A Risk-Based Framework for Assessing Real-time Lung Cancer Screening Eligibility That Incorporates Life-Expectancy and Past Screening Findings

Abstract

Background:

Current lung cancer risk-based screening approaches use a single risk-threshold, disregard life-expectancy, and ignore past screening findings. We address these limitations with a comprehensive analytical framework, the individualizEd luNG cAncer screeninG dEcision (ENGAGE) tool, that aims to optimize lung cancer screening for U.S. ever-smokers under dynamic risk assessment by incorporating life expectancy and past screening findings over time.

Methods:

ENGAGE employs a partially observable Markov decision process framework that integrates published risk prediction and disease progression models, to dynamically assess the trade-off between the expected health benefits and harms associated with screening. ENGAGE evaluates lung cancer risk annually and provides real-time screening eligibility that maximizes the expected quality-adjusted life-years (QALYs) of ever-smokers. We compare ENGAGE against the 2013 U.S. Preventive Services Task Force (USPSTF) lung cancer screening guideline and single-threshold risk-based screening paradigms.

Results:

Compared to the 2013 USPSTF guidelines, ENGAGE expands screening coverage among ever-smokers (ENGAGE: 78%, USPSTF: 61%) while reducing the number of screening exams per person (ENGAGE:10.43, USPSTF:12.07, P<0.001), yields higher effectiveness in terms of increased lung cancer-specific mortality reduction (ENGAGE:19%, USPSTF:15%, P<0.001) and improves screening efficiency (ENGAGE:696, USPSTF:819 screens per death avoided, P<0.001). When compared against a single-threshold risk-based screening strategy, ENGAGE increases QALY requiring 30% fewer screens per death avoided (ENGAGE:696, single-threshold:889, P<0.001), and reduces false-positives by 40%.

Conclusion(s):

ENGAGE provides a comprehensive framework for dynamic risk-based assessment of lung cancer screening eligibility by incorporating life-expectancy and past screening findings that can serve to guide future policies on the effectiveness and efficiency of screening.

Lay Summary

We develop a novel decision-analytical screening framework for lung cancer, the individualizEd luNG cAncer screeninG dEcision (ENGAGE) tool to provide personalized screening schedules for ever-smokers. ENGAGE captures the dynamic nature of lung cancer risk and incorporates life expectancy into the screening decision-making process. ENGAGE integrates past screening findings and changes in smoking behavior of individuals and provides informed screening decisions that outperform existing screening guidelines and single-threshold risk-based screening approaches. We find that a personalized lung cancer screening program facilitated by a tool such as ENGAGE could enhance the efficiency of lung cancer screening.

Precis

We developed a comprehensive framework, the individualizEd luNG cAncer screeninG dEcision (ENGAGE), that delivers personalized lung cancer screening schedules. ENGAGE changes the lung cancer screening paradigm from population-level to individual-level screening decisions and can serve to guide future policies on the effectiveness and efficiency of screening.

Introduction

Lung cancer remains the leading cause of cancer related death in the United States.¹ Existing lung cancer screening guidelines define screening eligibility based on age and smoking history, using an annual screening interval for all eligible individuals.^2–4 The 2013 U.S. Preventive Services Task Force (USPSTF) guideline, which is actively being reconsidered,⁵ recommends annual screening with low-dose computed tomography (LDCT) for smokers aged between 55–80 years, who have at least 30 pack-years smoking history, and currently smoke or are within 15 years from smoking cessation.² However, a growing number of experts are advocating for risk-based lung cancer screening, arguing that selecting individuals based on their personal lung cancer risk improves the effectiveness of screening.^4,6,7 Current risk-based programs define screening eligibility using a single, model-specific risk threshold (hereafter called “single-threshold risk-based screening”),⁸ with the individual risk assessment ascertained by a variety of lung cancer risk calculators.^9–13 Single-threshold risk-based screening has been shown to yield higher mortality reduction than the USPSTF guideline, but prescribes screening to an older population, and thereby is associated with increased overdiagnosis and marginally increased life-years gained.⁸

At issue is that single-threshold risk-based programs select individuals solely based on their lung cancer risk and ignores individuals’ life expectancy. Although clinicians may consider life expectancy when making screening recommendations to patients, the development of guidelines that define screening eligibility based on the lung cancer risk and life expectancy of individuals needs to be formalized. Augmenting screening eligibility criteria with life-expectancy and developing a comprehensive risk calculation that consolidates prescreening risk and past screening findings could improve the effectiveness of lung cancer screening.^14–17 Personalized risk-based lung cancer screening that dynamically incorporates life-expectancy, which is adaptive upon acquisition of screening findings has been regarded as the most effective and efficient screening approach.^18,19 We present an analytical approach and initial results for such a personalized risk-based screening framework that weighs the benefits and potential harms associated with lung cancer screening and makes decisions in an uncertain environment.

We apply principles from Markov decision processes to introduce a comprehensive framework that provides analytically optimal personalized risk-based lung cancer screening schedules for U.S. ever-smokers. Using our comprehensive patient-centered lung cancer screening framework, individual screening eligibility is assessed annually, depends on the individual’s age-specific lung cancer risk, and is informed by: (i) life-expectancy, (ii) past screening exams, and (iii) smoking history. In this study, we provide the analytical framework for a personalized lung cancer screening program and assess the potential net effects in benefits and harms from changing the lung cancer screening paradigm from population-level to individual-level screening decisions.

