Artificial Intelligence Benchmarking of Neonatal Jaundice Screening from Smartphone Photographs Under Unmeasured Melanin-Dependent Chromatic Calibration Drift: A Healthcare Fairness-Transfer Evaluation

Smartphone-based neonatal jaundice screening promises to extend bilirubin estimation beyond the nursery, but the fairness of such systems is rarely characterised under the conditions that actually obtain in deployment: unmeasured, melanin-dependent drift in the chromatic calibration of the camera pipeline. the present analysis treats that drift as a structural measurement problem rather than a nuisance and asks what a benchmarking protocol must estimate if it is to report fairness-transfer honestly. The analysis is a design and power study of a before-after-control-impact (BACI) evaluation of a screening model under calibration drift. No observational data are used or claimed. Under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 56 clusters to detect a 5-percentage-point fairness-transfer effect at 80% power. Fairness-transfer benchmarking is a cluster-level inferential problem, that the intraclass correlation of 0.64 dominates the design, and that reporting a single pooled fairness metric without a drift-stratified, cluster-aware design is uninterpretable. The contribution is a formalisation of melanin-dependent calibration drift as a latent stratum and a design-theoretic account of what a credible fairness-transfer benchmark must measure.