Artificial Intelligence Adversarial Robustness Optimisation of Diabetic Retinopathy Screening Classifiers Under Variable Fundus Camera Illumination Drift: A computed Healthcare Fairness Evaluation

Diabetic retinopathy screening classifiers trained on fundus photographs are known to degrade when camera illumination drifts, but the fairness consequences of that degradation across patient subgroups have not been characterised with a design that separates illumination-induced distribution shift from ordinary sampling variability. The framework formalises illumination drift as a bounded perturbation of the image-formation operator and derives a robust training objective that minimises worst-case subgroup loss over an ambiguity set defined by a Wasserstein ball around the empirical illumination distribution. Operating characteristics were estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, under a declared variance structure with between-cluster SD of 6.0, residual SD of 4.5 and intraclass correlation of 0.64. The design achieved 80% power to detect a 5 percentage point fairness gap at 54 clusters. Illumination drift produces subgroup-specific error inflation that is invisible to aggregate accuracy metrics, and that the robust objective trades a small aggregate accuracy cost for a substantial reduction in worst-subgroup error. The contribution is a formal bridge between adversarial robustness and algorithmic fairness evaluation for screening classifiers, together with a design-theoretic account of how many clusters a fairness evaluation requires.