Oculomics: Can Retinal AI Predict Heart and Kidney Disease?

Hitarth , B. Tech Computer Science & Engineering

• October 14, 2026 • 7 min read • 0 views

Last verified: October 24, 2026 Next review: April 24, 2027

Oculomics: Can Retinal AI Predict Heart and Kidney Disease?

Oculomics uses AI to read signs of whole-body disease in retinal images. Research models can estimate cardiovascular risk, flag chronic kidney disease and type 2 diabetes, and detect Alzheimer's dementia in study data. As of October 2026, none is FDA-cleared, and none is part of routine US care or reimbursed.

At a glance

Landmark study: Poplin et al., Nature Biomedical Engineering 2018: models trained on 284,335 patients predicted major adverse cardiac events with an AUC of 0.70
Chronic kidney disease: AUC 0.911 internally, 0.733 and 0.835 in external cohorts (Sabanayagam et al., Lancet Digital Health 2020)
Kidney disease and type 2 diabetes: AUC 0.85 to 0.93 from 115,344 fundus photographs (Zhang et al., Nature Biomedical Engineering 2021)
Alzheimer's dementia: AUC 0.93 internally, 0.73 to 0.91 externally, from 12,949 photographs (Cheung et al., Lancet Digital Health 2022)
Foundation model: RETFound trained on 1.6 million unlabelled retinal images (Nature, September 2023)
US regulatory status: No FDA-cleared device predicts systemic disease from a retinal image (FDA 510(k) and De Novo databases, October 2026)
Authorised outside the US: Reti-CVD: Korea MFDS, CE per manufacturer; CLAiR: CE and UKCA per Toku; CLAiR holds FDA Breakthrough designation (2023)

On this page

What Is Oculomics?
What Does the Research Show So Far?
Heart and blood vessels
Kidneys and diabetes
Brain
Foundation models
Research under way
Is Any Oculomics Tool Cleared?
Where Is the Evidence Still Weak?
What Can Your Practice Do Now?
How Should You Judge an Oculomics Claim?
Our Recommendation
Sources

Oculomics is real science, but it is not yet a clinical service you can offer in the United States. Deep learning models trained on retinal photographs can estimate cardiovascular risk, detect chronic kidney disease and type 2 diabetes, and separate people with Alzheimer's dementia from those without, in research datasets. As of October 2026 we found no device cleared by the FDA to predict systemic disease from a retinal image, although two cardiovascular-risk tools are authorised outside the US. This guide sets out what the studies show, what is cleared where, and what your practice can sensibly do now.

What Is Oculomics?

Oculomics is the use of eye data, mostly retinal images, to find markers of disease elsewhere in the body. The retina is the one place where a clinician can see blood vessels and nerve tissue directly, without an incision or a dye. Optometrists have always spotted hypertensive and diabetic changes this way.

What is new is artificial intelligence. A deep learning model, a type of AI trained on many labelled examples, can find patterns in a fundus photograph that no human grader has described. That is a different job from the diabetic retinopathy tools you may already know. The systems in our comparison of AI retinal screening software detect an eye disease in the eye. Oculomics models use the eye to estimate the risk of heart, kidney or brain disease.

What Does the Research Show So Far?

The figures below are as each study reported them. The AUC (area under the curve) measures how well a model separates people with and without a condition: 0.5 is chance and 1.0 is perfect.

Heart and blood vessels

The landmark paper came from Google researchers. Poplin and colleagues, in Nature Biomedical Engineering in 2018, trained models on data from 284,335 patients and validated them on two independent datasets of 12,026 and 999 patients. From fundus photos alone, the models estimated age to within 3.26 years, sex with an AUC of 0.97 and systolic blood pressure to within 11.23 mmHg. They predicted major adverse cardiac events with an AUC of 0.70, which is modest for a clinical decision.

Reti-CVD, developed by researchers in Singapore and South Korea, sorts people into three cardiovascular risk tiers from one retinal photograph. In a UK Biobank validation in BMC Medicine (2023) of 48,260 people with no history of cardiovascular disease, the high-risk tier carried a 13.1% 10-year risk. Adding Reti-CVD to the QRISK3 score raised its C-statistic by only 0.013 to 0.023. A pivotal study of 1,106 patients in JAMIA (2023) supported its authorisation in Korea.

Kidneys and diabetes

Sabanayagam and colleagues (Lancet Digital Health, 2020) developed a model on 5,188 Singapore patients to detect chronic kidney disease. Its AUC was 0.911 in internal validation but fell to 0.733 and 0.835 in two external groups from Singapore and Beijing. In those external groups, a simple model using age, sex, ethnicity, diabetes and hypertension did as well or better, at 0.829 and 0.887. Zhang and colleagues (Nature Biomedical Engineering, 2021) trained on 115,344 fundus photographs from 57,672 patients and reported AUCs of 0.85 to 0.93 for chronic kidney disease and type 2 diabetes.

Brain

A multicentre study by Cheung and colleagues (Lancet Digital Health, 2022) used 12,949 retinal photographs from 648 people with Alzheimer's disease and 3,240 without. The model reached an AUC of 0.93 in internal validation and 0.73 to 0.91 in external test sets. It was a case-control design, which compares people already diagnosed with healthy controls. That design tends to flatter accuracy.

Foundation models

RETFound, published in Nature in September 2023, was trained on 1.6 million unlabelled retinal images. It can be adapted to new tasks with fewer labelled examples, and its authors report better prediction of heart failure and myocardial infarction than comparison models. Foundation models matter because labelled data is the main bottleneck in this field.

