Machine Learning Prediction Of Visual And Keratometric Outcomes Following Corneal Cross-Linking In Keratoconus

Corneal cross-linking (CXL) is the standard treatment for halting keratoconus progression; however, predicting individual postoperative visual and tomographic outcomes remains challenging. This study aimed to develop and evaluate machine learning models to accurately predict postoperative maximum keratometry (Kmax), mean keratometry (Kmean), and corrected distance visual acuity (CDVA) at 6 months following CXL using preoperative clinical and tomographic parameters.

Department of Ophthalmology, Bezmialem Vakif University Hospital, a tertiary academic referral center in Istanbul, Turkey.

This retrospective study included 170 eyes of 127 keratoconus patients treated with CXL. Preoperative clinical and Pentacam tomographic parameters were used to predict postoperative 6-month Kmax, Kmean, and CDVA. Data were split into training (75%) and testing (25%) sets. Extreme Gradient Boosting (XGBoost) and CatBoost machine learning models were developed. Model performance was evaluated using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Feature importance and correlation heatmap analyses were performed to identify key predictive variables and enhance model interpretability.

Machine learning models demonstrated high predictive performance. The CatBoost model achieved R² = 0.936 (RMSE 1.05 D, MAE 0.74 D) for Kmean and R² = 0.926 (RMSE 1.68 D, MAE 1.28 D) for Kmax prediction. The XGBoost model yielded R² = 0.96 (RMSE 0.78 D) for Kmean and R² = 0.94 (RMSE 1.55 D) for Kmax. For CDVA, the optimized XGBoost model achieved R² = 0.81 (RMSE 0.083), compared to R² = 0.78 (RMSE 0.091) using all variables. Feature importance analysis identified preoperative CDVA, True Net Power P-Max, and corneal keratometric parameters among the most influential predictors. Correlation heatmap analysis demonstrated relationships between tomographic parameters and postoperative outcomes.

Machine learning models enabled prediction of postoperative visual and keratometric outcomes following corneal cross-linking using preoperative clinical and tomographic parameters. These findings indicate that preoperative data may be used to estimate postoperative Kmean, Kmax, and CDVA. Such predictive models may support clinical decision-making, patient counseling, and individualized prognosis assessment. Further validation in larger and multicenter cohorts is warranted to confirm generalizability and clinical applicability.