---
title: "Risk models incorporating NLST screening history to stratify lung cancer diagnosis timing"
id: "plos-one-8-novel-risk-models-based-on-screening-history-results-and-timing-of-lung-cancer"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-8-novel-risk-models-based-on-screening-history-results-and-timing-of-lung-cancer"
content_type: "clinical_feed_article"
specialty: "Oncology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0363145"
published_at: "2026-11-18T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Risk models incorporating NLST screening history to stratify lung cancer diagnosis timing
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-8-novel-risk-models-based-on-screening-history-results-and-timing-of-lung-cancer
- **Specialty:** [Oncology](https://medichelpline.com/clinical-feed/oncology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0363145)
- **Published At:** 2026-11-18T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- Investigators performed four case-control analyses using the CT arm of the **NLST**, comparing 26,038 controls with four lung cancer case groups defined by screening history and timing. The four case groups were: 270 prevalence cancers, 44 interval cancers, 206 screen-detected cancers with a baseline positive screen, and 164 screen-detected cancers with a baseline negative screen. - The study evaluated 48 candidate covariates spanning demographics, smoking metrics, alcohol use, occupational exposures, medical history, family history of **lung cancer**, and personal cancer history; several NLST variables were re-categorized (age quartiles, WHO BMI categories, smoking onset dichotomized at age 16). - For each case-control comparison, univariable screening identified significant covariates that were then entered into stepwise backward selection to derive parsimonious logistic models. Model discrimination used area under the curve (AUC); internal validation used 500 bootstrap replications repeating variable selection to produce optimism-corrected AUCs. Calibration was assessed with calibration slope and calibration-in-the-large. - Prevalence model (AUC = 0.711) retained age, pack-years, BMI, smoking status, smoking onset age, personal cancer history, family history of lung cancer, alcohol consumption, and milling occupation; optimism-corrected AUC = 0.676. - Interval model (AUC = 0.734) included age, smoking status, smoking onset age, cigar smoking, marital status, and asbestos occupation; optimism-corrected AUC = 0.644. - Baseline-positive screen-detected model (AUC = 0.685) included age, pack-years, BMI, emphysema, chemicals/plastics exposure, and milling occupation; optimism-corrected AUC = 0.640. - Baseline-negative screen-detected model (AUC = 0.701) included age, pack-years, BMI, smoking status, emphysema, sarcoidosis, and sandblasting occupation; optimism-corrected AUC = 0.652. - LASSO sensitivity analyses largely confirmed the covariates selected by stepwise models. Calibration slopes across models ranged roughly from 0.67 to 0.81. - Authors noted that, beyond age and smoking (screening inclusion criteria), additional clinical and occupational factors contributed to discriminatory performance; they described these results as exploratory and hypothesis-generating and called for external validation and further study.
