Autonomous Diabetic Eye Screening: Strong Trial Results, Thin Evidence on Real-World Follow-Through

The use case

In a Baltimore trial that ran from November 2021 to June 2022, all 81 young people with diabetes who were offered a camera exam graded by software completed an eye check within six months, against 18 of 82 who were given a referral to an eye specialist (measured in a randomised controlled trial)¹. That gap is the case for autonomous eye screening, and also the reason its headline number needs careful reading.

In this use case a nurse or technician photographs the back of the eye with a retinal camera. Software built as a deep-learning model (a neural network trained on large sets of labelled images) grades the photographs for diabetic retinopathy, damage to the retina's blood vessels caused by diabetes, and returns a result without a specialist reading the images first. In the Baltimore trial the result came within 60 seconds¹. A positive result triggers referral to an eye specialist. In the United States the Food and Drug Administration (FDA) granted the IDx-DR system marketing authorisation through its De Novo route on 11 April 2018⁸. A 2026 ophthalmology society magazine lists three such systems described as FDA-approved: LumineticsCore (2018), EyeArt (2020) and AEYE-DS (2022)⁹.

Scope note. Only machine-learning image grading counts as AI here. Human graders in conventional screening programmes, and any automated grading whose method is not stated, are not credited as AI. Studies in which a human always reads the images, and AI for other eye diseases such as glaucoma or macular degeneration, are out of scope. Most of the evidence concerns "referable" disease, meaning disease severe enough to warrant specialist review.

Where it is running

These are studies, not a census. The sources read give no count of sites or screenings in routine use for any of the three FDA-listed systems, so the scale of everyday use could not be established. Three of the four settings above collected data more than three years before this paper (Thailand in 2018 to 2020, the United States and India in 2021 to 2022); only the Belgian study is more recent. In the United Kingdom, automated grading is not part of the national diabetic eye screening programme: the UK National Screening Committee (UK NSC) "does not currently recommend" it and opened a consultation in May 2026².

What the evidence shows

Randomised trial (strongest design). The ACCESS trial at Johns Hopkins randomised 164 young people with type 1 or type 2 diabetes to a point-of-care exam read by IDx-DR or to a standard referral plus an education handout; 163 were analysed¹. All 81 in the software arm (100%; 95% confidence interval 96% to 100%) completed an eye exam within six months, against 18 of 82 (22%; 14% to 32%) in the control arm, a difference of 78 percentage points (measured in a randomised controlled trial)¹. The mechanism is built into the design: the software arm was examined during the same clinic visit, so completion required nothing further from the family, whereas the control arm depended on a separate appointment elsewhere. Of 25 software-arm participants whose result showed disease, 16 (64%) then saw an eye specialist, against 18 of 82 (22%) in the control arm who completed a specialist visit (measured in the same trial)¹. Against a retina specialist reference, the software showed sensitivity of 100% and specificity of 78.9% (measured in the same trial)¹. Sensitivity is the share of people with disease that a test catches; specificity is the share without disease that it correctly clears.

National programme cohort. In nine primary care sites in Thailand, between December 2018 and March 2020, a deep-learning system was run in real time on 7,940 people screened, of whom 7,651 were eligible (prospective cohort, abstract)³. For vision-threatening disease the abstract reports sensitivity of 91.4% (87.1% to 95.0%) and specificity of 95.4% for the system, against 84.8% (79.4% to 90.0%) and 95.5% for retina specialist over-readers (measured without a control group, abstract)³. The comparison concerns accuracy on the same images, not patient outcomes. The study was funded by Google and Rajavithi Hospital³.

Independent single-clinic study. In Brussels, between January and April 2024, a Belgian-developed system, not one of the three listed in⁹, was used by nurses and a medical student on one non-dilated image per eye. Among 353 analysed patients, sensitivity for referable disease was 88.9% (77.4% to 95.8%) and specificity 98.7% (96.7% to 99.7%) (measured without a control group, against human graders)⁵. There were 6 false negatives and 4 false positives, and 52 of the 405 patients screened were excluded from the primary analysis⁵.

Weaker result in a public-health setting. In the validation phase in India (March to June 2021, 250 people with diabetes), the abstract reports sensitivity of 68.42% (59.71% to 76.05%) and specificity of 96.01% (93.24% to 97.72%) (measured without a control group)⁴. A sensitivity near 68% means roughly one in three people with referable disease would be missed. A second phase with 343 participants followed in 2022⁴.

