A multiview video–text model detects obstructive coronary stenosis and predicts 1-year cardiovascular events from angiography

Invasive coronary angiography is the reference standard for diagnosing coronary artery disease, yet assessment is subjective, and existing artificial-intelligence tools typically assess stenosis in individual frames or videos and do not integrate the full multiview angiographic study. Here we show that DeepCORO-CLIP, trained on 203,808 angiographic videos and reports from 28,117 patients, interprets whole multiple-video angiograms at once. It detects obstructive stenoses (≥70%), with an area under the receiver operating characteristic curve (AUROC) of 0.89 internally and externally, and identifies chronic total occlusion (AUROC = 0.90 to 0.95), intra-coronary thrombus and calcification (AUROC = 0.90 to 0.91). It also predicts 1-year cardiovascular events (AUROC = 0.79, 95% CI of 0.74 to 0.84, versus 0.65 for a model derived using clinical variables) and left ventricular ejection fraction (mean absolute error 7.3% versus 8.8% for a previous model), both achieved using coronary angiography alone. In a blinded review of 365 evaluations across 288 angiograms, its reports were clinically usable in 76% of evaluations versus 88% for cardiologists, with comparable false positives (17% versus 14%, P = 0.38); residual missed stenoses reflected an adjustable threshold rather than an absent signal.