School-level prediction and management of myopia in children and adolescents.

backgroundPrevious Artificial Intelligence (AI) models are mainly based on hospital data to predict myopia progression in myopic children. However, school-level AI models for predicting myopia onset in non-myopic children are lacking. There is a need for more precise and comprehensive tools for full myopia management, from the onset of myopia to its progression.

methodsThis study was conducted in 870 centers dispersed across seven cities in China from September 2019 to December 2021, with participants observed for two years. Machine learning models were trained and internally validated on datasets from Shenzhen, and then externally tested from the other six cities. Of 1,123,602 children and adolescents aged 4-18 years old, 1,105,271 individuals were confirmed eligible. After two-year follow-up, 915,991 individuals were included in analysis. Data were analyzed from June 2022 to July 2023. Main outcomes and measures are the occurrence of myopia, and the change of spherical equivalent refraction in 2 years.

resultsOf 915,991 participants, 405,784 (44.3%) were identified as myopic at baseline. In two years, the occurrence of myopia was 45.5% and the overall myopic shift was - 0.97 ± 1.32D. The optimal machine learning models were established for predicting myopia occurrence in non-myopic individuals; and predicting progression of any amount (defined as an annual progression of < - 0.25D) in myopic individuals. The performance in predicting occurrence (AUC 0.962, 95% CI 0.956-0.968) and progression (AUC 0.923, 95% CI 0.912-0.934) were acceptable in the external test set. The interactions between age and other variables were revealed by the algorithms, and hence the final prediction models were established based on age segmentation. In the external test set, the performance of the models for both 4-11-years (occurrence prediction: AUC 0.978, 95% CI 0.972-0.984; progression prediction: AUC 0.957, 95% CI 0.951-0.963) and 11-18-years age groups (occurrence prediction: AUC 0.966, 95% CI 0.960-0.972; progression prediction: AUC 0.928, 95% CI 0.917-0.939) were improved. Finally, user-friendly software was developed, which makes this model accessible for school use.

conclusionsThe algorithm developed in the current study for predicting myopia occurrence and progression showed excellent performance in school settings using real-world data, indicating its potential application in school-level myopia management. Trial registration chictr.org Identifier: ChiCTR2200057391.