A preoperative Artificial Intelligence model to estimate cancer-specific mortality in nonmetastatic kidney cancer patients

Surgical resection is the standard treatment for nonmetastatic renal cell carcinoma, yet survival outcomes vary significantly among patients. Current prognostic models lack precision and cannot be applied preoperatively. Here we show the development and validation of a preoperative, interpretable machine learning model to estimate cancer-specific mortality. Using real-world clinical data from 2536 patients and an independent external validation cohort of 580 patients, we combine random survival forests with white-box models to ensure clinical transparency. Our survival tree model relies on exactly eight preoperative features, including tumor size, lymph node involvement, and performance status. We demonstrate that this model outperforms the established GRANT model, achieving a C-index of 0.88 and a Brier score of 0.02 on the external cohort, with notable accuracy in the first year postsurgery. Finally, we provide this tool as a web-based application to facilitate personalized, preoperative risk stratification. Current prognostic models for renal cell carcinoma (RCC) cannot be applied preoperatively. Here, the authors develop a preoperative, interpretable machine learning-based TRIPOD type III model for RCC mortality prediction based on clinical and demographic data, which is deployed in real-world patient cohorts, outperforming the GRANT model.