Artificial Intelligence Screening for Diabetic Retinopathy Under Variable Mydriatic Drop Uptake: A Causal Transportability Evaluation of Healthcare Referral Equity

Diabetic retinopathy screening programmes increasingly deploy artificial intelligence (AI) grading, yet the referral equity of these systems depends on a clinical precondition that varies widely across African health facilities: whether mydriatic drops were instilled before fundus photography. the present analysis treats variable mydriatic uptake as a transportability problem rather than a nuisance covariate. Using a before-after-control-impact (BACIP) design evaluated across 2000 replications per design cell, with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, the analysis characterises the operating characteristics of a causal transportability estimator for referral-rate effects. Power was estimated for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design required 52 clusters to detect a 5-percentage-point effect at 80% power. The results show that the estimator is well calibrated for detecting transportability failure when cluster counts reach 24 or more, but that at 8 and 16 clusters the minimum detectable effect exceeds the clinically meaningful range. Referral equity evaluations of AI retinopathy screening must be powered for interaction, not main effects, because the harm of non-transportability is concentrated in under-mydriatic facilities. The contribution is a design-theoretic account of how transportability of AI screening effects should be evaluated, with explicit consequences for cluster allocation in African screening trials.