The current state of oculomics in optometry: Commonly asked questions answered

How can retinal vasculature act as a “window” into cardiovascular health?

The retina provides a uniquely accessible view of the body’s microvasculature, allowing a window to systemic vascular health in real time. Because retinal vessels share embryologic origin, structural characteristics, and regulatory mechanisms with the microvasculature throughout the brain, heart, and kidneys, changes in retinal vessel caliber, architecture, and perfusion often mirror broader cardiovascular pathology.

It has been demonstrated that retinal vascular signs—specifically arteriolar narrowing and venular widening—predict future coronary heart disease. Beyond simple vessel caliber, network‑level features such as reduced fractal dimension and increased tortuosity reflect microvascular remodeling and have been linked to higher risks of cardiovascular disease and heart failure.

In diabetes, the retina’s predictive power is even more pronounced. Both the presence and severity of diabetic retinopathy correlate with elevated rates of cardiovascular events, kidney disease progression, and all‑cause mortality. Emerging artificial intelligence (AI)–derived biomarkers, such as the retinal age gap, further strengthen this connection by quantifying microvascular aging; larger gaps consistently associate with worse cardiometabolic outcomes.

Advanced imaging modalities such as optical coherence tomography angiography (OCTA) deepen this relationship by capturing microvascular perfusion directly. Findings from OCTA studies show that vessel density loss and foveal avascular zone enlargement occur in hypertension, coronary artery disease, heart failure, and stroke. Longitudinal data indicate that sustained reductions in vessel density—even by a few percentage points—are associated with increased rates of major adverse cardiovascular events, suggesting that retinal perfusion may serve as a dynamic biomarker of cardiovascular risk.

AI has accelerated this field further. AI models trained on retinal photographs can predict cardiovascular risk factors, incident cardiovascular disease, and even mortality. When combined with clinical variables, AI information becomes even more powerful. This reinforces the concept that the retina encodes systemic vascular information that can be extracted computationally. In sum, the retina acts as a window to cardiovascular disease because its microvasculature reflects systemic vascular health, its perfusion changes track cardiometabolic dysfunction, and its imaging biomarkers—especially when enhanced by AI—offer powerful, noninvasive insights into cardiovascular risk long before clinical events occur.

Which retinal imaging modalities (OCT, OCTA, fundus photography, adaptive optics) are currently most useful for detecting systemic disease biomarkers?

There isn't one universal winner. Their usefulness depends on the biomarker being sought. And honestly, using them together to gain value is most useful.

Fundus photography

For systemic oculomics, color fundus photography currently has the strongest practical position because it is inexpensive, widely available, and relatively standardized and produces images that can be analyzed by AI. It has produced evidence for cardiovascular, metabolic, renal, and other systemic associations. Reviews of AI o­culomics have found that most studies about systemic disease to date have used fundus photographs.

OCT

OCT adds 3-dimensional structural information and is particularly valuable when systemic disease affects retinal or optic-nerve structure. Retinal nerve-fiber and ganglion-cell measurements have been investigated extensively in neurodegenerative and neurologic disease.

OCTA

OCTA is especially interesting for cardiovascular oculomics because it provides quantitative information about the retinal microcirculation without dye injection. However, device-to-device differences, acquisition protocols, segmentation errors, and analysis pipelines remain major problems when trying to compare imaging.

Adaptive optics

Adaptive optics can visualize individual photoreceptors, retinal pigment epithelial cells, and much finer vascular structures. Its limitation is practical: Equipment is specialized, acquisition and analysis are more demanding, and standardized clinical end points are still developing. So it is currently more valuable for mechanistic research and biomarker discovery than routine systemic screening.

The bottom line: For a primary care/optometry-facing oculomics program today, I'd rank fundus photography first, OCT/OCTA next, and adaptive optics primarily as an investigational modality. Multimodal systems ultimately outperform any single modality in most disease states (Table).

At what point should an optometrist refer a patient to a physician based on an oculomics-derived systemic risk flag vs continuing routine monitoring?

As oculomics tools enter clinical practice, one principle must remain central: An AI‑derived retinal risk flag is a prompt for medical evaluation, not a diagnosis. Retinal biomarkers can reveal systemic risk, but they do not replace conventional cardiovascular or cerebrovascular assessment. At present, there is no universally accepted threshold—no “AI cardiovascular risk score > x = refer to cardiology.” The evidence base is not mature enough to support such deterministic rules, and clinically actionable guidance for cerebrovascular risk remains sparse despite substantial research progress.

