Accurate Risk Prediction Models Identified for Selecting Smokers for Lung Cancer Screening
Four risk prediction models have been shown to accurately select the highest risk smokers for lung cancer screening, according to a study published in Annals of Internal Medicine (online May 14, 2018; doi:10.7326/M17-2701).
Current lung cancer screening guidelines suggest utilizing individualized risk models to refer ever-smokers for screening. However, different models select different screening populations and the performance of each model in selecting ever-smokers is relatively unknown.
Hormuzd A Katki, PhD, division of cancer epidemiology and genetics, National Cancer Institute (Bethesda, MD), and colleagues conducted a study to compare the United States screening populations selected by nine lung cancer risk models and to assess their predictive performance in two cohorts of patients. The risk models evaluated were the Bach model; the Spitz model; the Liverpool Lung Project (LLP) model; the LLP Incidence Risk Model (LLPi); the Hoggart model, the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial Model 2012 (PLCOM2012); the Pittsburgh Predictor; the Lung Cancer Risk Assessment Tool (LCRAT); and the Lung Cancer Death Risk Assessment Tool (LCDRAT).
The population-based prospective study assessed models that selected screening populations from the National Health Interview Survey from 2010-2012, including 337,388 ever-smokers in the National Institutes of Health-AARP Diet and Health Study as well as 72,338 ever-smokers in the CPS-II Nutrition Survey cohort.
Researchers utilized model calibration (ratio of model-predicted to observed cases) and discrimination (area under the curve) to measure model performance.
At a 5-year risk threshold of 2.0%, the models chose screening populations ranging from 7.6 million to 26 million ever-smokers. In both validation cohorts, four models (the Bach model, PLCOM2012, LCRAT, and LCDRAT) were well-calibrated (ratio range of model-predicted to observed cases, 0.92-1.12) and had higher area under the curves (range, 0.75-0.79) compared with the remaining five models that generally overestimated risk (ratio range of model-predicted to observed cases, 0.83-3.69) and had lower area under the curves (range, 0.62-0.75).
Additionally, researchers reported that the four best-performing models had the highest sensitivity at a fixed specificity and similar discrimination at a fixed threshold. These best-performing models showed better agreement on size of the screening population (7.6 million to 10.9 million) and achieved consensus on 73% of individuals chosen.
Dr Katki and colleagues concluded that while the nine lung cancer risk models chose widely different screening population, the Bach model, PLCOM2012, LCRAT, and LCDRAT most accurately predicted risk and performed best in selecting ever-smokers for screening.
Authors noted the lack of consensus on risk thresholds for screening as a limitation of the study.—Zachary Bessette