What if a brain scan could say whether a young adolescent’s brain looks older or younger than their actual age? This study tests whether that brain-age gap really tracks maturation—or only seems to. Brain-Age Gap Estimation, or BrainAGE, is presented as a tool that purports to evaluate brain maturity in adolescent populations.

The central problem is that it remains unclear whether BrainAGE tracks with other maturational metrics in adolescence. To test that, BrainAGE was related to pubertal and cognitive development using both a previously validated model and a novel model trained specifically on early adolescence.

Across both models and age bands, an older BrainAGE was related to more advanced pubertal development, while the relationship with cognition was less clear and conflicting across models. Longitudinal analysis also revealed moderate to high stability in BrainAGE across early adolescence.

BrainAGE is more commonly used in investigations of older adult populations, but researchers are beginning to use it in developmental cognitive neuroscience. Because BrainAGE was not originally developed as a measure of individual differences in adolescent brain maturation, its interpretation is less clear in adolescents than in older populations.

Differences between BrainAGE and chronological age have conventionally been interpreted as accelerated or decelerated brain maturation in children and adolescents, but that interpretation has not been validated with longitudinal data. Validation is an important step for any measure interpreted as reflecting maturation, and the Tanner stages provide an example of a validated developmental metric.

Developed in the nineteen-sixties, the Tanner stages describe and quantify physical changes associated with puberty, and their validation helped make them widely accepted in clinical and research settings. For adolescent BrainAGE, a similar relationship would be expected with other maturational metrics that develop over adolescence.

This study therefore examines BrainAGE in relation to two domains: pubertal development and cognitive development. Earlier research on cognition and BrainAGE in adolescence has produced mixed results. Some work linked a positive BrainAGE with faster processing speed, while other work associated a negative BrainAGE with better cognitive performance.

Other studies found very small or non-significant effects. The study examines BrainAGE and measures of pubertal and cognitive maturation within two narrow early-adolescent age bands: nine to eleven years and ten to thirteen years. Because the models use structural MRI data, the expectation is that BrainAGE will track maturational metrics reflected in brain-structure changes during these age ranges, including pubertal and cognitive development.

The study also examines BrainAGE stability in early adolescence through a longitudinal analysis. Three different BrainAGE models were used to compare models trained on wide and narrow age ranges, balancing specificity with replication. This strategy balances extensive data from the age range of interest against overly restricting training data and possible predictions.

Study 1 used an existing, previously validated BrainAGE model trained on a wide age range to predict BrainAGE in early adolescents. Study 2 trained and tested two new BrainAGE models specifically on the age ranges of interest: nine to eleven years and ten to thirteen years.

These models were designed to reflect the unique structural changes occurring during early adolescence and to capture changes in those ranges rather than dynamic changes across all adolescence. Although BrainAGE models include adolescents, none had sampled early adolescence specifically, despite its unique period of structural brain development.

Predictions from all models were examined in relation to youth-report pubertal development, parent-report pubertal development, and cognition scores. The ABCD Annual Release 4.0 included baseline data and one- and two-year follow-ups, but only baseline and the two-year follow-up included imaging data.

Baseline imaging data were available for eleven thousand eight hundred seventy-eight participants aged nine to eleven years, with eleven thousand four hundred two remaining after excluding missing or low-quality structural MRI data. At follow-up, imaging data were available for seven thousand eight hundred twenty-seven participants aged ten to thirteen, with seven thousand six hundred ninety-six remaining after the same exclusions.

Model training used the tidymodels framework and the XGBoost machine-learning algorithm. Scan age was predicted from one hundred eighty-nine features, including cortical and subcortical volume and area measurements. The model features included cortical gray matter volume and surface area measurements, along with bilateral global and subcortical volume measurements.

Cortical features came from the Freesurfer default Desikan-Killiany atlas, while bilateral global and subcortical volume measures came from Freesurfer output. ABCD data used for model testing were harmonized across sites with the longCombat package.

BrainAGE models can be susceptible to prediction bias toward the group mean. For BrainAGE, younger participants are more likely to be predicted as slightly older than their chronological age, while older participants are predicted to be slightly younger.

Bias correction required fitting a linear model to the validation set and extracting its intercept and slope. For a corrected estimate, the intercept is subtracted from predicted age and the value is then divided by the slope; age was also included as a covariate to further correct for age-related bias.

