Machine Learning–Based Prediction Of Visual Acuity After Corneal Cross-Linking In Keratoconus Using Corneal Tomographic Parameters

Keratoconus is characterized by progressive corneal thinning and irregular astigmatism, resulting in variable visual impairment. Although corneal tomography provides detailed structural information, predicting functional vision from tomographic parameters remains challenging. This study aimed to evaluate whether a machine learning approach using the XGBoost algorithm and Pentacam-derived corneal parameters can predict corrected distance visual acuity (CDVA) after corneal cross-linking operation in keratoconus patients, and to identify the tomographic features most strongly associated with visual outcomes

This retrospective single-center study was conducted at a tertiary ophthalmology center. Keratoconus patients who had undergone corneal cross-linking and corneal tomography were included. Demographic data, corrected distance visual acuity, and Pentacam-derived tomographic parameters were collected from clinical records.

In this study, keratoconus patients who underwent corneal tomography were included. CDVA was recorded and converted to logMAR. Corneal tomographic parameters were obtained using the Pentacam system (Oculus, Wetzlar, Germany). In addition to the automatically generated indices, corneal reflectivity values were measured and calculated from Scheimpflug images and included as input variables. An Extreme Gradient Boosting (XGBoost) regression model was used to predict CDVA from tomographic features. Model performance was evaluated using 5-fold cross-validation and assessed with mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²).

A total of 161 keratoconus eyes that had previously undergone corneal cross-linking were included (mean age 23.68 ± 5.09 years). A linear regression model yielded a MAE of 0.085 logMAR. The XGBoost model using all 60 tomographic parameters achieved an MAE of 0.078 logMAR. When restricted to the 15 most influential features, the XGBoost model demonstrated improved performance, with an MAE of 0.072 logMAR, a RMSE of 0.093 logMAR, and an R² of 0.67. Feature importance analysis identified Kmax, mean keratometry, corneal asphericity, and pachymetric deviation indices among the strongest predictors of visual acuity.

Machine learning models using corneal tomographic parameters can estimate visual acuity in keratoconus with reasonable accuracy. The XGBoost model demonstrated better performance than linear regression, and prediction accuracy improved after feature selection. These findings suggest that tomographic features are closely associated with visual function and may help account for variations in visual acuity among keratoconus patients.