#6880 prediction of graft survival prior to accepting an offer for living donor kidney transplant: an artificial intelligence approach

Abstract Background and Aims The current available models for evaluation of outcomes of living donor kidney transplant before accepting an offer are poorly developed, reported, validated and have small sample sizes. We aim to use Artificial Intelligence to build a model that can accurately predict death censored graft survival for living donor kidney transplant prior to accepting an offer. Method All living kidney transplant patients who were: registered in the UNOS database between 1/1/2007 and 1/6/2021, maintained on TAC/MMF immunotherapy were included in our analysis. We excluded patients with age<18 years old and ABO incompatible transplant. We divided the data randomly into training and testing dataset with ratio 80:20. We performed recursive feature elimination to select the important ones for prediction. Features were selected based on their Gini impurity scores. We performed Artificial Neural Network analysis (ANN). We evaluated the model using Harrell Concordance-time-dependent score (for discrimination), and Integrated Brier score (for calibration). We also assessed dynamic AUC for model performance. Results 54,110 living donor kidney transplant patients were included in the study. Harrell C-Statistic scores were 0.70 at 5 years post-transplant, 0.68 at 10 years post-transplant and 0.68 at 13 years post-transplant, indicating very high discrimination power. Integrated Brier Score was 0.08, indicating very high calibration score for our model. Dynamic AUC scores were 0.71 at 5 years post-transplant and 0.68 at 10- and 13-years post-transplant, indicating adequate performance for our model. The key players in our model were recipient age (variable importance = 0.26), donor age (variable importance = 0.17), donor ethnicity (variable importance = 0.90), followed by dialysis vintage pre-transplantation. Conclusion The ANN model had high discrimination, calibration, and performance indices for predicting death censored graft survival prior to transplant. It can aid the clinical decision for management of the transplant patients. We are currently developing a user-friendly web application that can be used to apply the ANN model for prediction. Our model can help ranking potential living kidney donors based on graft outcomes. Therefore, our model can help improve current outcomes of kidney paired exchange schemes.