A prognostic model integrating systemic inflammation and clinical features for overall survival in lung cancer: development, internal validation, and external evaluation

Conventional clinicopathological factors do not fully capture host-related influences on lung cancer survival. We developed and externally evaluated a prognostic model integrating systemic inflammatory markers with routine clinical characteristics for overall survival (OS). The development cohort comprised 272 single-center patients with pathologically confirmed lung cancer (60 deaths). A random survival forest (RSF) was built from variables retained in a multivariable Cox model. Internal validation used out-of-bag prediction, full-pipeline bootstrap optimism correction, and repeated cross-validation, with a Cox model and three machine-learning comparators on the same variables. The frozen Cox model, with its coefficients, baseline survival function, and development-derived cutoff, was applied without refitting to 74 patients (28 deaths) at a second institution; an RSF reconstructed from the development data alone was evaluated as an exploratory sensitivity analysis. Age, smoking history, AJCC stage, and the aggregate index of systemic inflammation (AISI) were retained in the final Cox model. Under full-pipeline correction the C-index fell from an apparent 0.776 to a corrected 0.698 and the 5-year AUC from 0.871 to 0.816 under the same design. The RSF conferred no advantage over Cox regression (cross-validated C-index 0.757 vs. 0.761), and no machine-learning method exceeded Cox on the cross-validated C-index. Externally, the frozen Cox model achieved a C-index of 0.823 (95% CI 0.739–0.897), 1- and 3-year AUCs of 0.798 and 0.868, and a calibration slope of 1.173 (0.752–1.593), compatible with 1 but imprecise. Absolute survival was nonetheless overestimated by 0.081 and 0.173 at 1 and 3 years. The development-derived Cox cutoff, applied unchanged, separated groups with markedly different survival (hazard ratio 42.48, 95% CI 5.73–314.71). The higher external discrimination reflected case mix (0.773 to 0.801 standardized). Integrating host inflammatory status with conventional staging provided modest but reproducible risk stratification for OS. Internally, the RSF performed comparably to, not better than, Cox regression. Externally, the frozen Cox model retained discrimination and risk-group separation, whereas the reconstructed forest gave concordant but exploratory findings. Because absolute survival was overestimated and both cohorts were small and retrospective, recalibration and larger multicenter confirmation are needed before clinical use.