Assessing Deceased-Donor Kidneys Through Posttransplant Survival Prediction Algorithms

Abstract

Rationale & Objective

The Kidney Donor Risk Index (KDRI) is widely used to rank the quality of deceased-donor kidneys and is integrated into the U.S. kidney allograft allocation system. However, the KDRI has modest predictive accuracy for allograft survival, and recent revisions to the KDRI, which removed donor race and hepatitis C virus status, also revealed model calibration problems. This study aimed to evaluate novel approaches for predicting posttransplant allograft survival.

Study Design

Retrospective cohort study using Organ Procurement and Transplantation Network data from May 1, 2007, through December 31, 2021.

Predictors

(1) Donor demographic and clinical variables (established predictors); (2) longitudinal laboratory data from the donor’s terminal hospitalization, such as serum creatinine (new predictors); and (3) recipient clinical variables (new predictors).

Setting & Participants

75,867 adult kidney recipients at U.S. centers.

Outcomes

The primary outcome was time to all-cause allograft failure over 3 years. A secondary outcome was delayed graft function, defined as dialysis in the first week after the transplant.

Analytical Approach

We implemented and compared machine-learning statistical models versus traditional modeling approaches (ie, proportional hazards for the primary outcome and logistic regression of the secondary outcome) that incorporated various combinations of predictors. The performance metrics used to assess discrimination were the integrated (time-dependent) area under the curve (AUC) for allograft survival and the AUC for delayed graft function. To assess calibration, we calculated Brier scores and visually compared the predicted outcomes with the observed ones. Predictive performance was assessed in a 20% testing data split.

Results

Neither machine-learning models nor the addition of longitudinal laboratory data from the donor hospitalization to traditional models improved discrimination. For the primary outcome, the final model (named the Kidney Allograft Survival Index) used a proportional hazards modeling approach. Adding recipient variables improved model discrimination (integrated AUC, 0.68) and achieved excellent calibration for the overall cohort and subgroups. The final model for delayed allograft function used logistic regression, included recipient variables, and had an AUC of 0.75 with acceptable calibration.

Limitations

No external validation.

Conclusions

Improving the discrimination and calibration of kidney allograft survival prediction models is achievable by including recipient characteristics. These enhanced models have potential to improve the system of kidney allocation.