XGBoost Outperforms Traditional Transplant Regulatory Models Based on Regression.

OBJECTIVE: To determine whether machine learning (ML) models, specifically Extreme Gradient Boosting (XGBoost), improve the prediction of graft and patient survival compared with Scientific Registry of Transplant Recipients (SRTR) regression models across heart, lung, liver, and kidney transplants. BACKGROUND: Accurate prediction of post-transplant outcomes is essential for organ allocation and regulatory oversight. Current SRTR models rely on regression techniques that may not capture complex donor-recipient interactions. ML offers the potential for improved predictive accuracy. METHODS: A retrospective cohort study analyzed United Network for Organ Sharing data from 1987 to 2023 for heart, lung, liver, and kidney transplants. Cox proportional hazards models (SRTR) were reconstructed and compared with XGBoost models. Outcomes included graft and patient survival at 1 and 3 years. Model performance was assessed using area under the curve and DeLong's test. RESULTS: XGBoost models outperformed SRTR models for 1-year graft survival: heart (area under the curve 0.698 vs 0.576), kidney (0.736 vs 0.649), liver (0.706 vs 0.616), and lung (0.612 vs 0.572). Similar improvements were observed for patient survival and 3-year outcomes, with statistical significance for most organs except lung graft survival at 1 year. CONCLUSIONS AND RELEVANCE: ML models provide superior predictive accuracy compared with traditional regression-based models for transplant outcomes. While current clinical utility remains limited, these findings support further development and integration of ML techniques to enhance regulatory evaluations and optimize patient care.