Artificial Intelligence Benchmarking of Diabetic Retinopathy Grading Ensembles Under Unmeasured Mydriatic Drop Instillation Delay: A Healthcare Calibration-Transfer Evaluation

Ensemble classifiers for diabetic retinopathy grading are benchmarked on mydriatic fundus images whose pupillary dilation is assumed to be complete, yet the instillation-to-capture interval is rarely recorded and varies across clinics. the present analysis treats that unmeasured delay as a calibration-transfer problem and characterises, by formal analysis rather than by observational data, how ensemble aggregation rules propagate delay-induced distribution shift into grading decisions. A before-after-control-impact design is specified with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), and its operating characteristics are evaluated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The design attains 80% power at 52 clusters for a 5-percentage-point effect. The analysis derives a bias bound for soft-voting ensembles under delay shift, shows that temperature scaling cannot recover the lost calibration because the shift is covariate-dependent, and proposes a delay-stratified recalibration protocol. The contribution is a design-level account of when ensemble benchmarking for retinopathy grading remains valid under unmeasured mydriatic delay, and a specification of the cluster counts and effect sizes at which calibration-transfer failures become detectable.