In a Tanzanian trial, an AI-supported eye-screening pathway increased attendance at specialist follow-up among people with referable diabetic retinopathy. The intervention combined an immediate screening result with face-to-face counselling. Its clearest reported outcome was getting people to the next appointment, rather than demonstrating that an algorithm alone preserved their vision.
That distinction is useful for health services considering AI. Finding a possible problem and ensuring that a person receives the next stage of care are separate achievements.
The result arrived before the person left
The published trial results summary describes 2,378 adults attending screening across 24 hospitals in Tanzania’s Kilimanjaro and Arusha regions. Participants were individually randomised between two pathways.
In the AI-supported pathway, retinal images were assessed immediately and the person received their result with face-to-face counselling before leaving. In the standard pathway, images were assessed by a human later, with results and counselling provided by telephone weeks afterwards.
The comparison therefore changed both how the image was assessed and when and how the result was discussed. It did not hold everything else constant while replacing a human grader with software.
For a service manager, that makes the intervention more concrete. The operational question includes whether a result is available during the encounter and whether staff can explain what happens next. Those are parts of a pathway, alongside the image-analysis technology.
Which denominator answers the referral question?
The primary outcome concerned participants classified by masked reference graders as having truly referable diabetic retinopathy. There were 904 such participants: 437 in the AI-supported group and 467 in standard care.
Within eight weeks, 157 of the 437 people in the AI-supported group attended the central ophthalmology clinic, compared with 85 of 467 in standard care. That is 35.9% versus 18.2%, an absolute difference of 17.7 percentage points. The reported 95% confidence interval for that difference was 12.0 to 23.4 percentage points.
The denominator is not everyone screened, and the outcome is not the proportion correctly diagnosed. Using either of those descriptions would change the meaning of the result.
It also matters that most people with referable disease still did not attend within that period, even in the intervention group. A substantial improvement in one stage of a pathway does not mean the access problem has been solved.
The trial was designed around the next step
The published protocol explains the rationale: non-attendance after referral can undermine a screening programme. Immediate results create an opportunity to discuss a finding while the person is still with the screening team.
The protocol specifies eight-week attendance among people with truly referable disease as the primary question. It also explains that the study was not powered to establish improved visual outcomes or treatment uptake. A protocol describes the planned design; the later results summary supplies the attendance finding.
Reading those documents together helps avoid moving between different claims. More referrals, more attendance, more treatment and less sight loss would each require their own evidence. They are connected stages, rather than interchangeable measures of success.
What this adds to an AI assessment
The LSHTM and Tanzanian collaborators’ account emphasises the opportunity for immediate explanation and the continuing need for access to specialist clinics and treatment. It concerns the same trial, rather than a separate confirmation of its findings.
The study was supported by the British Council for the Prevention of Blindness, Christian Blind Mission and the Sir Halley Stewart Trust. Its results describe this programme and population; they do not establish the same attendance effect in every country or referral system.
Our chest-X-ray AI explainer examines another case where an intermediate service measure needs to remain distinct from patient benefit. Here, the important contribution is a randomised test of a screening-to-referral pathway. It shows why evaluating AI in care includes what people are told, when they are told and whether they can reach the care that follows.