Diabetic retinopathy AI: what held outside the trial, and who came back

In an endocrinology clinic in Israel, a staff member with no ophthalmic training photographed the retinas of people with diabetes during their visit. AI software, AEYE-DS, gave a definitive result for 95.7 percent of 256 patients. Of the flagged patients who were then examined, four in 34 needed injections or laser treatment ^[1].

In 2018 the United States (US) Food and Drug Administration (FDA) authorised IDx-DR to read a retinal photograph for diabetic retinopathy with no clinician looking at the image. In its pivotal trial of 900 patients it reached a sensitivity of 87.2 percent and a specificity of 90.7 percent ^[2]. Grzybowski and colleagues later ran IDx-DR in two Polish cohorts of 807 and 758 patients. Sensitivity was 99 and 99.3 percent. Specificity was 68 and 69 percent ^{[3, 4]}. The Polish papers do not state the software version, used a different reference standard and left out patients whose images could not be read.

This issue follows a patient through an AI screening visit, from the side of the optometrist or ophthalmologist who receives the referral.

What the software does in a diabetes clinic

Four systems are cleared to make the decision. They are LumineticsCore from Digital Diagnostics, formerly IDx-DR, EyeArt from Eyenuk, AEYE-DS from AEYE Health, and iPredict-DR from iHealthscreen, which was cleared on 2 July 2026 ^{[5, 6, 7]}.

A member of the clinic staff photographs the retina, usually without dilating drops. The operator does not need to be an eye care professional.

The software reads the photographs during the visit and returns a decision rather than a grade. The threshold is more than mild diabetic retinopathy, which means moderate non-proliferative retinopathy or worse, or diabetic macular oedema, in either eye. Above it the patient is referred to an eye care professional. Below it no referral is made. If the images cannot be read, the patient is dilated and photographed again ^{[2, 8]}. Some programmes refer an ungradable patient to eye care instead ^[9].

This means two things for the optometrist. A negative result is not a normal retina, because mild non-proliferative retinopathy returns a negative by design. And the referral that reaches an eye care practice carries a finding made by software in another building, with no human grade attached.

England has tested a different use: whether software could take the first read of the stored images of the North East London Diabetic Eye Screening Programme and leave human graders only the images it flags ^[10]. That is triage, tested on past images rather than in a live service. Figure 1 sets the evidence from both uses against three steps of the visit.

Step one: can the photograph be read?

The IDx-DR version 2.3 clearance summary gives two figures for the same 850 participants. The first images were gradable in 64.9 percent without dilation. After a protocol that dilates when the first images fail, the figure was 95.2 percent ^[8]. Five reports from US adopters of LumineticsCore or EyeArt gave average gradability without dilation of 49 to 75 percent ^[5]. The pivotal trial reported 76.4 percent without dilation ^[2].

Other clinics did better without drops. In an Australian five-site study the AI found 13.4 percent of images ungradable and human graders found 15.6 percent ^[9].

So in a clinic that does not dilate, somewhere between one patient in twenty and one in two leaves without a result. That patient needs a second attempt, drops, or a referral for the photograph alone. The spread reflects the dilation protocol, the camera, the patients and operator training. It is not evidence that the software got worse.

Step two: is the reading right?

Where systems have been compared on the same images, sensitivity for moderate to severe disease was high. Rudnicka and colleagues ran the eight systems over the images from 202,886 screening encounters in London. Sensitivity for moderate to severe non-proliferative retinopathy was 96.7 to 99.8 percent. Depending on the vendor, sensitivity for any referable retinopathy ran from 83.7 to 98.7 percent. The false-positive rate ran from 4.3 to 61.4 percent ^[10]. In the Australian five-site study, sensitivity was 83.7 percent and specificity 96.1 percent in 863 participants ^[9].

Outside its pivotal trial, IDx-DR caught more disease and was less specific. Its pivotal specificity of 90.7 percent held at 90.07 percent when version 2.3 was rerun on the same data ^{[2, 8]}. In Quebec, a different, semiautomated system had a specificity of 66.2 percent against a full eye examination ^[14]. Low specificity outside a trial is not unique to one product. These studies cannot show why it falls.

