NEW ORLEANS — Almost one fourth of patients with no history of heart disease met the clinical guidelines for starting lipid-lowering therapy, according to a new prospective, multicenter randomized trial of an AI algorithm that analyzes retinal images to estimate cardiovascular risk.

The CLAiR AI trial used an investigative AI algorithm trained against the 10-year atherosclerotic cardiovascular disease (ASCVD) risk score of 7.5%, derived from the Pooled Cohort Equations introduced in the American College of Cardiology (ACC) and American Heart Association (AHA) 2013 cholesterol guidelines.

“In this cohort, 26% of those we enrolled had elevated cardiovascular risk, so it’s not like there’s a small percentage of people walking around at elevated risk,” Michael V. McConnell, MD, MSEE, a preventive cardiologist and professor at Stanford University in California, told Medscape Medical News.

Since McConnell’s group completed their study, presented at the American College of Cardiology (ACC) Scientific Session 2026, the 2026 ACC/AHA guideline recommended using the AHA PREVENT online risk calculator that lowered the 10-year ASCVD risk threshold to 5%.

Trial Design and Results

The trial enrolled people without a history of ASCVD who had retinal imaging at 10 eye care and primary care sites in the United States. Patients who were already on lipid-lowering therapy or had advanced eye disease were excluded. Other medical data, including blood pressure and lipid profiles, were collected to calculate the 10-year ASCVD risk score in 847 enrollees.

“We intentionally did not enroll people who were already on lipid-lowering therapy because we’re trying to help the more than 50% of people walking around who would benefit from preventive therapy but are not on it,” McConnell said.

The analysis found that 26.3% of enrolled patients had a 10-year ASCVD risk score of 7.5% or greater, McConnell said.

“The bottom line is that an AI system for analyzing retinal images has strong sensitivity and specificity for matching what had been until 2 weeks ago the standard risk score,” he said, referencing the Pooled Cohort Equations from the 2013 guidelines that had been supplanted by the PREVENT threshold in the 2026 guidelines.

The AI algorithm achieved a sensitivity of 91.1% (95% CI, 87.4-94.4) and a specificity of 86.2% (95% CI, 83.5-88.6), with positive and negative predictive values of 70.2% and 96.4%, respectively. The overall predictive performance of the algorithm, measured as area under the curve, was 96%, and upward of 94% of captured images across the 10 sites were suitable for evaluation.

What makes this study different is its prospective nature, according to McConnell. “The vast majority of all the AI studies around cardiovascular risk have been on datasets that have already been collected and then they run the AI,” he said.

The diversity of the study population was also noteworthy, he said. It was evenly divided between men and women. And 78.1% of enrollees were White, 19% were Black, and 26% were Hispanic. The mean BMI was 29.6.

Several studies have demonstrated the vessels in the back of the eye can be diagnostic for systemic disease. A German study this year demonstrated the accuracy of an FDA-approved retina imaging device that uses AI to diagnose and monitor diabetic retinopathy.

“One of the good, but challenging, things about AI for analyzing images, and in particular around cardiovascular risk, is there can be some obvious features to the human eye, butthere can be aspects that the AI is detecting that are not necessarily obvious to the human eye,” McConnell said.

Commentary

The prospective nature and multicenter design makes the study stand out from other studies using AI to analyze retina images to identify systemic disease, Dinesh Kalra, MD, chief of cardiovascular medicine at the University of Louisville in Louisville, Kentucky, who served as discussant of the study, told Medscape Medical News.

“Another big advantage that wasn’t publicized, but from my viewpoint is important, is that this study used a regular slit lamp camera that’s in an optometrist’s office, so you don’t need any dilating eye drops, and it takes about 20 seconds to get a picture of your eye,” Kalra said.

“It can be done during a routine office visit if a primary care doctor wishes to buy such a camera, or it can be done in an optometrist’s office. And most importantly, it doesn’t have any radiation and there’s no blood draw,” he added.

A larger study would be needed to further validate the use of AI-driven retina images for estimating the risk for ASCVD, Kalra said. “The output of the AI algorithm should not just be a risk calculator,” he said. “It should be the actual disease, so the output should estimate coronary plaque and, more important to patients, what is your likelihood of actually having heart attacks?”

The study was funded by Toku, developer of CLAiR AI. McConnell reported being the chief health officer at Toku. Kalra reported having no relevant financial relationships.

Richard Mark Kirkner is a medical journalist based in Philadelphia.