Development and validation of a competing-risk nomogram for predicting cancer-specific mortality in prostate cancer: a SEER population-based study

Abstract Prostate cancer prognosis is complicated by substantial competing risks of non-cancer death, particularly in older patients. Existing SEER-based nomograms for high-risk prostate cancer predominantly use standard Cox regression, which does not directly estimate the absolute probability of cancer-specific mortality (CSM) when non-cancer deaths are common. We therefore developed and validated a competing-risk nomogram based on the Fine-Gray subdistribution hazard model, and explored N stage interactions and conditional survival in a large, population-based cohort. Patients with high-risk or locally advanced prostate cancer (operationalized from EAU criteria) were identified from the SEER database (2010–2021). The derivation cohort (2010–2015, n = 137,073) was split into training (70%) and internal validation (30%) subsets; the 2018–2021 cohort (n = 105,391) served as an independent temporal validation set. Diagnoses from 2016 to 2017 were excluded a priori to create a temporal gap between the derivation and validation eras. A Fine-Gray competing-risk model was fitted and visualized as a nomogram. Model performance was evaluated using time-dependent AUC, calibration (plots and IPCW-based calibration intercept/slope with bootstrap 95% confidence intervals), IPCW prediction error, and decision curve analysis (DCA) with comparison against re-estimated versions of D’Amico risk stratification, TNM staging, and the Song 2020 nomogram. N stage interactions and conditional survival were explored as secondary analyses. A total of 242,464 patients were included. Competing events accounted for 72.9% of all-cause mortality, so absolute-risk estimates in this population depend critically on explicit handling of competing mortality. The Fine-Gray nomogram incorporated eight variables: age, race, marital status, T stage, N stage, Gleason score, PSA, and treatment. The time-dependent AUC was 0.745 (3-year) and 0.758 (5-year) in training, 0.766 and 0.775 in internal validation, and 0.630 and 0.648 in temporal validation; discrimination was nearly identical to a cause-specific Cox model with the same variables. In the temporal validation cohort, the IPCW calibration slope was 0.74 (95% CI 0.70–0.80) and the 5-year IPCW Brier prediction error was 0.041 (95% CI 0.039–0.042). On DCA, the nomogram’s net benefit was comparable with the best reference model in derivation cohorts and highest in the temporal validation cohort across clinically relevant thresholds (approximately 5–30%). The global N stage interaction test was not significant (P = 0.512); individual interaction tests were exploratory (no individual test reached nominal significance; the corrected T stage x N test yielded chi-square = 1.86, df = 3, P = 0.601), and stratified analyses suggested a less steep Gleason score-CSM association in N1 than in N0 patients (Gleason 5 + 5: HR 3.33 vs. 1.94), a hypothesis-generating finding. In this large, population-based study, a competing-risk nomogram provided directly interpretable absolute-risk estimates of CSM; calibration was strongest in the derivation and internal validation cohorts and moderately attenuated in the temporal validation cohort. Its net benefit was comparable with or higher than three reference models, most clearly in temporal validation. Because treatment was included as treatment received, treatment associations are non-causal and the tool must not be used to compare treatments. The exploratory stratified analyses support further study of N stage-informed prognostication but do not establish effect modification. An interactive online tool is available at https://prostate-cancer.shinyapps.io/high-risk_CSS/.