Abstract

Introduction: Renal transplantation is a critical treatment that can save the lives of individuals who are suffering from end-stage renal disease (ESRD), but graft failure remains a significant concern. Accurate prediction of graft survival after renal transplantation is crucial as it enables clinicians to identify patients at higher risk of graft failure. This study aimed to develop clinical prognostic models for predicting graft survival after renal transplantation and compare the performance of various statistical and machine learning models.
Methodology: The study utilized data from a retrospective cohort of renal transplant recipients at the Ethiopian National Kidney Transplantation Center from September 2015 to February 2022. Various statistical and machine learning models were evaluated based on their discrimination, calibration, and interpretability. The comparison of models included standard Cox, Lasso-Cox, Ridge-Cox, Elastic net-Cox, Random Survival Forest, and Stochastic Gradient Boosting. The prognostic predictors of graft survival were selected based on the significance and relative importance of variables in different models.
Results: The study analyzed a total of 278 completed cases and observed the event of graft failure in 21 patients. The median graft survival time was 33 months, and the mean hazard of graft failure was 0.0755. The results revealed that the 1-year, 3-year, and 5-year graft survival rates are 0.936, 0.924, and 0.914 respectively. The study found that the Random Survival Forest and Stochastic Gradient Boosting models demonstrated the best calibration and discrimination performance shown by an equal AUC of 0.97 and the overlapped calibration plots. On the other hand, the Cox proportional hazards model has the highest interpretability and established superior accuracy in estimating survival probabilities, as evidenced by its lowest Brier score of 0.000071. The current study indicates that an episode of chronic rejection, recipient residence, an episode of acute rejection, post-transplant urological complications, post-transplant nonadherence, blood urea nitrogen level, and number of post-transplant admissions were consistently identified as the top significant prognostic predictors of renal graft survival.
Conclusions: The Random Survival Forest and Stochastic Gradient Boosting models demonstrated superior calibration and discrimination performance, while the Cox proportional hazards model offered accurate estimation of survival probabilities and interpretability. Clinicians should consider the trade-off between performance and interpretability when choosing a model. Incorporating these findings into clinical practice can improve risk stratification, enable early interventions, and inform personalized management strategies for kidney transplant recipients.
Keywords: Renal Transplantation, Graft Survival, Prognostic Models, Statistical Models, Machine Learning Models.