AI Reads Whole Coronary Angiograms and Predicts Heart Risk a Year Ahead

An artificial intelligence system trained on more than 200,000 coronary angiography videos can now interpret an entire heart catheterization in one pass, detecting dangerously narrowed arteries, spotting chronic total occlusions, blood clots and calcification, and even forecasting a patient’s risk of a major cardiac event over the following year. The model, called DeepCORO-CLIP, was developed by researchers at the Montreal Heart Institute and collaborators at the University of California, San Francisco and Cedars-Sinai, and is described in a study published in Nature Cardiovascular Research. Its creators say it represents a shift away from narrow, single-task AI tools toward a general-purpose foundation model for one of medicine’s most commonly performed and most subjective procedures.

Coronary angiography, in which dye is injected into the heart’s arteries while X-ray video is recorded, remains the reference standard for diagnosing coronary artery disease, the world’s leading cause of death. Yet the interpretation of those videos is strikingly inconsistent. Visual assessment of stenosis, the narrowing of a vessel, varies between observers by anywhere from 7.6 to 22.5 percent, and the problem is worse when a single physician reads the study alone. Prior research cited by the team suggests that overestimation of stenosis severity occurs in more than a quarter of cases, potentially contributing to inappropriate coronary interventions in roughly 17 percent of bypass surgery cases and 10 percent of stent procedures. That variability has long motivated efforts to bring objective, reproducible measurement into the catheterization laboratory.

Earlier AI systems made progress but were limited in scope. CathAI, a previous model from some of the same investigators, achieved an area under the receiver operating characteristic curve, or AUROC, of 0.86 for detecting epicardial stenoses, while a follow-up video model called DeepCORO showed that temporal information across frames improves analysis. But these tools processed single videos or frames, relied on multistep pipelines rather than end-to-end training, and were optimized for one task at a time. DeepCORO-CLIP takes a fundamentally different approach, borrowing from vision-language models such as CLIP and their medical descendants like EchoCLIP and EchoPrime in echocardiography. Instead of learning from labeled examples of a single disease, it learns a shared representation between angiographic video and the free-text reports cardiologists write, allowing one encoder to support many downstream tasks.

The training data were enormous by the standards of cardiac imaging. The team screened 940,366 video-report pairs and, after excluding videos contaminated by procedural equipment, lacking contrast, or acquired after stenting or in patients with prior bypass surgery, assembled 203,808 videos from 28,117 patients, spanning 32,473 studies and 56,299 reports. The cohort, drawn from seven years at one of Canada’s highest-volume catheterization centers, had a mean age of 67.3 years and was 66.6 percent male, with a heavy burden of hypertension, diabetes and prior intervention. In the first training phase, a Multiscale Vision Transformer encoder learned to align each video with its paired report text using a contrastive objective. In the second phase, the pretrained encoder was fine-tuned with a multi-instance learning framework that uses gated attention to aggregate information across all of a study’s videos, generating predictions for 18 coronary segments across four tasks simultaneously.

The diagnostic results were strong. For detecting obstructive stenosis of 70 percent or greater, the model achieved a micro-averaged AUROC of 0.89 both internally and in external validation at UCSF, with sensitivity of 0.85 and specificity of 0.70 at the operating point chosen. It identified chronic total occlusion with AUROCs of 0.90 to 0.95, and detected intracoronary thrombus, a rare finding present in only 0.25 percent of segments, with an AUROC of 0.90. Moderate-to-severe calcification was graded with AUROCs of 0.90 to 0.91. Notably, the contrastively pretrained encoder outperformed a baseline video model initialized on the general-purpose Kinetics-400 dataset across every task, with the largest gains on rare phenotypes, and it beat a latent-predictive alternative encoder on stenosis, calcification and occlusion detection despite that competitor having ten times more parameters. Multiview attention pooling consistently improved accuracy over analyzing single videos or simply averaging their outputs.

