Development and Validation of a Multivariable Lung Cancer Risk Prediction Model That Includes Low-Dose Computed Tomography Screening Results: A Secondary Analysis of Data From the National Lung Screening Trial | Cancer Screening, Prevention, Control | JAMA Network Open | JAMA Network
[Skip to Navigation]
Sign In
Figure 1.  Schema Showing PLCO2012results Model Development Plan
Schema Showing PLCO2012results Model Development Plan

Lung-RADS indicates Lung CT Screening Reporting & Data System developed by the American College of Radiology; NLST, National Lung Screening Trial; PLCOm2012, risk prediction model described by Tammemägi et al6; PLCOm2012bu, PLCOm2012 model with predictors age, smoking duration in current smokers, and quit time in former smokers updated to the start of study follow-up (T3) by adding 3 years to baseline values (the PLCOm2012bu is estimated for a 3-year period, not the original 6-year period); T0, baseline screen; T1, first annual screen; and T2, annual screen at year 2.

Figure 2.  Cumulative Incidence of Lung Cancer in the National Lung Screening Trial Low-Dose Computed Tomography Group Occurring 1 to 4 Years After the Last Low-Dose Computed Tomography Screen Among 23 227 Participantsa
Cumulative Incidence of Lung Cancer in the National Lung Screening Trial Low-Dose Computed Tomography Group Occurring 1 to 4 Years After the Last Low-Dose Computed Tomography Screen Among 23 227 Participantsa

Stratified by Lung CT Screening Reporting & Data System screen results (positive vs negative) at baseline, 1-year, and 2-year annual screenings categorized into 4 groups.

aCompeting risks (ie, non–lung cancer deaths) were taken into account according to the method by Fine and Gray.24 Included in the analysis were 298 incident lung cancer cases and 735 competing-cause deaths during the 3-year follow-up period.

