Tohoku University develops AI model for predicting retinal age using fundus photographs

A research group led by Professor Toru Nakazawa and Part-time Lecturer Takahiro Ninomiya of Department of Ophthalmology at Tohoku University Graduate School of Medicine, has developed an AI model that can accurately estimate retinal age, which reflects the age of the person's body, from a single fundus photograph. The model is expected to provide a new approach for assessing the overall aging and health status of the body from a single fundus image. The results were published in Communications Medicine.

The retina is one of the few sites in the body where vascular and nerve status can be directly observed noninvasively. In recent years, oculomics is attracting interest for its capability to predict the overall health status of the body from fundus images. The difference between the age predicted from fundus photographs and the actual age is regarded as a potential indicator of biological aging. However, the prediction accuracy tends to drop when different imaging equipment is used or with different subjects. Development of versatile models that can be easily used in health checkups and clinical research is needed.

By training an AI model on 50,595 quality-assured fundus photographs taken at health checkups, the research group developed an AI model that predicts retinal age from a single fundus photograph and internally validated it on 7,288 additional images. In addition to age prediction, the model was designed to capture retinal aging patterns more robustly. This was achieved by combining multitask learning, in which the model was concurrently trained on HbA1c (an indicator of the average blood sugar level over the past two to three months), and ensemble learning, in which predictions from five AI models were averaged. HbA1c was used as auxiliary information during training, so only a single fundus photograph is needed in the actual retinal age prediction.

The model achieved mean errors of 2.78 years in internal validation, 3.39 years in an independent external cohort and 8.63 years in a foreign cohort (4,992 eyes). The errors were smaller than those achieved with existing methods (9.02 years).

It was also shown that the standard deviation of predictions from five AI models potentially serves as an indicator of certainty of the predictions. In internal validation, higher accuracy was achieved in a group with small variation, with a mean error of 2.46 years.

The validation in an independent external population also showed that better prediction accuracy was achieved with a group with small variation (a mean error of 2.87 years) compared to a group with large variation (a mean error of 3.91 years). Moreover, in an age- and sex-adjusted analysis, the retinal age gap was significantly larger in those with diabetes, cardiac disease or stroke.

The external validation in a foreign cohort was conducted as part of a research collaboration with Professor Pearse A. Keane's laboratory at University College London, United Kingdom, using data from AlzEye2018, a large ophthalmic image database in the UK.

In the future, it is expected that fundus photographs, which are routinely taken at health checkups or eye clinics, will be used as an auxiliary tool to evaluate the overall health status of the body without additional blood sampling.

Meanwhile, the present findings were based on cross-sectional analyses, and further validation will be needed to see how well the model can actually predict future disease onset.

Using the data obtained in the present study, the group plans to further investigate the association between changes in the retinal age gap and the development of systemic diseases.

Journal Information
Publication: Communications Medicine
Title: High-accuracy retinal age prediction via fundus-based multitask learning reveals the effect of systemic disease
DOI: 10.1038/s43856-026-01573-y