Artificial Intelligence in Predicting Ocular Hypertension After Descemet Membrane Endothelial Keratoplasty

Abstract

Purpose: Descemet membrane endothelial keratoplasty (DMEK) has emerged as a novel approach in corneal transplantation over the past two decades. This study aims to identify predisposing risk factors for post-DMEK ocular hypertension (OHT) and develop a preoperative predictive model for post-DMEK OHT.

Methods: Patients who underwent DMEK at Gangnam Severance Hospital between 2017 and 2024 were included in the study. Four machine learning models-XGBoost, random forest, CatBoost, and logistic regression-were trained to assess feature importance and develop a predictive classifier. An ensemble of these four models was used as the final predictive model. The ensemble model identified clinically significant patients for prediction or exclusion.

Results: A total of 106 eyes from patients who underwent DMEK were analyzed, with 31 eyes (29.2%) experiencing post-DMEK OHT. The final ensemble model achieved clinically significant classification for 61 eyes (57.5%) in the total patient population. Significant risk factors identified in all four models included angle recess area (ARA), best-corrected visual acuity, donor graft size, angle-to-angle distance, crystalline lens rise, and central corneal thickness. The average accuracy, precision, recall, area under the receiver operating characteristic curve, and area under the precision-recall curve values of the ensemble model obtained by a 5-fold cross-validation were 80.2%, 60.0%, 59.7%, 82.3%, and 68.0%, respectively.

Conclusions: This study identified significant risk factors for post-DMEK OHT and highlighted the importance of ocular topographic measures in risk assessment. The development of a final machine learning model to differentiate between clinically predictable patient groups demonstrates the clinical utility of the proposed model for predicting post-DMEK OHT.