Monday, November 30 | 1:30 p.m.–1:40 p.m. | M6-SSCA03-1 | Room E352
In this session, a study will show that deep-learning models applied to routine chest x-rays can predict cardiovascular risk as accurately as the American Heart Association' standard PREVENT score.
Jay Chandra, MD, of Harvard University in Cambridge, will present a multistage study drawing on data from nearly 40,000 patients across the National Lung Screening Trial and the Mass General Brigham Biobank. The researchers trained convolutional neural networks to extract coronary artery calcium (CAC) scores and cardiac radiomics features from chest x-rays, and then compared the resulting risk scores against PREVENT’s 10-year major adverse cardiovascular event (MACE) predictions.
CXR-based models identified patients with any CAC with an AUROC of up to 0.80. In predicting incident MACE, the CXR-derived composite radiomics score and CXR-predicted CAC achieved concordances of 0.65 and 0.64, compared with 0.66 for PREVENT alone. Combined models reached 0.71. Among patients PREVENT classified as low-risk (below 5% 10-year MACE probability) those in the highest quartile of the CXR composite radiomics score had a fourfold higher rate of MACE than those in the lowest quartile. For CXR-predicted CAC quartiles, the difference was threefold.
PREVENT requires serum cholesterol and other lab values that are often unavailable in emergency departments or for patients with only episodic care, which limits its reach in opportunistic risk screening, the group noted. Chest x-rays are routinely obtained in those settings, making them a natural substrate for a lab-free risk tool, they wrote.
“CXR-derived CAC and a CXR-derived composite radiomics score are comparable to and improve cardiovascular risk stratification beyond the clinical standard PREVENT risk score,” Chandra and colleagues concluded.
Check out this session if you're interested in tracking AI’s expansion into preventive cardiovascular care.
