Establishing and validating a predictive model for long-term control outcomes following orthokeratology lenses wear: a five-year cohort study.

Purpose: To develop and interpret a clinical prediction model for identifying children at risk of poor 5-year axial length (AL) control following orthokeratology (Ortho-K) lens wear, integrating traditional regression modeling with explainable machine learning. Methods: A total of 504 children with baseline myopia were included. The 5-year AL control outcome was defined as an AL increase of <1.0 mm (effective control, EC) or ≥1.0 mm (ineffective control, IC). Feature selection was performed using least absolute shrinkage and selection operator (LASSO), Boruta, and multivariable logistic regression. Machine learning (ML) model performance was evaluated across multiple algorithms, including logistic regression (LR), random forest (RF), support vector machine (SVM), artificial neural network (ANN), decision tree, light gradient boosting machine (lightGBM), and XGBoost. The best-performing model was visualized as a nomogram, dynamically deployed as a web-based risk calculator, and further interpreted using SHapley Additive exPlanations (SHAP) analysis. Results: Feature selection consistently identified a change in AL over the 3 years (Δ1, Δ2, Δ3), and flat E as the most stable predictors of 5-year AL control. Among all models, logistic regression achieved the best overall performance (F1 = 0.897, AUC = 0.969), while XGBoost showed the highest F1-score among ML methods. Enhancing clinical applicability and further simplifying the included parameters, the results Δ1, Δ3, and flat E still demonstrate strong predictive performance (F1 = 0.714, AUC = 0.949). The nomogram demonstrated good calibration and discrimination, with decision curve analysis confirming its clinical utility. SHAP interpretation revealed that Δ3 and Δ1 had the greatest influence on risk prediction, with a notable inflection point around 0.05 mm, beyond which the predicted risk of poor control increased sharply. Conclusion: A robust and interpretable predictive model was developed to estimate 5-year Ortho-K lens control efficacy using Δ1, Δ3, and flat E. The integrated SHAP analysis provided mechanistic insight and highlighted the clinical threshold (Δ3 = 0.05 mm) as a potential early warning indicator for suboptimal myopia control. The dynamic online nomogram enables individualized risk estimation and supports precision-guided intervention in pediatric myopia management.