Enhancing Expedited Kidney Allocation Through Machine Learning

Document Type

Conference Proceeding

Publication Date

1-15-2026

Publication Title

Am J Transplant

Abstract

Background: Expedited kidney allocation often suffers from inefficiencies due to the lack of predictive tools that align donor profiles with transplant centers most likely to accept. We developed a machine learning clustering approach that leveraged k-means algorithms to improve allocation efficiency and utilization by stratifying donor kidney profiles and matching them to centers with high acceptance likelihoods. Methods: Retrospective analysis of data from the Scientific Registry of Transplant Recipients (SRTR) on deceased donor kidney (DDK) offers between January 1 and June 30, 2024. The dataset was refined to include only accepted, adult, single-kidney transplants in the continental U.S. We applied an unsupervised k-means clustering algorithm optimized for numeric variables. Feature selection and cluster optimization were performed via iterative testing of 60 models, using average silhouette scores to determine fit. The final clusters were overlaid onto match-run data to calculate center-specific acceptance behavior, producing ranked lists of the top 20 centers most likely to accept kidneys from each cluster. A Power BI allocation tool was built for real-time operational use by allocation specialists at Gift of Life Michigan (GOLM). Results: The final model (Model 20; average silhouette = 0.43) incorporated three key donor variables: Kidney Donor Profile Index (KDPI), terminal creatinine (Cr), and donor age. A total of 9973 DDKs were classified into three distinct clusters representing organ quality profiles (Table 1). Cluster 2 (N = 4,916) comprised kidneys with favorable characteristics, Cluster 3 (N = 4,325) comprised kidneys with less favorable characteristics, and Cluster 1 (N = 732) reflected intermediate quality. Acceptance metrics were developed using a formula combining raw acceptance counts and relative rates, adjusted for each center’s overall volume. Across all clusters, 41 unique transplant centers appeared in the top 20 rankings, with five centers present across all clusters (Figure 1). The allocation tool provided dynamic accepting center recommendations for new kidneys — following input of UNOS donor ID by the allocation specialist, the tool queried GOLM’s internal database for real-time donor input, classified the kidney into its nearest cluster, and provided a ranked list of the top 20 centers most likely to accept. Conclusions: This study demonstrates the utility of machine learning and k-means clustering for stratifying donor kidney quality and predicting center-specific acceptance behavior. By integrating donor heterogeneity with center-level preferences, this approach provides a scalable, adaptive, and real-time solution to improve expedited allocation and enhance kidney utilization.

Volume

26

Issue

1

Find It @ Sladen

Share

COinS