## Clinical Analysis & Structured Key Points
Novel risk models based on screening history results and timing of lung cancer diagnosis: Post hoc analysis of the National Lung Cancer Screening Trial | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Background Our group previously reported that lung cancer (LC) screening history results and subsequent timing of diagnosis are associated with significant differences in survival outcomes. As a follow-up study, we sought to develop novel personalized risk models that considered screening history for incidence cancers, interval LCs, and prevalence LCs. Methods Using data from the CT-arm of the NLST, four independent case-control analyses were conducted to develop parsimonious risk models. Controls (n = 26,038) were those never diagnosed with LC. The four LC case groups were 270 prevalence LCs, 44 interval LCs, 206 screen-detected LCs (SDLCs) that had a baseline positive screen, and 164 SDLCs that had a baseline negative screen. For each case-control analysis, univariable analyses identified statistically significant covariates from 48 variables and then significant covariates were included into a stepwise backward selection approach to identify a model with the most informative covariates. Model discrimination was quantified by the area under the curve (AUC). Internal validation was performed using 500 bootstrap replications in which the entire variable-selection procedure was repeated, yielding optimism-corrected AUCs, and calibration was assessed using calibration slope and calibration-in-the-large. LASSO logistic regression was applied to all 48 covariates as a sensitivity analysis. Results For prevalence LCs, the model (AUC = 0.711) included age, pack-years smoked, BMI, smoking status, smoking onset age, personal history of cancer, family history of LC, alcohol consumption, and milling occupation. For interval LCs, the model (AUC = 0.734) included age, smoking status, smoking onset age, cigar smoking, marital status, and asbestos occupation. For baseline positive SDLCs, the model (AUC = 0.685) included age, pack-years smoked, BMI, emphysema, chemicals/plastics exposure, and milling occupation. For baseline negative SDLCs, the model (AUC = 0.701) included age, pack-years smoked, BMI, smoking status, emphysema, sarcoidosis, and sandblasting occupation. 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, and calibration slopes ranged from 0.67 to 0.81. LASSO sensitivity analyses retained largely consistent model covariates. Conclusions Besides smoking and age, which are inclusion criteria for screening, these models identified other risk factors that may inform personalized LC risk assessment. These findings should be regarded as exploratory and hypothesis-generating. Citation: Haddan SA, Waqas A, Rasool G, Schabath MB (2026) Novel risk models based on screening history results and timing of lung cancer diagnosis: Post hoc analysis of the National Lung Cancer Screening Trial. PLoS One 21(11): e0363145. https://doi.org/10.1371/journal.pone.0363145 Editor: Eugenio Paci, Centro per lo Studio e la Prevenzione Oncologica, ITALY Received: April 13, 2026; Accepted: November 3, 2026; Published: November 18, 2026 Copyright: © 2026 Haddan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: All de-identified NLST data files are available from the National Cancer Institute (NCI) Cancer Data Access System (CDAS). We did not generate any proprietary variables. As such, the same data used in our analyses are available to all researchers by submitting a project proposal application at CDAS. Based on the agreement signed with the NCI, the authors are not permitted to share the dataset and only received access to the data after completing a project proposal and receiving permission to access. To request access to the NLST data files, a project proposal must be submitted in the CDAS portal ( https://cdas.cancer.gov/learn/nlst/instructions/ ). Project proposals require users to identify their institution, a Principal Investigator(PI), a Signing Official, and project staff. Data-Only project proposals must be submitted by the PI. Funding: Funding support came from the National Cancer Institute Early Detection Research Network (U01-CA200464 to MBS). Competing interests: NO authors have competing interests. Introduction The National Lung Screening Trial (NLST) demonstrated a 20% relative reduction in lung cancer mortality for individuals screened by low-dose helical computed tomography (LDCT) compared with chest radiography in a high-risk population of 53,454 current and former smokers aged 55–74 years [ 1 ]. While lung cancer screening guidelines in the United States are based on age and smoking history [ 2 – 4 ], these criteria do not take into account other important lung risk factors such as comorbidities, family history, occupational exposure, or other smoking-related factors (e.g., age of initiation). While there are existing risk models [ 5 , 6 ] that incorporate additional risk factors beyond age and smoking history, and some lung cancer screening guidelines do encourage using such risk calculators [ 2 , 4 ], these risk models do not consider