Multiview AI model reads whole coronary angiograms, predicts 1-year cardiac events

A vision-language model trained on more than 200,000 coronary angiography videos can interpret an entire angiographic study — not just a single frame or projection — and use those videos alone to estimate a patient's risk of a cardiovascular event within a year. The work, published in Nature Cardiovascular Research, was led by Robert Avram at the Montreal Heart Institute, Université de Montréal, with first author Sarra Harrabi.

It matters because invasive coronary angiography is the reference standard for diagnosing coronary artery disease, yet how severely a vessel is narrowed is judged by eye, with substantial variation between readers and even within the same reader over time. That variability contributes to overestimated stenoses and, in some cases, to unnecessary stenting or bypass surgery. An automated second read that is consistent and near-instant could help standardize those judgments.

DeepCORO-CLIP was trained on 203,808 angiography videos paired with catheterization reports from 28,117 patients (32,473 studies) at one high-volume Canadian center. It learns by contrastive video–text alignment — matching what a video shows with what the report describes — and then pools information across the multiple views acquired during a study.

For obstructive stenosis of 70% or more, the model reached a micro-AUROC of 0.89 internally and the same 0.89 at an external validation site (UC San Francisco), where sensitivity was 0.84 and specificity 0.77 despite a lower disease prevalence. Its negative predictive value exceeded 98%, meaning a normal prediction reliably excluded obstructive disease. It also identified chronic total occlusions (AUROC 0.90–0.95), intracoronary thrombus (0.90) and moderate-to-severe calcification (0.90–0.91). Against core-laboratory quantitative adjudication across 662 studies, the model's stenosis estimates were closer to consensus than the clinical reports themselves (mean absolute error 13.6% versus 19.0%).

The prognostic result is the more novel claim. Fine-tuned on 1,047 patients and tested on 350, DeepCORO-CLIP predicted a composite of urgent revascularization, nonfatal myocardial infarction or complete occlusion at one year with an AUROC of 0.79 (95% CI 0.74–0.84), versus 0.65 for a model built from routine clinical variables. Risk stratification split patients into three groups with hazard ratios of 5.6 and 18.6 for intermediate- and high-risk compared with low-risk. It also estimated left ventricular ejection fraction from the same videos with a mean absolute error of 7.3%, better than a previous angiography-based model (8.8%).

The clinical-readiness test was more sobering. In a prospective, blinded reader study covering 288 angiograms and 365 evaluations, the model's automatically generated reports were rated clinically usable in 76% of evaluations versus 88% for cardiologist reports. Crucially, the gap came from omission, not invention: flagged hallucinations were no more common in AI than in operator reports (17% versus 14%, P = 0.38). Missed findings clustered in proximal vessels and traced to the deliberately conservative reporting threshold of 0.50, which favors precision; loosening it recovers most missed lesions but roughly halves the precision of what gets flagged. Because those operating points were derived and evaluated in the same test set, the achievable gains are likely optimistic.

The limitations are substantial and the authors state them plainly. The design is retrospective; external validation covered a single center; patients with prior bypass surgery were excluded; and the cohort was 66.6% men, limiting confidence in women, who more often have nonobstructive disease. Performance fell off sharply in distal and side-branch segments, where severity correlation was near zero. The reader study used two readers at one center with only 77 doubly read studies. The model was deployed in a PACS-integrated pipeline with roughly four-second latency, but no trial has yet shown that using it changes procedural decisions or patient outcomes.

What to watch next is a larger, prospective, multireader evaluation with prespecified thresholds — and whether risk scores derived from angiography alone hold up over longer follow-up than one year. Code, weights and deployment infrastructure were released publicly.

Reference: Harrabi, S., Wu, Y., Majwal, O. et al. A multiview video–text model detects obstructive coronary stenosis and predicts 1-year cardiovascular events from angiography. Nature Cardiovascular Research (2026). https://doi.org/10.1038/s44161-026-00890-8