How does a picture of your retina reveal heart risk?
The retina is the only place in the body where blood vessels can be seen directly without surgery. Changes in these tiny vessels—like narrowing, bleeding, or damage to the nerve fiber layer—often mirror damage happening in the heart and brain. Deep learning AI models can spot these subtle patterns that human eyes might miss. For example, one 2025 study showed that AI analyzing routine diabetic retinal photos predicted 10-year risk of heart attack or stroke as accurately as the standard clinical risk score (both had an AUC of 0.70) [1]. Another study found that localized defects in the retinal nerve fiber layer were linked to a 44% higher odds of having a high coronary artery calcium score (CACS ≥300), a strong marker of heart disease [6].
How accurate is retinal imaging for predicting heart disease?
Accuracy depends heavily on what you're trying to predict and which AI model you use. The best-case scenario: a 2025 multi-input deep learning model achieved 94% accuracy and an AUC of 0.96 for detecting CVD from retinal images in one dataset [5]. But that study used a case-control design (comparing known CVD patients to healthy controls), which tends to overestimate real-world performance. When predicting future events—the harder, more useful task—performance drops. A 2023 meta-analysis of 26 studies found that models predicting future heart attacks or strokes had AUCs between 0.68 and 0.81 [2]. The 2025 study in type 2 diabetes found an AUC of 0.70, matching the standard clinical score [1]. So while retinal imaging can work, it's not yet a magic bullet—it's roughly as good as current risk calculators, not dramatically better.
What are the caveats and limitations?
Several important limitations mean retinal imaging isn't ready to replace your doctor's risk assessment. First, most studies have been done in specific populations—like people with type 2 diabetes [1][3] or type 1 diabetes [4]—so results may not apply to everyone. Second, the technology is still evolving: a 2023 meta-analysis noted high variability in study designs and called for more real-world testing [2]. Third, not all retinal features are equally predictive. One 2025 study found that the standard atherosclerotic CVD risk score was NOT significantly associated with diabetic retinopathy severity [3], suggesting that different retinal changes signal different risks. Finally, even the best models still miss about 30% of events (sensitivity around 70%) [7], meaning many high-risk people would be missed. The largest and most cited study here—RETFound, trained on 1.6 million images—showed that AI can predict heart failure and heart attacks, but it was designed as a general-purpose model, not a dedicated CVD predictor [8].
About These Sources
This answer is built on 8 peer-reviewed studies — published from 2021 to 2025, 4 from 2024 or later, 3 in Q1 journals, collectively cited 737 times — selected as the most relevant from 11 studies that passed quality screening, drawn from 73 papers retrieved from a database of over 500 million.
Sources used in this answer
Deep-learning prediction of cardiovascular outcomes from routine retinal images in individuals with type 2 diabetes
In 6,127 people with type 2 diabetes, a deep-learning AI model predicted 10-year major cardiovascular events from routine retinal photos as accurately as the standard clinical risk score (both AUC 0.70), and combining them improved prediction (AUC 0.73) [1].
A Systematic Review and Meta-Analysis of Applying Deep Learning in the Prediction of the Risk of Cardiovascular Diseases From Retinal Images
A meta-analysis of 26 studies found that deep learning models predicting future CVD events from retinal images had AUCs between 0.68 and 0.81, while models detecting risk factors like diabetes achieved higher AUCs (0.80–0.96) [2].
Association between atherosclerotic cardiovascular disease risk and diabetic retinopathy in patients with type 2 diabetes mellitus
In 215 people with type 2 diabetes, the atherosclerotic CVD risk score was not significantly associated with diabetic retinopathy severity, though diabetes duration, HbA1c, and neuropathy were [3].
Associations of Microvascular Complications With the Risk of Cardiovascular Disease in Type 1 Diabetes
In the DCCT/EDIC study of type 1 diabetes, advanced microvascular disease (especially kidney disease with albuminuria) was associated with 2- to 3-fold higher risk of subsequent cardiovascular events over 35 years [6].
Eyes on the Heart: Detecting Cardiovascular Disease Using Retinal Images with Explainability
A multi-input CNN using retinal images and segmented maps of the optic disc, cup, and blood vessels achieved 94% accuracy and AUC 0.96 for detecting CVD in the EyePACS dataset [8].
Relationship between Retinal Nerve Fiber Layer Defects and Coronary Artery Calcium Score in Patients at Risk for Cardiovascular Disease
In 1,316 people without known CVD, localized retinal nerve fiber layer defects were associated with 44% higher odds of having a coronary artery calcium score ≥300, a marker of advanced atherosclerosis [9].
Predicting cardiovascular disease risk using retinal optical coherence tomography imaging.
Using retinal OCT images from 2,846 UK Biobank participants, a deep learning model predicted future heart attack or stroke within 5 years with AUC 0.75, sensitivity 0.70, and specificity 0.70 [10].
A foundation model for generalizable disease detection from retinal images
RETFound, a foundation model trained on 1.6 million unlabeled retinal images, outperformed comparison models in predicting heart failure and myocardial infarction from retinal photos, demonstrating generalizable AI for systemic disease detection [11].