Pooled evidence. A 2023 systematic review of AI screening for referable diabetic retinopathy in real-world settings pooled 15 studies from China (7), India (5) and Australia (3). Patient-level pooled sensitivity was 95.33% (90.60% to 100%) and specificity 92.01% (87.61% to 96.42%) across 10 datasets and 45,785 patients (pooled from published studies, searches run February 2023)⁶. The review covers AI systems in general, not only autonomous ones, and the upper limit of the sensitivity interval reaches 100%, which signals imprecise estimates.

What it takes: cost, integration, governance

Cost. No official price or payment rate could be verified in this research, and none of the studies read reports a cost-effectiveness analysis. A 2022 commentary describes benefits such as referral adherence, offloading of graders and lower cost as "implied but unproven"⁷. The UK NSC found no new UK value-for-money studies².

Integration. The same commentary lists the practical barriers: no universal electronic health record, slow internet, operator and camera problems, and no communications standards for retinal images or AI outputs⁷. The Baltimore trial avoided pupil dilation, which speeds the exam but may not carry over to adults¹.

People. The Belgian study used nurses and a medical student rather than ophthalmologists⁵. In Thailand, retina specialists still over-read the images to provide the comparison³. A positive result still depends on a specialist and a patient who attends.

Governance. In the United States the IDx-DR device is classed as Class II under the De Novo decision of 11 April 2018⁸. The ACCESS authors state that the system is not FDA-cleared for people aged 21 and under, so the trial used it outside that clearance¹. In the United Kingdom the UK NSC reviewed the topic in 2021 and found evidence insufficient, re-opened it in 2024, and in May 2026 consulted on a new evidence summary while still not recommending automated grading².

Where it falls short

Accuracy does not travel. Sensitivity ranged from 68.42% (India validation) to 91.4% (Thailand) and 88.9% (Belgium)³⁴⁵. The systems, populations, cameras and reference standards differ, so the figures cannot be ranked against each other. The spread does show that a result from one setting cannot be assumed for another. The UK NSC reaches the same view: performance varies by company and setting, some systems may vary by ethnicity or age, and most match humans for more severe disease but not for people without disease².

The trial measured whether an exam happened, not whether sight was preserved. The 78-point gain in completion comes from moving the exam into the clinic visit¹. It does not show fewer cases of vision loss, and the follow-up window was six months. Even with an immediate result, 16 of 25 people with a positive result (64%) reached a specialist, so more than a third did not within the window¹. The Thailand abstract reports that 2,412 of 7,651 eligible people (31.5%) were referred but does not report how many of those reached specialist care³.

Design limits. ACCESS was a single-centre trial of 164 young people, and its control arm received an education handout rather than active support to book an exam¹. The gain therefore combines the software with same-day, no-dilation convenience; the trial cannot separate them. Funding and competing-interest statements for the trial could not be read.

Independence. The Thailand study was funded by Google and Rajavithi Hospital³. Independent data are thinner: the Belgian study had no declared competing interests but analysed 353 patients at one clinic⁵.

Failures and withdrawals. No programme withdrawals, lawsuits or regulator enforcement actions were found in the sources read. The searches for them were limited, so their absence is not proof that none exist.

Answer to the challenge question. The usual presentation rests on the Baltimore completion figures. Those figures come from one trial in youth, enrolled in 2021 and 2022, with a control arm that relied on referral. The measured gain is real within that design, but how much comes from the software, and how much from placing the exam inside the visit, is unknown. Evidence on treatment uptake and sight outcomes is thin, and accuracy results vary between settings.

Phase: In early use — why

The row's hypothesis, In early use, is kept. Confirmed machine-learning systems are authorised in the United States and are being studied in four countries¹³⁴⁵⁸⁹. The sources read, however, give no figure for how many sites or patients use them routinely, and one national screening body, the UK NSC, still does not recommend automated grading². Where the scale of genuine use cannot be established, the lower of the plausible phases applies, so Pilot with results is the alternative on the evidence and In wide use is not supported.

What to watch next

• An updated UK NSC recommendation on automated grading in the diabetic eye screening programme, after a consultation that closed at 11.59pm BST on 3 August 2026; no outcome was found at the time of this research².
• Any announcement of tests of automated grading in real NHS services, which the UK NSC says could follow its finding that some systems are as good as the first human grader².
• Publication of a cost-effectiveness or value-for-money study, which the UK NSC found missing for the UK².

Questions to ask before adopting

• Was the system tested on a population, camera and staff mix like ours, given that sensitivity ranged from 68% to 91% across studies?
• What share of people with a positive result actually reach a specialist in our referral pathway, and what happens to the rest?
• Which part of any reported gain comes from the software and which from moving the exam into the visit?
• Is the system authorised for the ages and settings in which we would use it, noting the Baltimore trial used it outside its clearance for people aged 21 and under?
• What is the total cost per person screened, including cameras, integration, staff and onward referral, and who pays for it?