Performance of Lung Cancer Risk Prediction Models in Different Racial and Ethnic Groups in the United States: Results
Xiaoshuang Feng1, Florence Guida2, Aghiles Guenoun1
1Early Detection, Prevention, and Infections Branch, International Agency for Research on Cancer, Lyon, France (X.F., A.G., K.A., M.J., H.A.R.).
Background:
Racial and ethnic disparities are a concern in lung cancer screening.
Objective:
To investigate the performance of risk prediction models to define screening eligibility across 4 U.S. racial and ethnic groups.
Design:
Cohort study.
Setting:
United States, Lung Cancer Cohort Consortium.
Participants:
641 830 participants aged 50 to 80 years with a smoking history from 12 U.S. cohorts, including 6390 Asian, 9781 Hispanic, 39 872 non-Hispanic Black, and 585 787 non-Hispanic White participants.
Measurements:
Calibration and discrimination were quantified for 16 lung cancer prediction models. Then, screening-related metrics were calculated after applying model thresholds to select the same number of eligible participants as the 2021 criteria from the U.S. Preventive Services Task Force (USPSTF-2021). These included eligibility, sensitivity, and efficiency measured as estimated number needed to screen (NNS; the ratio between participants and lung cancer cases) for each strategy or prediction model in each racial and ethnic group.
Results:
General patterns across the 16 models included substantial underestimation of lung cancer risk in non-Hispanic Black participants (expected-observed ratio < 0.75 for 11 of 16 models), lower discrimination in Asian participants than all other groups (13 of 16 models), and lower discrimination in non-Hispanic Black than non-Hispanic White participants (15 of 16 models). When a same-sized screening-eligible population as USPSTF-2021 (38.0%) was enforced, all risk-based strategies achieved better average estimated screening efficiency and reduced racial and ethnic differences in efficiency compared with USPSTF-2021. The Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial Model 2012 (PLCOm2012) and Life Years gained From Screening-Computed Tomography model (LYFS-CT) performed best (mean estimated NNS, 36.5 [SD, 8.8] and 40.1 [SD, 8.2], respectively). However, no strategy could simultaneously optimize eligibility, sensitivity, and efficiency while also reducing racial and ethnic differences.
Limitation:
Smaller sample for Asian and Hispanic participants.
Conclusion:
To optimize efficiency and minimize its variation across racial and ethnic groups, risk-based strategies were superior to USPSTF criteria. Further optimization of prediction models for the diverse U.S. population is needed.
Primary Funding Source:
U.S. National Cancer Institute, Lung Cancer Research Foundation, and Cancer Research UK.
Related Concept Videos
01:25Statistical Methods for Analyzing Epidemiological Data
01:21Cancer Survival Analysis