A practical, defensible referral framework for optometrists can be organized into 3 tiers:

High‑risk or urgent findings → immediate medical assessment

Referral should not wait for routine follow‑up when an AI‑derived risk signal appears alongside clinical findings that already warrant urgent evaluation. These include markedly elevated measured blood pressure, optic disc edema, severe hypertensive retinopathy, retinal vascular occlusion, acute neurologic symptoms, chest pain, dyspnea, or other signs of end‑organ injury. Historically, severe hypertension accompanied by severe hypertensive retinopathy has been treated as a medical emergency requiring immediate systemic management. In these situations, the AI flag is not the driver of referral; it is an additional indicator reinforcing the need for urgent care.

Persistent systemic-risk signal → primary-care referral

For asymptomatic patients, an AI oculomics flag becomes more clinically meaningful when it is reproducible, credible, and contextually supported. Referral to primary care is reasonable when:

• The finding is consistent on repeat imaging.
• Image quality is adequate.
• The model used has external validation.
• The patient has additional cardiovascular or metabolic risk factors.
• The flagged pattern corresponds to a potentially actionable condition.
• The output falls outside the model’s expected uncertainty range.

In these cases, the optometrist’s role is to communicate risk without overstating certainty. A safe and clinically appropriate phrasing might be: “This retinal analysis suggests increased vascular risk. It does not diagnose cardiovascular disease, but I recommend that your primary‑care clinician review your blood pressure, lipids, glucose/HbA_1C [hemoglobin A_1C], and overall cardiovascular risk.”

This approach aligns with existing guidance. The American Optometric Association’s diabetes guideline, for example, recognizes that retinal disease can signal underlying vascular pathology and recommends considering cardiovascular disease, hypertension, and smoking status when retinopathy is present.

Low‑confidence or isolated AI signal → confirm and monitor

If the patient is asymptomatic, conventional measurements are normal, and the AI model produces an isolated, low‑confidence signal, immediate specialist referral is generally not warranted. In these cases, the most defensible approach is:

1. Repeat imaging to confirm the finding
2. Ensure routine primary‑care screening is up to date
3. Monitor for changes over time

This avoids overreferral based solely on algorithmic output, especially when the model’s uncertainty is high or the signal is not clinically corroborated.

The core principle

Across all scenarios, the guiding sequence should be: AI flag → clinician verification → conventional medical assessment → referral based on established clinical risk. Oculomics tools can enhance systemic risk detection, but they must be integrated into clinical reasoning—not replace it. As evidence matures, thresholds may become more standardized. For now, optometrists should treat AI‑derived risk signals as valuable prompts that inform, but do not dictate, medical referral.

What are the current limitations of AI-based oculomics tools in terms of accuracy, bias, and generalizability across diverse populations?

The biggest barrier: Accuracy can look better than it really is.

Perhaps the biggest barrier to widespread clinical oculomics is not whether AI can detect signals in retinal images but whether those signals remain reliable when the technology is used in the real world. Many published studies are retrospective, case-control studies or use highly selected patient populations. A systematic review of 31 deep-learning studies predicting systemic conditions from fundus images found that 96.8% were considered at high risk of bias. Only 16.1% used prospective external validation cohorts, and just 6.4% reported calibration.¹

For an oculomics system to be clinically useful, it needs more than good discrimination. It should also demonstrate good calibration, prospective validation, external validation in independent populations, clinically appropriate sensitivity and specificity, meaningful positive and negative predictive values, evidence that its predictions actually improve patient outcomes, and an assessment of referral burden, false positives, and downstream clinical consequences.

Data set shift: The real world is different

A model trained and tested in one population may encounter very different data in clinical practice. Performance can change because of differences in:

• ethnicity and ancestry
• age distribution
• socioeconomic status
• disease prevalence
• camera manufacturer
• image resolution
• field of view
• image illumination
• clinical setting
• patient selection

This problem, often referred to as data set shift, is particularly important for oculomics because retinal imaging is influenced by both biological and technical factors. Recent reviews of oculomics, therefore, emphasize the need for prospective, multicenter validation and systematic auditing of performance across different populations before widespread clinical adoption.

Bias as a clinical safety issue

If a demographic group is underrepresented in the training data, an AI system may perform differently in that population. The consequences can extend beyond questions of fairness.

