Binocular Non-Cycloplegic Ocular Parameters for Modeling Cycloplegic SER in Children

Purpose: The purpose of this study was to evaluate an interpretable binocular modeling approach for predicting cycloplegic spherical equivalent refraction (SER) in children while accounting for inter-eye correlation. Methods: Clinical data from children aged 4 to 17 years were collected at 2 centers. Participants were stratified into three age groups. The binocular approach constructed inter-eye features, including means, differences, and error-related metrics. A linear mixed-effects model (LMM) with subject-specific random intercepts was applied. Model performance was evaluated using R², mean absolute error (MAE), mean square error (MSE), and root mean square error RMSE. SHapley Additive exPlanations (SHAP) were used to quantify feature contributions. Results: In internal validation, the optimal binocular models achieved R² values of 0.882, 0.957, and 0.969, with MAE values of 0.399 diopter (D), 0.285 D, and 0.232 D in the 4 to 5, 6 to 12, and 13 to 17 years age groups, respectively. In external validation, the best binocular models achieved R² values of 0.834, 0.947, and 0.969, with MAE values of 0.488 D, 0.287 D, and 0.354 D in the 3 age groups. SHAP analysis revealed axial length/corneal curvature radius (AL/CR) and AL predominated in children aged 4 to 5 years; non-cycloplegic SER was the determinant in those aged 6 to 12 years; non-cycloplegic SER and its mean became dominant in adolescents aged 13 to 17 years. Conclusions: The binocular modeling approach shows potential for improving accuracy of refractive status prediction in children, and the integration of interpretability methods further supports its utility as an auxiliary tool in large-scale vision screening programs. Translational Relevance: This study translates ocular biometric modeling into clinical practice by providing an interpretable binocular approach to support risk assessment and preliminary screening in pediatric myopia management.