A predictive model for lung cancer screening nonadherence in a community setting health-care network

Abstract

Background: Lung cancer screening (LCS) decreases lung cancer mortality. However, its benefit may be limited by nonadherence to screening. Although factors associated with LCS nonadherence have been identified, to the best of our knowledge, no predictive models have been developed to predict LCS nonadherence. The purpose of this study was to develop a predictive model leveraging a machine learning model to predict LCS nonadherence risk. Methods: A retrospective cohort of patients who enrolled in our LCS program between 2015 and 2018 was used to develop a model to predict the risk of nonadherence to annual LCS after the baseline examination. Clinical and demographic data were used to fit logistic regression, random forest, and gradient-boosting models that were internally validated on the basis of accuracy and area under the receiver operating curve. Results: A total of 1875 individuals with baseline LCS were included in the analysis, with 1264 (67.4%) as nonadherent. Nonadherence was defined on the basis of baseline chest computed tomography (CT) findings. Clinical and demographic predictors were used on the basis of availability and statistical significance. The gradient-boosting model had the highest area under the receiver operating curve (0.89, 95% confidence interval = 0.87 to 0.90), with a mean accuracy of 0.82. Referral specialty, insurance type, and baseline Lung CT Screening Reporting & Data System (LungRADS) score were the best predictors of nonadherence to LCS. Conclusions: We developed a machine learning model using readily available clinical and demographic data to predict LCS nonadherence with high accuracy and discrimination. After further prospective validation, this model can be used to identify patients for interventions to improve LCS adherence and decrease lung cancer burden.