Annual lung cancer screening (LCS) with low-dose CT can save lives, but only reaches those who meet eligibility criteria. The current US Preventive Services Task Force (USPSTF) recommends screening for people aged 50-80 who have smoked at least 20 pack-years and currently smoke or quit within the past 15 years.
A pack-year is defined as smoking one pack per day for one year, so 20 pack-years could mean smoking half a pack daily for 40 years or two packs daily for 10 years. The problem is that this metric combines both the intensity and duration of smoking into a single number, which can systematically exclude people who smoked lightly but for a very long time.
Modeling studies suggest that the USPSTF criteria are not optimally effective or equitable because many people who would derive high benefit from screening are excluded. Additionally, the 2021 USPSTF criteria have not substantially reduced racial and ethnic disparities in screening eligibility compared to the older 2013 criteria.
Rather than simply identifying who is at high risk for lung cancer, this study uses a tool called the Life-Years Gained from Screening CT (LYFS-CT) model to measure who would actually benefit most from screening. LYFS-CT calculates the expected number of extra days of life a person would gain from three annual CT scans.
LYFS-CT incorporates two key submodels: one estimating individual all-cause mortality risk (accounting for other health conditions that might shorten life regardless of cancer), and another estimating individual lung cancer death risk over five years. This combined approach means people with serious other illnesses who are unlikely to live long enough to benefit from screening can be appropriately identified.
The threshold used in this study was 16.2 days of life gained from three annual screens - the level recommended by the American College of Chest Physicians to identify people for whom screening is strongly beneficial. While 16.2 days may sound small, it represents an average over an entire population where most people will not develop lung cancer; those who do develop lung cancer and have it detected by screening typically gain 1.5 to 3 years of life.
A key methodological innovation was the use of counterfactual eligibility for racial and ethnic minorities: life expectancy calculations were adjusted to remove the effect of systemic racism on health outcomes, so that Black patients would not be excluded from screening simply because structural inequities had shortened their expected lifespan.
The researchers used machine learning to develop what are called fast-and-frugal trees (FFTs), which are simple decision algorithms that can classify people with just a few yes-or-no questions. Unlike complex risk calculators that require entering many variables, FFTs are designed to be remembered and applied quickly in clinical practice.
The training dataset included 71,978 respondents from the National Health Interview Survey (NHIS) spanning 1997 to 2014, representing a nationally representative sample of US adults who had ever smoked. The FFTs were constrained to use only the same types of information already collected for USPSTF decisions: age, years of tobacco use, pack-years, and years since quitting.
Multiple FFT variants were generated and tested, with the best-performing tree selected for final evaluation. The winning FFT (called FFT16.2) uses just two decision paths, making it simple enough to be applied in any clinical setting without a calculator or computer.
Validation was conducted using three independent datasets: a NHIS test set (1997-2014 even years), a NHIS evaluation set (2015-2018), and the Behavioral Risk Factor Surveillance System (BRFSS) 2022 dataset. This multi-dataset validation approach confirms the results hold across different time periods and survey methodologies.
The final alternative screening criteria (FFT16.2) uses two simple paths: Path 1 selects anyone who has smoked for 40 or more years regardless of intensity; Path 2 selects people aged 60-80 who have accumulated 40 or more pack-years regardless of how recently they quit. Together these two paths identify a population of approximately 14.1 million people - nearly identical in size to the 14.1 million selected by USPSTF.
The first path (40 or more years of smoking) captured 89% of all FFT-selected individuals and was the primary driver of improved performance. This path is especially powerful because it can be determined with a single simple question - 'How many years have you smoked?' - without needing to calculate pack-years, which requires knowing average daily cigarette consumption over a person's entire smoking history.
The second path (age 60-80 with 40+ pack-years) captured the remaining 11%, primarily recovering high-intensity smokers who smoked heavily but for fewer than 40 years and are not captured by the first path. Adding this second path increased overall sensitivity from 85% to 91% with only a small decrease in specificity.
FFT16.2 substantially outperformed the USPSTF criteria for identifying high-benefit candidates: sensitivity improved from 78% to 91% (p less than 0.001) and specificity improved from 84% to 86% (p less than 0.001) in the main NHIS evaluation set. These gains were replicated in the independent 2022 BRFSS dataset (sensitivity 88% vs. 78%, specificity 83% vs. 81%).
