Risk-Targeted Lung Cancer Screening

The National Lung Screening Trial (NLST) demonstrated a 20% reduction in lung cancer deaths among persons screened with low-dose computed tomography (LDCT) versus chest radiography (1). The study participants were between the ages of 55 and 74 years and had a smoking history of at least 30 pack-years; former smokers had no more than 15 years of smoking abstinence (1). Most subsequent guidelines, including those from the U.S. Preventive Services Task Force (USPSTF), have largely adopted the NLST eligibility criteria when recommending who should receive lung cancer screening (24). A retrospective analysis of the NLST showed that 60% of participants at highest predicted risk for lung cancer death accounted for 88% of all screening-prevented lung cancer mortality (5). Other influential studies also have suggested that using several risk factors to selected participants with the highest predicted risk for either having or dying of lung cancer would greatly improve the efficiency and yield of lung cancer screening (610). A general limitation of these studies is that the benefits of screening with LDCT is measured in terms of reduced lung cancer mortality over the short term, generally the first 5 to 7 years per patient screened; these studies do not account for differences in long-term survival or costs in higher- versus lower-risk patients. In this analysis, we applied a novel multistate model to calculate the predicted lifetime benefits and costs of screening with LDCT versus chest radiography for each NLST participant. We used these estimates to examine the value of applying an individualized risk-targeted approach to selecting participants for screening compared with the broader NLST inclusion criteria at common willingness-to-pay (WTP) thresholds. Methods Model Overview Using data from the NLST (1), we derived a multistate regression model (11, 12) to predict health state transitions as a function of each participant's baseline characteristics and the lung cancer screening technology used (chest radiography or LDCT). The rationale for using a multistate model is described in the Appendix. This technique jointly estimates the baseline hazards and the effects of risk factors on 4 transitions among the 4 health states shown in Figure 1: alive without cancer, alive with lung cancer, dead as a result of other causes, and dead as a result of lung cancer. The resulting equation provides individualized predictions, under each screening condition, of the probability of being in each of the 4 states at any given time point. These predictions were used to risk stratify the population as well as estimate individualized lung cancer mortality benefits of LDCT screening at 7 years. To estimate individualized lifetime health benefits, extrapolations of the baseline hazards were used to project the probability of being in each of the 4 states beyond 7 years. Individualized costs were estimated by using utilization data from the NLST and linear regression prediction models combined with assumptions to estimate lifetime medical costs. For the purposes of presentation, we stratified participants into deciles (10% cohorts) on the basis of their 7-year risk for lung cancer mortality when screened with chest radiography. Finally, we calculated the incremental net monetary benefit (iNMB) for each participant to summarize the value of a risk-stratified screening strategy compared with the NLST inclusion criteria at common WTP thresholds of $50000 and $100000 per quality-adjusted life-year (QALY) (13, 14). Figure 1. Structure of the multistate model. All patients enter the multistate model as high-risk smokers defined by the NLST entry criteria. The model consists of 4 health states with 4 possible transitions. The figure describes the number of illnessdeath multistate transitions during the trial period. NLST = National Lung Screening Trial. Model Derivation and Validation The NLST randomly assigned 53454 participants to have 3 consecutive annual screenings from randomization (years 0, 1, and 2) with either LDCT (n= 26722) or the control screening method of chest radiography (n= 26732). Median follow-up was 6.5 years. Our multistate regression model included 8 baseline variables for each NLST participant (age, sex, race, family history in a first-degree relative, body mass index, smoking exposure [pack-years], years since smoking cessation, and self-reported history of emphysema). We selected these variables on the basis of a previously published model (5). We excluded 368 NLST participants for whom data were missing for any of these variables. Because the transition from smoker to lung cancer did not satisfy the proportional hazard assumption (as the difference in lung cancer diagnoses between screening conditions occurred mostly within the first year after randomization), participants were stratified on initial screening assignment. The proportional hazards assumption also was checked for all other included