Machine learning prediction of overall survival in patients with cT1b renal cell carcinoma after surgical resection using the SEER database

Accurate survival prediction is essential for guiding follow-up strategies in patients with cT1b renal cell carcinoma (RCC). Traditional AJCC TNM staging systems provide limited prognostic accuracy. Data from the SEER database were used, which included 22,426 patients with cT1b RCC who underwent surgical resection. The data were randomized into training and validation sets in a 7:3 ratioen, suring comparability using standardized mean differences (SMD < 0.1). A random survival forest (RSF) model was developed and compared with support vector machine (SVM) and extreme gradient boosting accelerated failure time (XGB-AFT) models. Model performance was assessed using AUC, sensitivity, specificity, and calibration, with 1000 bootstrap resamples. Shapley additive explanation (SHAP) values were calculated to explore variable importance and enhance interpretability. The RSF model achieved the highest discrimination for predicting 5- and 10-year overall survival (AUC: 0.746 and 0.742), outperforming AJCC TNM (AUC: 0.663 and 0.627), SVM, and XGB-AFT. The model demonstrated good calibration and clinical net benefit. SHAP analysis identified age, tumor size, grade, and marital status as the top contributors to survival prediction. The RSF model significantly improves survival prediction over conventional staging systems and other machine learning methods, with enhanced interpretability through SHAP analysis. While the lack of external validation and the use of overall survival (including non-cancer deaths) are limitations, the model shows strong potential for clinical implementation and may facilitate individualized follow-up planning. Future studies should validate the model prospectively and explore integration into clinical decision support systems.