Look Into My Eyes to See My Sole: A Retinal AI Score Tracks Amputation Risk #ActAgainstAmputation

Every so often a paper lands that makes us want to reach for the ophthalmoscope instead of the monofilament. This is one of them.

Park J, Yoon JS, Thakur S, Nam D and Hwang S have just published a study in Frontiers in Endocrinology asking a beautifully strange question: can a deep-learning score built to estimate coronary artery calcium from a retinal photograph tell us something about whether a person with diabetes is headed for an amputation?

The answer, apparently, is yes.

What they did

The team studied 392 people with type 2 diabetes receiving ophthalmic care at a Korean university hospital — 79 with a history of diabetes-related amputation and 313 without. Every fundus photograph was run through a validated convolutional neural network (Dr. Noon CVD) originally trained on more than 216,000 paired fundus photographs and coronary artery calcium scores. The output is a single number from 0 to 100.

Then they asked whether that number added anything to the ordinary stuff we already collect: age, sex, HbA1c, diabetes duration, and retinopathy status.

It did. Adding the retinal score to a basic clinical model moved the AUC by 0.146 (95% CI 0.046–0.249), with a continuous net reclassification index of 0.629 and an integrated discrimination improvement of 0.062. The full model reached an AUC of 0.791.

The part that made us sit up: the score still added value after diabetic retinopathy was already in the model (ΔAUC 0.081). In other words, the algorithm is not simply re-reading the retinopathy we can already see. It is pulling out something else — a systemic vascular signature hiding in the same pixels.

Feature importance in the full model: retinopathy (β = 1.64), the retinal AI score (β = 0.70), and HbA1c (β = 0.44) all carried independent weight. Age, sex, and diabetes duration contributed essentially nothing.

The numbers that matter clinically

At a prevalence of 27% (roughly what you’d expect in a diabetic foot clinic), the rule-out threshold delivered 87.5% sensitivity with a negative predictive value of 92.6%. The rule-in threshold gave 90.4% specificity with a positive predictive value of 56.3%.

Run the same model against the general diabetes population, where amputation prevalence is around 1.8%, and the positive predictive value collapses to 6%. The authors are refreshingly honest about this. This is not a population screening tool. It is a triage instrument for people who are already in trouble.

Why this is not as weird as it sounds

We have known for a long time that the eye and the foot are talking to each other. In a large Korean population study, retinopathy carried a 4.03-fold increased risk of amputation — a stronger signal than nephropathy. Foussard N and colleagues showed in Diabetes Care that adding retinopathy stage improved discrimination for major lower-extremity arterial disease. And a 2026 umbrella review of 21 systematic reviews found that diabetic nephropathy and diabetic retinopathy carried the most credible associations with diabetes-related foot complications of any factor examined — both graded Class I evidence, ahead of neuropathy, sex, and HbA1c.

The mechanistic story is plausible enough. Coronary calcium correlates with lower-extremity arterial calcification. Tibial and pedal arterial calcification predict major amputation. And the retina is the only place in the body where we can look directly at a capillary bed without cutting anything open. If the microvasculature is a canary, the retina is the cage we happen to have a camera pointed at.

The other eye story — and a long-overdue nod to Manchester

Before we get too excited about retinal vessels, it is worth remembering that the eye has been telling us about the foot for a couple of decades already — just through a different tissue.

Rayaz Malik, my friend and sometimes office-mate during my Manchester days and now at Weill Cornell Medicine-Qatar in Doha, spent much of his career establishing that the cornea — the most densely innervated tissue in the human body — can be imaged non-invasively to quantify small fibre nerve damage. Corneal confocal microscopy measures nerve fibre density, branch density, and length in a scan that takes minutes and hurts nobody.

Malik and colleagues published a review in 2013 that they had the good sense to title “Corneal confocal microscopy to assess diabetic neuropathy: an eye on the foot.” Which rather takes the wind out of my headline, and I will happily concede the point.

The body of work that followed is substantial. Corneal nerve loss predicts incident neuropathy, tracks severity, and detects nerve regeneration — enough evidence, Malik’s group has argued, to satisfy FDA criteria for a biomarker suitable as a primary endpoint in disease-modifying trials. A 2025 meta-analysis across 52 studies and 34 conditions confirmed that corneal nerve fibre length separates clinical and subclinical neuropathy from healthy controls.

