A photograph of the back of the eye, taken in seconds during a routine ophthalmology visit, contains far more information than most people realize. A study published in Nature Communications by researchers from the University of Lausanne, Erasmus University Medical Center, and several other European institutions demonstrates that a deep learning model trained on retinal images can predict a person’s chronological age with a mean absolute error of 2.85 years. More importantly, the gap between predicted age and actual age turns out to be a meaningful biological signal, one linked to cardiovascular disease, dementia, cancer, inflammation, cognitive performance, and all-cause mortality.
A Foundation Model Trained on 71,343 Eyes
To build this system, the research team fine-tuned RETFound, a foundation model designed for retinal image analysis, using color fundus photographs from 71,343 participants in the UK Biobank. The result is a model that does not simply estimate age. It produces what the researchers call a “retinal age gap”: the difference between the age the model predicts from the image and the person’s actual chronological age.
This gap is the real finding. A positive retinal age gap means the eye looks older than the person’s birth year would suggest. And that discrepancy, it turns out, correlates with a range of serious health conditions. Cardiometabolic traits, systemic inflammation, cognitive decline, incident cardiovascular disease, dementia, and cancer all show associations with this single number derived from a non-invasive eye scan. The retina, which is essentially an extension of the brain’s neural tissue, appears to carry visible traces of systemic aging processes that extend well beyond the eye itself.
Sex-Specific Signatures Hidden in the Same Image
Here is what most coverage of AI medical imaging tends to miss: the same model, applied to the same type of image, can reveal fundamentally different biological stories depending on the patient’s sex. The study’s sex-stratified analysis found that while the model performed consistently across males and females in terms of prediction accuracy, the underlying biological associations diverged sharply.
In males, a higher retinal age gap was more strongly linked to metabolic syndrome. In females, both the model’s attention patterns and the genetic signals pointed toward a greater involvement of the retinal vasculature, the network of blood vessels visible in the fundus image. This is not a minor technical footnote. It suggests that the same aging process leaves different anatomical and molecular fingerprints depending on sex, and that a model capable of detecting both is capturing something genuinely biological rather than a statistical artifact.
The menopausal transition adds another layer. Postmenopausal females in the study showed higher retinal age gap values than premenopausal females, and their clinical associations more closely resembled those observed in males. This pattern implies that hormonal changes associated with menopause may accelerate or redirect retinal aging in ways that shift the biological profile toward a different risk landscape.
Genome-wide analyses reinforced these findings. Genes identified through the retinal age gap were associated with longevity, metabolism, neurodegeneration, and age-related eye diseases, a convergence of biological pathways that positions the retina as a genuine window into systemic aging rather than a proxy for eye health alone.
Why This Matters Beyond the Clinic
Aging research has long struggled with a fundamental measurement problem. Chronological age is easy to record but a poor predictor of biological condition. Two people born in the same year can have dramatically different health trajectories, and conventional risk factors only partially explain why. What this study contributes is a candidate biomarker that is non-invasive, scalable, and grounded in a biological structure that is already routinely imaged in clinical settings.
Retinal fundus photography requires no blood draw, no radiation, and no invasive procedure. If the retinal age gap proves robust across broader and more diverse populations, it could become a practical tool for stratifying health risk in ways that go beyond what standard clinical assessments currently capture. The research team notes that the associations they found hold even after accounting for chronological age itself, which means the gap is not simply restating what a birth certificate already tells you.
For AI in medicine more broadly, this study illustrates a pattern worth understanding. Foundation models, large pretrained systems adapted to specific tasks through fine-tuning, can extract biological signal from existing clinical data in ways that were not anticipated when that data was originally collected. The UK Biobank was not built to study retinal aging as a systemic biomarker. Yet a model trained on its images surfaces genetic and clinical associations spanning cardiology, neurology, oncology, and endocrinology.
The human expertise required to interpret and validate those associations remains essential. The model identifies patterns; clinicians and researchers determine what those patterns mean and whether they are actionable.
In Short
A deep learning model trained on retinal photographs from 71,343 UK Biobank participants can predict chronological age within 2.85 years. The gap between predicted and actual age correlates with cardiovascular disease, dementia, cancer, inflammation, and mortality. Males and females show the same prediction accuracy but different biological signatures, with the menopausal transition visibly affecting retinal aging patterns in females. The retina, already a routine imaging target in clinical practice, may carry more systemic information about how a person is aging than previously understood.
Based on reporting from Nature: Machine Learning.