Deep learning model accurately predicts childhood myopia risk

A deep learning model using minimal baseline data can accurately predict an individual’s risks of myopia and high myopia, according to a recent study published in JAMA Network Open.¹

More than 1 billion individuals worldwide are expected to be impacted with high myopia by 2050, significantly increasing the risks of severe complications such as myopic maculopathy and permanent vision loss.² Early intervention has been highlighted as a vital strategy for managing this condition.¹

“While fundus images have been successfully used to diagnose severe retinal complications, current methods have not accurately identified children at high risk within 3 years of myopia progression or quantitatively predicted myopia progression,” wrote investigators.

Predicting myopia progression

The longitudinal cohort study was conducted to develop and validate a quantitative method for predicting myopia progression trajectory and high myopia risk in schoolchildren. Participants included grade 1 students aged 6 to 9 years and attending primary school in Anyang, China.

Exclusion criteria included past or current myopia treatment, amblyopia, and strabismus surgery. Fundus images were obtained from annual follow-up examinations at the same school as recruitment.

Spherical equivalent refraction (SER), defined as spherical power plus half cylinder from cycloplegic refraction, was included as a variable. Investigators defined myopia as SER of −0.5 diopter (D) or less, while high myopia was defined as SER of -0.6 D or less.

An increase in myopic SER greater than 0.75 D per year indicated rapid myopia progression. Slow myopia progression was defined as SER progression of less than 0.50 D per year.

Model development and myopia incidence

A nondilated fundus camera (Canon CR-2; Canon) was used to capture fundus images. Investigators used these images in a multiyear myopia prediction network developed using a model integration strategy to determine accuracy.

There were 2 models assessed, the first of which belonged to the encoder component while the second belonged to the decoding component. External validation was also performed using 2 independent cohorts. The baseline analysis included 3048 patients aged a mean 7.1 years, with 56.3% being female and 43.7% being male.

Myopia was reported in 5.7% of patients at baseline and high myopia in 0.5%. Of nonmyopic children, 56.3% developed myopia across the 5-year follow-up period. Additionally, high myopia was developed in 5% of children without the condition at baseline. The deep learning model included 16,211 total fundus images.

Model performance and predictive accuracy

An area under the receiving operating characteristic curve (AUC) of 0.941 was reported for the myopia risk prediction of the deep learning model and 0.985 for high myopia risk prediction. This indicated prediction accuracies of 0.870 and 0.979, respectively. The overall mean absolute error (MAE) for SER prediction was 0.322 D per year.

For 3-year myopia risk, the AUC and MAE were 0.888 and 0.235 D per year, respectively. When using only 1 year of data to predict myopic SER, the MAEs were 0.369 D, 0.554 D, 0.704 D, 0.906 D, and 1.098 D per year for 1-year, 2-year, 3-year, 4-year, and 5-year predictions, respectively.

Significant disparities in model performances were identified based on baseline myopic status, with an MAE of 0.44 D per year for children without baseline myopia vs 0.58 D per year for those with baseline myopia. Minimal differences were reported between male and female children.

This data validated the deep learning model for predicting childhood myopia progression using fundus images. However, investigators highlighted the limitation of predicting future progression rates in children with myopia at baseline.

“Overall, the model may be potentially used for large-scale screening and early- intervention efforts in resource-limited settings,” wrote investigators.

References

1. Kang M, 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
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