Related Experiment Video
Updated: Jul 15, 2026
Comparison of Agreement and Accuracy using Binocular Wavefront Optometer with Autorefractor and Phoropter
Published on: September 16, 2025
Assessing the reliability of non-cycloplegic refraction in children: a machine learning approach based on
Haoqiang Cui1,2, Jianning Huang3, Zhenbao Zhou4
1Department of Ophthalmology and Optometry, Fujian Medical University, Fuzhou, China.
Background:
Traditional cycloplegic refraction is the gold standard for pediatric vision screening but is often limited by low efficiency and poor compliance. This study aimed to develop a machine learning model using non-cycloplegic visual function and refractive parameters to evaluate the reliability of non-cycloplegic refraction in children and adolescents. The model aims to identify individuals at risk of significant refractive error deviation (defined as an absolute difference in spherical equivalent [DSE] > 0.25 D) and to determine the necessity for cycloplegic intervention.
Methods:
A total of 300 children and adolescents (547 eyes) were included. Eighteen demographic, refractive, and binocular vision variables were collected. Feature selection was performed using the univariate logistic regression and least absolute shrinkage and selection operator (LASSO). Five models (decision tree, logistic regression, support vector machine, random forest, and multilayer perceptron) were developed. Performance was evaluated using accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and decision curve analysis (DCA), and interpretability was assessed using SHapley Additive exPlanations (SHAP).
Results:
An absolute DSE greater than 0.25 D was associated with multiple accommodative and refractive parameters. Logistic regression showed the best performance, with an AUC of 0.871 (95% CI: 0.798-0.944), an accuracy of 0.809, a sensitivity of 0.771, and a specificity of 0.827. The most important predictors included the monocular estimation method (MEM), negative relative accommodation (NRA), positive relative accommodation (PRA), cylindrical, and accommodative facility. A nomogram was constructed to estimate the probability of DSE greater than 0.25 D.
Conclusion:
Machine learning models based on routine non-cycloplegic parameters can effectively identify children who require cycloplegic refraction, providing an interpretable and practical decision-support tool to reduce unnecessary cycloplegia and improve clinical efficiency.