Table 1.  Univariable and Multivariable Logistic Regression Model Odds Ratios for Predicting Lung Cancer Occurring 1 to 4 Years After the Third Screen, Stratified by 8-Level and 4-Level Screening Results, Prepared in LSS Data and Validated in ACRIN Data
Univariable and Multivariable Logistic Regression Model Odds Ratios for Predicting Lung Cancer Occurring 1 to 4 Years After the Third Screen, Stratified by 8-Level and 4-Level Screening Results, Prepared in LSS Data and Validated in ACRIN Data
Table 2.  Distributions of Lung Cancer Outcomes Occurring 1 to 4 Years After the Last Screen in National Lung Screening Trial Participantsa
Distributions of Lung Cancer Outcomes Occurring 1 to 4 Years After the Last Screen in National Lung Screening Trial Participantsa
Table 3.  Potential Implications of the Study Findings Regarding Offering the Next Annual Screen Based on PLCOm2012 and PLCO2012results Models
Potential Implications of the Study Findings Regarding Offering the Next Annual Screen Based on PLCOm2012 and PLCO2012results Models
1.
Brambilla  E, Travis  W. Lung cancer. In: Stewart  B, Wild  C, eds.  World Cancer Report 2014. Lyon, France: International Agency for Research on Cancer, World Health Organization; 2014:350-361.
2.
Aberle  DR, Adams  AM, Berg  CD,  et al; National Lung Screening Trial Research Team.  Reduced lung-cancer mortality with low-dose computed tomographic screening.  N Engl J Med. 2011;365(5):395-409. doi:10.1056/NEJMoa1102873PubMedGoogle ScholarCrossref
3.
Moyer  VA; U.S. Preventive Services Task Force.  Screening for lung cancer: U.S. Preventive Services Task Force recommendation statement.  Ann Intern Med. 2014;160(5):330-338. doi:10.7326/M13-2771PubMedGoogle ScholarCrossref
4.
Centers for Medicare & Medicaid Services (CMS). Decision memo for screening for lung cancer with low dose computed tomography (LDCT) (CAG-00439N). https://www.cms.gov/medicare-coverage-database/details/nca-decision-memo.aspx?NCAId=274&NcaName=Screening+for+Lung+Cancer+with+Low+Dose+Computed+Tomography+(LDCT)&MEDCACId=68&IsPopup=y&bc=AAAAAAAAAgAAAA%3d%3d&. Published February 5, 2015. Accessed August 31, 2015.
5.
Bach  PB, Mirkin  JN, Oliver  TK,  et al.  Benefits and harms of CT screening for lung cancer: a systematic review.  JAMA. 2012;307(22):2418-2429. doi:10.1001/jama.2012.5521PubMedGoogle ScholarCrossref
6.
Tammemägi  MC, Katki  HA, Hocking  WG,  et al.  Selection criteria for lung-cancer screening.  N Engl J Med. 2013;368(8):728-736. doi:10.1056/NEJMoa1211776PubMedGoogle ScholarCrossref
7.
Kovalchik  SA, Tammemägi  M, Berg  CD,  et al.  Targeting of low-dose CT screening according to the risk of lung-cancer death.  N Engl J Med. 2013;369(3):245-254. doi:10.1056/NEJMoa1301851PubMedGoogle ScholarCrossref
8.
Tammemägi  MC, Church  TR, Hocking  WG,  et al.  Evaluation of the lung cancer risks at which to screen ever- and never-smokers: screening rules applied to the PLCO and NLST cohorts.  PLoS Med. 2014;11(12):e1001764. doi:10.1371/journal.pmed.1001764PubMedGoogle ScholarCrossref
9.
Katki  HA, Kovalchik  SA, Berg  CD, Cheung  LC, Chaturvedi  AK.  Development and validation of risk models to select ever-smokers for CT lung cancer screening.  JAMA. 2016;315(21):2300-2311. doi:10.1001/jama.2016.6255PubMedGoogle ScholarCrossref
10.
Cressman  S, Peacock  SJ, Tammemägi  MC,  et al.  The cost-effectiveness of high-risk lung cancer screening and drivers of program efficiency.  J Thorac Oncol. 2017;12(8):1210-1222. doi:10.1016/j.jtho.2017.04.021PubMedGoogle ScholarCrossref
11.
Li  K, Hüsing  A, Sookthai  D,  et al.  Selecting high-risk individuals for lung cancer screening: a prospective evaluation of existing risk models and eligibility criteria in the German EPIC cohort.  Cancer Prev Res (Phila). 2015;8(9):777-785. doi:10.1158/1940-6207.CAPR-14-0424PubMedGoogle ScholarCrossref
12.
Weber  M, Yap  S, Goldsbury  D,  et al.  Identifying high risk individuals for targeted lung cancer screening: independent validation of the PLCOm2012 risk prediction tool.  Int J Cancer. 2017;141(2):242-253. doi:10.1002/ijc.30673PubMedGoogle ScholarCrossref
13.
ten Haaf  K, Jeon  J, Tammemägi  MC,  et al.  Risk prediction models for selection of lung cancer screening candidates: a retrospective validation study.  PLoS Med. 2017;14(4):e1002277. doi:10.1371/journal.pmed.1002277PubMedGoogle ScholarCrossref
14.
Tammemägi  MC, Schmidt  H, Martel  S,  et al; PanCan Study Team.  Participant selection for lung cancer screening by risk modelling (the Pan-Canadian Early Detection of Lung Cancer [PanCan] study): a single-arm, prospective study.  Lancet Oncol. 2017;18(11):1523-1531. doi:10.1016/S1470-2045(17)30597-1PubMedGoogle ScholarCrossref
15.
Katki  HA, Kovalchik  SA, Petito  LC,  et al.  Implications of nine risk prediction models for selecting ever-smokers for computed tomography lung-cancer screening.  Ann Intern Med. 2018;169(1):10-19. doi:10.7326/M17-2701PubMedGoogle ScholarCrossref
16.
Kavanagh  J, Liu  G, Menezes  R,  et al.  Importance of long-term low-dose CT follow-up after negative findings at previous lung cancer screening.  Radiology. 2018;289(1):218-224. doi:10.1148/radiol.2018180053PubMedGoogle ScholarCrossref
17.
Yousaf-Khan  U, van der Aalst  C, de Jong  PA,  et al.  Risk stratification based on screening history: the NELSON lung cancer screening study.  Thorax. 2017;72(9):819-824. doi:10.1136/thoraxjnl-2016-209892PubMedGoogle ScholarCrossref
18.
Maisonneuve  P, Bagnardi  V, Bellomi  M,  et al.  Lung cancer risk prediction to select smokers for screening CT: a model based on the Italian COSMOS trial.  Cancer Prev Res (Phila). 2011;4(11):1778-1789. doi:10.1158/1940-6207.CAPR-11-0026PubMedGoogle ScholarCrossref
19.
Patz  EF  Jr, Greco  E, Gatsonis  C, Pinsky  P, Kramer  BS, Aberle  DR.  Lung cancer incidence and mortality in National Lung Screening Trial participants who underwent low-dose CT prevalence screening: a retrospective cohort analysis of a randomised, multicentre, diagnostic screening trial.  Lancet Oncol. 2016;17(5):590-599. doi:10.1016/S1470-2045(15)00621-XPubMedGoogle ScholarCrossref
20.
American College of Radiology. Lung CT Screening Reporting & Data System (Lung-RADS). http://www.acr.org/Quality-Safety/Resources/LungRADS. Accessed April 27, 2017.
21.
Aberle  DR, Berg  CD, Black  WC,  et al; National Lung Screening Trial Research Team.  The National Lung Screening Trial: overview and study design.  Radiology. 2011;258(1):243-253. doi:10.1148/radiol.10091808PubMedGoogle ScholarCrossref
22.
Pinsky  PF, Gierada  DS, Black  W,  et al.  Performance of Lung-RADS in the National Lung Screening Trial: a retrospective assessment.  Ann Intern Med. 2015;162(7):485-491. doi:10.7326/M14-2086PubMedGoogle ScholarCrossref
23.
Royston  P, Sauerbrei  W.  Multivariable Model-Building: A Pragmatic Approach to Regression Analysis Based on Fractional Polynomials for Modelling Continuous Variables. Hoboken, NJ: John Wiley & Sons; 2008. doi:10.1002/9780470770771
24.
Fine  JP, Gray  RJ.  A proportional hazards model for the subdistribution of a competing risk.  J Am Stat Assoc. 1999;94(446):496-509. doi:10.1080/01621459.1999.10474144Google ScholarCrossref
25.
Tammemägi  MC, Berg  CD, Riley  TL, Cunningham  CR, Taylor  KL.  Impact of lung cancer screening results on smoking cessation.  J Natl Cancer Inst. 2014;106(6):dju084. doi:10.1093/jnci/dju084PubMedGoogle ScholarCrossref
26.
Barry  SA, Tammemägi  MC, Penek  S,  et al.  Predictors of adverse smoking outcomes in the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial.  J Natl Cancer Inst. 2012;104(21):1647-1659. doi:10.1093/jnci/djs398PubMedGoogle ScholarCrossref
27.
Woodward  M.  Epidemiology: Study Design and Data Analysis. 3rd ed. Boca Raton, FL: CRC Press; 2014.
28.
Harrell  FE Jr.  Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. 2nd ed. Basel, Switzerland: Springer Nature Switzerland AG; 2015. doi:10.1007/978-3-319-19425-7
29.
Vickers  AJ.  Decision analysis for the evaluation of diagnostic tests, prediction models and molecular markers.  Am Stat. 2008;62(4):314-320. doi:10.1198/000313008X370302PubMedGoogle ScholarCrossref
30.
Brown  LD, Cai  TT, DasGupta  A.  Interval estimation for a binomial proportion.  Stat Sci. 2001;16:101-133. doi:10.1214/ss/1009213286Google Scholar
31.
Miettinen  O.  Estimability and estimation in case-referent studies.  Am J Epidemiol. 1976;103(2):226-235. doi:10.1093/oxfordjournals.aje.a112220PubMedGoogle ScholarCrossref
32.
Cuzick  J.  A Wilcoxon-type test for trend.  Stat Med. 1985;4(1):87-90. doi:10.1002/sim.4780040112PubMedGoogle ScholarCrossref
33.
Collins  GS, Reitsma  JB, Altman  DG, Moons  KG.  Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement.  Ann Intern Med. 2015;162(1):55-63. doi:10.7326/M14-0697PubMedGoogle ScholarCrossref
34.
Field  JK, Duffy  SW, Baldwin  DR,  et al.  The UK Lung Cancer Screening trial: a pilot randomised controlled trial of low-dose computed tomography screening for the early detection of lung cancer.  Health Technol Assess. 2016;20(40):1-146. doi:10.3310/hta20400PubMedGoogle ScholarCrossref
35.
Gomes  M, Teixeira  AL, Coelho  A, Araújo  A, Medeiros  R.  The role of inflammation in lung cancer.  Adv Exp Med Biol. 2014;816:1-23. doi:10.1007/978-3-0348-0837-8_1PubMedGoogle ScholarCrossref
36.
ten Haaf  K, Tammemägi  MC, Bondy  SJ,  et al.  Performance and cost-effectiveness of computed tomography lung cancer screening scenarios in a population-based setting: a microsimulation modeling analysis in Ontario, Canada.  PLoS Med. 2017;14(2):e1002225. doi:10.1371/journal.pmed.1002225PubMedGoogle ScholarCrossref
Limit 200 characters
Limit 25 characters
Conflicts of Interest Disclosure

Identify all potential conflicts of interest that might be relevant to your comment.

Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.

Err on the side of full disclosure.

If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.

