AI Reads Your Eyes: Retinal Scans Reveal How Fast You Are Really Aging

A routine photograph of the back of the eye may soon tell you something no birthday can: how fast your body is actually growing old. In a large new study published in GeroScience, researchers in South Korea trained a deep learning model to estimate a person’s age directly from retinal fundus images and then measured the gap between that predicted retinal age and the person’s true chronological age. The resulting measure, called the retinal age gap, turned out to be closely linked to some of the most important determinants of health, including smoking, diabetes, and common eye diseases. The findings add momentum to a rapidly growing field that treats the eye as a window onto the aging body, and they raise the tantalizing prospect that a cheap, noninvasive imaging test already performed in millions of health checkups could one day help quantify biological aging at population scale.

The logic behind the approach rests on a biological quirk of the retina. As an extension of the central nervous system, the retina shares embryological origins and physiological characteristics with the brain, heart, and kidney, and it is the only place in the human body where clinicians can directly observe both living microvasculature and neural tissue without cutting into anything. Age leaves visible fingerprints there: arterioles narrow, vessel density declines, and the branching complexity of the vascular network diminishes in ways that parallel systemic vascular aging. Because fundus photography is fast, painless, inexpensive, and already embedded in routine screening programs, the retina offers something that molecular aging clocks and brain MRI scans cannot: a scalable, low-cost readout of biological age that requires nothing more than a camera and an algorithm.

Earlier studies had already hinted at the promise of this idea. A foundational analysis of UK Biobank participants found that each additional year of retinal age gap was associated with a roughly two percent increase in all-cause mortality risk, and subsequent work linked the measure to stroke, cardiovascular disease, chronic kidney disease, and Parkinson’s disease. But those models were trained almost exclusively on healthy individuals and validated mostly in European populations, leaving open questions about how well they would perform in real-world screening settings where disease is common and populations differ. The new study, led by researchers at Seoul National University Hospital, set out to close those gaps with a dataset drawn from nearly two decades of comprehensive health screenings at the institution’s Gangnam Center.

The technical achievement at the heart of the paper lies in how the model was built. The team started with 44,362 fundus images from 14,951 participants acquired between 2003 and 2019, all captured with a single standardized fundus camera to eliminate equipment-related artifacts. From this pool they constructed a development cohort of 29,530 images from 7,535 participants and two mutually exclusive evaluation cohorts: one with 5,606 participants for lifestyle and systemic disease analysis, and another with 1,810 participants for ocular disease analysis. Rather than training only on images from disease-free eyes, as most prior efforts had done, the researchers built a mixed dataset that included retinas with documented disease, on the reasoning that a model destined for real clinics should learn from real patients.

Two further innovations sharpened the model’s accuracy. First, the team used a multi-task architecture based on a fundus-specific foundation model with a ResNet-50 backbone, training the network to predict both age and sex simultaneously. Sex was chosen as an auxiliary task because well-documented differences in retinal morphology between men and women, such as optic disc size and vascular caliber, can otherwise introduce systematic bias into age predictions. Second, the researchers applied a post-hoc linear calibration to correct for regression to the mean, a notorious artifact of age prediction models in which young people’s ages tend to be overestimated and older people’s underestimated. The combination worked: the final bias-corrected multi-task model achieved a mean absolute error of 2.656 years and a correlation of 0.921 with chronological age in the first evaluation cohort, and 2.529 years and 0.938 in the second, figures comparable to or better than previously published retinal age models, which typically err by 2.79 to 3.55 years.

With a validated retinal age estimator in hand, the team turned to the central question: what does an older-appearing retina actually mean? In the lifestyle analysis, both former and current smokers stood out. After adjusting for age and sex, ex-smokers showed a retinal age gap roughly 0.46 years higher than never-smokers, and current smokers about 0.50 years higher, associations that survived the study’s strict statistical corrections for multiple comparisons. The similar magnitude in both groups suggests that even past tobacco exposure leaves a measurable imprint on the retinal microvasculature, consistent with a large body of evidence linking smoking to microvascular dysfunction and accelerated vascular aging. Interestingly, drinking frequency and household income showed no significant relationship with the retinal age gap.