Research under way

The cohorts behind the studies above come mainly from the UK, the US and East Asia. The Africa Oculomics Research Programme, a Wellcome Discovery Award made in 2025 and led from the London School of Hygiene & Tropical Medicine with Ugandan partners, will repeat eye imaging five years apart in a Ugandan population cohort and build AI models for African populations. Optometry is engaging too. The American Academy of Optometry devoted its 2026 Monroe J. Hirsch Symposium to oculomics, and its moderator, Andrew Pucker, OD, PhD, told Optometry Times that some of the work is "not quite ready for clinic".

Is Any Oculomics Tool Cleared?

Not in the United States. When we searched the FDA's 510(k) database and De Novo database in October 2026, we found no device authorised to predict cardiovascular, kidney or neurological disease from a retinal image. Every retinal AI device the FDA has cleared detects eye disease, chiefly diabetic retinopathy; our list of FDA-cleared AI devices for eye care tracks them.

Outside the US, the picture differs:

Reti-CVD, sold as DrNoon for CVD, holds marketing authorisation from Korea's Ministry of Food and Drug Safety and, according to its manufacturer, CE certification under the EU Medical Device Regulation. That is reported in a 2026 review in the Journal of Clinical Medicine, which adds that Korean patients pay for it out of pocket and that it has no FDA authorisation.
CLAiR from Toku estimates cardiovascular risk from a retinal image. The company states it has CE and UKCA marks and is not available in the United States. The FDA granted it Breakthrough Device designation in November 2023, which speeds review but is not a clearance.

So no oculomics test is part of routine US care, and none is reimbursed as a heart, kidney or brain screen. Bill retinal photography for the eye findings it documents, nothing more.

Where Is the Evidence Still Weak?

Performance drops outside the training data. The kidney model above lost ground in external cohorts, where a risk-factor model matched it. A 2026 scoping review in Frontiers in Digital Health found external validation limited and calibration inconsistently assessed in retinal AI studies of type 2 diabetes.
Bias. The same review found that analyses of fairness and of differences between cameras were rarely reported. A model trained on one population or one camera may misjudge another.
The added value is small so far. Reti-CVD improved QRISK3 discrimination by about 0.01 to 0.02, and the 2026 Journal of Clinical Medicine review notes that no randomised or outcome-based study has shown that retina-guided care improves outcomes.
Explainability. Poplin's team showed its models attended to the optic disc and blood vessels, but a clinician still cannot see why a given image scored high.
Consent and data governance. Research relies on large image archives linked to health records. Patients consented to an eye exam, not necessarily to their images training a commercial model.

What Can Your Practice Do Now?

You cannot offer oculomics yet, but you can make your images and records ready for the day a validated tool arrives.

Capture good images consistently. Models are trained on well-exposed, centred, gradable photos and perform worse on poor ones. Standardise your fields, such as macula- and disc-centred views, and record ungradable images. Our guide to digital retinal cameras for optometry compares the hardware.
Archive originals, not screenshots. Keep full-resolution images with the date, eye and camera recorded, in a standard format such as DICOM where your system supports it.
Record systemic data in structured fields. Blood pressure, diabetes status, smoking and medications belong in coded fields, not free-text notes.
Plan the route to primary care. A risk result is only useful if it reaches the patient's physician. Our article on EHR interoperability for optometrists covers how to exchange records with medical systems.
Review your consent language. If you might share de-identified images for research or product development, say so plainly in your intake forms and privacy notice, and take HIPAA advice first.

How Should You Judge an Oculomics Claim?

Expect heavy marketing as camera makers and software firms enter this field. Ask five questions before you buy or promote anything:

What is the exact cleared indication, in which country and with which camera? Breakthrough designation and a CE mark are not FDA clearance.
Was the model validated prospectively, on an external population like your patients?
How much does it add to a standard risk score built from blood pressure and cholesterol?
Who receives a high-risk result, and who acts on it?
Where are images processed and stored, and may the vendor use them to train its models?

Our Recommendation

Treat oculomics as research to follow, not a service to sell. If you want to be ready, invest in image quality, clean archives and a working link to primary care, because those pay off now. Revisit the question when a device gains FDA clearance with outcome evidence behind it. Until then, the proven AI for your practice is diabetic retinopathy screening, and our AI retinal screening comparison is the next article to read.

Sources

Frequently Asked Questions

An eye exam can show signs linked to cardiovascular risk, such as hypertensive changes in the retinal vessels. AI research models go further and estimate cardiovascular risk from a retinal photo, but none is FDA-cleared as of October 2026. A cardiovascular diagnosis still needs a physician.

Not in the US. No oculomics test is FDA-cleared, so none is reimbursed as a heart, kidney or brain screen. In Korea, Reti-CVD is offered as an out-of-pocket test. Retinal photography for eye disease is billed as usual.

Diabetic retinopathy AI detects an eye disease in the eye, and several such systems are FDA-cleared. Oculomics uses the eye to estimate the risk of disease elsewhere, such as heart or kidney disease. In the US that is still research.

Not yet, because no oculomics tool is available in the US. Most research models use standard colour fundus photographs. A good non-mydriatic camera and consistent, gradable capture are the best preparation.

Only with a lawful basis: properly de-identified images or the patient's written authorisation, plus a data use agreement. Tell patients in plain language how images may be used, and take HIPAA advice before you share anything.