Análise de sobrevivência e aprendizado de máquina na previsão do tempo de espera para transplante renal: uma abordagem otimizada

Abstract

The increasing demand for kidney transplants and the limited availability of organs make efficient management of waiting lists a critical challenge for the National Transplant System (SNT) in Brazil. This study aims to develop and compare predictive models to estimate waiting time for kidney transplantation, combining survival analysis and machine learning techniques, such as Cox Regression, Random Survival Forest (RSF) and XGBoost. Using a retrospective cohort of data from the Organ Allocation System of the State of São Paulo (SP-OAS)/Brazil, between 2000 and 2017, the research analyzes demographic, clinical and immunological variables, including age, comorbidities, time on dialysis, frequency of HLA-DR, -B and -A, and panel reactivity antibody (cPRA). The performance of the models was evaluated through robust metrics, such as the C-index, Brier Score and ROC curves, to measure accuracy and discrimination capacity.The expected results include the identification of the most relevant variables that impact the probability of transplantation and the validation of an optimized predictive model. The developed tool is expected to provide personalized estimates of waiting time, assisting physicians and patients in making informed decisions and managing expectations. The resulting predictive calculator can be implemented as a practical application, promoting a more efficient allocation of resources and contributing to the reduction of mortality on the waiting list. In addition, the study highlights the role of machine learning as an innovative and promising approach in predicting health events, with the potential to positively impact the transplant system in Brazil.