Cardiac Function Predicted From Single PA Chest Radiograph Using AI-Based Model

Background: Chest radiography is widely available and very quick to obtain. The cardiovascular silhouette is easily visible but often overlooked in the workup of patients with cardiac disease. Objective: To create and test a deep learning–based model aimed at detecting the presence of valvular disease and cardiac dysfunction from a single posteroanterior (PA) chest radiograph. Methods: Standing PA chest radiographs were taken within 14 days of echocardiography at 4 institutions during a 4-year period. Patients with prior valvular surgery were excluded. A total of 22,551 chest radiograph and corresponding echocardiograms were evaluated. The chest radiographs were then labeled using the echocardiogram reports with respect to valvular disease (mitral stenosis, mitral regurgitation, aortic stenosis, aortic regurgitation, tricuspid regurgitation, and pulmonary regurgitation), left ventricular ejection fraction, and inferior vena cava dilatation. Severity of disease was classified ac more...