Not all submitted comments are published. Please see our commenting policy for details.

Limit 140 characters
Limit 3600 characters or approximately 600 words
    Original Investigation
    Oncology
    March 1, 2019

    Development and Validation of a Multivariable Lung Cancer Risk Prediction Model That Includes Low-Dose Computed Tomography Screening Results: A Secondary Analysis of Data From the National Lung Screening Trial

    Author Affiliations
    • 1Department of Health Sciences, Brock University, St Catharines, Ontario, Canada
    • 2Department of Public Health, Erasmus MC University Medical Center, Rotterdam, the Netherlands
    • 3Department of Radiology, Stanford University School of Medicine, Palo Alto, California
    • 4Institute for Technology Assessment, Massachusetts General Hospital, Harvard Medical School, Boston
    • 5Department of Medicine, Stanford University School of Medicine, Palo Alto, California
    • 6School of Public Health, University of Michigan, Ann Arbor
    • 7Information Management Systems, Rockville, Maryland
    • 8Department of Epidemiology, University of Michigan, Ann Arbor
    JAMA Netw Open. 2019;2(3):e190204. doi:10.1001/jamanetworkopen.2019.0204
    Key Points español 中文 (chinese)

    Question  In this study of data from the National Lung Screening Trial (NLST), can a lung cancer risk model’s prediction be improved by inclusion of lung cancer screening results?

    Findings  In this secondary analysis of NLST data including 22 229 participants, a model incorporating a validated lung cancer risk prediction model, the PLCOm2012 model, with National Lung Screening Trial results (PLCO2012results) predicted future lung cancer significantly better than a model excluding results.

    Meaning  The PLCO2012results model estimates may improve stratification of patients being screened for lung cancer into high- and low-risk strata and may help guide decision making regarding screening interval.

    Abstract

    Importance  Low-dose computed tomography lung cancer screening is most effective when applied to high-risk individuals.

    Objectives  To develop and validate a risk prediction model that incorporates low-dose computed tomography screening results.

    Design, Setting, and Participants  A logistic regression risk model was developed in National Lung Screening Trial (NLST) Lung Screening Study (LSS) data and was validated in NLST American College of Radiology Imaging Network (ACRIN) data. The NLST was a randomized clinical trial that recruited participants between August 2002 and April 2004, with follow-up to December 31, 2009. This secondary analysis of data from the NLST took place between August 10, 2013, and November 1, 2018. Included were LSS (n = 14 576) and ACRIN (n = 7653) participants who had 3 screens, adequate follow-up, and complete predictor information.

    Main Outcomes and Measures  Incident lung cancers occurring 1 to 4 years after the third screen (202 LSS and 96 ACRIN). Predictors included scores from the validated PLCOm2012 risk model and Lung CT Screening Reporting & Data System (Lung-RADS) screening results.

    Results  Overall, the mean (SD) age of 22 229 participants was 61.3 (5.0) years, 59.3% were male, and 90.9% were of non-Hispanic white race/ethnicity. During follow-up, 298 lung cancers were diagnosed in 22 229 individuals (1.3%). Eight result combinations were pooled into 4 groups based on similar associations. Adjusted for PLCOm2012 risks, compared with participants with 3 negative screens, participants with 1 positive screen and last negative had an odds ratio (OR) of 1.93 (95% CI, 1.34-2.76), and participants with 2 positive screens with last negative or 2 negative screens with last positive had an OR of 2.66 (95% CI, 1.60-4.43); when 2 or more screens were positive with last positive, the OR was 8.97 (95% CI, 5.76-13.97). In ACRIN validation data, the model that included PLCOm2012 scores and screening results (PLCO2012results) demonstrated significantly greater discrimination (area under the curve, 0.761; 95% CI, 0.716-0.799) than when screening results were excluded (PLCOm2012) (area under the curve, 0.687; 95% CI, 0.645-0.728) (P < .001). In ACRIN validation data, PLCO2012results demonstrated good calibration. Individuals who had initial negative scans but elevated PLCOm2012 six-year risks of at least 2.6% did not have risks decline below the 1.5% screening eligibility criterion when subsequent screens were negative.

    Conclusions and Relevance  According to this analysis, some individuals with elevated risk scores who have negative initial screens remain at elevated risks, warranting annual screening. Positive screens seem to increase baseline risk scores and may identify high-risk individuals for continued screening and enrollment into clinical trials.

    Trial Registration  ClinicalTrials.gov Identifier: NCT00047385

    Introduction

    Lung cancer is the leading cause of cancer death in the world.1 The National Lung Screening Trial (NLST)2 demonstrated a 20% lung cancer mortality reduction for low-dose computed tomography (LDCT) screening of high-risk individuals compared with chest radiography screening. Consequently, the US Preventive Services Task Force3 and the Centers for Medicare & Medicaid Services4 have recommended annual screening for high-risk individuals, including current and former smokers, who quit less than 15 full years ago, smoked at least 30 pack-years, and are aged 55 to 80 years or 55 to 77 years, respectively. These criteria are identical to NLST enrollment criteria, except for increasing the upper age limit beyond 74 years. Lung cancer screening programs using these criteria are being implemented across the United States. Major concerns regarding lung cancer screening are the harms and costs of the high number of false-positive screens and excess radiation exposure.5 Eliminating nonbeneficial screens is expected to reduce the harms and improve cost-effectiveness. It has been demonstrated that, compared with NLST-like criteria, accurate lung cancer risk prediction models are more sensitive in selecting individuals who develop lung cancer, have higher positive predictive values, have a lower number needed to screen to avert 1 lung cancer death, and are more cost-effective.6-10 The PLCOm2012 is a lung cancer risk prediction model that has been validated by research teams in several countries, including the United States, Germany, Australia, and Canada.6,11-16 One appropriate PLCOm2012 risk threshold for identifying individuals for screening is at least 1.5% 6-year risk.8

    Evidence suggests that lung cancer screening results predict future lung cancer risk. The Dutch-Belgian Lung Cancer Screening (NELSON) trial demonstrated that the results of the initial screen were associated with subsequent incidence of lung cancer in 5.5 years of follow-up.17 The results of the first 3 screenings, especially the third screening round, were indicative of the risk of lung cancer detected in the fourth screening round, which occurred 2.5 years later. Data from the Continuous Observation of Smoking Subject (COSMOS) trial18 were used to develop a model that included nodule features from the initial screen that predicted lung cancer identified at subsequent screens. In the NLST,19 individuals with a negative initial screen result had lower subsequent lung cancer risks. We hypothesized that prediction of future lung cancer risk could be improved by combining information from screening results with risk prediction model estimates. Such updated risk estimates could help determine whether one might defer the next annual screen owing to low risk or more strongly recommend the next screen owing to high risk.

