Computer Vision for Diabetic Foot Ulcer Classification Under Variable Smartphone Lighting: A Domain-Generalization Evaluation of Artificial Intelligence Referral Safety

Smartphone-based computer vision for diabetic foot ulcer triage is typically validated under lighting conditions that do not match those of the clinics where referral decisions are made. the present analysis treats the resulting distribution shift as a domain-generalization problem and asks whether reported classification accuracy survives realistic illumination variation, and what that survival implies for referral safety. Because no primary image corpus was available, the study is a methodological and analytical evaluation: it formalises illumination shift as a transformation of the input distribution, derives a bound on the degradation of a classifier's referral sensitivity under covariate shift, and characterises the operating characteristics of a before-after-control-impact design for field validation 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 spanning 8, 16, 24 and 32 clusters and effects of 3, 5 and 8 percentage points; 54 clusters are required for 5 percentage points at 80% power. Lighting-induced shift can silently erode sensitivity in the very range where referral decisions are made, that aggregate accuracy conceals this erosion, and that the field designs commonly used to validate such systems are underpowered to detect it. The contribution is a formal account of why lighting variation is a safety problem rather than a nuisance variable, and a design framework for evaluating it.