Clinical validation of a deep-learning–based model to predict lung cancer risk from chest X-rays.

Abstract

10507 Background: Screening with chest CT prevents lung cancer death, but only 5% of eligible Americans are screened. The updated 2022 Centers for Medicare and Medicaid Services (CMS) criteria expand eligibility to 50-77 year-olds, current or former (<15 years) smokers, with a ≥20 pack-year history. However, smoking quit date and pack-years are often not available in the electronic medical record (EMR), stymying identification of high-risk smokers for lung cancer screening (LCS). We previously developed a free, open-source deep-learning model (CXR-LC) that identifies smokers at high risk of lung cancer better than the 2015 CMS criteria, based on a chest X-ray (CXR) image automatically extracted from the EMR. CXR-LC was developed in the PLCO trial, an asymptomatic LCS population. The present study externally validates CXR-LC using existing CXRs obtained through usual care and compares CXR-LC to the expanded 2022 CMS lung screening eligibility criteria. Methods: Current or former smokers who had a routine clinical CXR in 2013-2014 were included. Patients with known lung cancer or who were undergoing LCS were excluded. Smoking status and 2022 CMS LCS eligibility were determined by manual review of the EMR. The primary outcome was 6-year incident lung cancer, identified using International Classification of Disease (ICD) codes and confirmed through discharge summaries and pathology reports. The CXR-LC model and the 2022 CMS eligibility criteria were compared for prediction of 6-year incident lung cancer. Results: Among 10,784 patients (mean age 63.1± 6.8 years; 61.0% male), CMS eligibility could be determined in 45% (4,886/10,784), with 55% missing pack-year or quit date data. In this population, 3.9% (191/4,886) developed lung cancer. The 2022 CMS criteria missed 57% (109/191) of lung cancers, while CXR-LC missed 8.9% (17/191, p < 0.001). After adjusting for sex, race, and prevalent COPD, the hazard ratio for CXR-LC eligibility (3.6, 95% CI [2.2,6.0]) was greater than for CMS eligibility (2.7, 95% CI [2.0,3.7]). In the 5,897 patients for whom CMS eligibility could not be determined, CXR-LC eligible patients had a 3-fold higher rate of lung cancer than CXR-LC ineligible patients (3.8% vs. 1.3%, p < 0.001). All CXR-LC eligible groups had 6-year lung cancer incidence >3.5%, above accepted screening risk thresholds of >1.5% and >2%. Conclusions: Using routine CXRs from the EMR, CXR-LC identified high-risk smokers who may benefit from LCS with CT, complementary to the 2022 CMS eligibility criteria. [Table: see text]