Abstract
PURPOSE: The study aimed to develop a predictive model for refraction after cycloplegia by leveraging non-cycloplegia ocular parameters and focusing on lens-related features. METHODS: A total of 153 children 4 to 15 years old were enrolled in this study. This study randomized gender distribution. Sex, age, intraocular pressure (IOP), refraction before and after cycloplegia, and optical biometry (OB) parameters were collected. Four prediction models for spherical refraction were developed: a control group without lens-related features and three experimental groups incorporating lens-related features. Features such as lens diopter, anterior surface curvature radius, and lens thickness played significant roles. The models were evaluated using statistical measures: mean square error (MSE), Root mean square error (RSME), Mean absolute error (MAE) and r-square (r2). Least absolute shrinkage and selection operator (LASSO) regression and the L1 regularization term were used for feature screening and machine learning for extreme gradient enhancement. The extreme gradient boosting (XGBoost) method was used to develop the model. RESULTS: The predictive model incorporating lens-related features demonstrated superior performance in estimating refraction after cycloplegia compared to the model without such features. Among the models with lens-related features, the IOL of contact lens algorithm (IOLcl) group exhibited the highest efficacy, boasting an r2 of 0.964, MSE of 0.241, RMSE of 0.472, and MAE of 0.307. CONCLUSIONS: The study provided valuable insights into developing a robust predictive model for refraction after cycloplegia, emphasizing the importance of lens-related features and the morphological changes in the crystalline lens during accommodation. TRANSLATIONAL RELEVANCE: This predictive model has potential advantages in avoiding complications associated with cycloplegia and can be widely applied for clinic vision screening in optometry.