Enhancing short-term survival prediction in liver transplant recipients: comparative analysis of penalized Cox regression, random survival forest and XGBoost

Liver transplantation is the definitive treatment for end-stage liver disease (ESLD), yet accurately predicting post-transplant survival remains a significant challenge. This study investigates the effectiveness of machine learning models—Random Survival Forest (RSF) and Extreme Gradient Boosting (XGBoost)—compared to the traditional penalized Cox regression model for predicting short-term survival in liver transplant recipients. The present study analyzed data from 2,111 patients to predict 30-day and 90-day survival rates. Our results demonstrate that RSF significantly outperforms both XGBoost and the Cox model, achieving Harrell’s C-index scores of 0.80 for 30-day and 90-day survival versus 0.70 for XGBoost and 0.72 and 0.69 for the Cox model, respectively. In addition, RSF exhibited greater accuracy in survival time prediction, with a lower Root Mean Square Error (RMSE) of 19.08 for 30-day and 59.94 for 90-day survival, relative to 23.24 and 68.40 for XGBoost, and 20.96 and 69.64 for the Cox model. Notably, pre-transplant creatinine levels emerged as the most important prognostic factor. These findings suggest that machine learning approaches enhance accuracy and reliability, which may improve early postoperative risk stratification, postoperative monitoring, and individualized postoperative care.ت We recommend further research to validate these results across diverse patient populations and to refine predictive algorithms. Not applicable.