Give people their eye result before they leave the room and twice as many turn up

What happened

Diabetic retinopathy screening programmes in Africa have a problem that has nothing to do with cameras or algorithms. People get screened, the images go away to be graded, and weeks later someone phones to say they need to see an eye specialist. Most of them never go.

A team from the International Centre for Eye Health at the London School of Hygiene and Tropical Medicine and the Eye Department at Kilimanjaro Christian Medical Centre in Moshi decided to attack that step directly. They ran a randomised controlled trial across 24 hospitals in the Kilimanjaro and Arusha regions of Tanzania, enrolling 2,378 people with diabetes between March 2023 and August 2024.

Half got the standard pathway: photographs taken, images graded later by a human, result and counselling by telephone a few weeks on. Half got their result read by AI before they left the photography room, with face-to-face counselling on the spot.

Among the people who genuinely had sight-threatening retinopathy, 35.9 percent in the immediate-result arm attended the eye clinic within eight weeks. In the standard arm it was 18.2 percent.

The test

Two-arm, parallel-group, individually randomised controlled trial, 1:1, conducted at non-communicable disease or diabetic clinics. Eligible participants were adults with diabetes attending retinopathy screening.

In the intervention arm, images were submitted to the AI software and the participant was told the result – non-referable, referable, or ungradable – plus given face-to-face counselling before leaving, with a referral to the central ophthalmology clinic if the output was referable or ungradable.

The primary outcome was attendance at the central ophthalmology clinic within 8 weeks of screening, among participants classified by masked UK-certified graders as having true referable diabetic retinopathy. That last clause matters: the denominator was defined by independent human graders, not by the AI under test.

The numbers

Measure	AI-supported, immediate result	Standard of care
Randomised	1,192	1,186
With true referable retinopathy	437	467
Attended within 8 weeks	157 (35.9%)	85 (18.2%)
Absolute difference	17.7 percentage points (95% CI 12.0-23.4)
Odds ratio	2.52 (95% CI 1.86-3.42), p<0.0001
Hospitals	24, Kilimanjaro and Arusha regions
Assessed for eligibility	2,439; enrolled 2,378

The result

Attendance doubled. The absolute difference was 17.7 percentage points with a confidence interval from 12.0 to 23.4, and an odds ratio of 2.52 with a p-value below 0.0001.

The authors’ interpretation is that AI-enabled screening pathways have the potential to improve diabetic retinopathy screening programmes in Africa. The funders were the British Council for the Prevention of Blindness, Christian Blind Mission and the Sir Halley Stewart Trust.

The number that should stay with you, though, is 64. Even in the better arm, 64 percent of people with sight-threatening retinopathy did not reach the eye clinic within eight weeks.

How this compares to what came before

Nearly every published study of diabetic retinopathy AI measures whether the algorithm grades images as well as a human. Dozens have shown that it does, and several systems have regulatory clearance on that basis.

Grading accuracy was never the thing stopping Tanzanian patients from keeping their sight. The screening worked; the follow-up did not. This trial is one of very few in the entire clinical AI literature to take the endpoint one step further down the pathway, to the behaviour the programme was failing at, and to measure that instead.

The effect size is also worth comparing with what the accuracy literature promises. A better grader improves classification at the margin. Changing when and how the patient is told doubled attendance.

What it means

The lesson generalises well beyond eyes. When a care pathway loses people, the loss usually happens at a handover – between the test and the result, between the result and the appointment. AI is nearly always proposed for the measurement step, because that is where the technical problem is. The value here came from using AI to collapse a handover.

For African health systems specifically, this is an unusually useful piece of evidence. It was generated in the setting it applies to, with local infrastructure, by a partnership that includes a Tanzanian hospital. Most clinical AI evidence is imported.

It also sets a realistic expectation. Doubling attendance from 18 to 36 percent is a large relative gain and still leaves most patients unreached. Anyone presenting this as a solved problem is overselling it.

Business ideas from this paper

Point-of-screening result delivery as a product category. The hardware is a camera and a laptop; the software is a model plus a printed result and a script for the counsellor. Packaging that for screening programmes in other conditions – cervical, hypertension, hearing – is a direct translation.

Attendance as the contracted outcome. Screening vendors currently sell grading. A contract written against clinic attendance would align the vendor with the programme’s actual failure mode, and this trial gives a baseline to write it against.

Counselling scripts and training. If the human conversation is doing part of the work, there is a training product in it, and training is far cheaper to scale across 24 hospitals than devices are.

Transport and appointment logistics. 64 percent still did not attend. Whatever is stopping them – cost, distance, time off work – is the next bottleneck, and it is a logistics problem, not a medical one.

What could be wrong with it

The intervention is a bundle. Immediate result and face-to-face counselling were introduced together, so the trial cannot say which did the work. If it is the counselling, a trained human grader on site would achieve the same thing without any AI.

Ungradable images were referred along with referable ones, which inflates referrals in the intervention arm and may affect how the clinic experienced the change.

The result is from two regions of one country, with one AI system and one screening infrastructure.

Attendance within 8 weeks is a proxy. The trial does not report whether attending led to treatment or to preserved vision, which is the outcome that ultimately matters.

And the absolute level remains low. A programme adopting this should plan on reaching roughly a third of the people who need care, not most of them.

How sure can you be?

We score this 9 out of 10.

It is a peer-reviewed individually randomised controlled trial in The Lancet Global Health, run across 24 hospitals, with named institutions at the London School of Hygiene and Tropical Medicine and Kilimanjaro Christian Medical Centre, a primary outcome adjudicated by masked UK-certified graders independent of the AI, and ISRCTN registration. The trial is complete.

It loses a point because the data are not publicly deposited.

The effect on attendance is well established by this design. What caused it is not, and that is an open question the authors leave open too.

Do this today

If you run any screening programme, measure the gap between a positive screen and the follow-up appointment attended. That number, not your test’s sensitivity, is probably where your programme is losing people.

If you work in diabetes care in East Africa, this trial was run in your region with your infrastructure and the protocol is published. It is directly actionable rather than aspirational.

If you are designing an AI product for a low-resource setting, ask what handover it could remove. Removing a three-week wait for a phone call turned out to be worth more than any improvement in grading accuracy that has been published.

Source: Artificial intelligence-supported diabetic retinopathy screening in Tanzania: a randomised controlled trial, The Lancet Global Health, October 2026. PMID 42863478 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.