Charting Normative Brain Variability Across the Human Lifespan

Introduction

Since De Montbeillard characterized the first curve of his son’s height in the late 18th century, growth charts have become increasingly essential for pediatricians to describe group-level variability and monitor the over-development of individual-level growth [1]. Growth charts can provide a standardized statistical reference to estimate deviations in a child’s physiological measurements (e.g., height and weight). However, there was no such normative data for human brain development, nor for a longer period across the lifespan, which is crucial for evaluating typical and atypical brain maturation and aging. A recent study published in Nature generated the first-ever set of normative brain charts over the human lifespan [2].

Delineating such lifespan charts of the brain faces significant technical challenges. First, a normative lifespan chart requires the curation of extensive data samples from different age stages, which no single cohort has so far achieved. With limited samples, existing brain charts have characterized the developmental trajectories of brain morphology in school-aged children [3]. Aggregating sufficient samples to cover the entire life course would thus be the first critical step in building a lifespan brain chart. The second challenge is integrating data collected by different sites, that is, site effects. Data harmonization is essential for reducing site/scanner bias and increasing the generalizability of findings with multi-site cohorts. This is particularly crucial for creating a brain chart usable for future normative evaluation of new MRI data from ‘unknown’ sites or scanners.

Bethlehem and colleagues overcame these challenges and produced the lifespan chart of the human brain using the largest MRI samples by far—more than 120,000 MRI scans from >100 studies and 100,000 participants aged from 16 postconceptional weeks to 100 years, representing almost the whole human lifespan. They combined generalized additive models for location, scale, and shape with multi-site harmonized big data to delineate the normative trajectories of different brain tissue volumes and other morphometric phenotypes such as surface area and thickness. The lifespan curves show that brain volume generally increases throughout the fetal and neonatal period, infancy, and early childhood, peaking at 5.9 years for cortical gray matter, 14.4 years for subcortical gray matter, and 28.7 years for white matter (Fig. 1). In addition, the growth curve of surface area parallels brain volume (peaking at 10.97 years), while cortical thickness shows an earlier peak at 1.7 years (Fig. 1) [2]. Further important results show regional variability of growth peaks of gray matter volume: relatively later peaks are found in the insula, anterior cingulate cortex, and frontotemporal cortex, highlighting a prolonged maturation of associative regions of the higher-order salience network [4]. Intriguingly, this pattern of peak times resembles the pattern of brain volume alterations across common psychiatric disorders (e.g., schizophrenia and bipolar, depression) [5], suggesting that brain vulnerability to mental disorders might be associated with critical times in brain development.

Fig. 1.

Normative lifespan trajectories of common MRI phenotypes. The median (50th centile) for different global MRI phenotypes is plotted as a function of log-transformed age. This figure is adapted from Figure 3 of Bethlehem et al. [2].

It is crucial to scrutinize the time points that show the most rapid brain changes during development (i.e., developmental milestones). Bethlehem and colleagues present the maximum increase in the velocity of gray matter volume at ~5 months and white matter volume at 2.4 years. Importantly, two new developmental milestones have been identified: the maximum increasing velocity of cortical thickness at mid-gestation (−0.38 years) and the peak increase in the differentiation between gray matter volume and white matter volume, indicative of underlying myelination and synaptic proliferation [2].

An essential contribution of this lifespan brain chart is offering a normative reference for the individual variability of brain changes in the normal population. It can be used to monitor individuals’ development and its associations with brain disorder risks. To this end, the centile score is used, assessing to what extent an individual deviates from the normative distribution of reference samples with the same sex and similar age. The centile score outperforms the raw metrics by showing high longitudinal stability and heritability. With individualized centile scores, widespread differences in brain volume, cortical thickness, and surface area have been reported for various disorders. Using an overall score to summarize centile scores of all morphometric estimates shows the most considerable structural deviations from the normative population in Alzheimer’s disease, with mild cognitive impairment and schizophrenia ranked in the 2nd and 3rd places. Moreover, an interesting finding from clustering analysis on centile scores across disorders reveals clusters of neurodegenerative disorders (e.g., Alzheimer’s disease, frontotemporal dementia, and Lewy body dementia), neurodevelopmental disorders (e.g., autism and attention deficit hyperactivity disorder), as well as mood and anxiety disorders, in line with genomic findings [6].

Bethlehem and colleagues suggest a scheme based on maximum likelihood estimation to incorporate new datasets for future clinical translation [2]. Such a scheme would be able to provide centile scores with the ‘batch effect’ considered via their online platform (http://www.brainchart.io). Notably, the reliability of estimated centile scores of out-of-sample MRI data would be warranted with a sample size >100.

To sum up, the study by Bethlehem and colleagues represents the first set of solutions to chart normative lifespan changes in the human brain morphology. Several remarks are worth considering for future work. First, the current brain charts were generated with MRI data from populations with limited diversity, including a large proportion of European and North American populations. Although neuroimaging studies do not face the same population stratification problem as genetics, increasing data have shown differences in brain size and shape, as well as growth curves between different ethnicities [3]. These differences might be attributed to environmental factors such as cultural diversity and/or genetic factors. Including more populations from Asia, Africa, and South America is necessary to produce more representative brain charts. The rapid growth of neuroimaging data cohorts from China, such as the Chinese Color Nest Project and the undergoing China Brain Project (for review, see [7]), would significantly increase samples from the Chinese population and therefore raise the diversity of the global population in the future. Second, advances in data harmonization methodology may improve the accuracy of the current brain charts. Compared to the statistically-based univariable method used in [2], the state-of-the-art deep-learning-based harmonization framework has been shown to achieve better performance and be flexibly expanded for new datasets as well [8]. Third, the global measures of brain structure involved in the current brain charts are likely insufficient to characterize the brain alterations in several mental disorders, such as autism, major depressive disorder, and anxiety. Fusing the multivariate regional deviations of patients is crucial to delineate the heterogeneity within and across mental disorders. And sharing relevant data would also be largely beneficial for the research and clinical community. Fourth, we must be aware that structural changes in the brain are not necessarily equivalent to or related to brain disorders. Moreover, it might also be more informative for cognitive development to characterize developmental trajectories with brain metrics that directly reflect brain function such as language- or emotion-related functional activation [9], while their reliability and validity must be ensured before modeling normative charts [10].

Conflict of interest

The authors report no financial interests or potential conflicts of interest.