Artificial Intelligence Computer Vision for Sickle Cell Retinopathy Screening Under Variable Fundus Camera Flash Intensity Decay: A Photometric Normalization Evaluation of Healthcare Referral Equity

Fundus camera flash intensity decays with lamp age and power instability, producing illumination gradients that differ systematically between screening sites and over time. When artificial intelligence (AI) computer vision models for sickle cell retinopathy are deployed across such sites, photometric normalization is intended to remove this nuisance variation, but the evaluation designs used to certify normalization are rarely examined for their ability to detect the referral-equity effects they are meant to protect. the present analysis treats the before–after–control–impact (BACI) design as an evaluation instrument and characterises its operating characteristics under a declared variance structure. No observational data are used or claimed. The analysis assumed a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.640, and estimated power across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The computed intraclass correlation of 0.64 implies that cluster-level variation dominates residual variation, so the effective information in a BACI contrast is governed by the number of clusters rather than the number of images. Under these parameters, 54 clusters are required to detect a 5-percentage-point referral-equity effect at 80% power. The paper derives the BACI interaction estimator, states the conditions under which photometric normalization can be credited with an equity improvement, and argues that evaluation designs with fewer than 54 clusters cannot distinguish a normalization effect from site-level illumination drift. The contribution is a design-theoretic account of what a photometric normalization trial must contain before its referral-equity claims are credible.