Population impact of lung cancer screening in the United States: Projections from a microsimulation model

Background Previous simulation studies estimating the impacts of lung cancer screening have ignored the changes in smoking prevalence over time in the United States. Our primary rationale was to perform, to our knowledge, the first simulation study that estimates the health outcomes of lung cancer screening with explicit modeling of smoking trends for the whole US population. Methods/Findings Utilizing a well-validated microsimulation model, we estimated the benefits and harms of an annual low-dose computed tomography screening scenario with a realistic screening adherence rate versus a no-screening scenario for the US population from 2016–2030. The Centers for Medicare and Medicaid Services (CMS) eligibility criteria were applied: age 55–77 years at time of screening, history of at least 30 pack-years of smoking, and current smoker or former smoker with fewer than 15 years since quitting. In the screened population, cumulative mortality reduction was projected to reach 16.98% (95% CI 16.90%–17.07%). Cumulative mortality reduction was estimated to be 3.52% (95% CI 3.50%–3.53%) for the overall study population, with annual mortality reduction peaking at 4.38% (95% CI 4.36%–4.41%) in 2021 and falling to 3.53% (95% CI 3.50%–3.56%) by 2030. Lung cancer screening would save a projected 148,484 life-years (95% CI 147,429–149,540) across the total population through 2030. There were estimated to be 9,054 (95% CI 9,011–9,098) overdiagnosed cases among the 252,429 (95% CI 251,208–253,649) screen-detected lung cancer diagnoses, yielding an overdiagnosis rate of 3.59%. The limitations of our study are that we do not explicitly model race or socioeconomic status and our model was calibrated to data from studies performed in academic centers, both of which may impact the generalizability of our results. We also exclusively model the effects of the CMS guidelines for lung cancer screening and not any other screening strategies. Conclusions The mortality reduction and life-years gained estimated by this study are lower than those of single birth cohort studies. Single cohort studies neglect the changing dynamics of smoking behavior across generations, whereas this study reflects the trend of decreasing smoking prevalence since the 1960s. Maximum benefit could be derived from lung cancer screening through 2021; in later years, mortality reduction due to screening will decline. If a comprehensive screening program is not implemented in the near future, the opportunity to achieve these benefits will have passed.


Smoking History Generator
Additional detail on the Cancer Intervention and Surveillance Modeling Network's smoking history generator 3

Background
The National Cancer Institute's Cancer Intervention and Surveillance Modeling Network developed the smoking history generator (SHG) in order to provide smoking history and other cause mortality inputs for lung cancer models. The smoking history generator provides cohortspecific smoking histories, as well as other cause death rates to be used in simulation modeling analyses of lung cancer interventions, including screening and tobacco control. The SHG has been used to estimate the impact of tobacco control on U.S. smoking-related mortality since the publication of the 1964 Surgeon General's Report [1] and to estimate the health effects of raising the minimum purchase and sale age of tobacco products in the U.S. [2].

Determination of Parameters
The methods for deriving the model parameters and extrapolating the SHG's output to calendar year 2030 have been previously described [3]. The data sources of the SHG include the National Health Interview Survey (NHIS), the Human Mortality Database (HMD) life tables, and the Cancer Prevention Study (CPS)-I and II [3-6].
The parameters in the SHG were estimated by fitting the data sources with an age, period, and cohort model. The smoking data from the NHIS was smoothed to provide information by gender, age, and year on the mean and variability in smoking status (never, current, and former by years quit), intensity, and duration. The HMD life tables were used to derive life tables for other-cause mortality associated with specific levels of smoking. The smoking related parameters, including initiation and cessation rates, were fitted using generalized linear models. The cigarettes smoked per day were fitted using a cumulative logistic model with constrained splines for temporal effects. To extrapolate the smoking prevalence and intensity to year 2030, the SHG used the latest available data of birth cohort in 1979 and kept all model parameters constant.

Lung Cancer Policy Model
Additional detail on the Lung Cancer Policy Model 4

Background
The LCPM is a Monte Carlo microsimulation model of lung cancer development, detection, and treatment coded in C++. Simulated patients progress through five potential model states: 1) General Population, 2) Follow-Up, 3) Diagnosis & Staging, 4) Treatment, and 5) Death.

Overview of Process
Initially, the LCPM is populated with healthy, disease-free individuals who enter into the aforementioned model states based on monthly transition probabilities. The likelihood of developing lung cancer is tied to an individual's smoking history, while non-smokers are also capable of developing lung cancer from other causes. During each simulated month, individuals may develop lung cancer, experience growth of an existing cancer, or develop metastases and other symptoms.
Once a patient has developed lung cancer, it can be detected through three different means: clinical presentation due to symptoms, incidental imaging, or low-dose CT screening. Screendetected lung cancers will experience different behavior. Next, patients receive diagnostic testing and are evaluated for staging, possibly undergoing treatment based on the outcomes of those activities.

Simulating Lung Cancer Development and Progression
An individual can develop a maximum of three cancers and monthly probabilities of developing lung cancers are estimated using an independent logistic equation based on cancer cell type. For each histologic type, the logistic equation utilizes specific intercept, cell type-specific coefficients for age, age 2 , years of cigarette exposure (smoke-years, SY), an interaction term between SY and age 2 , average number of cigarettes smoked per day, and years since quitting smoking. Growth of malignant tumors is governed by a Gompertz function where doubling times decrease with an increase in tumor size. Cancer progression over time is modeled to include nodal involvement and/or distant metastases.

Model Calibration and Validation
Potential lung cancer cell types include non-invasive and invasive adenocarcinoma, large cell, squamous cell, small cell, and other. Natural history parameters were obtained from de-identified data from the National Lung Screening Trial (NLST) and Prostate, Lung, Colorectal and Ovarian (PLCO) screening trial participants, with outputs calibrated and validated to these studies [7]. A detailed description of the original LCPM is publicly available, as recorded within a designated National Cancer Institute website (http://www.cisnet.cancer.gov/lung/profiles.html).

Treatment & Survival
Death General Population