Methods

Overview and Study Design

We develop the individualizEd luNG cAncer screeninG dEcision (ENGAGE) tool to identify optimal, personalized, real-time screening eligibility decisions for ever-smokers age 50 and above. ENGAGE formulates the decision of whether or not to screen an individual at a given age as a partially observable Markov decision process (POMDP) – a decision-analytical method, which has been applied to optimize screening programs for other cancer sites.^20–23 A POMDP provides a framework well-suited to optimize cancer screening decisions because it weighs the benefits and harms of screening, while accounting for the uncertainty stemming from the unobservable health state of individuals, unknown future health trajectory, and imperfect information collected from screening modalities.^24,25 The ENGAGE screening decisions (i.e. “Screen” or “No Screen”) are based on a real-time assessment of an individual’s health state, lung cancer risk, and life-expectancy.

The health state of individuals contains information about their personal lung cancer-related state and smoking status (Supplementary Figure 1). At any given age, each individual can exhibit one of four smoking behaviors: former smoker, current light smoker (defined as a person smoking on average 1–10 cigarettes per day), current moderate smoker (11–20 cigarettes per day), and current heavy smoker (>20 cigarettes per day). We grouped smokers into the aforementioned 4 states using the half-pack (that is, 10 cigarettes) and one pack (20 cigarettes) per day average smoking intensity as reference points, to alleviate some of the challenges observed in clinical practice during the collection of individuals’ smoking history.²⁶ In addition to the smoking states, the individual can exhibit one of the following six lung cancer states: (1) cancer free, (2) undetected early stage (AJCC stages I and II), (3) undetected advanced stage (AJCC stages III and IV), (4) diagnosed early stage, (5) diagnosed advanced stage, and (6) death. The post-diagnosis and death states are completely observable – that is, the decision maker (which may be the physician and individual working together in a shared decision-making process) is 100% certain about the individual’s health state at the time of cancer diagnosis or death. In contrast, the cancer-free and undetected lung cancer states are not directly observable. However, the decision maker can infer the unobservable cancer state of the individual based on screening findings and underlying risk factors. Because every LDCT exam may induce potential harms, each screening decision necessitates an assessment of the tradeoffs between the benefits and harms associated with screening. ENGAGE provides an analytical framework that optimizes the lung cancer screening decision-making process by estimating the tradeoff between the potential QALYs gained from screening and QALYs lost due to harms associated with LDCT. A detailed description of the POMDP model underlying ENGAGE is provided in the Supplementary Material §1.

Lung Cancer Incidence, Disease Progression, and Overall Mortality

To inform ENGAGE, we rely on published risk prediction models⁹ and models developed within the Cancer Intervention and Surveillance Modeling Network (CISNET) Lung Working Group, used to inform the USPSTF recommendations on lung cancer screening.^27–29

We estimate smoking state transition probabilities using the smoking history generator (SHG) (Supplementary Material §2, Supplementary Figure 2).^30,31 The SHG is a validated microsimulation model developed within CISNET based on data from the National Health Interview Survey (NHIS), that provides sex- and birth cohort-specific smoking histories representative of the U.S. population. Smoking histories consist of the age at smoking initiation, smoking intensity measured by the average number of cigarettes smoked per day, and age at smoking cessation for former smokers. To account for the high relapse rates within the first two years of smoking cessation, the SHG considers an individual as a former smoker after two-years of continuous smoking abstinence and assumes no relapse back to smoking thereafter.

We estimate the lung cancer incidence and mortality from competing causes of death using the Bach et al. lung cancer and other cause mortality risk prediction models, respectively.⁹ While several risk models exist, we choose the Bach et al. model because it provides annual risk estimates, and is derived using a small number of predictors, making it interpretable. It integrates age, sex, smoking duration, smoking intensity, years since smoking cessation, and asbestos exposure, as independent predictors of lung cancer risk and is ranked among the best externally validated lung cancer risk prediction models.^6,7 Given the low prevalence of asbestos exposure in the U.S.,³² we assume zero asbestos exposure for all individuals.

We leverage a published natural history model of lung cancer to estimate transitions between cancer states (Supplementary Material §3, Supplementary Tables 1–2).^33,34 The use of well-established simulation models to estimate the cancer state transitions is necessary, because these transitions occur at the preclinical stage of the disease and no clinical data exist for untreated lung cancer patients. Our natural history model is one of the CISNET models used to inform the USPSTF recommendations on lung cancer screening.^27–29 It assumes exponential growth for the primary tumor and proportional (to the primary tumor) growth for the lethal metastatic burden. Patients with observable lethal metastatic burden at the time of diagnosis are considered as having advanced stage lung cancer. We estimate the all cause survival time from diagnosis using data from the Surveillance, Epidemiology, and End Results (SEER) registry and a microsimulation model (Supplementary Material §4, Supplementary Figure 3).³⁵

To validate ENGAGE, we compare the ENGAGE-predicted incidence and mortality rates versus observed incidence data obtained from the SEER program and U.S. mortality data, respectively (Supplementary Material §5, Supplementary Figure 4).

Dynamic Lung Cancer Risk Assessment

ENGAGE dynamically assesses the probability that an individual will be screen-detected with lung cancer if (s)he undergoes screening at that specific moment given that person’s smoking and screening histories; hereon called “ENGAGE-derived lung cancer prevalence.”

