AI Brain Age Clock Trained on Death Risk Predicts Dementia Differently in Men and Women

A brain age clock trained not on birthdays but on the risk of dying may capture what is really going wrong inside the aging brain, according to a large new study published in the Journal of Translational Medicine. Researchers led by Ke Jiang, Lei Lin, Tao Zhang, and colleagues at Sichuan University developed an artificial intelligence framework that reads magnetic resonance imaging scans and estimates how far an individual brain has drifted from healthy aging, then showed that this measure links everyday health traits to future dementia in strikingly different ways in men and in women. The findings offer a potential bridge between systemic health and neurodegeneration, and they suggest that a single number derived from a routine brain scan could help clinicians stratify dementia risk long before symptoms appear.

The research team started from a growing frustration in the brain aging field. Most existing brain age models are trained to predict chronological age from MRI features, and the difference between predicted age and true age, often called the brain age gap, is used as a marker of accelerated aging. But chronological age, the authors argue, may not be the most clinically meaningful target. What matters for patients is not whether a brain looks older than expected for its birth year, but whether it is aging in a way that signals disease, decline, and death. To capture that, the team trained their model on all-cause mortality risk instead of calendar age, using multimodal MRI data from more than 46,000 participants in the UK Biobank.

Technically, the approach combined T1-weighted structural imaging, T2-FLAIR sequences that highlight white matter damage, and diffusion tensor imaging, which probes the microscopic integrity of the brain’s white matter tracts. Rather than using a flexible deep learning architecture, the researchers built sex-specific models within a Cox-LASSO framework, a regularized survival analysis method that selects and weights the neuroimaging features most strongly associated with death from any cause. The least absolute shrinkage and selection operator, or LASSO, effectively prunes the model down to a compact set of imaging-derived phenotypes, prioritizing brain characteristics that carry genuine prognostic information over the thousands of candidate measures modern MRI can produce. Fitting separate models for men and women acknowledged from the outset that brain structure, aging trajectories, and mortality risk differ by sex.

From these mortality-trained models, the team derived a residual-based measure of brain age acceleration, or BAA, which reflects how much a person’s predicted brain health deviates from what would be expected after accounting for chronological age. A positive value means the brain appears further along a mortality-relevant aging pathway than expected. The model was then validated externally in two independent cohorts: the Alzheimer’s Disease Neuroimaging Initiative, a North American study that has become a standard testing ground for dementia biomarkers, and the West China Health and Aging Cohort Study, which extends the findings to an East Asian population. Replication across these datasets, drawn from different continents, scanners, and populations, strengthens the case that the measure captures something real rather than a quirk of one dataset.

The results were consistent and consequential. Across both sexes, each additional year of brain age acceleration was associated with a roughly 10 to 11 percent increase in all-cause mortality risk, with hazard ratios of approximately 1.10 to 1.11. Higher BAA was also linked to a broad range of chronic diseases and neuropsychiatric outcomes, and it was significantly associated with incident dementia and Alzheimer’s disease. In other words, a brain that reads as accelerated in aging on this mortality-calibrated clock is not merely statistically unusual; it belongs to a person who is more likely to die earlier and, among survivors, more likely to develop dementia. The measure performed as a general-purpose indicator of brain health while also discriminating specific neurodegenerative outcomes.

The most intriguing part of the study, however, lies in the sex-stratified analyses. When the researchers examined which health-related phenotypes were associated with accelerated brain aging, men and women told very different stories. In males, BAA was broadly tied to cardiometabolic factors, inflammatory markers, and socioeconomic circumstances, painting a picture in which cardiovascular strain, systemic inflammation, and deprivation collectively etch themselves into brain structure. In females, the associations were far more concentrated: smoking and central adiposity, measured as waist to hip ratio, dominated the picture. This divergence matters because it suggests that the pathways leading poor health to brain decline are not uniform across sexes, and that prevention strategies built on male-centric data may miss key risks in women.

To test whether accelerated brain aging actually lies on the causal path between health traits and dementia, rather than merely correlating with both, the team deployed longitudinal causal mediation analysis. The temporal design is critical: exposures such as blood markers, lifestyle factors, and body measurements were recorded at baseline, brain age acceleration was assessed at the imaging visit, and dementia diagnoses were ascertained only afterward. This ordering supports a mediational interpretation, in which unhealthy phenotypes first push the brain along an accelerated aging trajectory, and that accelerated aging then contributes to eventual dementia. Mediation analysis quantifies how much of the total effect of an exposure on dementia travels through the intermediate brain aging measure.

The mediated pathways differed sharply by sex. In men, leukocyte count, a marker of systemic inflammation, showed the largest mediated proportion at 48.4 percent, meaning nearly half of the association between elevated white cell counts and incident dementia flowed through accelerated brain aging. Smoking, liver enzymes such as alanine aminotransferase, and cardiorespiratory measures including forced expiratory volume and peak expiratory flow showed smaller but interpretable mediated effects, consistent with the idea that metabolic, inflammatory, and pulmonary health each contribute to brain decline through structural brain changes visible on MRI. In women, the significant mediated routes ran through smoking pack-years and waist to hip ratio, echoing the concentrated association pattern and pointing to tobacco exposure and abdominal fat as the dominant modifiable pathways linking systemic health to dementia risk in females.

The authors conclude that a mortality-trained brain age model may better capture clinically relevant brain aging than conventional chronological-age-trained approaches, and that BAA can serve as a sex-specific neuroimaging marker linking systemic health to dementia risk. If validated further, the implications are substantial. A brain MRI is already widely available, and a computed brain age score could, in principle, be added to routine scans to flag individuals whose brains are aging dangerously fast, guiding earlier and more targeted prevention. For men, that might mean aggressive management of cardiometabolic and inflammatory burden; for women, smoking cessation and central adiposity control. The measure could also enrich clinical trials by serving as a surrogate endpoint that responds to interventions years before cognitive symptoms emerge.

There are, of course, important caveats. The study population, though enormous, is drawn largely from the UK Biobank, a cohort known to be healthier than the general population, and observational mediation analysis can support but never prove causation. Residual confounding, imaging visit timing, and the evolving nature of the accepted manuscript all warrant caution. Yet the convergence of evidence across two external validation cohorts, the rigorous temporal ordering of exposures and outcomes, and the sheer scale of the analysis make this one of the most compelling demonstrations to date that the aging brain can be read, quantified, and perhaps protected. As dementia rates climb worldwide with aging populations, a sex-aware, mortality-calibrated brain age clock derived from a standard MRI scan may prove to be a deceptively simple tool with life-changing reach.

Subject of Research: Mortality-trained MRI brain age acceleration as a sex-specific neuroimaging marker linking health phenotypes to incident dementia

Article Title: Mortality-trained MRI brain age acceleration mediates sex-specific associations between health-related phenotypes and incident dementia

Article References: Jiang, K., Lin, L., Zhang, T., Xiao, J., Li, X., Wu, D., Zhu, R., Wang, S., Chen, L., Ye, Y., Ma, T., Zhao, X., Dui, X., Zhao, Q., Chen, X., Zhang, X., Yan, H., Fan, M., Long, L., … Li, J. (2026). Mortality-trained MRI brain age acceleration mediates sex-specific associations between health-related phenotypes and incident dementia. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08912-6

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08912-6

Keywords: brain age acceleration, dementia, MRI, UK Biobank, mortality, Alzheimer's disease, sex differences, mediation analysis, neurodegeneration, biomarkers, machine learning, smoking