For the optometrist, specificity is the share of patients without the disease who are not sent on. At 68 percent, about three in ten disease-free patients screened in those Polish clinics would have been referred to eye care. In London the software would only triage images for human graders, so the vendor would set how many healthy images a grader still reads.

Some of those referrals do matter. In the Israeli clinic, the four treated patients needed anti-vascular endothelial growth factor injections or panretinal laser, and the examinations found other eye conditions nobody had recorded ^[1].

Step three: does the patient come?

Getting the eye examination done is where the systems have done better than the old pathway. The ACCESS trial (AI for Children's diabetiC Eye ExamS) randomised young people with diabetes to an AI eye examination in their own diabetes clinic or to a scripted referral to eye care. Every one of the 81 in the AI arm completed the examination, against 22 percent of the 83 referred ^[11]. Adult health systems reported smaller gains ^[5].

Attendance after a positive result is less impressive. Across four series it ran from 44.7 to 70.2 percent. The four count different things: an in-clinic examination only in Israel, follow-through in ACCESS, any ophthalmologist within 90 days at Stanford, and a scheduled visit in New Orleans ^{[1, 11, 12, 13]}.

Two randomised trials have tested attendance itself. In Rwanda in 2021, 51.5 percent of patients told of a positive AI result immediately attended within 30 days, against 39.6 percent told after human grading days later ^[15].

The larger trial was published on 12 October 2026, after the search behind this series closed. Cleland and colleagues randomised 2,378 people with diabetes at clinics in 24 hospitals in northern Tanzania. In one arm the AI read the photographs and the patient was told the result, with face-to-face counselling, before leaving. In the other, human graders read the photographs later and the result came by telephone a few weeks afterwards. Of the 904 participants with referable retinopathy, 157 of 437 (35.9 percent) in the AI arm attended the eye clinic within 8 weeks, against 85 of 467 (18.2 percent) in the standard arm ^[16].

The AI arm roughly doubled attendance. However, nearly two in three of the people who needed the clinic still had not come within 8 weeks. The trial also tested the immediate result and the counselling together, so it cannot say how much of the gain came from the software.

Two pooled estimates both found that more patients attended after AI screening. They used different measures and disagree on the reason. Rahmati and colleagues pooled six studies and 20,108 patients and found an odds ratio for follow-up of 1.89, which they attributed to instant results ^[17]. Leigh and colleagues, from Lions Outback Vision, found a relative risk of 1.89 with a far wider interval and heterogeneity of 92 percent. They credited the redesign of the care pathway around the AI as well as the technology ^[18].

For an eye care practice, the patient who screened negative does not arrive at all. And between a third and two thirds of the patients the software found did not reach an examination. That gap existed before the software.

In my own practice the same step fails. The patient leaves with the referral letter, and I find out at the next visit whether it was used.

What is on the Australian register

The Therapeutic Goods Administration (TGA) keeps a list of AI-enabled devices in the Australian Register of Therapeutic Goods (ARTG). Updated 10 September 2026, it carries four entries that read a fundus photograph for diabetic retinopathy. None is described as autonomous. The four FDA-cleared systems are not on the list, and a name search of the ARTG found no entry for any of them ^{[19, 20]}. Issue 4 covers the full list and the exemption rules that change on 1 November 2026.

Conflict of interest, funding and how this was written

I own OptoGuide Pty Ltd, which sells OptoGuide, a clinical decision support product for optometrists. Its clinical recommendations come from fixed rules, and no AI model runs when a clinician uses it.

No one funded this issue. I have no financial relationship with Digital Diagnostics, Eyenuk, AEYE Health, iHealthscreen, Lions Outback Vision, Google, or any other maker or operator of a diabetic retinopathy system named here. I checked every figure against the paper or the register it is cited to, and the errors are mine. This series will cover the field without promoting OptoGuide, and where a finding bears on rule-based decision support I will say so.

Issue 4 maps where AI now sits in an eye examination and which tools are on the Australian register. The issue after that covers the rules for Australian optometrists using AI: the guidance from the Australian Health Practitioner Regulation Agency (Ahpra) and the Code of conduct, the TGA's treatment of software as a medical device, the privacy obligations, including the automated decision-making transparency requirement that starts on 10 December 2026, and where liability sits when a tool is wrong.