ENGAGE makes the first screening decision at age 50 based on the age-specific lung cancer incidence. Once the optimal action at a given time t, denoted α_t, is taken (i.e. “Screen” or “No Screen”), an observation, o_t, is realized (“normal screen,” “abnormal screen,” “presence/absence of clinical symptoms”). Before making the next decision, ENGAGE updates the smoking state of the individual, assimilates newly collected information from past findings, and calculates the ENGAGE-derived prevalence using Bayesian updating (Equation 1). The updated ENGAGE-derived prevalence is then used to inform the next decision and the process repeats itself until lung cancer diagnosis, death, or age 100, whichever occurs first (Figure 1, Supplementary Material §6).

$$b_t\left(s_{t+1}\right)=\frac{\sum _{s_t\in S^{PO}}{b}_t\left(s_t\right)\mathrm{*}P\left(o_t|s_t,a_t\right)\mathrm{*}P\left(s_{t+1}|s_t,a_t,o_t\right)}{\sum _{s_t\in S^{PO}}{b}_t\left(s_t\right)\mathrm{*}P\left(o_t|s_t,a_t\right)}$$ (1)

for all $s_{t+1}\in S^{PO}$, where s denotes the individual’s health state, $S^{PO}$ represents the set of unobservable health states, and b represents the ENGAGE-derived lung cancer prevalence.

Figure 1. ENGAGE-related Event Ordering.

At any point between two consecutive screening decisions the individual may transition to the death state. A screening exam is considered abnormal if it detects a Lung-RADS category 3 or category 4 nodule; otherwise it is considered normal. A false-positive screening finding is defined as an abnormal LDCT screening exam that that is not diagnosed as cancer in a diagnostic follow-up management program.

It is important to note that ENGAGE is not a risk prediction model. Rather, ENGAGE leverages an existing risk calculator to estimate the age-specific lung cancer risk and makes screening recommendations based on the age-specific ENGAGE-derived lung cancer prevalence leveraging the most up-to-date information about the individual’s health state.

Imperfect Screening Information

ENGAGE accounts for the imperfect nature of information collected from past screening exams and clinical symptoms, by incorporating the uncertainty introduced by these observations. The accuracy of LDCT is obtained from the literature and is represented in terms of smoking-status-specific sensitivity and specificity (Supplementary Table 3).³⁶ Between two consecutive screening decisions, the individual may develop clinical symptoms indicative of lung cancer. We consider hemoptysis to be the only clinical symptom with strong predictive value for lung cancer and obtain the corresponding sensitivity and specificity from the literature (Supplementary Tables 4–5).³⁷

Quality-Adjusted Life-Years

We use published utilities associated with lung cancer, smoking, and screening outcomes to adjust the remaining lifetime of ever-smokers for quality of life (Supplementary Tables 6–9).

We measure health benefits using the total expected QALYs at age 50 and discount future survival at a 3% annual discount rate.³⁸ We assess the impact of alternative discount rates on the optimal screening schedules via a sensitivity analysis by ranging the discounting factor between 0% and 5%. Survival time beyond age 100 is estimated using U.S life tables (Supplementary Table 10).³⁹

Outcome Measures

We compare the effectiveness of the ENGAGE-based screening schedules against the 2013 USPSTF and single-threshold risk-based strategies. We opt to use the 2013 USPSTF recommendations and not the 2020 draft recommendations issued by the USPSTF, because the latter has not yet been finalized. We select the Bach et al. risk-prediction model for the single-threshold risk-based screening alternative, to make it directly comparable to ENGAGE. Risk-based screening programs using other risk prediction models may not be directly comparable to ENGAGE due to differences in the absolute risk estimates between risk prediction models,^6,8 but ENGAGE can be adopted for use with other models. Because there is no established risk-threshold used to define screening eligibility for the Bach model, we adopt a 1.91% 6-year risk-threshold as it corresponds to the threshold that yields similar sensitivity as the 2013 USPSTF criteria in the control arm of the Prostate, Lung, Colorectal, and Ovarian (PLCO) cancer screening trial.⁸ We apply the above screening policies on a cohort of 100,000 simulated ever-smokers per sex and smoking status assuming perfect implementation and adherence to all screening programs. We aggregate sex- and smoking status-specific results to obtain population level outcomes to compare each screening program.

Primary outcomes include personalized lung cancer screening schedules and lung cancer specific mortality. Secondary outcome measures include the total expected QALYs at age 50, number of screening exams and false-positive findings.

Statistical Analysis

All analyses are conducted in julia v1.2 using the POMDPs.jl framework.^40,41 We perform two-sided hypothesis tests using the HypothesisTests package (v.0.8.0) with an alpha-error cutoff point of 0.05. We use the paired t-test to assess statistical significance of differences in outcomes between smoking groups and the Fisher’s Exact test to compare rates.

Results

ENGAGE can be applied to any birth cohort and smoking history. To demonstrate the use of ENGAGE, we consider specific subpopulations of ever-smokers born in 1950 who started smoking at age 18 and stopped smoking at age 45 for the former smokers, unless noted otherwise. We simulate outcomes for these subpopulations under 3 alternative strategies: (i) ENGAGE; (ii) the 2013 USPSTF recommendations; and (iii) a single-threshold risk-based program.

Consistently with the 2013 USPSTF guideline, ENGAGE recommends no screening for former light and moderate smokers; however, ENGAGE differs in all other risk groups yielding higher QALYs (Table 1). ENGAGE yields similar mortality reduction to the 2013 USPSTF strategy for male current moderate and heavy smokers while requiring significantly fewer screening exams per individual (male current moderate smoker: −2.704, 95% CI: [−2.736, −2.673]; P<0.001, and current heavy smokers: −1.120, 95% CI: [−1.132, −1.107]; P<0.001) (Supplementary Table 11). In all other subpopulations, ENGAGE yields higher mortality reduction, with the improvement ranging from 3 to 14 percentage points. Moreover, ENGAGE expands screening coverage to current light smokers, a subgroup ineligible for screening per the 2013 USPSTF recommendation, covering 77% and 86% of male and female current light smokers, respectively. ENGAGE recommends fewer screening exams for male smokers who are eligible for lung cancer screening per the 2013 USPSTF guidelines (Table 1a), and recommends more screening exams for eligible women (Table 1b).

