Development and validation to predict visual acuity and keratometry two years after corneal crosslinking with...

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Data Science

Background:

• Progressive keratoconus poses a significant risk to vision.
• Corneal crosslinking (CXL) is a standard treatment for progressive keratoconus.
• Predicting treatment outcomes is crucial for patient management.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in predicting visual acuity and keratometry changes.
• To validate ML model performance using a limited dataset for progressive keratoconus patients undergoing CXL.
• To assess the predictive utility of ML 2 years post-corneal crosslinking.

Main Methods:

• A retrospective analysis of 277 eyes from 195 patients with progressive keratoconus treated with CXL.
• Development and evaluation of three ML models using training and testing datasets.
• XGBoost model performance assessed for predicting corrected distance visual acuity (CDVA) and maximum keratometry (Kmax) changes.
• Validation using an independent set of 43 eyes.

Main Results:

• Baseline CDVA and K2/K1 ratio were associated with CDVA changes.
• Baseline Kmax/Kmean ratio was associated with Kmax changes.
• The XGBoost model demonstrated high predictive accuracy (R²=0.9993 for CDVA, R²=0.9888 for Kmax in testing; R²=0.8956 for CDVA, R²=0.8382 for Kmax in validation).

Conclusions:

• Machine learning, particularly XGBoost, significantly enhances prediction accuracy for visual and keratometric outcomes after CXL.
• Incorporating specific patient parameters improves the reliability of ML predictions for progressive keratoconus.
• ML offers a valuable tool for forecasting 2-year post-CXL changes in visual acuity and keratometry.