Exploring Machine Learning Algorithms to Revise the Kidney Donor Risk Index

Document Type

Conference Proceeding

Publication Date

1-1-2025

Publication Title

Am J Transplant

Keywords

creatinine, adult, artificial neural network, cause of death, cerebrovascular accident, cohort analysis, complication, conference abstract, controlled study, diabetes mellitus, elastic tissue, female, glomerulosclerosis, graft failure, graft recipient, Hepatitis C virus, human, kidney donor, kidney graft, learning algorithm, machine learning, major clinical study, male, middle aged, patient history of transplantation, peripheral vascular disease, pH, random forest, randomized controlled trial, survival analysis

Abstract

Introduction: The Kidney Donor Risk Index (KDRI), originally based on patients transplanted between 1995-2005, has been demonstrated as a poor predictor of graft failure (C-statistic: 0.62). Recent policy developments have removed race and Hepatitis C virus from the model with recalculation of the variable coefficients, but without re-analysis of the variables included. We sought to develop an updated KDRI in a modern cohort of kidney transplant recipients using both conventional and machine learning algorithms. Methods: Kidney transplant alone recipients transplanted between 2016-2023 were included. To minimize the impact of recipient factors on death-censored graft failure, recipients who were <35 or >65 years, had diabetes, peripheral vascular disease or prior transplant were excluded. The remaining recipients were randomized into development (80%) and testing (20%) cohorts. Regular Cox, Lasso Cox, Elastic Net Cox models, and Random Forest, XgBoost, and Neural Network models were fitted to predict the risk of 1-year death-censored graft failure with their respective C-statistics compared. Shapley plots were generated post-hoc from machine learning models for feature interpretability. Results: In this modern cohort, the original KDRI was a poor predictor of graft failure with a C-statistic of 0.594 but higher than the race-neutral KDRI (C-statistic 0.589). Regular Cox, Lasso Cox, Elastic Net Cox models, and Random Forest all performed equally well, with C-statistic of 0.62. XgBoost and Neural Net survival analysis performed less well (C-statistic 0.55 and 0.59, respectively). Variables included in the new model overlapped with original KDRI (donor age, height, weight, creatinine, diabetes, hypertension) but also included new variables (donor pH and >20% glomerulosclerosis on biopsy), while some models excluded CVA as a cause of death. Conclusion: The discriminatory power

Volume

25

Issue

1

First Page

S96

Find It @ Sladen

Share

COinS