteristics and data collection.