Development and validation of a National Database–derived nomogram for predicting postoperative survival in colorectal cancer.

Abstract

3669 Background: Prognosis after curative-intent surgery for colorectal cancer remains heterogeneous and incompletely explained by TNM staging alone. Existing prognostic tools often exclude important demographic, clinical, and biologic factors. Improved risk stratification is needed to support individualized postoperative prognostication. Methods: Using the National Cancer Database (NCDB), data from 88,696 patients diagnosed between 2018–2019 were used to develop and internally validate five prognostic models. All models were extensions of the MSK colon cancer nomogram and were fit using the full cohort. The full model incorporated all MSK predictors plus additional demographic, tumor, and treatment variables including perineural invasion and microsatellite instability status. Reduced models included MSK predictors and select variable subsets. Cox proportional hazards regression was used to estimate 5-year (5y) overall survival. Calibration was assessed by comparing deciles of predicted vs observed 5y Kaplan–Meier survival probabilities (expected-to-observed (E/O) ratios). Discrimination was evaluated using 5y area under the time-dependent receiver operating characteristic curve (AUC). Bias-corrected calibration curves and AUCs were generated to assess model generalizability to external populations. Model performance was compared directly with the MSK nomogram by comparing 5y AUCs calculated using a random sample of 1,000 NCDB-cohort patients. Category-free net reclassification improvement (NRI) was also calculated for patients with known vital status at 5y. Results: Calibration and discrimination were similar across all models. In the full model, E/O ratios demonstrated excellent calibration, with median E/O = 1.01 (range = 0.82-1.08). Full model discrimination was excellent with 5y AUC = 0.787 (95% CI, 0.782–0.793). Bias-corrected calibration curves and AUCs indicated stable performance supporting generalizability. In our 1,000 patient random sample, the full model 5y AUC was 0.822 (95% CI, 0.779–0.866), compared with 0.799 (95% CI, 0.753–0.844) for the MSK nomogram, with no statistically significant difference (p = 0.13). In our 1,000 patient NCDB sample, 397 patients had known 5y vital status (258 deaths, 139 alive, 603 censored prior to 5y) with the full model showing improved overall, event and nonevent discrimination compared with the MSK nomogram (overall 5y NRI = 0.35, event NRI = 0.12, and nonevent NRI = 0.24). Conclusions: We developed a postoperative survival nomogram for colorectal cancer that demonstrated good calibration and discrimination and meaningfully improved risk classification compared with the MSK model. Integration of additional routinely available variables enables more individualized postoperative prognostication.