Post-transplant survival prediction in liver transplantation: implications for organ acceptance

Abstract

Liver transplantation is the standard treatment for end-stage liver disease. However, the imbalance between organ demand and supply leads to difficult decisions regarding organ acceptance and allocation. Predicting long-term survival after transplantation remains a challenge, as traditional scoring systems are limited in their ability to represent the complexity of donor, recipient, and operative factors. This study aims to develop machine learning models designed to predict post-transplant survival to support clinical decisions about whether to accept or decline an organ offer. Using data from the United Network for Organ Sharing (UNOS) registry, this study analyzed 72,379 deceased-donor, adult, first-time liver transplants performed between 2014 and 2024 in the United States. After thorough data cleaning and a feature selection processes including Cox Proportional Hazards (PH) model, LASSO (Least Absolute Shrinkage and Selection Operator) regression, and clinical review, the initial set of 416 variables were reduced to 33 predictive variables regarding the recipient, donor and the operative factors. To address the severe imbalance between the classes (namely, alive and dead), the analysis compared three strategies: models trained with no imbalance correction, weighted models that increased the penalty for misclassifying death cases, and an undersampled bagging approach that created relatively more balanced training subsets. Across these settings, a range of algorithms were evaluated, including Logistic Regression, Regularized Regression, Decision Trees, Random Forests, XGBoost (Extreme Graident Boosting), SVMs (Support Vector Machines), and MARS (Multivariate Adaptive Regression Splines). Survival was modeled as a set of classification tasks for four commonly cited post-transplant periods: 1 year, 1–3 years, 3–5 years, and 5–10 years. Models were tuned using 10-fold cross-validation, and performance was assessed using AUC, sensitivity, and specificity as these metrics are more informative than overall accuracy in imbalanced datasets. Gradient Boosting methods (namely weighted XGBoost) produced the strongest results for all time periods, with AUC values ranging from 0.645 to 0.69 and outperformed the Cox model benchmark. Performance of all classification models improved in longer time horizons, due to both increased event rates and clearer separation between alive or censored patients and deaths. Analysis of variable importance showed that recipient age was the most important factor across all models, followed by serum creatinine. Cold ischemic time, recipient BMI (Body Mass Index), functional status, donor age and height, serum sodium, INR (International Normalized Ratio), and diagnosis categories were also found as significant variables in predicting post-transplant survival. The survival probability estimates generated by these models could also be integrated into clinical decision support systems to influence transplant surgeons' judgment when evaluating organ offers. Instead of than relying solely on experience and simple scoring systems, clinicians could access patient-specific predictions that describe complex interactions between recipient, donor, and operative factors. This thesis complemented the survival analysis with a simulation study to guide organ acceptance decisions. In this simulation model, donor characteristics were modeled using fitted probability distributions for each OPTN (Organ Procurement and Transplantation Network) region, revealing significant differences in organ quality. These donor distributions were then integrated into an acceptance model to estimate waiting times for a better-quality organ to help assess the trade-offs between accepting or rejecting the current offer. The findings in this thesis aim to assist transplant teams in evaluating organ offers and strengthening pre-transplant decisions. On a wider scale, better survival prediction could support more efficient resource use and help shape consistent organ allocation policies.