Retinal Age

Don't seek out an AI retinal-age reading. Get a full dilated eye exam on schedule, and accept an OCT scan (a cross-section of the retina) if it is offered as part of that exam.

What it is

The retina is the only place a doctor can look at blood vessels and nerve tissue without cutting. Deep-learning models trained on colour fundus photographs, the standard photo of the back of the eye, predict a person's age from the image alone. The difference between that prediction and the person's real age is called the retinal age gap. Related models use the same photos to estimate heart and stroke risk or to detect Alzheimer's disease.

None of these tools is validated or endorsed for screening healthy adults, no professional body uses them for risk stratification, and some of the work is still in unreviewed preprints. The wider eye picture is on Vision.

A small mortality signal

The best-known model estimates age to within about 3.55 years on average, against about 8.1 years for simply guessing the average age of the same UK Biobank population, or 6.7 years for guessing the average for each sex.^[1] In 35,913 UK Biobank participants, each year by which retinal age exceeded real age went with a 2% higher death rate from any cause (hazard ratio 1.02), with steeper increases in deaths from causes other than heart disease and cancer at the top of the range.^[2] The 95% confidence interval, the range of values the data leave plausible, has a lower end that touches no effect at all, so on its own this is a borderline result.

That accuracy is about as good as a DNA-methylation clock (3.3–5.2 years), from a 30-second clinic photo instead of a blood draw. It is an accuracy comparison, not a mortality one: no study has run a retinal clock head-to-head against a methylation clock for predicting death.

A second model, RetiAGE, scores the probability that a person is over 65 and uses it directly as a mortality score. People in its top quarter had 67% higher all-cause, 142% higher cardiovascular and 60% higher cancer mortality over 10 years than people in the bottom quarter, off a small base: 636 cardiovascular deaths among 56,301 participants.^[3]

Newer, larger models reach about 2.8 years of error on their own data.^[4] One of them, still an unreviewed preprint, found that a high retinal age gap tracked metabolic syndrome most strongly in men and the retinal blood vessels more strongly in women, though separate male and female models did not improve accuracy.^[5]

Heart, stroke and Alzheimer's

• Stroke. In 45,161 UK Biobank participants followed for a median of 12.5 years, 29 retinal blood-vessel measurements each tracked roughly 10–20% higher stroke risk per standard-deviation change. Adding them to standard risk factors barely improved prediction: the score that measures how well a model separates people who go on to have a stroke from those who don't (a 0-to-1 scale) rose only from 0.739 to 0.752.^[6]
• Coronary calcium. A model that estimates the likelihood of coronary artery calcium from retinal photos (score 0.742 on the same scale) could reclassify borderline- and intermediate-risk adults when added to a standard risk equation. Several of its authors are employees or shareholders of the company that owns it.^[7]
• Alzheimer's disease. A model that detects existing Alzheimer's dementia scored 0.93 on its own data but 0.73–0.91 across five outside test sets, and 0.68–0.86 against amyloid brain scans. Its case-control design flatters apparent accuracy.^[8]

Why it isn't a test for you yet

The models travel badly. One trained on 27,214 disease-free Japanese adults predicted age to within 2.78 years on its own data and 3.39 years in a second Japanese cohort, but was off by 8.63 years in a UK hospital dataset.^[9] RetiAGE fell from 0.97 in Korea to 0.76 in UK Biobank on the 0-to-1 scale above. Any number one of these tools gives you is only as good as the population it was calibrated in.

One person's number is noisy. In 41 healthy adults whose eyes were photographed twice in one sitting, the predicted age moved by about 2.1 years on average; in 26 photographed at two separate visits it moved by about 2.4 years, in one person by just over nine.^[10] The population-level signal survives that noise; a single person's reading largely does not.

What to do

1. Get a full dilated eye exam on the age-based schedule in Vision. It finds cataract, glaucoma, diabetic eye disease and macular degeneration, which you can act on.
2. Accept an OCT scan if it is part of that exam. There is no professional-society basis for asking for OCT-angiography or an AI ageing-clock reading as a screening test.
3. Don't act on a retinal-age number if a clinic or app gives you one: it can move by years between two photos.

Further reading

• Zhu Z et al. Retinal age gap as a predictive biomarker for mortality risk. British Journal of Ophthalmology 2023.^[11]
• Nusinovici S et al. Retinal photograph-based deep learning predicts biological age, and stratifies morbidity and mortality risk. Age and Ageing 2022.^[12]
• Grimbly MJ et al. Estimating biological age from retinal imaging: a scoping review. BMJ Open Ophthalmol 2024.^[13]
• Cheung CY et al. A deep learning model for detection of Alzheimer's disease based on retinal photographs: a retrospective, multicentre case-control study. Lancet Digital Health 2022.^[14]
• Zoellin JRT et al. Evaluating the reproducibility of a deep learning algorithm for the prediction of retinal age. GeroScience 2025.^[15]
• Ninomiya T et al. High-accuracy retinal age prediction via fundus-based multitask learning reveals the effect of systemic disease. Commun Med 2026.^[16]