An old idea, a new reality

The notion that the eye reflects the body’s overall health is not new. Hypertensive retinopathy, diabetic macular edema, optic disc pallor noted in multiple sclerosis: ophthalmologists have long appreciated that what they see through the slit-lamp is a reflection of pathology far beyond the orbit. What is new is the scale and the precision with which artificial intelligence is formalizing that relationship, and the speed with which it is rapidly evolving from research observation to clinical tool.

Oculomics – the use of ocular biomarkers analyzed by AI to detect and predict systemic disease - is no longer a speculative field. Mediwhale’s Reti-CVD, an AI system that stratifies cardiovascular disease risk directly from retinal fundus images has been approved in the UK, South Korea, Singapore, and Australia – with recent integration into cardiovascular risk management protocols (1). Similarly, Toku’s CLAiR platform evaluates cardiovascular risk from standard retinal photographs and has both CE and UKCA marks (1), while RetiSpec has developed a clinically validated AI system which predicts amyloid burden from retinal imaging (1) – allowing for screening for Alzheimer’s disease without lumbar puncture, PET scanning, or plasma biomarker assays. These are not research prototypes; these are controlled medical devices for active use.

What the retina reveals

The central question for ophthalmologists is not whether oculomics is coming, it is already here. The question is what it means for the specialty and whether we are ready for the answer. The scientific basis for oculomics rests on the retina’s anatomical properties; it is the only location in the body where the microvasculature and neural tissue can be visualized non-invasively, at high resolution, and in real-time. The microvascular architecture of the retina – vessel caliber, tortuosity, branching patterns, capillary density – is a mirror of systemic vascular health with a fidelity hitherto only available through invasive investigations.

AI models trained on population-scale retinal imaging datasets have now demonstrated clinically meaningful performance across a remarkable range of systemic conditions. In Cardiovascular medicine, deep learning analysis of fundus photographs can estimate coronary artery calcium scores, offering a radiation-free alternative to CT calcium scoring. A 2025 Theranostics review by Wang et al., applied radiomic features to multimodal retinal images (by combining OCT and OCTA parameters combining capillary density and retinal layer thickness), achieving near-perfect accuracy in predicting cardiovascular risk in type 1 diabetic populations (2).

A similar pattern is emerging for renal disease. A 2023 study by Joo et al., externally validated a deep learning-derived Reti-CKD score against the UK Biobank and a Korean diabetic cohort, and found that patients in the highest risk quartile were between three-and-a-half and nine times more likely to develop chronic kidney disease than those in the lowest, outperforming the eGFR based blood tests used in routine practice today (3). Similar to Reti-CVD, the same pattern holds: a retinal photograph doing the diagnostic work of a laboratory panel – in patients whose kidney function currently looks unremarkable on paper.

AI analysis of retinal microvascular in cases of neurodegenerative disease has shown the ability to detect early Alzheimer’s and Parkinson’s disease, through notable retinal amyloid deposition and retinal nerve fiber layer thinning, paving the way for potential validated biomarkers for conditions traditionally diagnosed by neuroimaging or cerebrospinal fluid analysis (4, 5). A 2023 paper by Zhou et al. published in Nature described a foundational model for generalizable disease detection from retinal images – demonstrating a single AI system is capable of detecting a broad spectrum of systemic conditions from fundus photographs, representing a qualitative shift in what retinal screening can do (6).

The implications for population health is substantial. In principle, a standard fundus photograph already taken in routine diabetic eye screening programs across the NHS could screen for diabetic retinopathy, cardiovascular disease, early Alzheimer’s disease, and chronic kidney disease. This requires no additional investigation, no additional data, nor any additional patient contact.

The ophthalmologist’s new position

If this potential is realized in clinical practice, it redefines the role of the ophthalmologist in the healthcare system; while the specialty has traditionally had a defined clinical domain, oculomics now extends this field to include systemic diseases reflected in the eye. This is not simply an expansion of scope – it is a change in the nature of our clinical encounters.