Table 1a.

Expected outcomes for male ever-smokers by smoking status when the age at smoking initiation is 18 years, and for former smokers age at smoking cessation is 45 years, under alternative screening strategies (2013 USPSTF, Single-Threshold, ENGAGE).

Outcome	Policy	Former Light Smokers6	Former Moderate Smokers7	Former Heavy Smokers8	Current Light Smokers6	Current Moderate Smokers7	Current Heavy Smokers8
Total Expected QALYs 2 , n (95% CI)	2013 USPSTF	14.06 (14.04–14.07)	13.68 (13.67–13.70)	13.16 (13.14–13.18)	13.21 (13.19–13.23)	12.61 (12.59–12.63)	11.83 (11.81–11.85)
	Single-threshold Risk-based1	14.06 (14.04–14.07)	13.67 (13.65–13.69)	13.16 (13.14–13.18))	13.20 (13.18–13.22)	12.60 (12.59–12.62)	11.83 (11.81–11.85)
	ENGAGE	14.06 (14.04–14.07)	13.68 (13.67–13.70)	13.18 (13.16–13.20)	13.21 (13.19–13.23)	12.62 (12.60–12.63)	11.84 (11.82–11.86)
Total Expected LYs 2 , n (95% CI)	2013 USPSTF	34.41 (34.35–34.47)	32.90 (32.84–32.96)	30.99 (30.93–31.06)	32.16 (32.09–32.22)	30.49 (30.42–30.55)	28.37 (28.31–28.43)
	Single-threshold Risk-based1	34.41 (34.35–34.47)	32.93 (32.86–32.99)	31.06 (31.00–31.12)	32.24 (32.18–32.31)	30.51 (30.44–30.57)	28.40 (28.34–28.46)
	ENGAGE	34.41 (34.35–34.47)	32.90 (32.84–32.96)	31.01 (30.95–31.08)	32.22 (32.15–32.28)	30.47 (30.41–30.54)	28.38 (28.32–28.44)
Lung Cancer Deaths 3 , n	2013 USPSTF	1,652	2,714	3,993	4,430	5,284	7,384
	Single-threshold Risk-based1	1,652	2,342	3,067	3,380	4,968	6,986
	ENGAGE	1,652	2,714	3,741	3,824	5,397	7,391
Mortality Reduction 4 , %	2013 USPSTF	0	0	4	<1	27	32
	Single-threshold Risk-based1	0	14	26	24	32	35
	ENGAGE	0	0	10	14	26	32
Ever-Screened Individuals 3 , %	2013 USPSTF	0	0	98.7	<1	77.7	98.0
	Single-threshold Risk-based1	0	82.3	92.4	89.3	95.8	98.9
	ENGAGE	0	0	94.3	76.9	97.6	99.1
LDCT Exams per Person Ever Screened, n (95% CI)	2013 USPSTF	NA	NA	5.84 (5.83–5.84)	< 1 (<1-<1)	13.96 (13.90–14.02)	18.05 (18.01–18.09)
	Single-threshold Risk-based1	NA	5.91 (5.89–5.93)	13.26 (13.22–13.29)	10.68 (10.64–10.72)	17.03 (16.99–17.07)	21.51 (21.46–21.56)
	ENGAGE	NA	NA	4.30 (4.29–4.31)	5.02 (5.00–5.04)	11.25 (11.21–11.28)	16.93 (16.88–16.98)
LDCT Exams per Death Avoided, n	2013 USPSTF	NA	NA	3,944	613	696	523
	Single-threshold Risk-based1	NA	1,590	1,234	1,016	734	559
	ENGAGE	NA	NA	1,077	827	595	491
False-Positive Findings 3,5 , n	2013 USPSTF	0	0	49,731	52	109,316	142,222
	Single-threshold Risk-based1	0	50,371	112,726	85,153	135,396	170,087
	ENGAGE	0	0	36,818	38,777	87,748	132,983
Screen Detected Lung Cancers 3	2013 USPSTF	0	0	334	2	4,380	7,763
	Single-threshold Risk-based1	0	840	2,371	2,349	5,071	8,667
	ENGAGE	0	0	924	1,390	4,416	8,879
Early Stage	2013 USPSTF	0	0	263	2	3,586	6,276
	Single-threshold Risk-based1	0	660	1,908	1,866	4,148	6,985
	ENGAGE	0	0	709	1,084	3,565	7,129
Advanced Stage	2013 USPSTF	0	0	71	0	794	1,487
	Single-threshold Risk-based1	0	180	463	483	923	1,682
	ENGAGE	0	0	215	306	851	1,750

USPSTF, United States Preventive Services Task Force; QALYs, quality-adjusted life-years; LYs, life-years; LDCT, low-dose computed tomography

¹risk-based screening strategy using the Bach lung cancer risk prediction model between ages 50–80 with a static 1.91% 6-year risk threshold

²per person at age 50

³expected cases per 100,000 individuals from the general population

⁴mortality reduction is estimated for the subpopulation of ever smokers that started smoking at age 18 and stopped smoking at age 45 for the former smokers

⁵a false-positive finding is an abnormal LDCT exam (i.e. Lung-RADS category 3 or 4A) that is not diagnosed as cancer in a diagnostic follow-up management protocol

⁶light smoker is defined as a person smoking on average 1–10 cigarettes per day

⁷moderate smoker is defined as a person smoking on average 11–20 cigarettes per day

⁸heavy smoker is defined as a person smoking on average >20 cigarettes per day

Table 1b.

