The study team restructured participants from the low-dose CT arm of the NLST into four analytic case groups based on screening history and timing of diagnosis. Controls were those never diagnosed with lung cancer (n = 26,038). Cases were split into 270 prevalence cancers (cancers identified after a baseline positive screen), 44 interval cancers (diagnosed after a negative screen and before next scheduled screen), 206 screen-detected cancers with a baseline positive screen, and 164 screen-detected cancers with a baseline negative screen. The authors reported that these groupings allowed development of tailored risk models aligned to differing screening trajectories and clinical outcomes.
Data indicated 48 candidate covariates were available in the NLST dataset. These included demographic factors, six smoking-related variables, alcohol use, fourteen occupational exposures, sixteen medical-history variables, family history of lung cancer, and a composite personal history of cancer variable. Age and BMI were re-categorized (age into quartiles; BMI per WHO categories with underweight and normal combined), and smoking onset age was dichotomized at the cohort mean (age 16).
The analytic approach used univariable testing to identify significant predictors from the 48 candidates, then entered those into a stepwise backward logistic regression to derive parsimonious models for each case-control comparison. The authors repeated the entire variable-selection procedure within internal validation (see below). Results from the univariable screens and categorized variable handling are reported in the source; the manuscript did not impute or invent additional variable definitions beyond what NLST provided.
Model discrimination was reported using the area under the receiver-operating characteristic curve (AUC). optimism-corrected discrimination was estimated using 500 bootstrap replications in which the full variable-selection process was repeated, producing adjusted AUCs that account for model selection optimism. Calibration was summarized using calibration slope and calibration-in-the-large.
The reported apparent AUCs were: prevalence model 0.711, interval model 0.734, baseline-positive screen-detected model 0.685, and baseline-negative screen-detected model 0.701. After bootstrap correction, optimism-corrected AUCs were 0.676, 0.644, 0.640, and 0.652 for the prevalence, interval, baseline-positive SDLC, and baseline-negative SDLC models, respectively. Calibration slopes across models ranged from approximately 0.67 to 0.81, as reported by the authors.
Prevalence cancers: The final prevalence model included age, pack-years smoked, BMI, smoking status (current/former), smoking onset age, personal history of cancer, family history of lung cancer, alcohol consumption, and milling occupation. The authors presented an apparent AUC of 0.711 and an optimism-corrected AUC of 0.676 for this model.
Interval cancers: The interval cancer model retained age, smoking status, smoking onset age, cigar smoking, marital status, and asbestos-related occupation. This model had the highest apparent AUC (0.734) of the four, but a lower optimism-corrected AUC of 0.644 after bootstrap adjustment.
Screen-detected cancers with baseline positive screen: For SDLCs arising after a baseline positive screen, the model included age, pack-years smoked, BMI, emphysema, chemicals/plastics exposure, and milling occupation. The apparent AUC was 0.685 and the optimism-corrected AUC was 0.640.
Screen-detected cancers with baseline negative screen: The model for SDLCs following a baseline negative screen retained age, pack-years smoked, BMI, smoking status, emphysema, sarcoidosis, and sandblasting occupation. Reported apparent discrimination was 0.701 and optimism-corrected discrimination 0.652.
Across these models, age and measures of tobacco exposure (pack-years, smoking status, smoking onset) consistently contributed to discrimination, but non-smoking variables — comorbidity indicators and specific occupational exposures — also entered the final models in group-specific ways.
The authors ran LASSO logistic regression across all 48 covariates as a sensitivity check. LASSO largely retained the same covariates identified by the stepwise selection procedure, according to the manuscript. No numeric LASSO shrinkage coefficients beyond the variable retention statement were invented here; the source reported concordance between selection approaches rather than detailed coefficient tables in the summary text.
The authors explicitly labeled these models as exploratory and hypothesis-generating. Internal validation used bootstrap resampling, but the source did not report external validation in an independent cohort. The NLST eligibility criteria restricted the cohort to older current and former heavy smokers, and therefore model performance may not generalize to populations outside those screening criteria. The manuscript also noted calibration slopes below 1.0 (approximately 0.67–0.81), indicating some overfitting or miscalibration that requires correction before clinical use.
Data availability is governed by the National Cancer Institute Cancer Data Access System; investigators must submit project proposals to obtain the same de-identified NLST files used in these analyses. The authors recommended external validation and further work to test whether incorporating screening history into risk calculators improves clinical decision-making, but the present report did not perform such impact or prospective implementation testing.
Limitations documented in the source include the post hoc nature of the analysis, the potential for selection bias within the NLST trial population, and the need for validation in independent datasets. The manuscript concluded with a call for validation rather than a clinical endorsement of these models.