A data-driven approach to predict waiting times in kidney transplantation using machine learning

Predicting time to deceased donor kidney transplantation can inform counseling, allocation, and planning, yet nationwide tools for Brazil are lacking. We aimed to develop and temporally externally validate a machine learning model that predicts time to deceased donor kidney transplant for candidates in Brazil. We conducted a retrospective cohort study of the Brazilian National Transplant System. All candidates listed January 2012 to December 2022 informed model development; temporal external validation used listings from January 1 to May 30, 2023. Candidate predictors comprised demographic, clinical, immunologic, and geographic variables, including age, panel reactive antibody, ABO group, HLA features, dialysis duration, prior transplant, serologies, and center region. The outcome was time from listing to deceased donor kidney transplantation. We used Penalized Cox regression and an oblique random forest tuned by 10-fold cross validation in a 75% training set. Among 118,617 candidates, 35% were transplanted and 55% remained listed. In validation, the model showed strong discrimination (C-index 0.773) and delivered individualized predictions; median waiting time varied widely across regions (8–28 months), and highly sensitized candidates (cPRA ≥ 90%) had substantially lower 60-month transplant probabilities than unsensitized candidates (37% vs. 77%). Important predictors included cPRA, HLA compatibility, age, and geographic region. Pediatric patients and those at high-performing centers had shorter waiting times; patients with cPRA ≥ 90% and blood type O had reduced transplant probabilities. An interactive calculator is available online https://transplantmodels.shinyapps.io/waitlist_prediction/ . An oblique random forest provided accurate, individualized estimates of waiting time and transplant probability and may support patient counselling and future policy evaluation, pending prospective assessment.