Artificial Intelligence Benchmarking of Federated Learning Aggregation Under Unmeasured Clinic-Specific Label Shift in Multi-Site Diabetes Retinopathy Grading: A Healthcare Fairness-Transfer Evaluation

Federated learning promises multi-site diabetic retinopathy grading without centralising patient images, yet its aggregation rules are typically benchmarked under the assumption that participating clinics share a common label distribution. When clinic-specific label shift is unmeasured, the fairness of the aggregated model across sites is not identified by the quantities that standard benchmarks report. The framework is operationalised as a design-based evaluation of the before-after-control-impact estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power 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 analysis characterises the operating characteristics of the design rather than reporting observational findings. The results indicate that 54 clusters are required for 80% power at a 5 percentage point effect, and that the intraclass correlation of 0.64 dominates the detectable effect at small cluster counts. The contribution is a fairness-transfer benchmark that separates aggregation-rule performance from label-shift identifiability, together with a design-theoretic account of why unmeasured clinic-specific shift makes cross-site fairness claims underpowered at realistic federation sizes.