Development and multicohort validation of an interpretable postoperative model for intravesical recurrence after radical nephroureterectomy in upper tract urothelial carcinoma.

PURPOSE: Intravesical recurrence (IVR) after radical nephroureterectomy (RNU) for upper tract urothelial carcinoma (UTUC) is frequent, but postoperative recurrence risk is heterogeneous. We developed and multicohort-validated an interpretable postoperative model for IVR risk stratification after RNU.

METHODS

We included 813 patients with pathologically confirmed UTUC treated with RNU across four predefined cohorts: retrospective development (n = 400), retrospective external validation (n = 173), internal prospective validation (n = 166), and external prospective validation (n = 74). Missing data were handled using cohort-specific multiple imputation. A Cox model integrating clinicopathological factors and neutrophil-to-lymphocyte ratio (NLR) was developed and locked in the development cohort, then transported unchanged to validation cohorts. Bootstrap confidence intervals, competing-risk analyses, benchmark comparisons, and sensitivity analyses were performed.

RESULTS

The locked Cox + NLR model retained tumor location, ureteroscopic manipulation, hydronephrosis, pathological stage, surgical margin status, lymphovascular/perineural invasion, history of bladder cancer, and NLR. Harrell's C-indices were 0.709, 0.746, 0.868, and 0.811 across the four cohorts. At the primary 12-month horizon, AUCs were 0.739, 0.771, 0.953, and 0.791, respectively. Low-risk and high-risk groups remained separated under the competing-risk framework. More complex survival machine-learning models did not show a consistent transportability advantage. The 24-month estimates and surveillance simulation were considered exploratory because prospective follow-up was limited.

CONCLUSION

This interpretable postoperative Cox + NLR model showed favorable multicohort performance for IVR risk stratification after RNU. It may support postoperative risk assessment, but prospective implementation studies are needed before surveillance schedules are changed.