screening history results, which may provide additional value to personalized risk assessment. In a prior publication from our group [ 7 ], we conducted an in-depth analysis of the CT-arm of the NLST to characterize patient characteristics and survival outcomes of lung cancers diagnosed based on screening history results over the first three screening intervals. A striking finding was that screening history is an independent prognostic factor whereby incidence lung cancers, diagnosed at the follow-up screening intervals with one or more antecedent negative screens are associated with inferior survival (47.6% 5-year survival) compared to incidence cancers in which only positive screens preceded the lung cancer diagnosis (65.7% 5-year survival). We speculate that the inferior survival associated with incidence lung cancers with one or more antecedent negative screens represent cancers that may be faster growing and more biologically aggressive, arising from a lung environment previously lacking a focal abnormality. However, this remains a hypothesis that has not been directly tested. We also found that the prevalence lung cancers, which are those who have a baseline positive screen that was determined to be cancer, had a 59.9% 5-year survival, and interval cancers, which are diagnosed after a prior negative screening but before the next scheduled screening, were associated with extremely inferior survival (17.3% 5-year survival). As a follow-up to this work, we sought to identify novel personalized risk models for lung cancer based on screening history results and the timing of lung cancer diagnosis, which, subject to validation, could inform personalized management by combining risk factor assessment with screening history. Materials and methods NLST Study Design, Study Population, and CT Screening Results This research was approved by the Advarra Institutional Review Board. De-identified NLST data were obtained through the National Cancer Institute Cancer Data Access System [ 7 ]. The study design and main findings from NLST have been described previously [ 1 , 8 ]. Briefly, the NLST was a multicenter randomized clinical trial comparing screening with LDCT versus chest radiography in high-risk individuals. Eligibility criteria included current or former smokers aged 55–74 years with a minimum 30 pack-year smoking history. Former smokers must have quit within the past 15 years. Based on the NLST criteria, a positive screening result was defined as having one or more non-calcified nodules or masses with axial diameter measuring ≥ 4 mm. Less commonly, other abnormalities such as adenopathy or pleural effusion can be defined as a positive screening result. Negative screens were defined as CT scans with no abnormalities, minor abnormalities not suspicious for lung cancer, or significant abnormalities not suspicious for lung cancer. Stable abnormalities across all three rounds could be classified as negative screens at the final screen at the discretion of the interpreting radiologist. Based on a prior publication from our group [ 9 ], we restructured the CT arm of the NLST ( Fig 1 ) based on screening results at baseline (T0) and the first (T1) and second (T2) follow-up screening intervals. Based on this schema, there were 10 lung cancer case groups that were collapsed into 4 groups for these analyses. Specifically, prevalence lung cancers (green box in Fig 1 ) were diagnosed with lung cancer following a T0 positive screening. Interval lung cancers (yellow boxes in Fig 1 ) were diagnosed after a negative screening at any time between the next scheduled screening interval. A screen-detected lung cancer was defined as an incidence cancer that was identified following a positive screen at the T1 or T2 screening intervals. For this analysis, screen-detected lung cancers were then categorized into two distinct lung cancer case groups according to the baseline screening result (negative or positive). Those with a baseline positive screening who were diagnosed with lung cancer after a follow-up screening interval (T1 or T2) were grouped in one lung cancer case group (blue boxes in Fig 1 ). Those with a baseline negative screening who were diagnosed with lung cancer after a follow-up screening interval (T1 or T2) were grouped into a separate lung cancer case group (red boxes in Fig 1 ). Download: PNG larger image TIFF original image Fig 1. Schema for screening history for all participants spanning three rounds for the LDCT-arm of the NLST. https://doi.org/10.1371/journal.pone.0363145.g001 Independent variables We considered 48 independent covariates available from the NLST dataset, including eight demographic variables, six smoking-related variables, two related to alcohol use, fourteen occupational exposure variables, sixteen medical history variables, one variable for family history of lung cancer, and one variable for personal history of any cancer. For this analysis, some of the original NLST variables were re-categorized. Age was categorized into quartiles, and BMI was categorized using the World Health Organization (WHO) guidelines. However, underweight and normal categories were combined