Randomised trial of AI-supported diabetic retinopathy screening in Tanzania
In a randomised trial at 24 Tanzanian hospitals, giving people an AI screening result during the visit with face-to-face counselling roughly doubled the share of those with genuinely referable diabetic retinopathy who reached the referral eye clinic within eight weeks, from 18.2% to 35.9%.
- Model
- Unknown
- Field
- Medicine
- Date
- 2026-10-12
- Human collaborators
- Charles R Cleland, Matthew J Burton
What was found
Non-attendance after referral is one of the main failure points of diabetic retinopathy screening programmes in Africa, and an AI grader can return a result during the screening visit rather than weeks later. This two-arm, parallel-group, individually randomised trial ran at non-communicable disease and diabetic clinics in 24 hospitals in the Kilimanjaro and Arusha regions between 8 March 2023 and 31 August 2024. Both arms had retinal photographs taken of both eyes; in the intervention arm the images went to the AI software and the participant was given a non-referable, referable or ungradable result with face-to-face counselling before leaving the photography room, and was referred if the output was referable or ungradable, while the standard-of-care arm received deferred human grading and a telephone result typically two to four weeks later. Of 2,378 participants enrolled, 1,192 were assigned to the AI arm and 1,186 to standard care, and 904 were classified by masked UK-certified graders as having true referable disease. Among those, attendance at the central ophthalmology clinic within eight weeks was 157 of 437 (35.9%) in the AI arm against 85 of 467 (18.2%) under standard care, an absolute difference of 17.7 percentage points (95% CI 12.0 to 23.4) and an odds ratio of 2.52 (95% CI 1.86 to 3.42), with .
Novelty check
Searched Europe PMC on 2026-10-22 ("diabetic retinopathy" AND "artificial intelligence" AND "randomised controlled trial" AND Africa, 43 hits). This trial and its own 2024 BMJ Open protocol are the only randomised comparison of an AI-supported screening pathway against deferred human grading in sub-Saharan Africa that the search returned; the same group's 2025 implementation report in Transactions of the Royal Society of Tropical Medicine and Hygiene (doi 10.1093/trstmh/trae132) describes putting an AI system into the Tanzanian programme but is not randomised. The trial is registered with the ISRCTN Registry as ISRCTN18317152 and has no ClinicalTrials.gov record; ISRCTN is not a slug in this registry's registries vocabulary, so it is recorded here rather than under registrations. The registry already carries 2018-08-28-idx-dr-autonomous-diagnostic, the pivotal trial of an autonomous diabetic retinopathy diagnostic in United States primary care, which measured diagnostic accuracy rather than whether referred patients reached a clinic.
Caveats and known objections
Funded by the British Council for the Prevention of Blindness, Christian Blind Mission and the Sir Halley Stewart Trust, with the Wellcome Trust and the British Council also recorded in Crossref's funder list. No AI vendor appears among the funders, and the grading software is not named in either the published abstract or the trial protocol, so neither the system nor any vendor role could be established for this entry: that is why model is Unknown, and it means a reader cannot tell which algorithm was tested or how it would be obtained. The endpoint is attendance at the referral clinic, an access outcome, not sight saved: no visual-acuity or disease-progression result is reported. The trial compares two whole pathways, an immediate AI result with face-to-face counselling against deferred human grading with a telephone result, so the effect cannot be attributed to AI grading alone rather than to same-visit delivery and in-person counselling. Participants were referred on an ungradable output as well as a referable one, and masked UK-certified graders rather than the AI defined true referable disease for the primary outcome, which is why autonomy is ai-led and not autonomous: the model's output drove the referral and a human reference standard checked it. The trial was single-masked and ran in two regions of one country, so nothing here establishes how the pathway behaves elsewhere. Sensitivity, specificity and the number of false-positive referrals were pre-specified secondary outcomes and are not in the abstract, and the full text was not readable for this entry, so those figures and the authors' own limitations section remain unchecked.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Randomised trial of AI-supported diabetic retinopathy screening in Tanzania. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-12-ai-retinopathy-screening-tanzania-rctBibTeX
@misc{whataifound-londonschoolofhygieneandtropicalmedicine-2026-rct,
title = {Randomised trial of AI-supported diabetic retinopathy screening in Tanzania},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by London School of Hygiene and Tropical Medicine / Kilimanjaro Christian Medical Centre. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2026-10-12-ai-retinopathy-screening-tanzania-rct}
}