AI Eye Screening in Tanzania Nearly Doubled Specialist Follow Up in a Randomized Trial

Diabetic retinopathy screening only helps if the people who need treatment actually reach a specialist. In Tanzania, that second step is where the system loses patients. A randomised controlled trial published in The Lancet Global Health tested whether adding artificial intelligence to screening changed that outcome, and the difference was large.

The trial enrolled 2,378 people attending a diabetic eye screening programme across 24 hospitals in Tanzania. Half received AI supported screening, and half followed the standard screening pathway. The study was led by researchers at the London School of Hygiene and Tropical Medicine with the Kilimanjaro Christian Medical Centre and the Tanzanian Ministry of Health.

The primary question was not whether an algorithm could grade a retinal photograph in a laboratory. The protocol for the trial, published in BMJ Open in 2024, defined the primary outcome in advance as the proportion of people with true referable diabetic retinopathy who attended the central ophthalmology clinic within 8 weeks of screening. That is a delivery outcome, measured in clinics and calendars rather than in benchmark scores.

On that outcome, the AI pathway nearly doubled follow up. Among participants whose screening showed a referable level of diabetic retinopathy, 35.9 percent of the AI group attended a specialist clinic within 8 weeks, compared with 18.2 percent of the standard care group. People in the AI group who did attend were also seen much sooner, an average of 10.9 days after screening, compared with 36.9 days in the standard group.

The mechanism matters, and it is not magic. In the AI pathway, patients received their result immediately at the point of screening, followed by face to face counselling while they were still in the room. In the standard pathway, retinal images were graded later by a health care worker at a nearby hospital, and results and counselling were delivered weeks later by phone. The trial therefore tested a bundle: instant grading plus instant conversation, against delayed grading plus delayed conversation.

That bundle is the honest reading of the result. The study does not isolate what the algorithm alone contributed, because the counselling changed at the same time as the grading speed. A system that returned results instantly without any AI, if one existed at this scale, might capture part of the same gain. What the trial does show is that removing the waiting period changed behaviour. Patients who understood their result on the day, and could ask questions in person, acted on the referral at close to twice the rate.

Context explains why the gap is so consequential. Diabetic retinopathy is a common complication of diabetes and a leading cause of blindness worldwide. It damages the small blood vessels of the retina, often without symptoms in its early stages, and it can progress before a patient notices any change in vision. Screening can catch those early changes, and timely treatment can prevent much of the sight loss. None of that chain matters if the referral appointment never happens. The research team estimates that 44 to 82 million people will be at risk of sight loss from the condition by 2045.

The response in Tanzania has already moved beyond the paper. The Ministry of Health, a collaborator on the study, has committed to integrating AI diagnostics into diabetic eye screening programmes in multiple regions of the country. The authors add the necessary caution: better screening has to be matched by better access to specialist clinics and to treatment, or the system will simply identify more people than it can treat.

The lesson travels well beyond eye care. Most medical AI papers stop at accuracy, as if a correct output were the finished product. This trial measured the step after the output, whether a person with a sight threatening condition walked into a specialist clinic in time. On that measure, speed was the intervention. A result delivered in minutes, with a human in the room to explain it, beat a result delivered weeks later by phone, and it was not close.