Analysis of prognostic factors for survival after curative resections of clear-cell renal cell carcinoma based on Bayesian networks

Background Accurate risk stratification following curative resection of non-metastatic clear cell renal cell carcinoma (nm-ccRCC) remains clinically challenging. This study aims to develop and validate Bayesian network (BN) models to predict survival in patients with nm-ccRCC. Methods A dual-center retrospective cohort was randomly stratified into a training set (N = 326) and an internal validation set (N = 140). External validation was performed using an independent external cohort (N = 170). BN models were constructed based on clinicopathologic characteristics that demonstrated statistical significance in univariate analyses of the training cohort and were assessed using receiver operating characteristic (ROC) curves, confusion matrices, calibration and decision curve analysis. Results The Bayesian analyses identified T stages as the principal determinant of 5-year overall survival (OS), while ISUP grades emerged as the dominant predictor for 3-year disease-free survival (DFS). The 5-year OS and 3-year DFS models achieved areas under the ROC curve (AUCs) of 0.878 and 0.867 in the internal validation cohort and 0.838 and 0.846 in the external validation cohort. Calibration slopes were close to unity with small intercepts across cohorts, and decision curve analysis showed net benefit exceeding that of treating all or no patients across a wide range of thresholds. Conclusions This study establishes BNs for the first time as a clinically actionable tool for postoperative risk prediction in nm-ccRCC, with promising discriminative and calibrated performance supporting their potential use in patient counseling, risk-stratified surveillance, and the selection of candidates for adjuvant therapy or clinical trials. Future prospective, multicenter studies are warranted to extend these predictions to a 5-year DFS endpoint and to refine the decision threshold for clinical translation.