And most directly relevant to us: Lim JZM, Malik RA and colleagues published thresholds for corneal nerve metrics in the neuropathic foot at risk of ulceration, showing corneal nerve fibre density distinguished type 1 diabetes with and without foot ulcer at an AUC of 0.93 — and that the same measure predicted incident cardiovascular and cerebrovascular events over three years.

Sit that next to the Park paper for a moment. One group looks at corneal nerves and finds the at-risk foot and the at-risk heart. Another looks at retinal vessels and finds the same two things. Nerve and vessel, cornea and retina, two entirely different imaging modalities converging on the same conclusion: the eye is an honest reporter of what is happening at the far end of the body.

The nerve story got there first, and it got there with prospective data. The vascular story is newer and, for now, thinner. What neither has yet done is combine them.

The caveats, which are real

This is retrospective and case-control-like. Patients were classified by whether an amputation happened, not followed forward for events. That means the paper demonstrates association, not prediction, and the discrimination figures will almost certainly look different in a prospective cohort.

The external dataset used for the sensitivity analysis (mBRSET, from Brazil) had diabetic foot outcomes rather than amputation outcomes, a different camera, a different population, and no HbA1c. And critically for our readership: the amputation dataset contained no neuropathy data, no peripheral arterial disease data, no ankle-brachial or toe pressures, no lipids, no BMI. A model that beats “age, sex, and duration” is not yet a model that beats a good vascular exam.

What we’d like to see next

Here is the version of this study we want to read. Take a diabetic foot clinic. Photograph every retina at intake — a handheld camera will do, as the Brazilian data suggest. Pair it against toe pressures, skin perfusion pressure, WIfI stage, and a real neuropathy assessment. Then follow forward. Does the retinal score identify the patient whose ulcer will not heal? Does it flag the CLTI that the ABI missed in a calcified vessel? Does it change who gets sent for angiography on a Tuesday instead of a month from Tuesday?

That last question is the one that matters. We have no shortage of risk scores. We have a considerable shortage of risk scores that change what happens in the next 48 hours.

And the study after that one is obvious: put the corneal camera and the retinal camera in the same room. Nerve and vessel, from the same patient, in the same visit, against the same outcome. Nobody has done it. Somebody should.

Still — there is something wonderful about the idea that the next thing that saves a leg might come from a camera pointed at an eye. The vasculature does not respect our specialty boundaries. Perhaps we should stop pretending it does.

Look into my eyes to see my sole.

The paper

Park J, Yoon JS, Thakur S, Nam D, Hwang S. Deep learning-derived retinal biomarker associated with diabetes-related amputation in type 2 diabetes. Front Endocrinol. 2026;17:1866694. https://doi.org/10.3389/fendo.2026.1866694

The corneal side of the story

Tavakoli M, Petropoulos IN, Malik RA. Corneal confocal microscopy to assess diabetic neuropathy: an eye on the foot. J Diabetes Sci Technol. 2013;7(5):1179-1189. https://doi.org/10.1177/193229681300700509
Petropoulos IN, Ponirakis G, Ferdousi M, et al. Corneal confocal microscopy: a biomarker for diabetic peripheral neuropathy. Clin Ther. 2021;43(9):1457-1475. https://doi.org/10.1016/j.clinthera.2021.04.003
Lim JZM, Burgess J, Ooi C, et al. Corneal confocal microscopy predicts cardiovascular and cerebrovascular events and demonstrates greater peripheral neuropathy in patients with type 1 diabetes and foot ulcers. Diagnostics. 2023;13(17):2793. https://doi.org/10.3390/diagnostics13172793
Petropoulos IN, Madani OA, Gad HY, et al. Corneal confocal microscopy to diagnose peripheral neuropathy: a systematic review and meta-analysis. Eur J Neurol. 2025;32(11):e70396. https://doi.org/10.1111/ene.70396
Tavakoli M, Begum P, McLaughlin J, Malik RA. Corneal confocal microscopy for the diagnosis of diabetic autonomic neuropathy. Muscle Nerve. 2015;52(3):363-370. https://doi.org/10.1002/mus.24553
Pritchard N, Edwards K, Shahidi AM, et al. Corneal markers of diabetic neuropathy. Ocul Surf. 2011;9(1):17-28. https://doi.org/10.1016/s1542-0124(11)70006-4