Predicting Success in Descemet Membrane Endothelial Keratoplasty Surgery Using Machine Learning

Abstract

Purpose: This study aimed to predict early graft failure (GF) in patients who underwent Descemet membrane endothelial keratoplasty based on donor characteristics. Methods: Several machine learning methods were trained to predict GF automatically. To predict GF, the following variables were obtained: donor age, sex, systemic diseases, medications, duration of stay in the intensive care unit, death-to-preservation time (DPT), endothelial cell density of the cornea, tightness of Descemet membrane roll during surgery, anterior chamber tamponade, tamponade used for rebubbling, and preoperative best corrected visual acuity. Five classification methods were experimented with the study data set: random forest, support vector machine, k-nearest neighbor, RUSBoosted tree, and neural networks. In holdout validation, 75% of the data were used in training and the remaining 25% used in testing. The predictive accuracy, sensitivity, specificity, f-score, and area under the receiver operating characteristic curve of the methods were evaluated. Results: The highest classification accuracy achieved during the experiments was 96%. The precision, recall, and f1-score values were 0.95, 0.81, and 0.90, respectively. Feature importance was also computed using analysis of variance. The model revealed that GF risk was related to DPT and the intensive care unit duration (P < 0.05). No significant relationship was found between donor age, endothelial cell density, systemic diseases and medications, graft roll, tamponades, and GF risk. Conclusions: This study shows a strong relationship between increased intensive care duration, DPT, and GF. Experimental results demonstrate that machine learning methods may effectively predict GF automatically.