Machine learning-driven prediction of cycloplegic refractive error in Chinese children.

OBJECTIVE

To develop and validate machine learning (ML) models for predicting cycloplegic spherical equivalent refraction (SER) using non-cycloplegic parameters, addressing challenges in pediatric ophthalmic assessments.

METHODS

A prospective cohort of 2,274 Chinese children (4,548 eyes) aged 3∼16 years was stratified into development (n = 1819) and validation (n = 455) datasets. Six ML models (linear regression, random forest, extreme gradient boosting, multilayer perceptron, support vector machine, and light gradient boosting machine) were trained on demographics, non-cycloplegic refractive error, and ocular biometrics. Model performance was evaluated using R 2 , mean error (ME), mean absolute error (MAE), and clinical accuracy (proportions within ±0.50 D/±1.00 D).

RESULTS

In the validation dataset, ML models predicted cycloplegic SER with high R 2 (0.920∼0.934), low ME (-0.004∼0.015 D) and MAE (0.385∼0.413 D). The multilayer perceptron model achieved the highest accuracy (R 2 = 0.934, MAE = 0.385 D), with 73.08% and 94.29% of predictions within ±0.50 D and ±1.00 D, respectively. Performance was optimal in children aged 7∼10 years (77.17∼79.70% within ±0.50 D) and those with low myopia (-3.00 to -0.50 D; 83.09∼83.56% within ±0.50 D). Non-cycloplegic measurements systematically overestimated myopia (mean difference: -0.39 ± 0.71 D, P < 0.001), particularly in younger children and hyperopic eyes.

CONCLUSION

ML models provide accurate estimates of cycloplegic SER using non-cycloplegic parameters, offering a practical alternative for pediatric refractive assessments when cycloplegia is infeasible.