Patients walking into a standard eye appointment may soon walk out with an early warning about their heart. A new class of AI-powered software can analyze a single retinal photograph, the kind already taken during routine diabetes screenings, and flag elevated risk of heart attack or stroke years before symptoms surface. Multiple peer-reviewed studies now back the concept, and at least two named AI tools, Dr. Noon CVD and Reti-CVD, have reached the stage of formal validation or regulatory engagement. The question is no longer whether the technology works in a lab. It is whether it works reliably enough, across diverse patient groups, to change how primary-care doctors screen for cardiovascular disease.
Retinal AI screening fills a gap standard risk calculators miss
Traditional cardiovascular risk scores, such as pooled-cohort equations used in U.S. primary care, rely on a handful of inputs: age, cholesterol, blood pressure, smoking status, and diabetes diagnosis. Those tools were built on large epidemiological datasets but can miscalibrate for patients who fall outside the original study populations, including younger adults, some racial and ethnic minorities, and people with atypical risk profiles. When estimates are off, patients can be undertreated or exposed to unnecessary medication.
Retinal imaging offers a fundamentally different data stream. The tiny blood vessels at the back of the eye reflect systemic vascular health, and deep-learning algorithms can detect subtle patterns of vessel caliber, tortuosity, and microvascular damage that are invisible to clinicians. In a foundational study in retinal deep learning, Poplin and colleagues showed that a neural network could infer multiple cardiovascular risk factors, including age, blood pressure, and smoking status, directly from fundus photographs. That work established the scientific basis for an entirely new screening channel, one that requires no blood draw and no fasting, just a camera already present in most optometry offices.
The real test, though, is whether retinal AI risk scores align more accurately with actual rates of major adverse cardiovascular events (MACE, defined as cardiovascular death, myocardial infarction, and stroke) than existing calculators, especially in patients who do not yet have a diabetes diagnosis or other obvious red flags. No head-to-head prospective registry trial comparing retinal AI scores against pooled-cohort equations in a broad, non-diabetic primary-care population has been published. Until that comparison exists, the strongest claim supporters can make is that retinal AI adds information on top of conventional tools, not that it replaces them.
Two AI tools and the clinical data behind them
Two named software products have advanced farthest toward clinical use. Dr. Noon CVD is an AI software-as-a-medical-device designed to assess cardiovascular risk from retinal images. A peer-reviewed study in the Canadian ophthalmology journal focused specifically on the tool’s repeatability and reproducibility, measuring whether the same patient’s scans produce consistent risk-category outputs across sessions and across different imaging conditions. That kind of reliability testing is a prerequisite for any screening tool expected to guide clinical decisions, because a score that shifts unpredictably between visits would erode physician trust and complicate treatment planning.
Reti-CVD, a separate deep-learning-based retinal biomarker, underwent what its developers describe as a regulated pivotal validation using data from the CMERC-HI cohort, as detailed in the Journal of Clinical Medicine. In that analysis, Reti-CVD’s risk stratification was benchmarked against established cardiovascular markers within a well-characterized population, allowing investigators to compare how effectively each signal separated low-, intermediate-, and high-risk groups. Performance metrics such as discrimination and calibration suggested that the retinal biomarker captured prognostic information that was at least comparable to traditional measures and potentially complementary when combined with them.
A separate study in Cardiovascular Diabetology extended the concept to a real-world type 2 diabetes cohort, applying deep-learning prediction to routinely obtained diabetic retinal screening photographs to estimate time to first MACE event. Because these images were collected as part of standard care rather than a tightly controlled trial, the results speak more directly to how such algorithms might perform in busy clinical programs where image quality, patient demographics, and follow-up intervals vary. The findings supported the idea that retinal AI scores can meaningfully stratify patients by future cardiovascular risk even when the only new data added to the chart is a single fundus photograph.
These studies share a common thread: they show that algorithms trained on retinal images can extract cardiovascular signals that correlate with real clinical outcomes. But each study used a different patient population, a different imaging protocol, and a different definition of success. One emphasized technical repeatability, another regulatory-grade validation, and another long-term event prediction in diabetes. That fragmentation makes it difficult to compare results across tools or to generalize findings to the broader population that would encounter these scans during a routine eye exam.
Standardization gaps and the regulatory path ahead
The National Heart, Lung, and Blood Institute convened a workshop that produced a formal roadmap on standardization for retinal imaging biomarkers in cardiovascular disease. That document identified concrete evidence gaps: inconsistent imaging hardware, variable image quality across clinical settings, and the absence of large-scale prospective trials that test these tools in the populations where they would actually be deployed. It also highlighted the need for harmonized outcome definitions and shared reference datasets so that developers can benchmark new algorithms against common standards rather than siloed institutional cohorts.
Without standardized protocols, a retinal scan taken at one clinic could yield a different risk score than the same scan processed at another, even using the same algorithm. Differences in camera resolution, pupil dilation practices, and technician training can all affect the raw images fed into AI systems. For a technology that aims to influence lifelong preventive therapy-such as starting or intensifying statins or blood-pressure medications-these sources of variability must be tightly controlled or explicitly accounted for in model design and validation.
On the regulatory side, the FDA’s Breakthrough Devices Program offers a faster interaction pathway for qualifying medical devices that may provide more effective treatment or diagnosis for life-threatening conditions. Designation under that program can accelerate feedback cycles and clarify evidentiary expectations, but it does not equal clearance or approval. It signals that the FDA considers the device category promising enough to warrant priority review, not that any specific product has been authorized for clinical use. Patients and physicians should understand that distinction clearly, particularly as marketing claims around “breakthrough” AI tools proliferate.
For developers of retinal AI, the regulatory path will likely hinge on demonstrating three pillars: analytical validity (the algorithm consistently produces the same output from the same input), clinical validity (the output correlates with meaningful cardiovascular outcomes), and clinical utility (using the tool to guide care actually improves patient health compared with standard practice). Existing studies speak mostly to the first two pillars. The third-showing that integrating retinal risk scores into primary care changes prescribing decisions, patient behavior, or event rates-will require prospective implementation trials and careful attention to health equity.
Those equity questions loom large. If training datasets underrepresent certain racial or socioeconomic groups, the resulting models may systematically overestimate or underestimate risk for those patients. Because retinal imaging is already more common in diabetes programs than in general practice, there is also a risk that early deployments primarily benefit people who are already in specialty care, rather than those who seldom see a doctor but do visit an optometrist for glasses. Addressing these gaps will mean recruiting diverse cohorts, publishing subgroup performance metrics, and building workflows that allow primary-care teams to act on high-risk findings in timely, patient-centered ways.
For now, retinal AI for cardiovascular risk sits at an inflection point: technically impressive, backed by growing clinical evidence, but not yet woven into everyday care. If ongoing standardization efforts succeed and regulators gain confidence in the robustness of these tools, a routine eye photograph could become a powerful, low-friction gateway to heart-health screening-turning a window into the retina into an early warning system for the rest of the body.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
