Development and Internal Validation of a Clinical Cox Model for Predicting Overall Survival in Patients with Lung Cancer: Real-World Evidence from Seven Hospitals in China

Abstract Background In routine practice, complete TNM staging is often unavailable, limiting the prognostic utility of staging alone. We investigated whether a pragmatic clinical-variable–only Cox model based on routinely collected data could provide robust short- to mid-term risk stratification for overall survival (OS) compared with a simple stage-only model using an ordinal three-level system (early/mid/late). Methods We undertook a multicentre retrospective cohort study across seven hospitals in China. Consecutive adults with pathologically confirmed lung cancer diagnosed between 2011 and 2023 were screened; after prespecified exclusions, 865 patients were included and split into training (n=584) and internal validation (n=281) cohorts. Median follow-up was 8 months (range, 1–114). Prespecified predictors were smoking status, white blood cell count (WBC), lymphocyte count (LYM), prothrombin activity (PTA), D-dimer level, receipt of chemotherapy in the initial treatment window, and age. The comparator was a staging-only Cox model (three-level staging). The primary outcome was OS (diagnosis to death; censored at last contact). Performance was evaluated using time-dependent AUCs at 6/12/18 months and the C-index; calibration intercept/slope and plots; decision-curve analysis (DCA); and categorical and continuous net reclassification improvement (NRI). Risk-stratified Kaplan–Meier curves assessed separation. Results Multivariable Cox regression showed independent associations with OS for smoking (HR 1.57, 95% CI 1.26–1.96), WBC (HR 1.015, 95% CI 1.000–1.030), LYM (HR 0.98, 95% CI 0.966–0.993), PTA (HR 0.986, 95% CI 0.981–0.992), D-dimer (HR 1.036, 95% CI 1.019–1.053), chemotherapy (HR 0.434, 95% CI 0.273–0.688), and age (HR 1.038, 95% CI 1.027–1.048). Discrimination at 6/12/18 months was acceptable to good (training AUCs 0.729/0.761/0.761; validation AUCs 0.804/0.789/0.803), with overall good calibration (close alignment with the ideal line in the 0.30–0.80 range). Versus the staging-only model, validation AUCs for the comparator were ~0.512/0.516/0.525, and the clinical model achieved greater net benefit across DCA thresholds ~0.10–0.80; time-dependent discrimination substantially favored the clinical model (ΔC, fit2−fit1 = −0.253; 95% CI −0.292 to −0.204; p=6.56×10⁻³⁰). Reclassification improved at each horizon (categorical NRI 0.839/0.473/0.473; continuous NRI 0.879/0.884/0.884). Kaplan–Meier curves showed clear, monotonic separation of low-, intermediate-, and high-risk groups. Conclusions In this seven-centre real-world cohort, a clinical variable–only Cox model built from routinely available data outperformed a staging-only approach for predicting 6–18-month OS, showing superior discrimination, acceptable calibration, greater net benefit, and substantial reclassification gains. These findings support the use of readily obtainable clinical data for short- to mid-term risk stratification and shared decision-making when detailed TNM information is scarce.