    The American College of Radiology developed the Lung CT Screening Reporting & Data System (Lung-RADS),20 which has become commonly used in screening practice in North America. Lung-RADS 1 (no nodule or nonsuspicious nodule) and Lung-RADS 2 (nodule present with benign appearance or behavior, having very low likelihood of becoming a clinically active cancer) are described as negative in that they do not trigger additional surveillance imaging or clinical diagnostic investigation. Lung-RADS 3 indicates the presence of a probably benign nodule, which has a low likelihood of becoming a clinically active cancer, and a 6-month surveillance scan is recommended. Lung-RADS 4A, 4B, and 4X indicate a screening result that is suspicious for lung cancer, which requires further investigation that can range from a 3-month surveillance scan to tissue biopsy. In the present study, we use NLST screening results classified into dichotomized Lung-RADS screening results as positive (Lung-RADS 3 or 4) vs negative (Lung-RADS 1 or 2). In NLST data, we evaluate the association of PLCOm2012 risk and 3 annual screening results with lung cancer risk (incidence) from 1 to 4 years after the last screen, and we develop and validate a lung cancer risk prediction model that describes these associations. Of interest was the association of screening results with future lung cancer risk to help guide future screening decision making.

    Methods
    Study Design

    The NLST (ClinicalTrials.gov identifier NCT00047385) design has been described.2,21 Briefly, the NLST was a randomized clinical screening trial comparing the effect of LDCT vs chest radiography screening on lung cancer mortality. Between August 2002 and April 2004, a total of 53 452 participants were enrolled and scheduled to receive screens at baseline (T0), 1 year (T1), and 2 years (T2). Participants were followed up until December 31, 2009. The present prediction modeling analysis took place between August 10, 2013, and November 1, 2018. Epidemiological data were collected at study entry by structured questionnaires. The NLST institutional review board approval was obtained at each participating center and the National Cancer Institute (NCI), and informed written consent was obtained from all participants. This study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)33 reporting guidelines.

    The NLST was a collaborative effort, including the Lung Screening Study (LSS) component (10 sites administered by an NCI Division of Cancer Prevention contract) and the American College of Radiology Imaging Network (ACRIN) component (23 sites administered through an NCI Division of Cancer Treatment and Diagnosis grant). The LSS sites enrolled 34 612 participants (64.8%), and ACRIN enrolled 18 840 participants (35.2%). An Annual Study Update (ASU), a self-reported questionnaire, updated participant information on smoking each year. The ACRIN and LSS components of the NLST collected ASU data differently, and we carried out a sensitivity analysis using LSS data only as part of this study.

    In the present secondary analysis of data from the NLST, screening results are based retrospectively on the Lung-RADS definition.20 A comparison of the NLST criteria and the Lung-RADS classification for a positive screen, which considers Lung-RADS 2 to be negative, in NLST data has been published,22 and we use identical coding. The present analysis included individuals who completed all 3 LDCT screens and were lung cancer free 1 year after the third screen (23 574 of 26 722 [88.2%]) (eFigure 1 in the Supplement).

    Statistical Analysis

    The present study evaluates the association of lung cancer screening results with subsequent lung cancer risk, adjusted for other important lung cancer risk predictors. The latter were estimated jointly by risks estimated using the PLCOm2012 model.6,8 Complete-cases multivariable fractional polynomial logistic regression models23 were used to predict lung cancers occurring 1 to 4 years after the last screen (T3 to T6) using PLCOm2012-estimated risks with and without screening results as predictors. Eight combinations of the dichotomous Lung-RADS screening results in the 3 annual screens were evaluated as predictors (described later in the Screening Findings and Prediction Models subsection of the Results section). Lung cancers occurring during the NLST screening period (T0 to T3) were excluded because we were interested in studying the additional information screening results provide on subsequent lung cancer risk beyond 3 screens, not whether a specific abnormal screen represented lung cancer. Also, lung cancers detected more than 1 year after the last screen were primarily clinically detected and avoid screening-associated overdiagnosis.

    Limited follow-up remained between the start of T3 and the trial’s completion. Most individuals did not have 6 years of follow-up from the start of T3, as is used by PLCOm2012, but did have 3.5 years of follow-up. For our study, follow-up was truncated at 6 years since randomization, providing 3 years of follow-up. A schema demonstrating the study design used for model preparation is shown in Figure 1. Cumulative incidences of lung cancer from T3 to T6 were estimated, stratified by screening results, taking competing risks into account.24

    Selected PLCOm2012 predictor variable data were updated from baseline to T3, this study’s follow-up onset. Age, smoking duration in baseline current smokers, and smoking quit time in baseline former smokers were updated to T3 by adding 3 years to the T0 baseline values. The PLCOm2012-estimated risks were converted from 6-year risk to 3-year risk by dividing by 2 to match this study’s follow-up period. The label PLCOm2012bu identifies these baseline-updated 3-year PLCOm2012 risks. This updating potentially introduces some error in smoking duration and quit time at T3 because some baseline current smokers may have stopped smoking by then and some baseline former smokers may have relapsed. The NLST LSS ASU information25 indicates that 24.1% of baseline current smokers had quit by T3; in the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial,26 smoking relapse occurred in 3.3% of baseline former smokers after 8.5 years of follow-up. In LSS data, a subset analysis was conducted updating smoking information using ASU information. These ASU-updated data did not improve prediction and are thus not used in the present analysis for consistency across the LSS and ACRIN data sets.

    The LSS data were used to develop a PLCOm2012bu-based model incorporating screening results, and ACRIN data were used to validate the model. The ACRIN data were used for validation instead of a random sampling of the entire NLST LDCT group because they replicate a true external validation sample having different geographies, populations, practices, and radiologists.

    Overall model prediction was assessed using the Brier score (zero indicates perfect prediction, and 0.25 indicates random prediction).27 Discrimination, the ability to correctly classify individuals who develop lung cancer, was assessed using the receiver operating characteristic area under the curve (AUC). Calibration, how closely predicted probabilities match observed probabilities, was assessed by estimating the 50th and 90th percentiles of absolute error (difference between the observed and predicted).28 Calibration was also evaluated by plotting the observed vs predicted risks and by evaluating the P value for the Spiegelhalter statistic (significance indicates some inconsistency). For Brier scores and AUCs, 95% CIs were estimated using bias-corrected percentile intervals in 1000 bootstrap resamplings.

    Decision curve analysis (DCA) comparing the net benefit of PLCOm2012bu plus screening results vs PLCOm2012bu only was carried out.29 This is described in the eAppendix in the Supplement.

    Regarding descriptive statistics, for tests of proportions and rates, 95% CIs and P values were prepared using methods described by Brown and colleagues30 and by Miettinen.31 Testing differences in continuous variables with normal distributions used t test, not assuming equal variances, in skewed continuous variables used nonparametric test of trend,32 and for proportions used the χ2 test or Fisher exact test.

