Autonomous AI diabetic retinopathy screening means a retinal camera and algorithm in a primary care clinic deliver a diagnosis without an eye specialist reading the images. The FDA authorised the first system in 2018 after a 900-patient trial showed 87.2% sensitivity and 90.7% specificity. Its success came from redesigning who takes the photo, where and what happens next.
Diabetic eye disease is one of the most preventable causes of blindness, and one of the most under-screened. The test itself is not hard: photograph the back of the eye, look for signs of damage, refer the people who need treatment. The failure happens between the steps. A primary care doctor writes a referral to an eye specialist. The patient has to book a separate appointment, take time off work, get their pupils dilated, find a way home with blurred vision. Many never go.
In April 2018 the FDA authorised the first autonomous AI diagnostic system in any field of medicine, a diabetic retinopathy test then called IDx-DR and now marketed as LumineticsCore. It is often told as a story about an algorithm that could read retinal photos as well as specialists. That is true and not very interesting. Plenty of algorithms can read retinal photos.
The more useful story is that the first autonomous AI worked because its designers redesigned the workflow around it: who operates the camera, what counts as a usable image, where the result is delivered and what happens after a positive test. The model was necessary. The workflow was what made it a medical service.
What is autonomous AI for diabetic retinopathy screening?
“Autonomous” means the system produces a clinical result without a clinician reviewing the images. A trained operator, often a medical assistant in a primary care clinic, takes photographs of both eyes with a retinal camera. The software checks whether the images are good enough, analyses them and returns one of a small set of outputs: more-than-mild diabetic retinopathy detected, refer to an eye care professional; negative, rescreen in 12 months; or insufficient image quality.
The pivotal trial, published in npj Digital Medicine in August 2018 by Michael Abrà moff and colleagues, enrolled 900 people with diabetes and no history of retinopathy at primary care clinics. The reference standard was demanding: widefield stereoscopic photography and optical coherence tomography graded by the Wisconsin Fundus Photograph Reading Center. The AI exceeded all prespecified endpoints, with sensitivity of 87.2%, specificity of 90.7% and an imageability rate of 96.1%. About 28.6% of participants were African American, which matters for a disease that disproportionately affects minority communities.
Two details from that trial are easy to skip and should not be. Imageability was a primary endpoint, not a footnote. And the operators, who were not eye specialists, underwent a standardised training protocol before the study began. The trial tested a workflow, not just an algorithm.
Why did the first autonomous AI succeed where others stalled?
Four design choices stand out, and none of them is about neural network architecture.
It moved the test to where the patient already was. The biggest loss in diabetic eye screening is the referral itself. Putting the camera in the primary care clinic turned a two-appointment process into a one-appointment process.
It defined the output narrowly. The system does not attempt a full eye exam. It answers one question, refer or do not refer, at a threshold chosen for screening. That narrowness made validation tractable and made the result easy to act on, the pattern we described in AI diagnostics: the evidence is strongest where the question is narrowest.
It treated image quality as part of the diagnosis. A system that cannot tell a bad photo from a healthy eye is dangerous. Building an explicit “insufficient quality” output, and measuring how often it occurred, made the failure mode visible.
It came with a way to get paid. From 2021, US billing included a dedicated CPT code, 92229, for point-of-care retinal imaging with autonomous analysis. Without a payment route, a primary care clinic has little reason to buy a camera and train staff, however good the algorithm.
Does autonomous AI actually increase diabetic eye exams?
The clearest evidence comes from a randomised trial in young people, the group where screening rates are often worst.
The ACCESS trial, published in Nature Communications in January 2024 by Risa Wolf and colleagues at Johns Hopkins, randomised youth aged 8 to 21 with type 1 or type 2 diabetes at a paediatric diabetes centre. The intervention group received an autonomous AI eye exam at their diabetes visit. The control group got the usual approach: a scripted referral to an eye care provider plus education.
Within six months, 100% of the AI group completed a diabetic eye exam, against 22% of the control group. Among the 25 participants in the AI arm with an abnormal result, 64% went on to see an eye care provider, compared with 22% follow-through in the control arm.
That is an enormous effect, and it is almost entirely a workflow effect. The AI did not find disease that human graders would have missed. It made the exam happen.
What goes wrong when health systems deploy it?
The real-world picture is more mixed, and it is mixed in a revealing way.
A review published in Ophthalmology Science in September 2025 by Teng and colleagues examined the three FDA-cleared autonomous systems in the US, LumineticsCore, EyeArt and AEYE Diagnostic Screening, and interviewed ophthalmologists who had led implementation at academic health systems. It found adoption still relatively limited after five years of commercial availability. Adopters reported sensitivity of 87% to 100% and specificity of 60% to 91%. But average gradability without dilating the pupil ranged from 49% to 75%.
Compare that with the 96.1% imageability in the pivotal trial. In routine practice, a quarter to a half of patients in some programmes did not get a usable result from the AI at the point of care. Those patients fall back into the old referral pathway, the one that fails most often.