Expected outcomes for female ever-smokers by smoking status when the age at smoking initiation is 18 years, and for former smokers age at smoking cessation is 45 years, under alternative screening strategies.

Outcome	Policy	Former Light Smokers6	Former Moderate Smokers7	Former Heavy Smokers8	Current Light Smokers6	Current Moderate Smokers7	Current Heavy Smokers8
Total Expected QALYs 2 , n (95% CI)	2013 USPSTF	14.37 (14.35–14.38)	14.03 (14.02–14.05)	13.56 (13.54–13.58)	13.62 (13.60–13.64)	13.05 (13.03–13.06)	12.31 (12.30–12.33)
	Single-threshold Risk-based1	14.37 (14.35–14.38)	14.02 (14.01–14.04)	13.56 (13.54–13.58)	13.61 (13.59–13.63)	13.04 (13.03–13.06)	12.31 (12.30–12.33)
	ENGAGE	14.37 (14.35–14.38)	14.03 (14.02–14.05)	13.58 (13.56–13.59)	13.62 (13.60–13.63)	13.05 (13.03–13.07)	12.32 (12.30–12.34)
Total Expected LYs 2 , n (95% CI)	2013 USPSTF	38.29 (38.23–38.36)	36.72 (36.66–36.78)	34.71 (34.64–34.77)	36.06 (36.00–36.13)	34.24 (34.18–34.31)	31.93 (31.86–32.00)
	Single-threshold Risk-based1	38.29 (38.23–38.36)	36.74 (36.68–36.81)	34.79 (34.72–34.85)	36.16 (36.10–36.23)	34.27 (34.21–34.34)	31.97 (31.90–32.04)
	ENGAGE	38.29 (38.23–38.36)	36.72 (36.66–36.78)	34.76 (34.70–34.82)	36.15 (36.08–36.21)	34.26 (34.20–34.33)	31.98 (31.92–32.05)
Lung Cancer Deaths 3 , n	2013 USPSTF	2,107	3,383	5,323	5,705	7.144	9,935
	Single-threshold Risk-based1	2,107	3,099	4,489	4,787	6.842	9,499
	ENGAGE	2,107	3,383	4,850	5,048	6.868	9,121
Mortality Reduction 4 , %	2013 USPSTF	0	0	2	0	20	24
	Single-threshold Risk-based1	0	8	18	16	23	27
	ENGAGE	0	0	11	12	23	30
Ever-Screened Individuals 3 , %	2013 USPSTF	0	0	99.1	<1	76.2	98.4
	Single-threshold Risk-based1	0	86.9	94.1	92.0	96.8	98.8
	ENGAGE	0	0	98.1	85.8	99.2	99.7
LDCT Exams per Person Ever Screened, n (95% CI)	2013 USPSTF	NA	NA	5.89 (5.88–5.89)	< 1 (<1-<1)	14.40 (14.34–14.46)	19.00 (18.96–19.04)
	Single-threshold Risk-based1	NA	5.70 (5.68–5.72)	13.46 (13.43–13.49)	10.92 (10.88–10.96)	17.94 (17.90–17.98)	21.95 (21.91–21.99)
	ENGAGE	NA	NA	7.72 (7.71–7.73)	6.69 (6.66–6.72)	15.40 (15.36–15.44)	21.37 (21.32–21.42)
LDCT Exams per Death Avoided, n	2013 USPSTF	NA	NA	4,396	NA	823	606
	Single-threshold Risk-based1	NA	2,006	1,390	1,189	875	615
	ENGAGE	NA	NA	1,272	1,019	761	542
False-Positive Findings 3,5 , n	2013 USPSTF	0	0	50,140	9	113,695	150,720
	Single-threshold Risk-based1	0	114,654	112,726	87,823	143,784	174,907
	ENGAGE	0	0	65,960	52,217	121,501	169,706
Screen Detected Lung Cancers 3	2013 USPSTF	0	0	320	0	4,219	7,589
	Single-threshold Risk-based1	0	714	2,337	2,235	4,955	8,635
	ENGAGE	0	0	1,566	1,663	5,058	10,002
Early Stage	2013USPSTF	0	0	238	0	3,107	5,569
	Single-threshold Risk-based1	0	505	1,719	1,631	3,643	6,342
	ENGAGE	0	0	1,078	1,166	3,654	7,264
Advanced Stage	2013 USPSTF	0	0	82	0	1,112	2,020
	Single-threshold Risk-based1	0	209	618	604	1,312	2,293
	ENGAGE	0	0	488	497	1,404	2,738

USPSTF, United States Preventive Services Task Force; QALYs, quality-adjusted life-years; LYs, life-years; LDCT, low-dose computed tomography

¹risk-based screening strategy using the Bach lung cancer risk prediction model between ages 50–80 with a static 1.91% 6-year risk threshold

²per person at age 50

³expected cases per 100,000 individuals from the general population

⁴mortality reduction is estimated for the subpopulation of ever smokers that started smoking at age 18 and stopped smoking at age 45 for the former smokers

⁵a false-positive finding is an abnormal LDCT exam (i.e. Lung-RADS category 3 or 4A) that is not diagnosed as cancer in a diagnostic follow-up management protocol

⁶light smoker is defined as a person smoking on average 1–10 cigarettes per day

⁷moderate smoker is defined as a person smoking on average 11–20 cigarettes per day

⁸heavy smoker is defined as a person smoking on average >20 cigarettes per day

In comparison to the single-threshold risk-based screening program, ENGAGE yields higher QALYs and reduces the number of screening exams in every subgroup, but averts fewer deaths and yields fewer life-years gained (Supplementary Table 12).