due to a low number of participants in the underweight BMI category (e.g., zero participants in this category among interval cancers). The personal history of any cancer was generated by combining the personal history of 15 specific cancer diagnoses. Participants with a history of one or more of these cancers were coded as a “yes,” while a “no” indicated no personal history for any of these cancers. Smoking onset age was categorized into two groups based on the average smoking onset age of 16. For marital status, widowed, separated, and divorced were combined into one category, while never married, married, and not applicable (N/A) were kept as independent categories. The original education variable was reduced to four categories: (i) high school graduate/ GED, 9–11 th grade, and 8 th grade or less; (ii) associate degree/ some college, and post-high school training (excluding college); (iii) graduate school or bachelor’s degree; (iv) other or N/A. Missing data All analyses were complete case-control analyses, where participants with a missing value for a covariate retained in a given model were excluded from that model, and no imputation was performed. The number and percentage of participants with missing data for each of the 48 candidate covariates, overall and by case group, are reported in S3 Table . Entries of “NC” (non-calculable) in Table 2 and in S1 and S2 Tables denote subgroups in which an odds ratio could not be estimated due to sparse sample sizes. The potential consequences of missingness is addressed in the Discussion. Statistical Analyses All statistical analyses were performed using Stata/MP 14.2 for Mac (64-bit Intel). The lung cancer case groups for this analysis were prevalence lung cancers, interval lung cancers, screen-detected lung cancers with a T0 positive screen, and screen-detected lung cancers with a T0 negative screen. The control subjects were participants who were never diagnosed with lung cancer through the original follow-up of the NLST. To assess potential differences in demographic characteristics, Fisher’s exact test was used to test for differences between the control subjects and each of the four lung cancer case groups for categorical variables, and the Student’s t-test was used to test for differences for continuous variables. To identify unique and personalized models associated with lung cancer risk, four independent case-control analyses were performed based on the four different case groups. For each case-control analysis, univariable logistic regression analyses were used to identify statistically significant covariates (p < 0.05) from the 48 covariates. Then, the statistically significant covariates were included into a stepwise backward selection model, using a threshold of 0.1 for inclusion into the model, which was used to identify the most informative covariates for each case-control analysis. For each final model, we also adjusted for potential confounding, including gender, race, ethnicity, smoking status, and pack-years smoked, where applicable. Adjustment was implemented by force entry of these covariates into the final model, irrespective of their statistical significance, so that they were retained regardless of the backward-elimination criterion. “Where applicable” indicates that a potential confounder was force-entered only if it had not already been retained by the backward-elimination procedure for that case group. For example, smoking status and pack-years smoked were already retained in the prevalence lung cancer model and were therefore not additionally force-entered, whereas race and ethnicity, which were retained in none of the four models, were force-entered in all four. The adjusted models are presented in S2 Table and the corresponding unadjusted models in Table 2 . Area under the receiver operating characteristic curve (AUC) was generated for each model to measure model discrimination. Because model development and performance assessment were conducted in the same dataset, these apparent AUCs are optimistic. The internal validation and calibration procedures used to address this are described below. Control subjects were the 26,038 CT-arm participants never diagnosed with lung cancer throughout NLST follow-up, and all four case-control analyses used unmatched logistic regression. This analytical design was selected for three reasons. First, all candidate covariates were baseline data ascertained at trial enrollment so exposure information shares a common temporal origin for every participant. Second, the lung cancer case groups are defined by screening history and by the timing of diagnosis relative to the three original screening rounds (T0, T1, and T2) rather than by continuous time since entry. Risk sets in an incidence-density design would therefore be indexed by screening round, and, because all participants were offered the same three annual screens over a common interval, such risk sets closely approximate the unmatched comparison performed here [ 10 ]. Third, follow-up in the NLST was comparatively uniform where accrual occurred over the defined enrollment period, all participants were offered at least three