The systemic disease findings were even more striking. Using a rigorous propensity-score matching technique that paired each participant with diabetes to an otherwise comparable control of the same age and sex, the researchers found that clinical diabetes was associated with a retinal age gap roughly 2.52 years higher than matched controls, meaning the retinas of people with diabetes looked more than two years older than their birthdays implied. The association proved remarkably robust, persisting through every sequential adjustment the researchers tried, including body mass index, smoking, and alcohol use, and it appeared with similar strength in participants both taking and not taking diabetes medication, indicating it was not an artifact of treatment. Pre-diabetes, by contrast, showed no association at all, suggesting the effect emerges only once disease is clinically established. Notably, hypertension and hyperlipidemia were not associated with the retinal age gap in this cross-sectional analysis, a result that differs from some earlier longitudinal findings and may partly reflect the normalizing effects of antihypertensive and statin therapy on retinal vessels.

The ocular disease analysis added another layer of insight. In eyes with age-related macular degeneration, the retinal age gap was about 0.60 years higher than in normal eyes, and in eyes with cataract it was a substantial 1.86 years higher, both associations surviving correction. Glaucoma showed no significant difference. The cataract result comes with an important caveat that the authors themselves acknowledge: cataract was defined as visible opacity on fundus photography, and such opacity degrades image quality, so the older-appearing retina in cataract eyes may partly reflect blurry images rather than accelerated aging of the tissue itself. Still, the finding that a model trained partly on diseased eyes could detect these signals while remaining well calibrated across the age range addresses a concern raised in a recent scoping review about retinal age models trained solely on healthy cohorts, which may generalize poorly to clinical populations.

Perhaps the most sobering part of the study is its honest assessment of what the retinal age gap cannot yet do. The behavioral and socioeconomic associations, on the order of 0.3 to 0.5 years, are small relative to the model’s prediction error of roughly 2.5 to 2.7 years, making them meaningful only at the population level rather than for any individual. The study was retrospective, single-center, and limited to Korean participants undergoing voluntary health screening, a group likely to be more health-conscious than the general public. Lifestyle variables were self-reported, the analysis was cross-sectional and therefore cannot establish cause and effect, and the retinal age gap was never directly compared against established biological aging biomarkers such as DNA methylation clocks. The authors conclude that the measure currently suits population-level characterization far better than individual risk stratification, and that longitudinal validation in independent cohorts is essential before it could inform clinical decisions.

Even with those caveats, the study marks a meaningful step toward a future where aging itself becomes a measurable, monitorable target of medicine. The geroscience hypothesis holds that the fundamental processes of aging drive deterioration across organ systems, and that targeting aging directly could delay many chronic diseases at once. To act on that idea, medicine needs biomarkers of biological age that are cheap, noninvasive, and deployable at scale, and a retinal photograph fits that bill better than almost any alternative. If future longitudinal studies confirm that the retinal age gap tracks true biological aging and responds to interventions, the humble eye exam could evolve into a routine barometer of aging, flagging patients whose eyes look older than they should and guiding early efforts to slow the clock. For now, the message from Seoul is clear: your eyes may know your age better than you do, and scientists are learning to read them.

Subject of Research: Deep learning-based retinal age prediction and its associations with lifestyle, systemic, and ocular health factors

Article Title: Deep learning–derived retinal age gap and its associations with lifestyle, systemic, and ocular health in a health screening cohort

Article References: Jang, B., Oh, R., Lee, T.-H., Yoon, C. K., Choi, H. J., Choi, J., Kim, Y.-G., & Bae, K. (2026). Deep learning–derived retinal age gap and its associations with lifestyle, systemic, and ocular health in a health screening cohort. GeroScience. https://doi.org/10.1007/s11357-026-02538-8

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02538-8

Keywords: retinal age gap, biological aging, deep learning, fundus photography, diabetes, smoking, age-related macular degeneration, cataract, geroscience, biomarkers, health screening, multi-task learning