AI reads one eye photo to predict worsening myopia, study finds

A single picture of the back of the eye, taken without dilating the pupil, is enough for an algorithm to estimate how a patient's nearsightedness will change over the coming year. That is the finding presented in October at the annual meeting of the American Academy of Ophthalmology in New Orleans, where a team from Beijing Tongren Eye Center in China reported its work.

The team, led by Ang Ji and Yin Guo, built its model in two stages. It was first pretrained on 150,213 cross-sectional images, single snapshots of many eyes, and then fine-tuned on a longitudinal cohort of 16,384 children, the same young patients followed over time. That second stage is where the model learned to link what a photo shows with what happens to a child's vision later.

The accuracy figures are the core of the result. The area under the receiver operating characteristic curve, or AUC, measures how well a tool separates those who will progress from those who will not, with 1 meaning perfect separation. For progression to moderate myopia, the score reached 0.89 on internal validation, 0.89 on a first external cohort of 2,309 participants and 0.88 on a second of 2,783. For high myopia, the algorithm did better: 0.95, 0.96 and 0.91 across the same three groups.

Predicting that a child would develop myopia within one year was a weaker result. The AUC there was between 0.68 and 0.73, which the authors describe as moderate. The tool is therefore more reliable at spotting a worsening condition than at anticipating who will first become nearsighted.

For the children whose risk would be flagged early, and for the doctors who would decide what to do with that flag, the promise is a warning a year ahead, drawn from a photograph that is already part of an eye examination. The authors frame the tool as support for early clinical intervention. But the study gives no date for use in clinics, and it does not describe any trial in a clinical setting. Until those steps are reported, the result is a validated prediction on research cohorts, not a tool in daily practice.