    Statistical software (Stata 15.1 MP; StataCorp LP) was used for the analyses. Hypothesis testing used 2-sided P values, and an α error less than .05 determined statistical significance.

    Results

    Overall, 22 229 individuals (14 576 LSS and 7653 ACRIN) were studied. The mean (SD) age of participants was 61.3 (5.0) years, 59.3% were male, and 90.9% were of non-Hispanic white race/ethnicity. eFigure 1 and eTable 1 in the Supplement summarize study participant selection and characteristics, respectively. Overall, during follow-up from T3 to T6, a total of 298 lung cancers (202 LSS and 96 ACRIN) were diagnosed in the 22 229 individuals (1.3%) (14 576 in LSS and 7653 in ACRIN) with complete data for modeling. With a total cumulative incidence of 0.0134 (95% CI, 0.0119-0.0150) and with 64 921 person-years of follow-up, the incidence rate was 459 (95% CI, 408-514) per 100 000 person-years, including 474 (95% CI, 410-544) in LSS and 431 (95% CI, 349-526) in ACRIN (P = .45).

    Screening Findings and Prediction Models

    In LSS data, univariable and multivariable logistic regression models were developed predicting lung cancer 1 to 4 years after the last screen using Lung-RADS results with and without adjustment for PLCOm2012bu scores as predictors (Table 1). These models were prepared using Lung-RADS results in 8 groupings of dichotomous screen results over the 3 annual screens. In addition, based on odds ratio magnitudes and reasoning regarding the associations of multiple vs single positive results and number of individuals per result group, the screening Lung-RADS results were categorized into the following 4 groups (Table 1): (1) all 3 screens negative (− − −), (2) 1 positive screen and last negative (+ − − or − + −), (3) 2 positive screens with last negative or 2 negative screens with last positive (+ + − or − − +), and (4) at least 2 positive screens with last positive (+ − + or − + + or + + +). Adjusted for PLCOm2012 risks, compared with participants with 3 negative screens, participants with 1 positive screen and last negative had an odds ratio (OR) of 1.93 (95% CI, 1.34-2.76), and participants with 2 positive screens with last negative or 2 negative screens with last positive had an OR of 2.66 (95% CI, 1.60-4.43); when 2 or more screens were positive with last positive, the OR was 8.97 (95% CI, 5.76-13.97) (Table 1). The distributions of screening results classified in the 4 groups in LSS and ACRIN are listed in eTable 2 in the Supplement. The distributions of screening results and lung cancer rates were similar in LSS and ACRIN. As the screening results group increases in cumulative risk of lung cancer (0.9%, 2.0%, 3.6%, and 10.1% for groups 1 through 4 overall, respectively), the frequency of occurrence of each results group declined in LSS, ACRIN, and overall (Figure 2 and eTable 2 in the Supplement). Generally, lung cancer risk increased with the number and recency of positive screens.

    Considering first screens only, of the 22 167 NLST LDCT participants who had a Lung-RADS negative first screen, 264 (1.2%) had lung cancer diagnosed between T3 and T6. Of the 3156 individuals who had a positive first screen, 109 (3.5%) were diagnosed as having lung cancer in the same period, regardless of subsequent screening results.

    In LSS, the PLCOm2012bu risk-adjusted model with Lung-RADS results in 8 groupings had similar predictive performance as the model with results in 4 categories (Table 1). Overall prediction measured by the Brier score indicated good prediction for both models (0.013; 95% CI, 0.012-0.015 for both). Discrimination, measured by the AUC, was good for both models. For the adjusted 8-category risk model, the AUC was 0.772 (95% CI, 0.743-0.799); for the 4-category model, the AUC was 0.769 (95% CI, 0.741-0.797) (P = .07). The nonsignificant P values for Spiegelhalter statistics indicated good calibration for both models. In subsequent analyses, we used screening results in the 4 groupings and refer to the model that included PLCOm2012bu and 4-group Lung-RADS results as the PLCO2012results model.

    In ACRIN validation data, the PLCO2012results model had high overall prediction (Brier score, 0.012; 95% CI, 0.010-0.014), high discrimination (AUC, 0.761; 95% CI, 0.716-0.799), and good calibration as indicated by the nonsignificant Spiegelhalter statistic, as well as small 50th and 90th percentiles of absolute error (0.0018 and 0.0030, respectively). The observed and predicted mean probabilities were 1.25% and 1.49%, indicating that in the validation data the PLCO2012results model overestimates risk. This may in part be due to the ACRIN participants having a lower lung cancer incidence rate than LSS participants, in whom the model was developed. The plot of the observed and predicted probabilities (eFigure 2 in the Supplement) shows that the overestimation primarily occured above the 12.0% risk estimate, which is well above the range of risk thresholds considered for selecting individuals for lung cancer screening.8,34 Only 0.9% of the NLST LDCT sample had PLCO2012results risks exceeding 12.0%.

    In ACRIN validation data, compared with the PLCOm2012bu model, which excludes screening results, the PLCO2012results model demonstrated superior overall prediction (Brier score, 0.012 vs 0.013), discrimination (AUC, 0.761; 95% CI, 0.716-0.799 vs 0.687; 95% CI, 0.645-0.728) (P < .001) (eFigure 3 in the Supplement), and calibration (observed vs predicted, 0.84 vs 0.57). This analysis was also carried out applying the NLST criteria for positivity (≥4-mm nodules), and results were similar.

    The sensitivity, specificity, and positive predictive values of the PLCOm2012bu and PLCO2012results models for identifying lung cancers 1 to 4 years after the last screen using 0.8%, 1.0%, and 1.5% 3-year risk thresholds, stratified by LSS and ACRIN groups, are listed in eTable 3 in the Supplement. Those values indicate that the calibration of PLCOm2012bu and PLCO2012results differed. For example, the sensitivities, specificities, and positive predictive values for PLCO2012results with a threshold of 1.0% were similar to those of PLCOm2012bu with a threshold of 1.5% (85.4%, 48.9%, and 2.1% vs 84.4%, 47.2%, and 2.0%, respectively.

    Decision curve analysis demonstrated that between risk thresholds 0% to 6.0% the net benefit for model PLCO2012results exceeded that of PLCOm2012bu (eAppendix and eFigure 4 in the Supplement). In total, 99.1% of the overall NLST LDCT sample had PLCO2012results 3-year risks of 6.0% or less.

    Discussion
    Application of the PLCO2012results Model

    Lung cancer screening results may predict future lung cancer risk. The PLCO2012results model can provide information to help guide decision making regarding whether to undertake or omit the next annual screen. An easy-to-use spreadsheet calculator that estimates PLCO2012results risks for individuals with recent lung cancer screening results is available online (https://brocku.ca/lung-cancer-risk-calculator). For a given risk threshold used to select individuals for lung cancer screening, applying the PLCO2012results can lead to reassignment of PLCOm2012 selection classification.

