AI fundus model predicts 1-year high myopia progression in children

A deep learning algorithm that analyzes a single nondilated fundus photograph predicted which children would progress to moderate or high myopia within 12 months, according to research presented at the American Academy of Ophthalmology (AAO) 2026 annual meeting.¹ For progression to high myopia, the model's performance scores ranged from 0.91 to 0.96 across development and external validation cohorts, on a scale where 1.0 represents perfect prediction.¹

The findings speak to a practical dilemma in pediatric eye care: deciding which newly myopic children warrant early myopia control therapy, since available treatments are most effective before substantial axial elongation occurs.¹

Deep learning model design and myopia prediction performance

Investigators Ang Ji, MBBS, and Yin Guo, MD, of Beijing, China, built the model in 2 stages.¹ They first trained the algorithm on 150,213 retinal images to estimate spherical equivalent refraction. They then refined it in a longitudinal cohort of 16,384 children so it could predict change over the following 12 months.¹ The model was externally validated in 2 independent cohorts of 2309 and 2783 children whose images were not used in training.¹

For progression to moderate myopia (−3.00 D or worse), scores ranged from 0.88 to 0.89 across all 3 cohorts. For high myopia (−6.00 D or worse), they ranged from 0.91 to 0.96.¹ The model performed less well at predicting new-onset myopia in children with normal refraction, with scores of 0.68 to 0.73, which the investigators described as moderate.¹ The tool therefore appears better suited to risk-stratifying children who are already myopic than to forecasting onset.

Childhood myopia prevalence and high myopia complications

A meta-analysis of 145 studies projected that 49.8% of the global population will be myopic and 9.8% will have high myopia by 2050, up from 22.9% and 2.7%, respectively, in 2000.² Rates are highest in East Asia but have also risen in the US and Europe.¹ High myopia increases the risk of retinal detachment, glaucoma, cataract, and myopic maculopathy, and that risk rises with the degree of refractive error.¹

Myopia control options for children and timing of treatment

Several interventions can slow progression. A 3-year, double-masked randomized trial enrolled 144 children aged 8 to 12 years. In that trial, MiSight 1-day dual-focus soft contact lenses reduced change in spherical equivalent refraction by 59% and axial elongation by 52% vs single-vision control lenses.³ The FDA has also approved spectacle lenses for children aged 6 to 12 years that correct central vision while defocusing peripheral vision.¹

Low-dose atropine reduces progression by roughly 30% to 60%, depending on concentration. However, it remains off-label for myopia control in the US, and the evidence supporting it is stronger in East Asian children than in Western populations.¹

All of these therapies carry costs and work best when started early. As a result, clinicians must choose between treating every myopic child and risking delayed treatment for those who progress fastest.¹ Implementation may be relatively simple, because nonmydriatic fundus cameras are already common in pediatric clinics and school screening programs.¹

Related AI research on fundus-based myopia prediction

The results add to other recent work from China. In the Anyang Childhood Eye Study of 3048 children aged 6 to 9 years, a deep learning model using fundus images and baseline refraction achieved an AUC of 0.985 for predicting high myopia risk. Its performance held in external cohorts from Beijing and Lhasa.⁴

References

1. American Academy of Ophthalmology. A single photo of the eye may reveal which children are headed for severe nearsightedness. News release. Newswise. October 6, 2026. Accessed October 12, 2026. https://www.newswise.com/articles/a-single-photo-of-the-eye-may-reveal-which-children-are-headed-for-severe-nearsightedness

2. Holden BA, Fricke TR, Wilson DA, et al. Global prevalence of myopia and high myopia and temporal trends from 2000 through 2050. Ophthalmology. 2016;123(5):1036-1042. doi:10.1016/j.ophtha.2016.01.006

3. Chamberlain P, Peixoto-de-Matos SC, Logan NS, Ngo C, Jones D, Young G. A 3-year randomized clinical trial of MiSight lenses for myopia control. Optom Vis Sci. 2019;96(8):556-567. doi:10.1097/OPX.0000000000001410

4. Kang MT, Hu Y, Wang N, et al. Deep learning prediction of childhood myopia progression using fundus image and refraction data. JAMA Netw Open. 2026;9(1):e2553543. doi:10.1001/jamanetworkopen.2025.53543