Validation of a Deep Learning-Based Model to Predict Lung Cancer Risk Using Chest Radiographs and Electronic Medical...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Lung cancer screening with CT scans can reduce mortality, but participation rates are low (<5%).
• Automated identification of high-risk individuals is crucial for improving screening uptake.
• Existing electronic medical record (EMR) data and chest radiographs offer a potential resource for risk assessment.

Purpose of the Study:

• To validate the CXR-LC deep learning tool for identifying high-risk individuals for lung cancer screening.
• To compare CXR-LC performance with the 2022 US Centers for Medicare & Medicaid Services (CMS) screening guidelines.
• To assess the utility of CXR-LC in complementing existing lung cancer screening eligibility criteria.

Main Methods:

• A prognostic study compared CXR-LC risk estimates with CMS guidelines using EMR data from a large US hospital system.
• Participants included current or former smokers with chest radiographs between 2013-2014, excluding those with prior lung cancer or CT screening.
• CXR-LC inputs included chest radiograph images, age, sex, and smoking status; data analysis occurred between May 2021 and June 2022.

Main Results:

• CXR-LC identified high-risk patients, with those eligible by both CXR-LC and CMS criteria showing a significantly higher lung cancer rate (8.5%) compared to CMS alone (2.8%).
• Even when CMS eligibility was unknown (57.4% of cases), CXR-LC identified patients with a 5-fold higher lung cancer incidence.
• The tool demonstrated effectiveness across various subgroups, including female patients and Black individuals.

Conclusions:

• CXR-LC, utilizing routine chest radiographs and EMR data, effectively identifies individuals at high risk for lung cancer.
• The tool can automate the identification of patients who may benefit from lung cancer screening CT.
• CXR-LC shows promise in enhancing lung cancer screening participation and complementing current guidelines.