    Having at least 1 positive screen was associated with increased PLCOm2012 risk. So, individuals who entered into the screening program based on a PLCOm2012 risk threshold would continue to have PLCO2012results model risks above the threshold and should continue with annual screening.

    Having consecutive negative screens was associated with a lower lung cancer risk than indicated by the PLCOm2012 model, which does not account for screening results. Eligible individuals who were close to the screening threshold at the first screen and had consecutive negative screening results may have had a lower postscreening estimated risk than the eligibility threshold and could omit screening for 1 year. In contrast, there is a PLCOm2012 ceiling risk level at or above which the PLCOm2012 risk could not drop below a selected screening eligibility threshold even after 1, 2, or 3 consecutive negative screens. Individuals at or above the ceiling probability should continue annual screening. For example, if the eligibility threshold is 1.5%, individuals with PLCOm2012 risks of at least 2.6% should continue to receive annual screens because negative screens cannot lower 6-year risks below 1.5%. If the PLCOm2012 enrollment threshold is 2.00%, individuals with PLCOm2012 risks of at least 3.4% should continue screening independent of the negative results. Individual assessments would be required for those with risks above the threshold but below the ceiling and who have had negative screens.

    In NLST, 92.1% of individuals with an initial negative screen had all 3 screens negative. Individuals with an initial negative screen who have PLCOm2012 risks below the ceiling probability for 1.5% or 2.0% enrollment thresholds have low 3-year lung cancer risks of 0.3% and 0.5%, respectively (Table 2), which might justify delaying the next screen by 1 year. In contrast, those individuals with initial negative screens who have PLCOm2012 risks above the ceiling probabilities are at high 3-year risks, 1.8% and 1.9%, warranting continued annual screening. With a 1.5% 6-year eligibility risk threshold, 56.0% of individuals with an initial negative screen should continue with annual screening because their baseline risks were at or above the ceiling threshold. Similarly, with a 2.0% threshold, 42.1% should continue with annual screening. Table 3 summarizes potential implications of the study findings of PLCOm2012 and PLCO2012results model on screening schedules.

    The NLST screening results were associated with postscreening lung cancer risk. We modified the PLCOm2012 risk prediction model to incorporate past screening results and predict subsequent 3-year lung cancer risk. The PLCO2012results model was associated with improved overall prediction, discrimination, and the net benefit compared with the analogous model excluding screening results (PLCOm2012).

    Patz and colleagues19 studied the association of a normal initial screen with incidence of lung cancer in the NLST. In contrast, the present study evaluated the association of all 3 NLST screens with risk (adjusted for other risk factors using a validated risk prediction model), excluded cancers detected during the screening period to develop accurately calibrated risk estimates, removed overdiagnosis bias, and developed a highly predictive risk model incorporating screening results.

    The findings by Patz and colleagues19 might suggest that high-risk individuals who have a normal initial screen may not need subsequent screening. Our findings show that individuals with a negative initial screen who have 1 or more subsequent positive screens can remain at substantially elevated risk (Table 1). Moreover, some individuals start with such elevated risks that repeated negative screens do not lower their risk estimates below eligibility thresholds. Lung cancer risks in those with initial negative screens are heterogeneous. Overall, 46.1% of individuals with a Lung-RADS T0-negative screen had PLCO2012results 3-year risks exceeding 1.0%, and 183 of 9016 of them (2.0%) were diagnosed as having lung cancer during T3 to T6. Overall, 55.2% (174 of 315) of lung cancers occurring during the follow-up period occurred in individuals who had 3 negative screens (eTable 2 in the Supplement). The decision whether to undertake or omit the next annual screening can be personalized by using a risk model incorporating lung cancer risk factors and screening results over more than 1 screening, when available. If selection of individuals for a biennial screen was properly done based on accurate risk estimation and appropriate low-risk threshold, then there would be minimal or no excess deaths expected (Table 3).

    Having 3 consecutive negative screens was associated with lower lung cancer risk below that estimated by PLCOm2012. However, for any given eligibility threshold used to select patients for screening, there is a corresponding ceiling probability of PLCOm2012 risk that is so high that 3 negative screens will not bring the results-adjusted PLCOm2012 score below the screening eligibility threshold. Individuals who have an initial negative screen who have PLCOm2012 scores above the ceiling will remain at high enough risk that they should be offered annual screening. The PLCO2012results calculator can help determine the ceiling for a screening eligibility threshold and thus can help guide decision making after a negative screen. Potential implications of the study findings on screening interval are summarized in Table 3.

    Applying the PLCO2012results for planning screening regimens can lead to revision of decisions based on the PLCOm2012. Screening decision based on PLCO2012results estimates should be given priority over the parallel PLCOm2012 estimates because of more accurate predictions.

    The PLCO2012results model can also be used for selection of high-risk samples for research purposes, such as evaluation of biomarkers for early detection and prevention trials. The mechanism explaining the association between screening results and increased risk is unknown. One possible explanation is that positive screens are a marker of lung inflammation, which is recognized to be associated with lung carcinogenesis.35

    Strengths and Limitations

    Strengths of the study include that the findings are based on prospective data. In addition, the results represent outcomes from a large, well-conducted trial (the NLST2) that focused on lung cancer screening.

    An apparent study limitation of this study might be that lung cancer risks were estimated after 3 screenings, not just 1 or 2. This study found that conclusions drawn after 1 screening are simplifications that do not hold in all circumstances. Moreover, the 20% lung cancer mortality reduction observed in the NLST2 was estimated for a 3-screen program, so this seems the natural point at which to evaluate if further screens are warranted accounting for previous screening results.

    The findings of the present study need to be validated in prospective, high-quality data in different settings. It needs to be determined if subsequent screen-detected lung cancers that follow previous positive screens are predominantly early stage. The mechanism that leads to increased risk in those with false-positive (non–lung cancer) screens needs explanation. The cost-effectiveness of using PLCO2012results to guide screening decisions requires evaluation.

    Conclusions

    The study findings are expected to have clinical and public health implications. Including screening result information was associated with improved lung cancer risk prediction. A primary application of the PLCO2012results model is to help guide the decision when to undertake the next screen once an individual has participated in lung cancer screening. In a cost-effectiveness analysis of lung cancer screening in Ontario, Canada, the primary driver of costs was the computed tomography examinations.36 The use of PLCO2012results may lead to fewer screens, which may improve cost-effectiveness and reduce radiation exposure and other screening harms, such as false-positive findings and overdiagnosis.

    Back to top
    Article Information

    Accepted for Publication: January 13, 2019.

    Published: March 1, 2019. doi:10.1001/jamanetworkopen.2019.0204

    Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2019 Tammemägi MC et al. JAMA Network Open.

