Predicting onset of myopic refractive error in children using machine learning on routine pediatric eye examinations only

Myopia is increasingly prevalent among children, making routine eye exams crucial. This study develops machine learning (ML) models to predict future myopia development. These models utilize easily accessible, non-invasive data gathered during standard eye clinic visits, deliberately excluding more complex measurements such as axial length or corneal curvature. We used patient records from our pediatric ophthalmology clinic (2010-2022), including only those with at least two visits and no initial myopia. We created three prediction models: whether a patient will develop myopia at some point based on their first visit, be diagnosed in the subsequent visit, or be diagnosed with myopia within a year. We employed Random Forest and Gradient Boosting Tree algorithms for analysis. The dataset included 7814 visits from 2437 patients (average age 5.7 years, range 4 months to 21 years). Among them, 429 (11%) developed myopia. The models predicted myopia with up to 77% sensitivity and 92% specificity. This study introduces an AI-based method to identify children at higher risk for the onset of myopic refractive error, enabling personalized follow-up and treatment plans. Our approach offers caregivers an advanced screening tool to detect myopia risk using readily available data.