• Massachusetts General Hospital Cardiovascular Imaging Research Center, Brigham and Women's Hospital Program for Artificial Intelligence in Medicine, and Harvard Medical School, Boston, Massachusetts (M.T.L., V.K.R., U.H.).
• Harvard Medical School, Boston, Massachusetts,and Stralsund University of Applied Sciences, Stralsund, Germany (T.M.).
• Brigham and Women's Hospital Program for Artificial Intelligence in Medicine, Massachusetts General Hospital Cardiovascular Imaging Research Center, and Harvard Medical School, Boston, Massachusetts (H.J.A.).

BACKGROUND: Lung cancer screening with chest computed tomography (CT) reduces lung cancer death. Centers for Medicare & Medicaid Services (CMS) eligibility criteria for lung cancer screening with CT require detailed smoking information and miss many incident lung cancers. An automated deep-learning approach based on chest radiograph images may identify more smokers at high risk for lung cancer who could benefit from screening with CT.

OBJECTIVE: To develop and validate a convolutional neural network (CXR-LC) that predicts long-term incident lung cancer using data commonly available in the electronic medical record (EMR) (chest radiograph, age, sex, and whether currently smoking).

DESIGN: Risk prediction study.

SETTING: U.S. lung cancer screening trials.

PARTICIPANTS: The CXR-LC model was developed in the PLCO (Prostate, Lung, Colorectal, and Ovarian) Cancer Screening Trial (n = 41 856). The final CXR-LC model was validated in additional PLCO smokers (n = 5615, 12-year follow-up) and NLST (National Lung Screening Trial) heavy smokers (n = 5493, 6-year follow-up). Results are reported for validation data sets only.

MEASUREMENTS: Up to 12-year lung cancer incidence predicted by CXR-LC.

RESULTS: The CXR-LC model had better discrimination (area under the receiver-operating characteristic curve [AUC]) for incident lung cancer than CMS eligibility (PLCO AUC, 0.755 vs. 0.634; P < 0.001). The CXR-LC model's performance was similar to that of PLCO_M2012, a state-of-the-art risk score with 11 inputs, in both the PLCO data set (CXR-LC AUC of 0.755 vs. PLCO_M2012 AUC of 0.751) and the NLST data set (0.659 vs. 0.650). When compared in equal-sized screening populations, CXR-LC was more sensitive than CMS eligibility in the PLCO data set (74.9% vs. 63.8%; P = 0.012) and missed 30.7% fewer incident lung cancers. On decision curve analysis, CXR-LC had higher net benefit than CMS eligibility and similar benefit to PLCO_M2012.

LIMITATION: Validation in lung cancer screening trials and not a clinical setting.

CONCLUSION: The CXR-LC model identified smokers at high risk for incident lung cancer, beyond CMS eligibility and using information commonly available in the EMR.

PRIMARY FUNDING SOURCE: None.