**Background:** Diabetic retinopathy (DR) is a leading cause of blindness globally, and sub-Saharan Africa is projected to see a 129% increase in diabetes prevalence from 2021 to 2045. Tanzania has the highest age-adjusted diabetes prevalence in Africa at 12.3%. Screening for DR can prevent blindness in 95% of cases if treatment is timely, but current services in low- and middle-income countries face major challenges: insufficient specialist staff, limited quality control, and poor follow-up rates. In the Kilimanjaro Diabetic Programme, only 42% of referred patients attended the central ophthalmology clinic. Artificial intelligence (AI) systems that automate DR grading and provide immediate results could address these barriers by reducing reliance on specialists and enabling point-of-screening counselling. This paper presents the rationale and design of a randomised controlled trial (RCT) testing whether AI-supported DR screening increases attendance at the central ophthalmology clinic among those with true referable DR.
**Methods:** This is a single-masked, parallel-group, individually randomised controlled trial conducted in the Kilimanjaro and Arusha regions of Tanzania. A total of 2364 participants will be randomised in a 1:1 ratio to either AI-supported screening or standard of care. Inclusion criteria: adults (≥18 years) with diabetes attending diabetic clinics in the study regions, willing to consent and be randomised. Exclusion criteria: unable/unwilling to consent, <18 years, already attending the central ophthalmology clinic or had a diabetic eye examination in the previous 12 months. After consent, participants undergo demographic and clinical assessment (including visual acuity, blood pressure, random blood sugar, and fundus photography with two retinal photographs per eye after pupil dilation). In the AI-supported arm, retinal images are analysed immediately by the SELENA+AI software (EyRIS), which provides a referral decision and is followed by face-to-face counselling using a script developed from focus group discussions. In the standard of care arm, images are graded later (typically 2–4 weeks) by an ophthalmology resident, and results are delivered by phone or text message with counselling. All images are also graded by UK-certified graders as the reference standard. The primary outcome is the proportion of persons with true referable DR (defined by reference standard) who attend the central ophthalmology clinic within 8 weeks of screening. Secondary outcomes include proportion attending out of all referred, sensitivity and specificity of AI and local graders, false positive referrals, treatment uptake, gradable vs ungradable images, time to presentation, cost-effectiveness (incremental cost per QALY gained), and acceptability/fidelity. Sample size: assuming 22.5% of screened have true referable DR, 266 per arm, with 90% power to detect a 33% increase in attendance (from 42% to 56%). Analysis will be by intention-to-treat using logistic regression, with adjustment for potential confounders (gender, education, distance). The trial is registered as ISRCTN18317152.
**Key Results:** This is a protocol paper; no results are reported. The trial began enrolment on 8 March 2023 and is currently recruiting. The SELENA+AI system was selected based on regulatory approval (FDA/CE marking), offline capability, performance in an African population (sensitivity 92.25%, specificity 89.04% in a prior Zambia validation study), and camera compatibility. The counselling script was developed from two focus group discussions with persons living with diabetes.
**Clinical Implications:** If the AI-supported screening pathway increases follow-up attendance, it could significantly improve the effectiveness of DR screening programmes in Tanzania and similar low-resource settings. By providing immediate results and counselling, AI-supported screening may address the critical bottleneck of poor referral adherence (currently 42%). The study also evaluates sensitivity, specificity, cost-effectiveness, and acceptability, providing comprehensive evidence for potential scale-up. The results will inform whether AI can be a viable solution to the projected surge in diabetes-related eye disease in sub-Saharan Africa.