A computer model can extract information from an MRI scan that is associated with later health outcomes. That does not mean it knows an individual’s definitive biological age, or that ordering the scan improves their health. A new study illustrates both the possibilities of prediction and the boundary between a prognostic marker and clinical utility.
What does “brain age” actually mean?
The label sounds intuitive. If we know our chronological age, perhaps a brain scan reveals a second, more meaningful age. In practice, the result depends on the task the model was trained to perform. Some models estimate chronological age from imaging features. Training against a different outcome can produce a different indicator. These are not interchangeable thermometers of one directly observed process.
Jiang and colleagues investigated an MRI-based model trained in relation to mortality, and also examined associations with subsequent dementia. Accurately estimating chronological age from an image and ranking the risk of a later event are different tasks. The label alone does not tell us which task has been evaluated. The study was reported in the Journal of Translational Medicine.
For patients, this distinction matters before interpreting a result as “older” or “younger”. A model-derived score is not a personal deadline, and an age-like presentation does not make it a direct measurement of remaining years of life.
A large study, but not one interchangeable participant count
The full cohort comprised 46,361 people. Model development and internal testing involved 43,425: 34,741 in the training set and 8,684 in the test set. External datasets included 964 participants from ADNI and 146 from WCHAC. These figures describe distinct parts of the work, not a single undifferentiated validation sample. See the methods and population descriptions.
Why not simply advertise “more than 46,000 participants”? Because each stage answers a different question. Training data help determine the model. Internal testing examines performance beyond the examples used for training. Data from another source ask whether findings transfer to a different population and data-collection setting.
This separation is valuable. Standardisation was determined in the training data, according to the existing source review. That is a relevant methodological safeguard, not proof that every possible form of information leakage has been excluded. Likewise, the presence of external datasets does not establish that every claimed capability was validated against the same endpoint in each dataset.
Validation must be attached to a defined task, population and outcome. External replication of an age-related association is not automatically external validation of mortality prediction, and neither establishes that deploying the model reduces mortality.
What does the C-index tell us?
In the internal test of mortality discrimination, the C-index was 0.739 for men, with a 95% confidence interval of 0.685–0.793, and 0.694 for women, with an interval of 0.609–0.780. These are measures of discrimination: how well the model orders risk. They are not percentages of correctly diagnosed patients. See the model results.
A C-index of 0.739 must not become “73.9% certainty” that a particular person will become ill or die. It is not a ready-made individual probability for someone sitting in a consulting room. Interpreting absolute risk requires, among other things, a specified time horizon and evaluation of how predicted probabilities agree with observed event rates.
That agreement is calibration, a different question from ranking. The authors describe age calibration: a linear transformation onto an age scale using training-derived parameters, applied to other datasets. This is not calibration of event probabilities. In the article and supplements reviewed, we did not identify a report comparing predicted and observed event probabilities over a specified horizon. This does not establish that the authors never performed such an analysis.
The confidence intervals also matter. They remind us that performance is estimated rather than a fixed property of a device. Choosing the larger decimal value without considering uncertainty, case mix or endpoint can make a comparison look more decisive than it is.
Added predictive value was assessed
A fair reading should not claim that the authors ignored whether their indicator added information beyond an existing model. Table 2 reports incremental discrimination against a comparator including age, APOE and an earlier brain-age-gap measure. For dementia, the increase in the C-index was 0.0079 in women and 0.0124 in men. These comparisons appear in Table 2.
This is more informative than simply saying that the model “works”. However, a small numerical improvement in discrimination does not by itself show that a clinician makes a better decision. Risk ranking can improve without changing management. Management can change without improving patient outcomes. Those steps require separate evaluation.
These downstream outcome analyses used the full UK Biobank cohort, including participants involved in model development. Table 2 should therefore not be presented as an independent external test. ADNI and WCHAC examined other aspects of the model, not prospective validation of every event risk.
In practice, we would want to know who receives the result, what action follows, and what benefits or harms that pathway produces. Without an explicit pathway, a more detailed prognosis may remain additional information. It should not be presented as treatment or as proof that a screening programme is needed.
Representativeness matters
The authors discuss limitations associated with UK Biobank participation, including healthy-volunteer selection and survival to the time of assessment. The small Chinese external sample limits confidence in broad generalisation. Mean follow-up was 5.40 years. These issues are addressed in the discussion and limitations.
A dataset does not become representative of everyone merely by being large. People referred to a clinic with symptoms may differ from participants in a population research project. Imaging conditions, exclusion criteria and the frequency of disease may also affect performance. Before implementation, “How many participants?” should be followed by “How closely do they resemble our patients?”
Sex-specific results are important, but they do not settle every subgroup question. A service considering use would still need to examine the populations it intends to serve rather than assume that an overall result covers all relevant differences.
What must not be inferred?
This study does not establish that routine MRI scanning followed by calculation of this indicator prevents dementia. Nor does it demonstrate that lowering the model’s score through an intervention extends life. An association between a marker and a later event does not guarantee that changing the marker changes the event.
Similar caution applies to mediation analyses. They examine relationships under specified assumptions. A statistical mediation model should not be treated as direct observation of a biological mechanism or as instructions for treatment. An appealing biomarker name makes this distinction especially important.
Questions before implementation
- Which health outcome does the model predict, and over what time horizon?
- Was the relevant task tested in a population resembling the intended patients?
- Are probabilities appropriately calibrated, rather than merely well ranked?
- Which decision changes when the result becomes available?
- Has that decision pathway improved patient outcomes?
- How will costs, data protection and responsibility for interpretation be addressed?
These are editorial questions for assessing implementation, not a list of benefits demonstrated by this paper. Unanswered implementation questions do not erase the model’s research value. They identify the additional evidence needed before clinical use can be justified.
For a patient-facing report, the explanation should state what the number represents, which comparison supports it and what remains uncertain. A clear explanation is preferable to turning a complex prediction into a dramatic age label that suggests a diagnosis the study did not establish.
The takeaway for patients and clinicians
The model provides interesting prognostic information, and its added predictive value was assessed. It has not established that using it in care improves patients’ health. Brain age is best understood as a method-dependent indicator, not a personal verdict or a number of years that can be subtracted from life. A research finding is not, by itself, an indication for an MRI scan.
Source and transparency
Jiang et al. Mortality-trained MRI brain age acceleration mediates sex-specific associations between health-related phenotypes and incident dementia. Journal of Translational Medicine, Article in Press, 2026. DOI: 10.1186/s12967-026-08912-6. Funding: National Key R&D Program of China, grant 2022YFC3600600. The authors declare no competing interests. The article is licensed under CC BY 4.0, subject to separate third-party rights; the illustration above is original.
Educational content, not individual diagnostic advice.