The retina may offer far more than a window onto ocular health. A new study published in Nature Communications suggests that artificial intelligence can use color fundus photographs to estimate biological aging – and that the resulting “retinal age gap” may reveal sex-specific links with cardiovascular, metabolic, inflammatory, cognitive, and ocular disease.
Researchers fine-tuned RETFound, a Vision Transformer-based foundation model, to predict chronological age from 130,360 fundus images belonging to 71,343 UK Biobank participants. The model achieved a mean absolute error of 2.85 years when predictions from available images were averaged at the participant level, with a correlation of 0.90 between predicted and chronological age.
The investigators then calculated the retinal age gap (RAG): the difference between an individual’s predicted retinal age and their chronological age. In simple terms, a positive RAG indicates that the retina appears “older” than expected. The researchers corrected this measure for the regression-to-the-mean effect commonly encountered in age-prediction models.
Higher RAG was associated with a broad range of adverse health indicators, including smoking exposure, HbA1c, waist-to-hip ratio, BMI, leukocyte count, C-reactive protein, glucose, triglycerides, and systolic blood pressure. It was also associated with cataract, glaucoma, macular degeneration, diabetes, metabolic syndrome, cancer, and hypertension.
Importantly, the signal appeared to extend beyond cross-sectional associations. During approximately 15 years of UK Biobank follow-up, higher RAG was associated with all-cause mortality and incident cardiovascular disease, COPD, stroke, dementia, thrombosis, ischemic heart disease, and cancer after adjustment for established risk factors and previous disease events. RAG ranked among the five largest contributors to models predicting mortality, COPD, stroke, dementia, and cancer.
The study also uncovered marked sex-related patterns. Although age-prediction accuracy was similar between male- and female-specific models, the biological associations differed. Males showed generally stronger relationships between RAG and metabolic risk factors. In females, however, retinal aging appeared to shift across the menopausal transition: postmenopausal women had higher RAG values than premenopausal women, and their cardiometabolic and inflammatory associations became more similar to those observed in men.
Model interpretation offered clues to what the algorithm was “seeing.” Saliency and attention maps highlighted the macula, temporal optic disc, and retinal vasculature. Vascular arches were particularly prominent in older female participants, while latent model features correlated with vessel density, bifurcation count, and vessel diameter.
Genetic analyses added another layer, identifying loci associated with metabolism, longevity, retinal vascular traits, neurodegeneration, and age-related eye disease, while sex-stratified analyses revealed partially distinct genetic signatures.
For ophthalmologists, the findings strengthen the case for retinal imaging as a potential source of systemic biomarkers. But the study authors caution that, while RETFound was pre-trained using “data with some ethnic diversity”, both cohorts in this study were predominantly White European and relatively healthy, and RAG may also reflect a mixture of normal aging, disease-related retinal change, and non-aging image characteristics.
Rather than positioning retinal age as a ready-made clinical risk score, the study points toward a broader possibility: routine fundus photography may contain measurable biological information extending well beyond the eye, pointing towards a “sex-specific phenotype” that is linked to numerous age-related diseases.