Clinical Computer Vision for Sickle Cell Retinopathy Screening Under Variable Fundus Camera Resolution: A Super-Resolution Domain-Adaptation Evaluation of Referral Grading Safety

Sickle cell retinopathy screening in African primary care is constrained by fundus cameras of unequal resolution, and the safety of artificial-intelligence referral grading under that heterogeneity has not been characterised. The framework treats the referral decision as a thresholded function of a latent retinopathy severity score, decomposes the resolution-induced distribution shift into a recoverable component and an irreducible information-loss component, and derives a bound on the change in referral sensitivity under adaptation. A before-after-control-impact design with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) was used to characterise the operating characteristics of the evaluation design across 2000 replications per design cell, with cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters to detect a 5-percentage-point effect at 80% power. The framework proposes that adaptation safety should be reported as a resolution-stratified sensitivity profile rather than a single aggregate metric, and identifies the conditions under which super-resolution adaptation preserves or degrades referral safety. The contribution is a design-level and decision-level formalisation that makes resolution heterogeneity an explicit object of evaluation rather than an uncontrolled nuisance.