variables over each of the 4 transitions by examining the correlation of the Schoenfeld residuals with time. Proportionality in the 7-year follow-up data generally was not rejected, nor was any evidence found of a consistent nonsignificant trend for diminishing or strengthening predictor effects over time. To evaluate model performance, we constructed calibration plots under each screening condition for the observed proportions of outcome events against the predicted risks for individuals grouped by quintiles (20% cohorts) of risk. We used the concordance statistic (c-statistic) to assess model discrimination. All analyses were performed by using R, version 3.3.2 (The R Foundation), and SAS, version 9.3 (SAS Institute). Benefits We estimated benefits for each participant as the incremental model-projected health outcomes after LDCT compared with chest radiography screening. We used the multistate model to project potential outcomes for each subject through year 7 after randomization. We extrapolated subsequent survival conditional on survival to year 7 by using a Weibull distribution (Appendix Figure 1). Health outcomes included life expectancy and quality-adjusted life expectancy under each screening condition for each person. We computed the expected QALYs as the sum of each health state's contribution. Each state's contribution was equal to the product of the expected years lived in that state and the state's utility weight. The utility weights used were 1.0 (equivalent to 1 year of perfect health) for the initial state of being alive without lung cancer, 0.77 for alive with lung cancer (the average utility for metastatic and nonmetastatic lung cancer reported in a meta-analysis [15]), and 0.0 (equivalent to death). Appendix Figure 1. Cumulative baseline hazard of extrapolated multistate model and sensitivity analysis for transition 4. CT = computed tomography; LDCT = low-dose computed tomography. Top and Middle. Cumulative baseline hazard of extrapolated multistate model. Healthy to cancer (transition 1): The transition from healthy smoker to cancer did not satisfy the proportional hazard assumptions. To model this transition, the observed cumulative hazard by LDCT and chest radiography was used up to year 7 after randomization. A Weibull extrapolation, stratified by LDCT and chest radiography, was used beyond year 7 after randomization. When the cumulative hazard of a diagnosis of lung cancer by chest radiography exceeded that of LDCT, a Weibull extrapolation based on screening with chest radiography alone was used. Healthy to other death (transition 2): The transition from healthy smoker to death from nonlung cancerrelated causes was modeled by using the observed cumulative hazard up to year 7 after randomization, including a proportional CT-versus-radiography effect. A Weibull extrapolation, with a doubled shape parameter, based on the cumulative hazard of death from nonlung cancerrelated causes among patients screened with chest radiography was used beyond year 7 after randomization. Cancer to other death (transition 3): The transition from lung cancer to death from nonlung cancerrelated causes was modeled by using the observed cumulative hazard up to year 7 after randomization, including a proportional CT-versus-radiography effect. A Weibull extrapolation based on the cumulative hazard of death from nonlung cancerrelated causes among patients screened with chest radiography alone was used beyond year 7 after randomization. Cancer to cancer death (transition 4): The transition from lung cancer to death from lung cancer was modeled by using the observed cumulative hazard up to year 7 after randomization, including a proportional CT-versus-radiography effect. A Weibull extrapolation based on the cumulative hazard of death from lung cancerrelated causes among patients screened with chest radiography alone was used beyond year 7 after randomization. Bottom. Sensitivity analysis for transition 4. Our base case assumes that the lung cancer mortality benefits seen with LDCT screening are present only during the median follow-up of the trial, that is, 7 years. The hazard ratio of death from lung cancer in the subsequent Weibull extrapolation is the same as that in patients screened with chest radiography. Our sensitivity analysis tested this assumption by extrapolating ongoing improved survival of screening with LDCT past year 7. Costs Cost components included the initial screening; medical care use, including diagnosis and management after a positive screening result; and age-stratified background medical costs (Appendix Tables 1 and 2) (14). We estimated initial screening costs and costs associated with a positive screening as the product of utilization data and unit prices (by using 2016 Medicare reimbursement [16]). Utilization estimates came directly from the NLST, which recorded diagnostic and procedure codes pertinent to the screening, diagnosis, and treatment of lung cancer for each participant. Our analysis focused on co