The infrastructure to make this practical already exists. The NHS Diabetic Eye Screening Programme alone captures fundus images from several million people with diabetes every year: images that are already stored, quality-assured, and reviewed by a trained grader. Overlaying an oculomics algorithm onto that existing pipeline requires no new equipment and no new appointment – what it requires is a workforce that understands what the algorithm is telling them, and a service that knows what to do with the answer.

If we consider a patient who attends a routine diabetic eye screening – an oculomics-enabled AI system examining the same images obtained could theoretically detect a high cardiovascular score and a pattern compatible with early neurodegeneration.

Fora patient presenting for a routine eye check, however, the ophthalmologist or grader assessing the images in conventional NHS screening pathways now has information with important implications for a patient’s cardiological and neurological management. Numerous questions arise from this scenario: Where does professional responsibility lay? Who communicates this to the patient? Who follows it up? What is the medico-legal position if a red flag is not acted upon?

There is also a consent question that has received little attention thus far. Patients attending a diabetic eye check have consented to eye screening, not necessarily to systemic disease risk profiling. What that distinction matters in practice, and how it should be explained before the photograph is even taken, is a conversation the speciality has barely started.

These issues are not yet clearly addressed in UK clinical governance frameworks and their absence is itself a form of risk. While the technology is regulated, the clinical workflow around it is not.

What ophthalmology must do, and what must it resist

The appropriate response to oculomics is neither uncritical enthusiasm nor defensive caution. In order to maintain structural engagement, several things are required.

First, ophthalmology leads should drive the discussion on oculomic governance, not cede it to cardiologists, neurologists, or technology companies. The retina and imaging infrastructure fall under the central domain to the ophthalmologist, and by extension the relationships with patients who attend screening programs. If governance frameworks are developed for oculomics-derived systemic disease findings without active ophthalmological input, they will be developed in ways that do not reflect the realities of how eye care is truly delivered.

Some of this work is already underway, though not yet with ophthalmology in the room per se. MHRA’s AI Airlock, a regulatory sandbox for AI-based medical devices launched in 2024 (7), is designed to test exactly these kinds of real-world deployment questions before products reach full market authorization. Oculomics tools are a natural fit for that kind of scrutiny, and ophthalmologists should be pushing to shape how it is applied, not waiting to be consulted as an after-measure.

Second, the specific area must have clear referral and communication pathways (for example, a positive Reti-CVD risk flag from a diabetic screening image needs to reach the patient’s GP with the same reliability and speed as a referral for sight-threatening retinopathy), rather than sit in a report that nobody knows how to action. Arguably, it is cheaper and less damaging to develop these pathways now – prior to wide-scale deployment – than to develop them in response to adverse events.

Third, and perhaps most importantly, ophthalmologists and trainees need to develop literacy in the capabilities and limitations of oculomics. It’s widely evidenced that AI models trained predominantly on European or East Asian datasets could perform less well on South Asian and Afro-Carribean populations, the very groups vulnerable to the conditions oculomics is designed to detect (8).

This is not a hypothetical concern. Reti-CKD’s own validation – while encouraging as it is – rested on a UK biobank cohort that is itself healthier and less ethnically diverse than the population it was drawn from, and a Korean cohort with a different demographic and risk profile (3). It is important to note: a tool that performs well in both may still behave unpredictably in the more diverse population actually attending NHS diabetic eye screening.

If a cardiovascular model consistently performs poorly in a high-risk population, it is not a screening tool – it is a source of false reassurance. Understanding these limitations is not an optional exercise for ophthalmologists; it is a fundamental clinical duty for anyone who will be exposed to oculomics outputs in practice.

The window is open – who is looking through it?

The retina has long been thought of as a window to the body. What has now changed is that AI has opened that window wide enough to see conditions we could only previously diagnose by looking elsewhere. This is a great scientific achievement, and will translate into lives saved and earlier diagnoses than would never have been possible before oculomics. But a window without a frame is simply a hole in the wall. Oculomics requires the clinical governance, referral infrastructure, training frameworks, and the demographic validation to function equitably and safely at scale. Building those frameworks aren’t someone else’s jobs; it is ours.