Machine Learning-Based Prediction of Keratoconus Progression After Accelerated Corneal Cross-Linking Using the Thinnest Corneal Thickness-to-Maximum Keratometry Ratio.

PURPOSE: To develop and internally validate an interpretable model of keratoconus progression after accelerated corneal cross-linking (A-CXL), emphasizing the thinnest corneal thickness-to-maximum keratometry ratio (TCT/Kmax). METHODS: This retrospective cohort included 170 eyes treated with A-CXL; progression was an increase in maximum keratometry of at least 1.00 D within 1 year. Twenty-six baseline clinical, treatment, and Pentacam-derived tomographic predictors were evaluated using fully nested stratified cross-validation with 10 outer and five inner folds. Scaling, categorical one-hot encoding, 1-SE LASSO selection, hyperparameter tuning, and fitting were confined to each outer training set. Eight algorithms were compared using pooled out-of-fold (OOF) predictions. Discrimination, calibration, clinical utility, and OOF SHAP explanations were assessed. RESULTS: Of 170 eyes, 42 (24.7%) progressed. TCT/Kmax was selected in all 10 outer folds, whereas Km front and BAD-D were each selected in seven. The fold-specific logistic regression (LR) pipeline had the highest candidate-specific OOF AUC (0.772; 95% CI, 0.692-0.844), with an AUPRC of 0.528 and a Brier score of 0.156. Its calibration intercept was -0.19 (95% CI, -0.51 - 0.20) and slope was 1.15 (95% CI, 0.74-1.65). Confidence intervals overlapped across algorithms. TCT/Kmax had the largest mean absolute OOF SHAP value. A separate post-validation three-factor LR model was fitted after validation for interpretation and a research-only calculator prototype. CONCLUSIONS: Leakage-controlled internal validation identified TCT/Kmax as the most stable predictor and supported LR as a parsimonious candidate pipeline. The fixed three-factor model and accompanying calculator are exploratory research tools and require independent external validation before use in clinical decision-making.