With outcomes aggregated across all individuals (not just subpopulations), the ENGAGE-based screening strategy, as compared to the 2013 USPSTF policy, yields statistically significantly higher lung cancer-specific mortality reduction (ENGAGE: 19%, USPSTF: 15%, P<0.001), prescribes 14% fewer screening exams per person (ENGAGE: 10.43, USPSTF: 12.07, P<0.001), and requires 15% fewer exams per lung cancer death avoided (ENGAGE: 696, USPSTF: 819, P<0.001) (Table 2, Supplementary Table 13). ENGAGE screens an additional 17% of the population of ever-smokers who started smoking at age 18 (ENGAGE: 78.0%, USPSTF: 61.2%), thus increasing by 10% the number of LDCT exams and false-positive findings per 100,000 ever-smokers. In comparison to the single-threshold risk-based screening program (see Methods), ENGAGE yields lower mortality reduction (ENGAGE: 19%, Single-threshold: 25%, P<0.001), but requires 31% fewer screening exams per person (ENGAGE: 10.43, Single-threshold: 15.11, P<0.001), 22% fewer screening exams per death averted (ENGAGE: 696, Single-threshold: 889, P<0.001), and produces 40% fewer false-positive findings (Table 2, Supplementary Table 13).

Table 2.

Aggregated population-level outcomes for ever-smokers with age at smoking initiation 18 years, and age at smoking cessation 45 years for former smokers, under alternative screening strategies

Outcome	2013 USPSTF	Single-Threshold Risk-Based1	ENGAGE
Total Expected QALYs 2 , n (95% CI)	13.17 (13.16–13.17)	13.16 (13.16–13.17)	13.18 (13.17–13.18)
Total Expected LYs 2 , n (95% CI)	32.86 (32.84–32.88)	32.91 (32.89–32.93)	32.89 (32.87–32.90)
Lung Cancer Deaths 3 , n	5,183	4,584	4,917
Mortality Reduction 4 , % (95% CI)	15 (14–16)	25 (23–26)	19 (18–20)
Ever-Screened Individuals 3 , n (%)	61,243 (61.2)	88,284 (88.3)	77,953 (78.0)
LDCT Exams 3	739,130	1,333,974	812,788
LDCT Exams per Person Ever Screened, n (95% CI)	12.07 (12.04–12.10)	15.11 (15.03–15.19)	10.43 (10.40–10.45)
LDCT Exams per Death Avoided, n (95% CI)	819 (817 to 821)	889 (884 to 894)	696 (694 to 697)
False-Positive Findings 3,5 , n	59,420	108,999	65,125
Screen Detected Lung Cancers 3	2,073	3,457	2,901
Early Stage	1,618	2,682	2,205
Advanced Stage	455	775	696

USPSTF, United States Preventive Services Task Force; QALYs, quality-adjusted life-years; LYs, life-years; LDCT, low-dose computed tomography

¹risk-based screening strategy using the Bach lung cancer risk prediction model between ages 50–80 with a static 1.91% 6-year risk threshold

²per person at age 50

³expected cases per 100,000 individuals from the general population

⁴mortality reduction is estimated for the subpopulation of ever smokers that started smoking at age 18 and stopped smoking at age 45 for the former smokers

⁵a false-positive finding is an abnormal LDCT exam (i.e. Lung-RADS category 3 or 4A) that is not diagnosed as cancer in a diagnostic follow-up management protocol

Unlike single-threshold risk-based methods, ENGAGE produces dynamic risk thresholds that increase with age, accounting for changes in life-expectancy. We illustrate this feature of ENGAGE in Figure 2, which provides the ENGAGE-derived age-specific lung cancer prevalence for ever-smokers who never experience a positive screening exam or develop clinical symptoms, stratified by smoking intensity (light, moderate and heavy) and smoking status (former and current). For example, for a male current moderate smoker over age 80, ENGAGE recommends increased intervals (>1 year) between screenings, even though the individual is at a higher lung cancer risk than his risk between ages 65 to 79 years, when ENGAGE recommends annual screening (Figure 2A). This is due to the reduction in life-expectancy of older individuals, which depreciates screening. In a similar manner, ENGAGE identifies the optimal ages to start and stop screening for each individual. For example, for male and female former heavy smokers, ENGAGE initializes screening at age 61 and 57, respectively (Figures 2 A and B, top right-most panel), whereas per the 2013 USPSTF guidelines these individuals would start screening at age 55. For current heavy smokers, ENGAGE recommends initiating screening earlier than the 2013 USPSTF recommendation’s start age of 55 years (Figure 2A, bottom right-most panel). In general, as the smoking intensity increases, ENGAGE recommends initiating screening at younger and terminating at older ages. For women ever-smokers, ENGAGE recommends screening schedules that start earlier and stop later as compared to the corresponding screening schedules for men with the same smoking and screening histories (Figure 2B).

