Deep Learning Prediction of Childhood Myopia Progression Using Fundus Image and Refraction Data.

Importance: Childhood myopia is a global health concern with escalating prevalence and can lead to severe irreversible visual impairment. Early prediction of myopia progression is crucial for timely intervention to prevent high myopia and associated complications. Objective: To develop and validate a quantitative method, based on a deep learning method and using only fundus images and baseline refraction data, to predict both myopia progression trajectory and high myopia risk in schoolchildren. Design, Setting, and Participants: This longitudinal school-based cohort study (Anyang Childhood Eye Study) was conducted from February 2012 to May 2018, with annual follow-up examinations from February 2013 to May 2018. Grade 1 students, aged 6 to 9 years, were recruited from 11 randomly selected urban primary schools in Anyang, Henan Province, China. Children who received myopia control treatments, had amblyopia, or underwent strabismus surgery were excluded. Two independent external validation cohorts were used: the Lhasa cohort (predominantly consisting of Tibetan children) and the Beijing cohort (predominantly consisting of Han Chinese children). Data analysis was performed from July 2024 to July 2025. Main Outcomes and Measures: Performance of a novel deep learning model, created from combining convolutional neural network (34-layer residual network) and recurrent neural network (long short-term memory network). Area under the curve (AUC) was used to assess myopia and high myopia risk prediction, and mean absolute error (MAE) was used to assess spherical equivalent refraction (SER) prediction. Myopia was defined as SER of -0.5 D or less, and high myopia was defined as -6.0 D or less, using cycloplegic autorefraction. Results: Among the 3048 children (mean [SD] age, 7.1 [0.4] years; 1716 females [56.3%]) included, the baseline prevalence rate of myopia was 5.71% (174) and high myopia was 0.5% (15). The deep learning model achieved AUC scores of 0.941 (95% CI, 0.936-0.946) for myopia risk prediction and 0.985 (95% CI, 0.982-0.988) for high myopia risk prediction, with an overall MAE of 0.322 D per year for SER prediction. External validation in the Beijing cohort (n = 130; mean [SD] age, 9.9 [3.7] years; 82 males [63.1%]; 128 Han Chinese [98.5%], 2 Man Chu [1.5%]) and the Lhasa cohort (n = 1039; mean [SD] age, 6.8 [0.5] years; 536 males [51.6%]; 29 Han Chinese [2.8%], 1007 Tibetan [96.9%], 3 [0.2%] from other ethnic minority groups) demonstrated maintained cross-ethnicity performance by the model, with an MAE of 0.355 D and 0.261 D per year, respectively. Conclusions and Relevance: In this cohort study, a deep learning model, using minimal baseline data, provided highly accurate prediction of myopia and high myopia risk. This deep learning approach suggests the utility of large-scale screening and early-intervention efforts in resource-limited settings.