    Corresponding Author: Martin C. Tammemägi, PhD, Department of Health Sciences, Brock University, 1812 Sir Isaac Brock Way, Walker Complex S, Room 306, St Catharines, ON L2S 3A1, Canada (martin.tammemagi@brocku.ca).

    Author Contributions: Dr Tammemägi had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

    Concept and design: Tammemägi, ten Haaf, Toumazis, Han, Meza.

    Acquisition, analysis, or interpretation of data: All authors.

    Drafting of the manuscript: Tammemägi, Toumazis, Kong, Commins.

    Critical revision of the manuscript for important intellectual content: Tammemägi, ten Haaf, Kong, Han, Jeon, Riley, Meza.

    Statistical analysis: Tammemägi, ten Haaf, Han.

    Administrative, technical, or material support: Tammemägi, Commins, Riley, Meza.

    Supervision: Kong, Meza.

    Conflict of Interest Disclosures: Dr Tammemägi reported developing the PLCOm2012 risk prediction model (free to all noncommercial users) and reported that Brock University owns the rights to sublicense use of the PLCOm2012 to commercial users who financially profit from the use of the PLCOm2012 model (parts of those proceeds are to come to Dr Tammemägi; to date, zero financial returns have been made to Dr Tammemägi regarding any use of the PLCOm2012). Dr ten Haaf reported receiving grants from the National Institutes of Health/National Cancer Institute; receiving nonfinancial support from the International Association for the Study of Lung Cancer strategic screening advisory committee; receiving grants from Sunnybrook Health Sciences, Toronto, Ontario, Canada; receiving grants from the University of Zurich, Switzerland; receiving grants and nonfinancial support from the Dutch-Belgian Lung Cancer Screening (NELSON) trial; being a member of the Cancer Intervention and Surveillance Modeling Network (CISNET) lung working group; currently working on a number of studies regarding the cost-effectiveness of lung cancer screening and the effectiveness of risk-based lung cancer screening; and being a member of the Dutch-Belgian Lung Cancer Screening (NELSON) trial. Dr Meza reported receiving grants from the National Institutes of Health. No other disclosures were reported.

    Funding/Support: Although the National Cancer Institute funded the National Lung Screening Trial (NLST), there were no funding sources for the present ancillary study. Drs ten Haaf, Toumazis, Kong, Han, Jeon, and Meza acknowledge support from lung grant U01CA199284 from the Cancer Intervention and Surveillance Modeling Network (CISNET). Dr ten Haaf is affiliated with the Dutch-Belgian Lung Cancer Screening (NELSON) trial.

    Role of the Funder/Sponsor: The funder/sponsor had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

    Disclaimer: The statements contained herein are solely those of the authors and do not represent or imply concurrence or endorsement by the National Cancer Institute.

    Additional Contributions: We thank the National Cancer Institute for access to their data collected by the National Lung Screening Trial (NLST). We thank the NLST screening center investigators and staff members, as well as the staffs of Information Management Services, Inc and Westat, Inc. We thank the trial participants for their contributions that made this study possible.

