Development and validation of a clinically interpretable risk scoring system for predicting one-year all-cause mortality in esophageal cancer: a population-based study.

Background: The conventional American Joint Committee on Cancer (AJCC) staging system primarily relies on anatomical features and may not capture multidimensional prognostic factors, limiting individualized prediction of one-year mortality in esophageal cancer. We developed and internally validated a Surveillance, Epidemiology, and End Results (SEER)-based, clinically interpretable risk score and quantified its incremental value beyond AJCC staging, accompanied by an online calculator. Methods: We identified patients with esophageal cancer diagnosed between 2004 and 2022 from the SEER database and defined one-year all-cause mortality as the endpoint. The cohort was randomly split into development and validation sets (7:3), stratified by the outcome. Candidate predictors were explored using univariable analyses and selected using least absolute shrinkage and selection operator (LASSO) logistic regression with cross-validation. A multivariable logistic model was fitted, and coefficients were scaled to derive an integer-based score; performance loss due to discretization/rounding was assessed. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), Brier score, calibration intercept/slope, Hosmer-Lemeshow test, standardized mortality ratio, and decision curve analysis. Incremental value beyond AJCC staging was assessed using DeLong's test, net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Results: A total of 20,056 patients were included, with a one-year mortality rate of 45.14%. The final model incorporated staging information, treatment variables (surgery, chemotherapy), and clinicopathologic factors (age, sex, grade, tumor size, histology, tumor site), as well as diagnosis-to-treatment interval, which may partially reflect triage-related factors. The risk score achieved an AUC of 0.778 [95% confidence interval (CI): 0.766-0.790] in the validation cohort, outperforming the AJCC combined stage model (AUC 0.709; P<0.001). Calibration was good (Brier score 0.190; Hosmer-Lemeshow P=0.40), with no evidence of systematic calibration drift across prespecified demographic, clinicopathologic, tumor-site, and surgical-status subgroups. Compared with AJCC staging, the score improved reclassification (NRI 0.258; IDI 0.110) and provided stable net benefit across threshold probabilities of 0.25-0.75. Score-defined strata also showed persistent gradients in three- and five-year cumulative all-cause mortality. Notably, it revealed marked within-stage heterogeneity in AJCC Stage III, distinguishing groups with observed one-year mortality of 16.4% versus 91.4%. An interactive online calculator and a bedside scorecard were developed for clinical use. Conclusions: This SEER-based risk score provides robust discrimination, good calibration, and clinical utility, and may serve as an early-course adjunctive tool for short-term risk refinement beyond AJCC staging.