Screening algorithms are only as equitable as the data and assumptions baked into them. When the tools used to decide who gets a lung cancer scan perform differently across racial groups, entire populations can be quietly excluded from life-saving early detection — and that gap is rarely visible without a study designed specifically to look for it.
Analyzing data from 641,830 smokers aged 50 to 80 drawn from 12 U.S. cohorts, researchers evaluated 16 lung cancer risk prediction models across four racial and ethnic groups. A striking pattern emerged: 11 of the 16 models substantially underestimated lung cancer risk in non-Hispanic Black participants, with expected-to-observed ratios falling below 0.75 — meaning models predicted far fewer cases than actually occurred. Discrimination, the ability to correctly rank who will develop cancer versus who will not, was also lower for Black participants compared to White participants in 15 of 16 models, and lower for Asian participants than all other groups in 13 of 16 models. When model-based thresholds were calibrated to select a screening-eligible pool matching the size defined by current USPSTF-2021 criteria, these performance gaps translated directly into measurable differences in sensitivity and screening efficiency across groups.
This analysis carries significant practical weight. Lung cancer remains the leading cause of cancer death in the United States, and early detection through low-dose CT scanning meaningfully improves survival. Yet most risk models were developed predominantly on non-Hispanic White cohorts, embedding demographic blind spots into clinical decision-support tools now applied universally. The finding that calibration errors for Black Americans exceed 25% in the majority of tested models is not a marginal statistical concern — it represents a structural flaw in screening policy infrastructure. A key limitation is the observational cohort design, which cannot establish causality, and cohort sizes for Asian and Hispanic subgroups remain relatively small. Still, with 16 models evaluated across a population exceeding 600,000, this is among the most comprehensive equity audits of lung cancer screening tools to date — and its implications for model recalibration and USPSTF guideline refinement are hard to overstate.