Both participants and their parents completed the Pubertal Development Scale, or PDS. The PDS is a five-item scale that measures pubertal stage. Three items are asked of every participant and concern growth in height, skin changes, and body hair growth, while two depend on assigned sex at birth.

The female version asks about menstruation and breast growth, while the male version asks about facial hair and vocal changes. Cognition was measured with the NIH Toolbox Cognition Battery. The Toolbox uses seven cognitive tasks covering domains such as cognitive control, working memory, set shifting, and reading ability.

In the original validation, the existing model had a corrected mean absolute error of one point ninety-eight years, within the one-to-two-year range frequently reported for prior adolescent BrainAGE models. On ABCD data, the existing model had a mean absolute error of two point thirty-two before age-bias correction and one point forty-five after correction at baseline.

At follow-up, its uncorrected mean absolute error was one point three, and its corrected mean absolute error was one point fifty-eight. A slight age bias was observed, and the BrainAGE correlation with chronological age was negative zero point one at baseline and negative zero point zero one at follow-up.

In the baseline wave, covering ages nine to eleven, a higher or more positive BrainAGE was related to more advanced youth-report and parent-report pubertal development. For youth-report pubertal development, the coefficient was zero point twenty-five, with a standard error of zero point zero five and a p-value below zero point zero zero one.

For parent-report pubertal development, the coefficient was zero point twenty-four, with a standard error of zero point zero four and a p-value below zero point zero zero one. A higher or more positive BrainAGE was also related to lower cognition scores, with a coefficient of negative zero point zero zero three and a p-value below zero point zero zero one in both samples.

Figure two shows distributions of brain-age gaps in the baseline sample from a previously validated model. Panels A and B group participants by youth- and parent-reported pubertal stage, with vertical lines marking each group’s mean; the authors report positive associations between brain-age gap and both measures of pubertal development.

Panel C compares the lowest and highest cognition quartiles, with their means also marked, and the accompanying analysis reports an association between brain-age gap and cognition. At the two-year follow-up, a higher or more positive BrainAGE was related to more advanced youth-report and parent-report pubertal development.

For youth-report pubertal development, the coefficient was zero point fifty, with a standard error of zero point zero four and a p-value below zero point zero zero one. For the novel baseline model, the best cross-validation mean absolute error was zero point fourteen, and the hold-out validation mean absolute error was zero point forty-nine.

In the analysis sample, the uncorrected mean absolute error was zero point forty-nine and the corrected mean absolute error was zero point eighty-six. Performance improved over existing models, including the model used in Study 1, although the model still showed slight age bias.

Corrected BrainAGE had a negative zero point twenty correlation with chronological age. For the novel follow-up model, the best cross-validation mean absolute error and the hold-out validation mean absolute error were both zero point fifty-three.

In the analysis sample, the uncorrected mean absolute error was zero point fifty-two and the corrected mean absolute error was two point two. The novel follow-up model performed with similar accuracy to existing adolescent models and showed slight age bias.

Corrected BrainAGE had a correlation of zero point zero four with chronological age. Figure seven shows distributions of Brain Age Gap across pubertal stages and cognition groups in the baseline model. Panels A and B group participants by youth- and parent-reported pubertal development, with vertical dashed lines marking each group’s mean; the authors report positive associations with both measures.

Panel C compares the lowest and highest cognition quartiles, showing their overlapping distributions and mean markers, consistent with the reported association between Brain Age Gap and cognition. Figure nine tracks individual brain-age-gap trajectories across chronological age.

In panel A, both timepoints use the same prediction model, while panel B uses a model specific to each timepoint; colors indicate whether each person’s estimated gap increased or decreased. The reported test–retest consistency is reflected in intraclass correlations of zero point seven for Study 1 and zero point five three for Study 2, showing why model choice matters when interpreting longitudinal change.

The study used three BrainAGE models: one previously validated across a wide age range and two trained on specific, narrow early-adolescent age ranges. Using the previously validated model, BrainAGE was positively related to both youth-report and parent-report pubertal development across the nine-to-eleven and ten-to-thirteen-year age ranges.

That association was replicated with novel models created from independent samples of nine-to-eleven-year-olds and ten-to-thirteen-year-olds. BrainAGE consistently tracked more advanced pubertal development, but its relationship with cognition was inconsistent.

That makes it a promising biological maturation measure, but a questionable stand-in for broader development.