At AAO in New Orleans this month, artificial intelligence wasn’t a track; it was core to the meeting. The opening keynote was about generative AI and the future of medicine, and the theme carried through the symposia, the hallway conversations and the exhibit hall. Oculomics had its own moment with the debut of Oku Digital Health, Topcon’s new company built to turn routine retinal exams into systemic disease screening within a connected care network.
A few years ago, the work was convincing people the retina could tell us something about the heart, the brain and the kidneys. That case has largely been made, and the conversation in New Orleans had moved from whether to how. Implementation is where the rubber meets the road.
Regular readers know my bias here: it’s not just whether the algorithm works, but whether it fits the workflow. Oculomics, though, has a problem that sits underneath clicks and EHR integration, and it comes down to arithmetic.
The trap
Take an algorithm that is 99% sensitive and 99% specific, better than anything on the market. Run it on 10,000 people in a population where the disease has a 1% prevalence. It catches 99 of the 100 people who have it, but it also flags 99 of the 9,900 who don’t. This means half of everyone you call back is healthy.
Now use numbers closer to practice. Primary open-angle glaucoma affects roughly 2% of Americans over 40. An algorithm with 90% sensitivity and 90% specificity, run on 1,000 patients over 40, finds 18 true positives and generates 98 false ones. This means more than five healthy patients are recalled for every one you actually help.
This is Bayes’ theorem, and a better model doesn’t change it. When a disease is uncommon in the people you test, even a strong test produces mostly false alarms. In oculomics, the stakes rise. Each false alarm is a conversation with a patient about her heart or her brain, an anxious week and a referral to a cardiologist or neurologist who didn’t ask for one. Repeated across a retail optical network, that isn’t a screening program. It is a referral stream the downstream system will learn to ignore.
The practical question, then, is how to deploy a good test without being overwhelmed by its false positives. In my view, the answer lies in changing who you test, in what order and what you do with the result. I think of this process as three Rs.
One: Risk-stratify
Rather than running every model on every eye, use what you know before the image is taken to identify the patients in whom the disease is actually likely. Factors like age, ancestry, diabetes, hypertension, smoking, family history or the medication list. In glaucoma, Black and Latino patients carry roughly three times the prevalence of non-Latino white patients, and prevalence climbs steeply with age.
The math responds quickly. If the same 90/90 algorithm is applied to a group where intake has raised prevalence to 10% instead of 2%, one in two positives is real instead of one in six, with no change to the algorithm or the camera. The least expensive performance gain in oculomics usually comes not from a better model but from a better denominator.
Two: Restrict
Treat the objective structural tests, OCT and widefield fundus imaging, as gatekeepers. They’re fast, inexpensive per scan and well suited to AI. The costlier steps—functional testing, a dilated specialist exam, a referral out of the building—can then be reserved for the patients the first layer flags.
Each tier hands the next a population in which the disease is more common. Each subsequent test performs better than it would on its own. Run the flagged patients from that risk-stratified group through a second, independent 90/90 test, and nine in ten of the patients you ultimately refer have the disease. Testing in series does buy specificity at the expense of sensitivity, because every gate is another opportunity to miss someone. In the same example, you catch 81% of disease instead of 90%.
The word “independent” carries a lot of weight in that example. Two algorithms reading the same fundus photograph, often trained on overlapping datasets, tend to struggle with the same tilted discs and the same poor-quality images. Because of this, stacking them yields far less than the arithmetic promises. Real independence comes from a different kind of information, and increasingly that information will come from outside eyecare.
Of course, every specialty is using AI, not just ours. We can expect a growing range of intelligent diagnostic devices to weigh in on the odds. This includes voice and behavioral testing, smartwatches and video analysis to detect neurological conditions, each of which might supplement an eye-based AI flag or be prompted by one in the first place. A retinal model that raises concern for atrial fibrillation, followed by two weeks of wearable rhythm monitoring before anyone calls cardiology, is a chain in which each link sees something the other cannot. The retina serves as the inexpensive, high-reach first filter, and the confirmation comes from the specialty that owns the disease. The whole chain requires end-to-end validation and dedicated ownership, as devices were only tested on separate populations.
Every added gate also costs some sensitivity, which is why the third R carries the most weight.
Three: Rationalize
The referral threshold shouldn’t be a property of the algorithm. It should be property of the disease, and specifically of what it costs to miss it.
For high-harm, irreversible conditions, the threshold should favor sensitivity. It should accept more false positives and referring, as a missed case costs a life or a visual field, while a false alarm costs a visit.
For low-harm, slow-moving findings, the threshold should favor specificity, keeping the patient in your clinic in a longitudinal monitoring loop. Imaging again next year and following the trend lets a real finding declare itself, while one that isn’t real resolves without ever reaching someone else’s schedule.
That second half is the part I’d underline for optometry. Eyecare already holds the longitudinal relationship, seeing the same patients year after year on the same devices. And for many oculomic findings, monitoring is the appropriate outcome rather than a lesser one than referral.
The first half matters just as much. For many patients, a visit for glasses or contacts is their only contact with the healthcare system for years. Catching a Hollenhorst plaque that points to carotid disease, or, as the models mature, retinal signs associated with atrial fibrillation, can be lifesaving.
The real question
Risk stratification carries its own cost, because intake forms reflect who knows their own history. The patient most likely to have undiagnosed diabetes or undetected atrial fibrillation is the one who doesn’t check the box. A filter built on self-reported risk can quietly concentrate screening among people who already have a doctor. Whoever designs these pathways has to watch who is being filtered out, not only who is being caught.
None of this works without someone who owns the result. When the intake flag, the AI read, the wearable and the monitoring loop sit in different systems, findings fall between them unless one clinician is responsible for the synthesis.
Autonomous diabetic retinopathy screening offers the precedent. It has had FDA clearance and a billing code for years, and uptake remains a fraction of what it could be. This isn’t because the algorithm fell short but because the program around it did.
Oculomics can learn from that experience or repeat it. The outcome won’t be decided by AUROC. But by who gets tested, in what order and what we do when the result comes back.