Figure 2. Schedule of ENGAGE-derived screening (dashed verticals), age-specific lung cancer prevalence assuming following ENGAGE screening recommenation (red curve) vs no screening (dashed blue curve), for ever-smokers who never experience a positive screening exam or develop clinical symptoms, stratified by smoking intensity (“light smoker” on left panel, “moderate smoker” on middle panel and “heavy smoker” on right panel) and smoking cessation age (top panels: 45 years for “former smokers”; bottom panels: none for “current smokers”), for males (A) and females (B).

Note that the age-specific ENGAGE-derived prevalence under “No Screening” is computed by enforcing the “No Screen” action and calculating the ENGAGE-derived prevalence of lung cancer assuming that no clinical symptoms were developed between two consecutive screening decisions.

The impact of screening findings on the ENGAGE-derived lung cancer prevalence and the resulting adaptive screening schedules are also depicted in Figure 2 for specific subpopulations. For example, for male current light smokers, ENGAGE recommends the first screening exam at age 59. If the screening exam is negative, ENGAGE recommends no screening till age 62 and biennially thereafter till age 80. For current heavy and moderate smokers, the ENGAGE-derived prevalence monotonically increases despite consecutive negative screening exams (e.g. female current moderate smokers ENGAGE between ages 61–82), and are the only subpopulations for which ENGAGE recommends consistently annual screening exams, similarly to the 2013 USPSTF recommendation.

Figure 3 illustrates the ENGAGE recommendations for smokers who start smoking at age 18 and quit smoking at age 60 as a function of smoking intensity. Upon smoking cessation, ENGAGE recommends fewer future LDCT exams compared to the 2013 USPSTF recommendations. For example, for a female former smoker who quits smoking at age 60, ENGAGE recommends biennial screening between the ages of 61 to 79. Per the 2013 USPSTF recommendation, this woman upon smoking cessation would need to undergo 15 annual screening exams with the final exam occurring at age 75. Hence for this particular example, relative to the 2013 USPSTF guidelines, ENGAGE reduces the number of screening exams upon smoking cessation by 33% (ENGAGE: 10, USPSTF: 15) and recommends screening at older ages when the risk of developing lung cancer is higher. The adaptive screening framework offered by ENGAGE along with the impact of past screening findings and changes in the smoking behavior of individuals on the optimal ENGAGE screening schedule are presented through an example of a representative female smoker (Supplementary Table 14).

Figure 3. Effect of smoking cessation at age 60 on the ENGAGE screening schedule, stratified by smoking intensity (“light smoker” on left panel, “moderate smoker” on middle panel and “heavy smoker” on right panel), for males (A) and females (B).

The age-specific ENGAGE-derived prevalence under No Screening is computed by enforcing the “No Screen” action and calculating the ENGAGE-derived prevalence of lung cancer assuming that no clinical symptoms were developed between two consecutive screening decisions.

ENGAGE incorporates patient’s preferences into the screening decision. When the harmful effects of screening on quality of life are higher or lower than the base-case estimates (Supplementary Table 9), ENGAGE recommends fewer or more, respectively, screening exams (Supplementary Figure 5).

The impact of the discount rate for future health benefit on the optimal screening schedules is minimal for both sexes, maintaining the same number of screening exams, but occasionally moving the screening age by 1–2 years (Supplementary Figure 6).

Discussion

Precision health initiatives have called for personalized lung cancer screening.^6–8,14,42 To answer this call, we developed ENGAGE, a lung cancer screening framework that provides personalized schedules for U.S. ever-smokers by simultaneously accounting for the individual’s lung cancer risk and life-expectancy over time. ENGAGE provides screening decisions by reassessing the trade-off between benefits and harms associated with screening in real-time, leveraging up-to-date information collected from past screening exams and changes in smoking behavior.

We demonstrate that ENGAGE is more effective and more efficient than the 2013 USPSTF lung cancer screening recommendation and a single-threshold risk-based strategy, assuming perfect implementation and compliance to all programs. In comparison to the 2013 USPSTF recommendations, ENGAGE yields higher lung cancer-specific mortality reduction while requiring fewer screening exams per death avoided and fewer screening exams per person screened. In comparison to a single-threshold risk-based screening program, ENGAGE yields higher expected QALYs, reduces the number of LDCT exams per person, improves screening efficiency, and produces fewer false-positive results. Single-threshold risk-based approaches yield a higher mortality reduction, but offer a modest increase in life-years because they ignore life-expectancy.^8,14 ENGAGE incorporates life-expectancy by dynamically optimizing the risk threshold over an individual’s lifetime. At younger ages, ENGAGE uses a low-risk threshold to determine screening eligibility given the higher life-expectancy. As individuals age their lung cancer risk increases, but because their life-expectancy decreases, ENGAGE increases the risk threshold for screening eligibility. Despite the increased risk threshold, ENGAGE recommends screening at ages older than 80 years among individuals who are healthy to undergo curative treatment.

ENGAGE provides a comprehensive framework that can guide decision makers in designing more effective and efficient screening guidelines for diverse subpopulations. ENGAGE recommends expanding screening coverage to current smokers younger than 55 years of age who smoke at least 10 cigarettes per day (that is, current moderate and heavy smokers). This recommendation is consistent with the proposed changes to the 2013 USPSTF guidelines highlighted in the 2020 USPSTF draft statement, which lowers screening start age from 55 to 50 years, and pack-years from 30 to 20. Furthermore, the 2020 draft recommendations, pending final approval, provide coverage to former smokers with 20 pack-year smoking history who are within 15 years from smoking cessation. ENGAGE also recommends screening for former smokers who used to smoke at least 20 cigarettes per day on average, but offers screening exams at ages beyond 15 years from smoking cessation. Former heavy smokers have a significant risk of developing lung cancer even after 15 years since cessation.¹³ In fact, the Framingham Heart Study found that 41% of the former smokers diagnosed with lung cancer were diagnosed more than 15 years after smoking cessation.⁴³ In terms of screening frequency, the 2013 and 2020 draft USPSTF guidelines recommend annual LDCT exams for eligible individuals, whereas ENGAGE recommends regular annual screening exams for current moderate and heavy smokers, but not for former smokers. This suggests that former smokers may be over-screened by the USPSTF guidelines while within 15 years since smoking cessation.

