Why This Article Matters
Summary
This prospective longitudinal cohort study aimed to develop a deep learning model for predicting childhood myopia progression using baseline fundus photographs and spherical equivalent refraction (SER) data obtained at the initial visit. The study included 3048 Chinese schoolchildren aged 6-9 years from the Anyang Childhood Eye Study, who were followed annually over 6 years, with a total of 16 211 fundus images analyzed. The authors developed a multiyear prediction network combining retinal image analysis with temporal progression modeling to predict future refractive outcomes from baseline imaging and refractive measurements. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) for risk prediction and mean absolute error (MAE) for quantitative prediction of refractive progression. At baseline, myopia prevalence was approximately 6%, while high myopia prevalence was 0.5%. During follow-up, 56% of initially nonmyopic children developed myopia, whereas 5% developed high myopia. The model showed high discriminative ability, achieving an AUC of 0.941 (95% CI, 0.936-0.946) for myopia prediction and 0.985 (95% CI, 0.982-0.988) for high myopia prediction, with an MAE of 0.322 D/year for predicting future refractive progression. However, predictive performance decreased in children already myopic at baseline and over longer forecasting intervals. Similar performance was observed in the Beijing and Lhasa external validation cohorts. Overall, the study suggests that deep learning analysis of fundus photographs combined with baseline SER evaluation may represent a minimally invasive strategy for early identification of children at higher risk of myopia progression.
Commentary
Limitations
Despite external validation, both validation cohorts remained geographically and ethnically close to the derivation cohort, limiting the assessment of model generalizability. The Beijing cohort was relatively small, and the reported near-perfect discrimination for high myopia prediction raises concerns about potential overfitting or instability of the model. Moreover, the relatively low prevalence of high myopia may also have inflated performance estimates. In addition, predictive performance decreased in children already myopic at baseline, suggesting lower reliability once myopia progression is already established. Another important limitation concerns variable selection: the model intentionally relied on minimal baseline data (SER and fundus photographs) and therefore did not incorporate several clinically relevant predictors of myopia progression, including axial length, parental myopia, outdoor exposure, near-work activity, socioeconomic factors, and treatment status. Considering the multifactorial and polygenic nature of myopia, the exclusion of these variables may limit the biological and clinical comprehensiveness of the predictive model.
Clinical relevance
Fundus photography and baseline SER combined with artificial intelligence may represent a practical and minimally invasive screening approach for identifying children at higher risk of myopia progression.
Comparison with existing evidence
Previous machine-learning models for myopia prediction (Zadnik et al. JAMA Ophthalmol 2015 and Lin et al. PLoS Med 2018), mainly relied on longitudinal clinical and refractive data collected over repeated visits, limiting their applicability in large-scale screening settings. More recently, Foo et al. (NPJ Digit Med 2023) demonstrated that deep learning analysis of fundus photographs could predict future high myopia risk with relatively high accuracy. The present study extends these findings by showing that baseline fundus photographs and SER data alone may provide meaningful prediction of future refractive progression. However, similarly to previous literature (Mutti DO et al. BMC Ophthalmol 2023), predictive performance decreased in children already myopic at baseline, supporting the concept that myopia progression remains highly variable and influenced by multiple factors.
Unanswered questions
Future studies should evaluate whether integrating axial length, choroidal biomarkers, genetic information, and environmental exposure further improves predictive performance.
Key Take-Home Messages
- Deep learning models may predict childhood myopia progression using baseline fundus photographs and SER data, particularly in children before myopia onset. - AI-assisted fundus analysis could support large-scale screening strategies for identification of children at risk of myopia progression.
Conflict of Interest Statement
The author declares no conflict of interest related to this article.
Strengths
Large prospective cohort, long longitudinal follow-up, standardized cycloplegic refraction measurements, and availability of more than 16,000 fundus photographs for model training and validation.