annual screens, and active follow-up continued to a common administrative censoring date. As such, between-group variation in observation time is minimal. We nonetheless agree that differential follow-up could in principle bias the odds ratios, and we therefore performed a sensitivity analysis using incidence-density (risk-set) sampling, in which controls were sampled at random from the set of participants alive, under observation, and free of lung cancer at the screening round at which each case was diagnosed, matched on screening round and on age at randomization, followed by conditional logistic regression. Because the outcome is uncommon in this trial, the resulting odds ratios closely approximate incidence rate ratios. Internal validation was performed by bootstrap optimism correction [ 11 , 12 ]. For each of the four case-control analyses, 500 bootstrap samples were drawn with replacement. The complete model-building procedure, univariable screening at p < 0.05 followed by backward elimination at p < 0.10, was repeated in full within every bootstrap sample, so that the variable-selection step itself was subject to resampling. For each replication, the AUC of the bootstrap model was computed in the bootstrap sample and in the original sample. The mean difference across the 500 replications was taken as the optimism, and the optimism-corrected AUC was obtained by subtracting this optimism from the apparent AUC of the corresponding final model. The proportion of the 500 replications in which each covariate was retained (the bootstrap selection frequency) was recorded as a measure of variable-selection stability. The same procedure yielded an optimism-corrected calibration slope for each model, which we used as a uniform shrinkage factor: the coefficients of each final model were multiplied by that model’s slope, and the intercept was re-estimated with the shrunken linear predictor as an offset [ 11 , 13 ]. The odds ratios presented in Table 2 and S2 Table are the original, unshrunken maximum-likelihood estimates, which we retain as the primary endpoints because the goal of these analyses is to characterize the covariates associated with each case-control analysis rather than to develop a risk calculator, and because interval estimates are not defined for uniformly shrunken coefficients without resampling the entire selection-and-shrinkage procedure. The shrunken coefficients for all four models are reported in Supplemental Table 7 , where the shrinkage factors quantify the degree to which the unshrunken coefficients are too extreme. Calibration was assessed graphically by plotting observed against predicted probabilities using a loess (i.e., locally estimated scatterplot) smoother and, separately, by deciles of predicted risk, and numerically by the calibration slope, calibration-in-the-large (calibration intercept), the Brier score, and the Hosmer-Lemeshow goodness-of-fit test [ 13 ]. Because these are case-control analyses in which the ratio of cases to controls does not reflect the underlying incidence of lung cancer, the absolute predicted probabilities are not interpretable as absolute risks. Calibration was therefore evaluated within the case-control sample, and the calibration slope is the primary calibration measure of interest, as it quantifies the degree to which the estimated regression coefficients are too extreme. Calibration plots are presented in S1 Fig and calibration metrics in S4 Table . Univariable analyses followed by stepwise backward elimination was retained as the primary model-building strategy because the objective of this analysis was to derive, for each of four clinically defined case groups, a small and directly interpretable set of the most informative covariates from among 48 candidates, and this is the approach was successfully applied in the development of several established lung cancer risk models [ 5 , 14 – 17 ]. We acknowledge that stepwise procedures may produce unstable v
## Related Clinical Research

- [Early STAT+ Results for Alterome’s KRAS Inhibitor in Pancreatic, Lung, and Colorectal Cancer](https://medichelpline.com/clinical-feed/stat-news-1-stat-early-results-for-alterome-s-kras-drug-for-pancreas-lung-and-colon-cancer.md)
- [COMMD1 in Cancer: Dual Roles, Mechanisms, and Translational Targets](https://medichelpline.com/clinical-feed/pubmed-43059391.md) (DOI: 10.3892/or.2026.9224)
- [LCatlas: Point-of-Care Prediction Tool for Immunotherapy Response in NSCLC](https://medichelpline.com/clinical-feed/medrxiv-19-lcatlas-a-point-of-care-tool-for-predicting-immunotherapy-outcomes-in-non-small.md)
- [FDA leadership void after Overton withdrawal; FDA rejects lung cancer drug and STAT ranks China’s](https://medichelpline.com/clinical-feed/stat-news-1-fda-s-search-for-a-new-leader-and-china-s-largest-drugmakers.md)
- [Stage distribution and stage-specific survival drive lung cancer survival gaps for Māori in Aotear](https://medichelpline.com/clinical-feed/pubmed-43056629.md) (DOI: 10.26635/6965.7656)

## Navigation
- [← Back to Oncology Feed](https://medichelpline.com/clinical-feed/oncology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.