ENGAGE supports expanding screening coverage to current light smokers, defined as individuals who currently smoke on average less than 10 cigarettes per day. This is a group that, although possible, is unlikely to be eligible for lung cancer screening under the 2020 USPSTF draft recommendations because of the 20 pack-year eligibility criterion.

Our findings indicate that screen-eligible men and women based on the 2013 USPSTF guidelines are being over- and under-screened, respectively. ENGAGE accounts for sex-specific differences in lung cancer and addresses them by recommending that women should start screening at younger ages and stop screening at older ages than men with the same smoking history. Results from the NELSON trial and published cost-effectiveness analyses concluded that lung cancer screening is more effective for women as compared to men, thus verifying our findings.^35,44,45

ENGAGE provides screening schedules that adapt to screening findings and changes in smoking behavior. It incorporates new information to dynamically assess individuals’ risk and adjusts future screening decisions, accordingly, allocating screening resources efficiently. The value of past screening findings is supported by existing literature reporting that individuals with a negative screen have lower lung cancer risk.^16,46

ENGAGE implicitly consolidates personalized screening schedules with a smoking cessation program. It presents the positive effects of smoking cessation – that is, the reduction in future lung cancer risk and number of screening exams – which can be used to facilitate discussion about smoking cessation and amplify smoking cessation rates. Supplementing screening with a smoking cessation program has been shown to enhance the effectiveness and cost-effectiveness of lung cancer screening.^47–49

The feasibility of implementing a real-time decision tool on whether or not to screen an individual is beyond the scope of this study. For those who may consider the possibility of exploring this option, we find that in all subpopulations (except for male current moderate smokers) the screening frequency of the recommended schedules changes at most twice throughout the lifetime of an individual, despite allowing unlimited number of changes in screening frequency. This is important because well-established screening guidelines for other cancer sites (e.g. breast, cervical, and colorectal cancers) recommend differing screening frequencies as individuals age. In the case of ENGAGE, this change will be individualized. We would like to emphasize that ENGAGE relies on patient information that is already collected in clinical practice. More importantly, ENGAGE does not require the explicit number of cigarettes smoked per day as existing guidelines, because it groups individuals into light, moderate and heavy smokers. This could mitigate some of the challenges encountered by the health systems when collecting information regarding the smoking history of individuals, because individuals only need to provide their smoking intensity relative to the thresholds used by ENGAGE to define the smoking categories (one half and full pack of cigarettes per day).²⁶

While ENGAGE provides a novel framework for real-time decision-making on whether or not to screen, it has limitations. First, in the current version of ENGAGE, we define a former smoker as someone who has not smoke for two years accounting for the high relapse rates within the first years, and assume no relapse after two years. Hence, we ignore a small percentage of smokers who relapse back to smoking after 2 years, overestimating the proportion of former smokers at any given age. Second, we assess the lung cancer risk of ever-smokers using only a subset of the established lung cancer risk factors: age, sex, smoking duration and smoking intensity. Hence, we do not consider other risk-factors such as race, COPD, emphysema, and family history. Consequently, we underestimate the benefit of personalizing lung cancer screening. Incorporating these additional risk factors into ENGAGE will enhance the personalization of the proposed screening schedules and consequently their effectiveness and efficiency. However, considering additional risk factors is not trivial as it would require development and validation of natural history models that incorporate such covariates. It is important to note that ENGAGE is flexible and could be easily adapted to incorporate additional risk factors conditioned on the availability of well-validated risk prediction and natural history models that integrate these factors. Third, while we account for competing causes of death, we do not explicitly model the comorbidity status of individuals. Incorporating the comorbidity status of individuals into ENGAGE will ensure that individuals who are unfit to receive curative treatment and/or have limited life expectancy will not undergo screening, thus enhancing the efficiency and effectiveness of ENGAGE. Also, because we do not consider the impact of non-lethal procedure-related complications, we may be overestimating the QALYs accrued from screening. Fourth, we assume perfect adherence to all screening programs and do not consider potential implementation challenges for the risk-based screening programs. We recognize that effective implementation of a personalized screening program necessitates overcoming several challenges, including the practicality of the screening program, effective communication of personalized risk, and adherence to the screening program, among others. To address and mitigate the aforementioned difficulties, we recognize that a user-friendly decision aid tool based on ENGAGE is required. Furthermore, appointment of trained lung nurse navigators tasked with the shared decision-making sessions and personalized risk communication may improve the quality of the shared decision-making sessions. However, the development and quality assessment of the decision aid tool is beyond the scope of this study. The objective of this study is to provide an analytical framework for personalized lung cancer screening and apply this framework to explore the net benefits and harms of personalized screening relative to other strategies.

In summary, we have introduced ENGAGE, a novel lung cancer screening framework that optimizes screening decisions at a personalized-level based on the lung cancer risk and life-expectancy of ever-smokers. ENGAGE offers adaptive screening schedules that are more efficient than existing guidelines and single-threshold risk-based screening. Implementation of a personalized risk-based screening program warrants further evaluation.