Predicting recurrence and survival in patients with non-metastatic renal-cell carcinoma after nephrectomy: a prospective population-based study with multicenter validation

Accurate prognostication of oncological outcomes is crucial for the optimal management of patients with renal cell carcinoma (RCC) after surgery. Previous prediction models were developed mainly based on retrospective data in the Western populations, and their predicting accuracy remains limited in contemporary, prospective validation. We aimed to develop contemporary RCC prognostic models for recurrence and overall survival (OS) using prospective population-based patient cohorts and compare their performance with existing, mostly utilized ones. In this prospective analysis and external validation study, the development set included 11 128 consecutive patients with non-metastatic RCC treated at a tertiary urology center in China between 2006 and 2022, and the validation set included 853 patients treated at 13 medical centers in the USA between 1996 and 2013. The primary outcome was progression-free survival (PFS), and the secondary outcome was OS. Multivariable Cox regression was used for variable selection and model development. Model performance was assessed by discrimination [Harrell's C-index and time-dependent areas under the curve (AUC)] and calibration (calibration plots). Models were validated internally by bootstrapping and externally by examining their performance in the validation set. The predictive accuracy of the models was compared with validated models commonly used in clinical trial designs and with recently developed models without extensive validation. Of the 11 128 patients included in the development set, 633 PFS and 588 OS events occurred over a median follow-up of 4.3 years [interquartile range (IQR) 1.7-7.8]. Six common clinicopathologic variables (tumor necrosis, size, grade, thrombus, nodal involvement, and perinephric or renal sinus fat invasion) were included in each model. The models demonstrated similar C-indices in the development set (0.790 [95% CI 0.773-0.806] for PFS and 0.793 [95% CI 0.773-0.811] for OS) and in the external validation set (0.773 [0.731-0.816] and 0.723 [0.731-0.816]). A relatively stable predictive ability of the models was observed in the development set (PFS: time-dependent AUC 0.832 at 1 year to 0.760 at 9 years; OS: 0.828 at 1 year to 0.794 at 9 years). The models were well calibrated and their predictions correlated with the observed outcome at 3, 5, and 7 years in both development and validation sets. In comparison to existing prognostic models, the present models showed superior performance, as i