    References
    1.
    Brambilla  E, Travis  W. Lung cancer. In: Stewart  B, Wild  C, eds.  World Cancer Report 2014. Lyon, France: International Agency for Research on Cancer, World Health Organization; 2014:350-361.
    2.
    Aberle  DR, Adams  AM, Berg  CD,  et al; National Lung Screening Trial Research Team.  Reduced lung-cancer mortality with low-dose computed tomographic screening.  N Engl J Med. 2011;365(5):395-409. doi:10.1056/NEJMoa1102873PubMedGoogle ScholarCrossref
    3.
    Moyer  VA; U.S. Preventive Services Task Force.  Screening for lung cancer: U.S. Preventive Services Task Force recommendation statement.  Ann Intern Med. 2014;160(5):330-338. doi:10.7326/M13-2771PubMedGoogle ScholarCrossref
    4.
    Centers for Medicare & Medicaid Services (CMS). Decision memo for screening for lung cancer with low dose computed tomography (LDCT) (CAG-00439N). https://www.cms.gov/medicare-coverage-database/details/nca-decision-memo.aspx?NCAId=274&NcaName=Screening+for+Lung+Cancer+with+Low+Dose+Computed+Tomography+(LDCT)&MEDCACId=68&IsPopup=y&bc=AAAAAAAAAgAAAA%3d%3d&. Published February 5, 2015. Accessed August 31, 2015.
    5.
    Bach  PB, Mirkin  JN, Oliver  TK,  et al.  Benefits and harms of CT screening for lung cancer: a systematic review.  JAMA. 2012;307(22):2418-2429. doi:10.1001/jama.2012.5521PubMedGoogle ScholarCrossref
    6.
    Tammemägi  MC, Katki  HA, Hocking  WG,  et al.  Selection criteria for lung-cancer screening.  N Engl J Med. 2013;368(8):728-736. doi:10.1056/NEJMoa1211776PubMedGoogle ScholarCrossref
    7.
    Kovalchik  SA, Tammemägi  M, Berg  CD,  et al.  Targeting of low-dose CT screening according to the risk of lung-cancer death.  N Engl J Med. 2013;369(3):245-254. doi:10.1056/NEJMoa1301851PubMedGoogle ScholarCrossref
    8.
    Tammemägi  MC, Church  TR, Hocking  WG,  et al.  Evaluation of the lung cancer risks at which to screen ever- and never-smokers: screening rules applied to the PLCO and NLST cohorts.  PLoS Med. 2014;11(12):e1001764. doi:10.1371/journal.pmed.1001764PubMedGoogle ScholarCrossref
    9.
    Katki  HA, Kovalchik  SA, Berg  CD, Cheung  LC, Chaturvedi  AK.  Development and validation of risk models to select ever-smokers for CT lung cancer screening.  JAMA. 2016;315(21):2300-2311. doi:10.1001/jama.2016.6255PubMedGoogle ScholarCrossref
    10.
    Cressman  S, Peacock  SJ, Tammemägi  MC,  et al.  The cost-effectiveness of high-risk lung cancer screening and drivers of program efficiency.  J Thorac Oncol. 2017;12(8):1210-1222. doi:10.1016/j.jtho.2017.04.021PubMedGoogle ScholarCrossref
    11.
    Li  K, Hüsing  A, Sookthai  D,  et al.  Selecting high-risk individuals for lung cancer screening: a prospective evaluation of existing risk models and eligibility criteria in the German EPIC cohort.  Cancer Prev Res (Phila). 2015;8(9):777-785. doi:10.1158/1940-6207.CAPR-14-0424PubMedGoogle ScholarCrossref
    12.
    Weber  M, Yap  S, Goldsbury  D,  et al.  Identifying high risk individuals for targeted lung cancer screening: independent validation of the PLCOm2012 risk prediction tool.  Int J Cancer. 2017;141(2):242-253. doi:10.1002/ijc.30673PubMedGoogle ScholarCrossref
    13.
    ten Haaf  K, Jeon  J, Tammemägi  MC,  et al.  Risk prediction models for selection of lung cancer screening candidates: a retrospective validation study.  PLoS Med. 2017;14(4):e1002277. doi:10.1371/journal.pmed.1002277PubMedGoogle ScholarCrossref
    14.
    Tammemägi  MC, Schmidt  H, Martel  S,  et al; PanCan Study Team.  Participant selection for lung cancer screening by risk modelling (the Pan-Canadian Early Detection of Lung Cancer [PanCan] study): a single-arm, prospective study.  Lancet Oncol. 2017;18(11):1523-1531. doi:10.1016/S1470-2045(17)30597-1PubMedGoogle ScholarCrossref
    15.
    Katki  HA, Kovalchik  SA, Petito  LC,  et al.  Implications of nine risk prediction models for selecting ever-smokers for computed tomography lung-cancer screening.  Ann Intern Med. 2018;169(1):10-19. doi:10.7326/M17-2701PubMedGoogle ScholarCrossref
    16.
    Kavanagh  J, Liu  G, Menezes  R,  et al.  Importance of long-term low-dose CT follow-up after negative findings at previous lung cancer screening.  Radiology. 2018;289(1):218-224. doi:10.1148/radiol.2018180053PubMedGoogle ScholarCrossref
    17.
    Yousaf-Khan  U, van der Aalst  C, de Jong  PA,  et al.  Risk stratification based on screening history: the NELSON lung cancer screening study.  Thorax. 2017;72(9):819-824. doi:10.1136/thoraxjnl-2016-209892PubMedGoogle ScholarCrossref
    18.
    Maisonneuve  P, Bagnardi  V, Bellomi  M,  et al.  Lung cancer risk prediction to select smokers for screening CT: a model based on the Italian COSMOS trial.  Cancer Prev Res (Phila). 2011;4(11):1778-1789. doi:10.1158/1940-6207.CAPR-11-0026PubMedGoogle ScholarCrossref
    19.
    Patz  EF  Jr, Greco  E, Gatsonis  C, Pinsky  P, Kramer  BS, Aberle  DR.  Lung cancer incidence and mortality in National Lung Screening Trial participants who underwent low-dose CT prevalence screening: a retrospective cohort analysis of a randomised, multicentre, diagnostic screening trial.  Lancet Oncol. 2016;17(5):590-599. doi:10.1016/S1470-2045(15)00621-XPubMedGoogle ScholarCrossref
    20.
    American College of Radiology. Lung CT Screening Reporting & Data System (Lung-RADS). http://www.acr.org/Quality-Safety/Resources/LungRADS. Accessed April 27, 2017.
    21.
    Aberle  DR, Berg  CD, Black  WC,  et al; National Lung Screening Trial Research Team.  The National Lung Screening Trial: overview and study design.  Radiology. 2011;258(1):243-253. doi:10.1148/radiol.10091808PubMedGoogle ScholarCrossref
    22.
    Pinsky  PF, Gierada  DS, Black  W,  et al.  Performance of Lung-RADS in the National Lung Screening Trial: a retrospective assessment.  Ann Intern Med. 2015;162(7):485-491. doi:10.7326/M14-2086PubMedGoogle ScholarCrossref
    23.
    Royston  P, Sauerbrei  W.  Multivariable Model-Building: A Pragmatic Approach to Regression Analysis Based on Fractional Polynomials for Modelling Continuous Variables. Hoboken, NJ: John Wiley & Sons; 2008. doi:10.1002/9780470770771
    24.
    Fine  JP, Gray  RJ.  A proportional hazards model for the subdistribution of a competing risk.  J Am Stat Assoc. 1999;94(446):496-509. doi:10.1080/01621459.1999.10474144Google ScholarCrossref
    25.
    Tammemägi  MC, Berg  CD, Riley  TL, Cunningham  CR, Taylor  KL.  Impact of lung cancer screening results on smoking cessation.  J Natl Cancer Inst. 2014;106(6):dju084. doi:10.1093/jnci/dju084PubMedGoogle ScholarCrossref
    26.
    Barry  SA, Tammemägi  MC, Penek  S,  et al.  Predictors of adverse smoking outcomes in the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial.  J Natl Cancer Inst. 2012;104(21):1647-1659. doi:10.1093/jnci/djs398PubMedGoogle ScholarCrossref
    27.
    Woodward  M.  Epidemiology: Study Design and Data Analysis. 3rd ed. Boca Raton, FL: CRC Press; 2014.
    28.
    Harrell  FE Jr.  Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. 2nd ed. Basel, Switzerland: Springer Nature Switzerland AG; 2015. doi:10.1007/978-3-319-19425-7
    29.
    Vickers  AJ.  Decision analysis for the evaluation of diagnostic tests, prediction models and molecular markers.  Am Stat. 2008;62(4):314-320. doi:10.1198/000313008X370302PubMedGoogle ScholarCrossref
    30.
    Brown  LD, Cai  TT, DasGupta  A.  Interval estimation for a binomial proportion.  Stat Sci. 2001;16:101-133. doi:10.1214/ss/1009213286Google Scholar
    31.
    Miettinen  O.  Estimability and estimation in case-referent studies.  Am J Epidemiol. 1976;103(2):226-235. doi:10.1093/oxfordjournals.aje.a112220PubMedGoogle ScholarCrossref
    32.
    Cuzick  J.  A Wilcoxon-type test for trend.  Stat Med. 1985;4(1):87-90. doi:10.1002/sim.4780040112PubMedGoogle ScholarCrossref
    33.
    Collins  GS, Reitsma  JB, Altman  DG, Moons  KG.  Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement.  Ann Intern Med. 2015;162(1):55-63. doi:10.7326/M14-0697PubMedGoogle ScholarCrossref
    34.
    Field  JK, Duffy  SW, Baldwin  DR,  et al.  The UK Lung Cancer Screening trial: a pilot randomised controlled trial of low-dose computed tomography screening for the early detection of lung cancer.  Health Technol Assess. 2016;20(40):1-146. doi:10.3310/hta20400PubMedGoogle ScholarCrossref
    35.
    Gomes  M, Teixeira  AL, Coelho  A, Araújo  A, Medeiros  R.  The role of inflammation in lung cancer.  Adv Exp Med Biol. 2014;816:1-23. doi:10.1007/978-3-0348-0837-8_1PubMedGoogle ScholarCrossref
    36.
    ten Haaf  K, Tammemägi  MC, Bondy  SJ,  et al.  Performance and cost-effectiveness of computed tomography lung cancer screening scenarios in a population-based setting: a microsimulation modeling analysis in Ontario, Canada.  PLoS Med. 2017;14(2):e1002225. doi:10.1371/journal.pmed.1002225PubMedGoogle ScholarCrossref
    ×