That is the workflow lesson again, from the other direction. Trial conditions included trained operators working to a controlled protocol. In busy clinics, with staff turnover, older patients with cataracts and pressure to keep visits short, image quality drops. The algorithm has not changed. The workflow has.
A worked example: where the patients actually go
Hypothetical, with illustrative numbers. A primary care network looks after 10,000 adults with diabetes. It installs autonomous AI cameras and offers the exam at routine visits. Over a year, 8,000 patients are offered it.
Suppose gradability without dilation is 60%, in the middle of the real-world range. That gives 4,800 patients a result on the day. If about a fifth are positive, roughly 960 are referred to eye care, and if follow-through matches the ACCESS trial’s 64%, about 615 are seen.
The other 3,200 patients get “insufficient image quality.” They are referred on, just as before AI. If only a minority attend, the network has lost more patients through ungradable images than through positive results.
The fix is not a better model. It is operational: allow dilation in the clinic for failed images, retrain operators, track gradability by site and by operator, and send ungradable patients to a fast-track pathway rather than a standard referral. In this example, raising gradability from 60% to 85% adds 2,000 same-day results, which is worth more than any plausible improvement in the algorithm’s sensitivity.
Isn’t this really a telemedicine success, not an AI one?
The strongest counterargument is that autonomy is not the magic ingredient. Retinal screening with cameras in primary care and remote human graders has existed for years. National programmes that invite everyone with diabetes, photograph them locally and grade centrally achieve good coverage without any autonomous AI. On this view, the AI story is really a camera-in-the-clinic story, and the algorithm is a cost-saving substitute for human graders.
That is largely right, and it strengthens rather than weakens the thesis. The benefit comes from the redesigned pathway. What autonomy adds is specific: a result during the visit, when the clinician can explain it and book the referral before the patient leaves, and no need for a grading workforce that many regions do not have. The ACCESS follow-through numbers suggest that same-visit result matters. But anyone who thinks buying the algorithm will fix eye screening, without fixing imaging, referral and follow-up, is buying the least important part.
What this means for health systems, clinicians and developers
For health system leaders, budget for the workflow, not the software. That means operator training, time in the visit, dilation protocols, gradability monitoring and a named owner for the referral pathway. Measure the share of patients with diabetes who get a usable result, not the number of scans performed. The same logic applies to cancer screening, where we argued that the gap is about the second visit.
For primary care clinicians, autonomous AI shifts responsibility for eye screening into your clinic. The benefit is a completed exam. The duty is making sure positive and ungradable results turn into eye care appointments.
For eye care services, the effect of successful AI screening is more referrals of patients who actually need treatment. Capacity for those referrals needs to be planned before the cameras go in.
For developers of other autonomous AI, the lesson is that regulators, payers and clinics bought a redesigned service, not an accuracy figure. Imageability, operator training and a clear action for every output belong in the product, not the implementation guide.
For people with diabetes, regular diabetic eye exams remain important even without symptoms. An AI exam in a primary care clinic is a legitimate way to get one, and an abnormal or ungradable result should be followed up with an eye care professional.
Frequently asked questions
Is autonomous AI as accurate as an eye doctor for diabetic retinopathy?
For the narrow task of detecting more-than-mild diabetic retinopathy, the pivotal trial showed 87.2% sensitivity and 90.7% specificity against a rigorous reading-centre standard. Real-world studies report sensitivity of 87% to 100% and specificity of 60% to 91%. The AI does not replace a full eye exam, which can find other conditions such as glaucoma.
Which autonomous AI systems for diabetic retinopathy are FDA-cleared?
As of a September 2025 review in Ophthalmology Science, three systems were FDA-cleared in the US: LumineticsCore, formerly IDx-DR and the first authorised in 2018; EyeArt; and AEYE Diagnostic Screening. They differ in cameras, workflow and published evidence, and no study had compared them head-to-head on the same patients.
What happens if the AI cannot read the images?
The system reports insufficient image quality rather than guessing. The patient then needs another route to an exam, often dilation and repeat imaging or a referral to eye care. Real-world gradability without dilation has ranged from about half to three-quarters of patients, so the fallback pathway matters a great deal.
How often should people with diabetes have an eye exam?
Most guidelines recommend regular diabetic eye screening, commonly every year, with the interval adjusted by a clinician based on findings and risk. People with diabetes should ask their care team when they are due and what type of exam is appropriate, and report any change in vision promptly.
Is there a billing code for AI diabetic eye exams?
In the US, CPT code 92229, introduced in 2021, covers retinal imaging with point-of-care autonomous analysis and report. It was the first CPT code for an autonomous AI service. Coverage and payment levels vary by payer, and health systems should confirm them before building a programme around the test.
The first autonomous AI in medicine did not succeed because it outperformed specialists. It succeeded because someone decided the eye exam should happen in the room the patient was already sitting in, and designed everything else around that.
Related on the GHC blog: Why Universal Health Coverage Stalls After the Pilot Phase and Paying for Outcomes Rents Behaviour. It Doesn’t Buy It..
If you run diabetic eye screening, what share of your AI exams produce a usable image, and what happens to the rest?