Perhaps the most striking finding is prognostic. When fine-tuned on 1,047 patients with follow-up data and tested on 350 others, the model predicted composite major adverse cardiovascular events, including urgent revascularization, nonfatal heart attack and complete coronary occlusion, within one year with an AUROC of 0.79, compared with 0.65 for a logistic-regression model built from conventional clinical variables. Risk stratification separated patients into three groups with hazard ratios of 5.6 and 18.6 for the intermediate and high-risk groups relative to the low-risk reference. Remarkably, adding clinical variables to the model’s risk score did not improve discrimination, suggesting that the angiographic video itself, capturing disease extent, lesion severity and territorial distribution, already contains most of the prognostic signal. The model also estimated left ventricular ejection fraction from coronary angiograms alone with a mean absolute error of 7.3 percent, bettering the 8.8 percent of a previous dedicated model.

Against expert consensus, the machine held its own. In a core laboratory adjudication of 662 studies and 965 lesions, where two interventional cardiologists independently quantified stenosis using quantitative coronary angiography, DeepCORO-CLIP’s predictions deviated from the adjudicated values by a mean of 13.6 percentage points, compared with 19.0 points for the original clinical reports. The clinical reports turned out to be systematically conservative, missing nearly half of QCA-positive lesions, while the model’s errors were more evenly distributed. In error analysis across nearly 4,800 test catheterizations, the model correctly flagged at least one treated vessel in 85.8 percent of patients who underwent stenting, and a normal prediction reliably excluded obstructive disease, with negative predictive value exceeding 98 percent across all segment categories. Its main weakness was distal and branch vessels, where small caliber limits visualization and severity correlation collapsed toward zero.

The team also tested the model in the messiest setting of all: writing reports. In a prospective, blinded reader study, two independent cardiologists evaluated 365 interpretations across 288 angiograms, mixing AI-generated reports with those written by the operating physicians. Readers judged the AI reports clinically usable in 76 percent of evaluations versus 88 percent for physicians, with the gap driven almost entirely by omissions rather than fabrication. Hallucinated findings appeared no more often in AI reports than in human ones, 17 percent versus 14 percent, a difference that was not statistically significant. The unacceptable AI reports almost always involved missed proximal stenoses, and a threshold analysis showed those misses reflected the deliberately conservative 0.50 probability cutoff used by the report generator rather than a blind spot in the model itself. Loosening the threshold recovered most missed lesions but roughly halved precision, a trade-off the researchers have left as a user-configurable parameter in the released code.

Deployment is already underway. Using a bespoke platform called PACS-AI, angiography studies stream from the catheterization laboratory to the hospital’s picture archiving system, where DeepCORO-CLIP runs automatically with an end-to-end latency of 4.2 seconds, fast enough to inform decisions during the procedure itself. The model weights are freely available under an open license, and the code is released for noncommercial research. The authors are careful about limitations: the study was retrospective, external validation covered a single additional center, patients with prior bypass surgery were excluded, and the cohort was predominantly male. Follow-up for event prediction was also short. Still, the trajectory is clear. A single model, trained on routine clinical data without task-specific labels, now reads a complete multiview angiogram, quantifies disease, characterizes lesions, estimates heart function, predicts one-year risk and drafts a report in seconds. Whether it ultimately reduces inappropriate interventions or simply becomes a tireless second reader, the era of the general-purpose angiography interpreter has arrived.

Subject of Research: A multiview video-text AI model for interpreting coronary angiography to detect stenosis and predict cardiovascular events

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

Article References: A multiview video–text model detects obstructive coronary stenosis and predicts 1-year cardiovascular events from angiography. (n.d.). https://doi.org/10.1038/s44161-026-00890-8

Image Credits: AI Generated

DOI: 10.1038/s44161-026-00890-8

Keywords: coronary angiography, artificial intelligence, vision-language model, coronary artery disease, stenosis detection, cardiovascular risk prediction, deep learning, interventional cardiology, chronic total occlusion, left ventricular ejection fraction, foundation model, clinical AI deployment