AI Brain-Age Tool Adds Almost Two Years to Healthy Children’s Brains, Study Warns

A widely used artificial intelligence tool that estimates how old a brain looks from an MRI scan appears to be telling healthy children that their brains are nearly two years older than they really are. A new study of typically developing children aged six to nine years has found that the BrainStructureAges pipeline, a free online service hosted on the volBrain platform, systematically overestimates brain age in young patients, raising urgent questions about how such biomarkers should be used in pediatric medicine. The research, published in Pediatric Radiology, was led by Ekaterina Korotkova and colleagues at Ural Federal University in Yekaterinburg, Russia, together with collaborators from the Higher School of Economics, a children’s clinical hospital, an MRI clinic, and Ural State Medical University.

Brain age has become one of the most talked-about biomarkers in modern neuroimaging. The idea is elegantly simple: machine learning algorithms are trained on thousands of MRI scans to predict a person’s chronological age from the structure of their brain. If the predicted age is much higher than a person’s actual age, the gap, known as the brain delta or brain-predicted age difference, is interpreted as a sign of accelerated aging or abnormal brain maturation. Studies have linked elevated brain age to a range of conditions, and earlier work has even shown that brain age predictions correlate with mortality risk in adults. In children, the same logic is applied in reverse: a brain that looks older than expected might be assumed to be maturing too quickly, or to have been damaged by illness.

But the approach rests on a hidden assumption that is rarely tested in pediatric settings: that the underlying model was trained on data that adequately represent children’s brains. Many of the large datasets used to build brain age models, such as the Open Access Series of Imaging Studies, consist primarily of young, middle-aged, and older adults. When a model trained mostly on adult brains is asked to estimate the age of a child’s brain, it may simply not have learned what a healthy developing brain looks like, and the resulting predictions could be biased in a consistent direction. This is precisely the concern that motivated the new study.

The research team set out to test whether predicted brain age is systematically overestimated in typically developing children when using the BrainStructureAges pipeline, a relatively new tool developed by researchers at LaBRI, the CNRS and University of Bordeaux laboratory in France, and made available through the volBrain online platform, which is operated by the ITACA Institute at the Universitat Politècnica de València in Spain. The pipeline estimates not just a single whole-brain age but ages for 133 individual brain structures, offering a level of regional detail that few other tools provide. It builds on AssemblyNet, a large ensemble of convolutional neural networks designed for three-dimensional whole-brain MRI segmentation, and was originally introduced as a new biomarker for multi-disease classification.

To evaluate the tool fairly, the researchers assembled a carefully screened sample of 105 brain MRI scans from healthy, typically developing children aged six to nine years, all with well-documented medical histories. The scans were acquired on a Magnetom Amira 1.5-Tesla scanner manufactured by Siemens Healthineers in Erlangen, Germany, using T1-weighted MPRAGE sequences, a standard protocol for structural brain imaging. The study was conducted in accordance with the Declaration of Helsinki, approved by the Local Research Ethics Board of Ural Federal University, and carried out with appropriate parental and guardian consent. By restricting the sample to children with no neurological or developmental concerns, the team created what should have been a best-case scenario for the algorithm: a normative group in which predicted brain age should, on average, match chronological age.

The results were striking. The mean predicted brain age across the sample was 9.3 years, with a standard deviation of 1.2 years, while the children’s mean chronological age was just 7.7 years, with a standard deviation of 0.4 years. That translates to a mean brain delta of 1.6 years, meaning the algorithm judged the average child’s brain to be more than a year and a half older than the child actually was. Even more telling, positive brain delta values, indicating overestimation, appeared in 95.2 percent of participants, and the pattern held across all of the 133 brain structures analyzed. In a truly unbiased model, such deltas should cluster around zero, with roughly as many children appearing brain-young as brain-old.

The overestimation was not uniform across the brain. The largest deviations were concentrated in specific regions: the left cerebellar exterior and the left occipital fusiform gyrus each showed an average overestimate of 3.3 years, followed by cerebellar lobules VI and VII at 2.9 years, the left inferior occipital gyrus at 2.9 years, and the right frontal pole at 2.8 years. The prominence of the cerebellum and occipital regions is notable, since these areas undergo substantial developmental change during childhood, and prior research has flagged difficulties with automatic cerebellar segmentation in pediatric populations using volBrain software. Regional biases of this magnitude suggest that the discrepancy is not random noise but a structured error embedded in how the model interprets immature brain anatomy.

Importantly, the bias did not differ between boys and girls. The researchers found no significant sex differences in chronological age, predicted brain age, or brain delta, with boys showing a mean delta of 1.55 years and girls 1.61 years, a difference that was statistically indistinguishable (t=0.27, P=0.79). This uniformity strengthens the conclusion that the overestimation reflects a property of the algorithm rather than a genuine biological difference in the sample. It also means that any clinical interpretation based on this pipeline would be equally misleading for both sexes in this age range.

The findings arrive at a moment of growing scrutiny for pediatric brain age research. A comprehensive systematic review of artificial intelligence-based brain age prediction in children published in Frontiers in Neuroinformatics catalogued the rapid expansion of the field, and a 2025 commentary in Nature Communications highlighted the current challenges and future directions for brain age prediction in children and adolescents, noting that limited pediatric training data remain a central obstacle. The new results also echo earlier work by some of the same Russian researchers, who reported a surprising increase in predicted brain age among people who had suffered a stroke in childhood and questioned whether such findings represent real biology or pitfalls in MRI analysis. The present study suggests that at least part of the answer lies in the tools themselves.

What should clinicians and researchers take away from this? First, the authors urge caution in clinical interpretation: a brain delta of 1.6 years in a healthy child should not be read as evidence of accelerated maturation or pathology when the measurement tool itself carries a systematic bias of that size. Second, the study underscores the need for bias-adjustment procedures, which have been developed for adult brain age frameworks but require normative pediatric reference data to be applied meaningfully in children. Third, it highlights the value of validating any automated pipeline against a local, well-characterized normative sample before deploying it in research or clinical workflows. As brain age biomarkers move from the laboratory toward the clinic, this brief report serves as a reminder that an algorithm’s confidence is not the same as its accuracy, and that in pediatric neuroimaging, the healthiest brains can end up looking surprisingly old.

Subject of Research: Systematic overestimation of predicted brain age by the BrainStructureAges MRI pipeline in typically developing children

Article Title: Methodological limitations of the BrainStructureAges pipeline from the volBrain platform: a brief report of the normative pediatric group

Article References: Korotkova, E., Kulikova, S., Lvova, O., Tarasov, D., Meshkov, A., & Kovtun, O. (2026). Methodological limitations of the BrainStructureAges pipeline from the volBrain platform: a brief report of the normative pediatric group. Pediatric Radiology. https://doi.org/10.1007/s00247-026-06789-7

Image Credits: AI Generated

DOI: 10.1007/s00247-026-06789-7

Keywords: brain age, pediatric neuroimaging, MRI, volBrain, BrainStructureAges, algorithmic bias, machine learning, brain age delta, brain maturation, pediatric radiology, biomarker validation, cerebellum