Community health workers (CHWs) are critical to healthcare delivery in low-resource settings but often lack formal clinical training, limiting their decision-making. Large language models (LLMs) could provide real-time, context-specific support to improve referrals and management plans. This study aims to evaluate the potential utility of LLMs in assisting CHW decision-making in Rwanda.

This is a prospective, observational study conducted in Nyabihu and Musanze districts, Rwanda. Audio recordings of CHW-patient consultations will be transcribed and analysed by an LLM to generate referral decisions, differential diagnoses and management plans. These outputs, alongside CHW decisions, will be evaluated against a clinical expert panel’s consensus. The primary outcome is the appropriateness of referral decisions. Secondary outcomes include diagnostic accuracy, management plan quality, and patient and user perceptions to ambient recording of consultations. Sample size is set at 800 consultations (400 per district), powered to detect a 15–20 percentage point improvement in referral appropriateness.

Ethical approval has been obtained from the Rwandan National Ethics Committee (RNEC) (Ref number: RNEC 853/2025) in June 2025, recruitment started in July 2025 and results are expected in late 2025. Results will be disseminated via stakeholder meetings, academic conferences and peer-reviewed publication.

PACTR202504601308784.

To investigate whether quantitative retinal markers, derived from multimodal retinal imaging, are associated with increased risk of mortality among individuals with proliferative diabetic retinopathy (PDR), the most severe form of diabetic retinopathy.

Longitudinal retrospective cohort analysis.

This study was nested within the AlzEye cohort, which links longitudinal multimodal retinal imaging data routinely collected from a large tertiary ophthalmic institution in London, UK, with nationally held hospital admissions data across England.

A total of 675 individuals (1129 eyes) with PDR were included from the AlzEye cohort. Participants were aged ≥40 years (mean age 57.3 years, SD 10.3), and 410 (60.7%) were male.

The primary outcome was all-cause mortality. Quantitative retinal markers were derived from fundus photographs and optical coherence tomography using AutoMorph and Topcon Advanced Boundary Segmentation, respectively. We used unadjusted and adjusted Cox-proportional hazards models to estimate hazard ratios (HR) for the association between retinal features and time to death.

After adjusting for sociodemographic factors, each 1-SD decrease in arterial fractal dimension (HR: 1.54, 95% CI: 1.18 to 2.04), arterial vessel density (HR: 1.59, 95% CI: 1.15 to 2.17), arterial average width (HR: 1.35, 95% CI: 1.02 to 1.79), central retinal arteriolar equivalent (HR: 1.39, 95% CI: 1.05 to 1.82) and ganglion cell-inner plexiform layer (GC-IPL) thickness (HR: 1.61, 95% CI: 1.03 to 2.50) was associated with increased mortality risk. When also adjusting for hypertension, arterial fractal dimension (HR: 1.45, 95% CI: 1.08 to 1.92), arterial vessel density (HR: 1.47, 95% CI: 1.05 to 2.08) and GC-IPL thickness (HR: 1.56, 95% CI: 1.03 to 2.38) remained significantly associated with mortality.

Several quantitative retinal markers, relating to both microvascular morphology and retinal neural thickness, are associated with increased mortality among individuals with PDR. The role of retinal imaging in identifying those individuals with PDR most at risk of imminent life-threatening sequelae warrants further investigation.