Evaluation of early donor renal function recovery after living-donor nephrectomy as a predictor of allograft failure after kidney transplantation: a longitudinal cohort and machine learning study

Background Kidney transplant recipients continue to face a significant long-term risk of death-censored allograft failure (DCGF). We aimed to evaluate early donor renal function recovery following living-donor nephrectomy and investigate their relationship with long-term allograft outcomes. Methods We conducted a longitudinal cohort and machine learning study by including consecutive living-donor kidney transplantations between January 1, 2007, and December 31, 2022 from West China Hospital. Early renal function recovery was defined as the donor creatinine relative change (dCRC), calculated as the ratio of the difference between pre-donation creatinine and post-donation creatinine to the pre-donation creatinine. We examined the risk of 10-year DCGF from electronic medical records. We employed Cox regression, XGBoost, and Random Forest models to determine whether dCRC could enhance the predictive value of the established iBOX score. Model performance was assessed using Area Under the Receiver Operating Characteristic (ROC) curve (AUC). Findings Among 3241 transplantations screened, 2575 cases were included (mean recipient age 32.5 years). After a median follow-up of 85.5 months after transplantation, 113 recipients (4.4%) experienced DCGF. The median dCRC were 21.1% (IQR: 7.1%-36.8%). Receiver Operator curves reported stronger associations between dCRC and DCGF compared with predonation creatinine (AUC: 0.689 vs. 0.504, P = 0.0082). Each 10% increase in dCRC corresponded to 36.7% higher DCGF risk (HR 1.367, 95% CI 1.262-1.479, P < 0.0001), with a J-shaped dose-response relationship. This association persisted across pre-donation eGFR and post-donation eGFR (<90 vs. >= 90 mL/min/1.73 m2). Last, incorporating dCRC significantly improved the prediction of 10-year DCGF beyond the iBOX alone by multivariable Cox analysis (AUC: 0.868 [95% CI: 0.816-0.920] vs. 0.765 [95% CI: 0.682-0.848], P = 0.0017), XGBoost (AUC: 0.840 [95% CI: 0.765-0.914] vs. 0.727 [95% CI: 0.632-0.822], P = 0.0029), and Random Forest (AUC: 0.857 [95% CI: 0.792-0.921] vs. 0.701 [95% CI: 0.608-0.794], P < 0.0001) in the testing cohort. Interpretation Early donor renal function recovery represents an alternative predictor of death-censored graft failure after living-donor kidney transplantation. This index may enhance risk prediction when integrated with established models. The predominantly young recipient population may limit generalizability to older transplant candidates. Copyright (c) 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 2025;84: Published https://doi.org/10. 1016/j.eclinm.2025. 103278