Clinical Computer Vision for Diabetic Foot Ulcer Infection Grading Under Variable Smartphone Color Constancy: A Test-Time Adaptation Evaluation of Artificial Intelligence Sepsis Escalation Safety

Smartphone-based diabetic foot ulcer (DFU) photography is increasingly proposed as a triage input for infection grading and sepsis escalation in low-resource clinics, yet the colour constancy of such images varies with illuminant, sensor and white-balance pipeline in ways that convolutional grading models were not trained to tolerate. the present analysis treats the safety of test-time adaptation (TTA) under that shift as a design problem rather than a model-selection problem. It formalises DFU infection grading as a risk-controlled prediction task in which the clinically consequential error is under-escalation of a limb-threatening infection, and it derives the conditions under which unsupervised TTA can be certified not to increase that error. The analysis is methodological and theoretical: no clinical images, patients or model outputs are reported. Using a before-after-control-impact design with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), the paper characterises the operating characteristics of the evaluation that a TTA safety claim would require, and shows that 54 clusters are needed to detect a 5-percentage-point change at 80% power. The central argument is that colour-constancy shift and sepsis-escalation safety are coupled through a single quantity, the calibration of the under-escalation probability, and that TTA should be admitted only when that quantity is bounded by a pre-registered margin.