Abstract
INTRODUCTION: Cycloplegic refraction is the gold standard for assessing refractive error in children. However, logistical constraints hinder its implementation in large-scale surveys. METHODS: Data obtained from a nationwide ocular health survey conducted in ten provincial-level administrative divisions in China were analyzed (2020-2024). Participants aged 5-18 years underwent standardized non-cycloplegic and cycloplegic autorefraction, axial length (AL), corneal radius (CR), and AL/CR measurements. Random forest and XGBoost models were trained to predict the cycloplegic spherical equivalent (SE) using non-cycloplegic SE, uncorrected visual acuity (UCVA), and biometric parameters. Performance was evaluated using R(2), root mean square error (RMSE), and Bland-Altman analysis. RESULTS: Both models exhibited strong predictive performance. In the test set, random forest achieved R(2)=0.88 and RMSE=0.55 diopter (D), whereas XGBoost achieved R(2)=0.89 and RMSE=0.54 D. Non-cycloplegic SE, AL/CR ratio, AL, and UCVA were consistently the top predictors. The predicted SE exhibited strong agreement with the cycloplegic SE, with minimal residual bias. CONCLUSION: Machine learning models incorporating noncycloplegic SE and ocular biometrics accurately estimate cycloplegic SE in children and adolescents, providing a practical alternative for large-scale refractive-error surveillance when cycloplegia is impractical.