Evaluation of Sociomedical Factors on Corneal Donor Recovery Using Machine Learning.

Abstract

PURPOSE To evaluate co-morbid sociomedical conditions affecting corneal donor endothelial cell density and transplant suitability. METHOD(S) Corneal donor transplant information was collected from the CorneaGen eye bank between June 1, 2012 and June 30, 2016. A natural language processing algorithm was applied to generate co-morbid sociomedical conditions for each donor. Variables of importance were identified using four machine learning models (random forest, Glmnet, Earth, nnet), for the outcomes of transplant suitability and endothelial cell density. SHAP (SHapley Additive exPlanations) values were generated, with beeswarm and box plots to visualize the contribution of each feature to the models. RESULTS With a total of 23,522 unique donors, natural language processing generated 30,573 indices, which were reduced to 41 most common co-morbid sociomedical conditions. For transplant suitability, hypertension ranked the top overall variable of importance in two models. Hypertension, chronic obstructive pulmonary disease, history of smoking, and alcohol use appeared consistently in the top variables of importance. By SHAP feature importance, hypertension (0.042), alcohol use (0.017), ventilation of donor (0.011), and history of smoking (0.010) contributed the most to the transplant suitability model. For endothelial cell density, hypertension was the sociomedical condition of highest importance in three models. SHAP scores were highest among the sociomedical conditions of hypertension (0.037), alcohol use (0.013), myocardial infarction (0.012), and history of smoking (0.011). CONCLUSION In a large cohort of corneal donor eyes, hypertension was identified as the most common contributor to machine learning models examining sociomedical conditions for corneal donor transplant suitability and endothelial cell density.