Beyond Screening Populations: Context-Dependent Performance of Lung Cancer Risk Assessment Strategies in a Diagnostically Enriched Cohort

Background/Objectives: Lung cancer screening with low-dose computed tomography has been shown to reduce disease-specific mortality, but optimal strategies for identifying high-risk individuals remain debated. This retrospective, single-center cohort study examined the context-dependent behavior of two individualized lung cancer risk prediction models, PLCOm2012 and the Bach model, and National Comprehensive Cancer Network (NCCN) screening criteria when retrospectively applied to hospitalized patients undergoing diagnostic evaluation. The study was not designed as an external validation of these approaches for their intended screening purpose, but rather to assess their ability to discriminate between patients with and without prevalent lung cancer in a diagnostically enriched clinical population. Methods: Among 254 patients in the overall cohort, 86 had histopathologically confirmed primary lung cancer. For the primary head-to-head comparison, a common analytic cohort of 133 current or former smokers for whom all three approaches could be evaluated was used; 56 of these patients had lung cancer. Results: PLCOm2012 demonstrated high sensitivity (87.5%) but low specificity (15.6%), whereas the Bach model, at the cohort-derived exploratory threshold of ≥0.49%, showed lower sensitivity (42.9%) and higher specificity (57.1%). NCCN criteria yielded the highest sensitivity (96.4%) but very low specificity (3.9%). Likelihood ratios were close to unity for all three approaches, indicating limited discrimination in this diagnostically enriched population. Paired analyses demonstrated significant differences in sensitivity and specificity between several of the threshold-based approaches, whereas continuous discrimination did not differ significantly between PLCOm2012 and the Bach model; overall discrimination remained limited in this diagnostically enriched cohort. Conclusions: These findings highlight the context-dependent performance of screening-oriented risk assessment tools and indicate that results derived from asymptomatic screening populations should not be directly extrapolated to hospitalized patients undergoing diagnostic evaluation.