The clinical implications of the nonoverlapping populations reveal the real-world impact: for every 1,000 people screened under FFT16.2 but not USPSTF, 37.4 life-years would be gained. For every 1,000 people screened under USPSTF but not FFT16.2, only 24.4 life-years would be gained. This means FFT16.2 is directing screening resources toward more high-benefit individuals.
The FFT16.2 nonoverlapping population also prevents substantially more lung cancer deaths: projecting to the US population, FFT16.2 (but not USPSTF) screening would prevent an estimated 13,511 lung cancer deaths over 5 years, compared to only 5,448 deaths prevented by screening the USPSTF-only population.
The population uniquely identified by FFT16.2 differs meaningfully from the USPSTF-only group: 76.7% have 40+ years of tobacco use, 70.1% have fewer than 20 pack-years (who are excluded by the USPSTF minimum), and 31.2% quit more than 15 years ago (also excluded by USPSTF). These are precisely the groups the pack-year threshold was failing to protect.
The most striking finding is the impact on racial and ethnic minority populations. The current USPSTF criteria have only 56% sensitivity for identifying high-benefit Black candidates for screening. FFT16.2 raises this to 83% (p less than 0.001), a dramatic improvement that directly addresses a major racial health inequity in cancer prevention.
For Hispanic and Asian populations, FFT16.2 achieves sensitivities of 95% and 94% respectively, compared to 73% and 68% for USPSTF - gains of 22 and 26 percentage points. While these differences did not reach statistical significance due to smaller sample sizes, the consistent direction and magnitude are clinically meaningful.
The reason for these disparities is well-documented: Black individuals are more likely to develop lung cancer with fewer pack-years of smoking, and they have historically smoked cigarettes with different characteristics. One analysis found that among people diagnosed with lung cancer, the average pack-year history was 25.3 for Black patients vs. 49.0 for White patients - meaning USPSTF's 20-pack-year minimum disproportionately excludes high-risk Black smokers.
The first path of FFT16.2 (40+ years of smoking) is particularly effective at including minority populations, because it captures long-duration light smokers who are overrepresented among high-benefit Black, Hispanic, and Asian individuals. In the FFT16.2 nonoverlapping Black population, 88.5% had 40+ years of tobacco use and 86.0% had fewer than 20 pack-years.
A central scientific insight from this study is that smoking duration appears to be a stronger lung cancer risk factor than smoking intensity. This is consistent with established epidemiological evidence showing that the duration of tobacco exposure drives carcinogenic cellular changes more powerfully than the amount smoked per day.
Pack-years are also notoriously difficult to measure accurately. Patients often cannot reliably recall their average daily cigarette consumption across decades of smoking, and electronic medical records frequently contain inconsistent pack-year documentation. In contrast, asking 'How many years did you smoke?' is a simpler, more reliably answerable question.
The USPSTF criterion requiring tobacco cessation within the previous 15 years is also challenged by this study. Lung cancer risk continues to increase past 15 quit-years, especially for older individuals - a finding that led the 2023 American Cancer Society guideline update to eliminate the quit-year criterion entirely. FFT16.2 independently recovers approximately 1 million high-benefit long-time quitters in the US population.
The simple two-path structure of FFT16.2 is designed for real-world clinical implementation. Both paths use information already routinely collected in clinical settings - years of smoking and pack-years (or age) - and the criteria can be remembered without a calculator. This contrasts with complex risk models that, while more precise, face significant barriers to widespread adoption.
FFT16.2 selects more older individuals and those with more comorbidities than USPSTF, which means clinicians must still exercise judgment about individual prognosis and whether the benefits of screening outweigh potential harms for each patient. Shared decision-making with prediction tools remains important for borderline cases.
The researchers demonstrated that FFT16.2 showed consistent performance even in the more recent 2022 BRFSS dataset, suggesting it remains applicable as smoking patterns in the US population evolve. This resilience to changing epidemiology is an important practical strength for a criterion that may be in clinical use for years.
The methodology used here - training simple decision rules on outputs from validated prediction models - is broadly applicable beyond lung cancer screening. The same approach could be used to develop simple, equitable eligibility criteria for colorectal cancer screening, breast cancer screening, or other preventive interventions where complex models exist but simple clinical rules are needed for implementation.