If false-negative rates are higher in a particular group, patients from that population could be less likely to receive appropriate investigation or referral. In other words, bias in medical AI can become a patient safety problem. Fairness in medical imaging AI has therefore become an important area of research, with systematic reviews highlighting substantial methodological challenges in measuring and addressing demographic differences in model performance.

Shortcut learning: What is the AI actually seeing?

Another concern is shortcut learning. An AI system may identify a feature that happens to correlate with a disease without detecting the underlying biological signal. It might learn characteristics associated with a particular imaging device, hospital, clinical workflow, or patient population rather than the disease itself.

This raises a fundamental question: Is the model detecting disease biology—or recognizing characteristics of the data set in which the disease was diagnosed?

Interpretability methods can provide useful clues, but they are not a definitive answer. Even commonly used approaches such as saliency maps can be misleading, and research in oculomics has shown that conventional interpretability methods may themselves introduce erroneous conclusions.

Generalizability remains the test

The emergence of large foundation and multimodal models is one of the most exciting developments in oculomics. These models may eventually improve performance across multiple diseases and clinical settings.

But larger models should not be confused with universal generalizability. The real test is whether a model continues to perform reliably when it is taken outside the population, imaging system, hospital, and clinical workflow in which it was developed. For oculomics to move from promising research to routine clinical medicine, the field will need to demonstrate not simply that AI can extract information from the retina but that this information is accurate, calibrated, equitable, reproducible, and clinically useful across real-world populations.

What data privacy or informed consent considerations arise when patient retinal images are used to train or validate AI models?

Retinal images are not simply anonymous photographs. They are health data, and when linked with an electronic health record (EHR), demographics, diagnoses, or identifiers, they can constitute protected/identifiable health information.

In the United States

Under the Health Insurance Portability and Accountability Act (HIPAA), identifiable health information held by a covered entity can be protected health information (PHI). The US Department of Health and Human Services (HHS) permits research use of appropriately deidentified information, whereas identifiable PHI generally requires an appropriate authorization or an applicable institutional review board/privacy board waiver.²

For AI research, therefore, institutions should address:

• whether images are identifiable
• whether accompanying EHR data are identifiable
• how images are deidentified
• who holds the reidentification key
• where data are stored
• who can access them
• whether data will be shared with outside companies
• whether images will be used for future research
• whether trained models themselves could leak information

Importantly, deidentification reduces but does not necessarily eliminate reidentification risk. HHS explicitly notes that the residual risk is not zero.³

Consent should be specific enough: A particularly important issue for oculomics is secondary use.

A patient may consent to: “Your retinal photograph will be taken to assess your eye health.” That does not necessarily communicate: “Your image and associated clinical information may be stored indefinitely and used to train commercial AI systems for diseases unrelated to your eye examination.” For research repositories, HHS guidance recognizes broad consent as one possible mechanism for future research involving identifiable data, but the consent should describe the types of research, information that may be used, potential sharing, types of researchers/institutions involved, and anticipated storage/use duration.⁴

HIPAA authorization and informed consent in research are also not identical concepts: HIPAA authorization concerns use/disclosure of PHI, whereas informed consent concerns participation in the research itself. They can, however, be combined into one appropriate document.⁵

References

1. Li Y, Zhang R, Dong L, et al. Predicting systemic diseases in fundus images: systematic review of setting, reporting, bias, and models’ clinical availability in deep learning studies. Eye. 2024;38:1246–1251. https://doi.org/10.1038/s41433-023-02914-0

2. Research. US Department of Health and Human Services. Updated July 26, 2013. Accessed August 18, 2026. https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/research/index.html

3. Guidance regarding methods for de-identification of protected health information in accordance with the Health Insurance Portability and Accountability Act (HIPAA) privacy rule. US Department of Health and Human Services. Updated February 3, 2025. Accessed August 18, 2026. https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html

4. Attachment C - updated FAQs on informed consent for use of biospecimens and data. US Department of Health and Human Services. Updated April 13, 2018. Accessed August 18, 2026. https://www.hhs.gov/ohrp/sachrp-committee/recommendations/attachment-c-faqs-recommendations-and-glossary-informed-consent-and-research-use-of-biospecimens-and-associated-data/index.html

5. Do the HIPAA Privacy Rule’s requirements for authorization and the Common Rule’s requirements for informed consent differ? US Department of Health and Human Services. Updated January 9, 2023. Accessed August 18, 2026. https://www.hhs.gov/hipaa/for-professionals/faq/313/does-the-hipaa-requirement-for-authorization-differ-from-the-common-rule/index.html