Faithfulness of Integrated Gradients for Computer Vision Retinopathy Graders Under Mydriatic Pupil Dilation Variability: A Healthcare Explanation Stability Estimation

Integrated gradients (IG) is widely used to attribute diabetic retinopathy (DR) grader decisions to pixels, yet its completeness axiom is defined relative to a fixed baseline image. When mydriatic pupil dilation varies between acquisitions, the baseline itself becomes a random variable, and the faithfulness guarantee no longer holds in the form usually assumed. the present analysis formalises that problem and estimates the stability of IG explanations under dilation-induced baseline perturbation. The attribution map is treated as a functional of the baseline, and a variance decomposition separates the contribution of baseline shift from that of model sensitivity. A before-after-control-impact design is used to characterise the operating characteristics of the stability estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power 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. The analysis is methodological: no clinical images or patient records were used, and no empirical attribution maps are reported. The contribution is a formal account of why IG faithfulness is dilation-conditional, a design for estimating explanation stability at the grader level, and a demonstration that 54 clusters are required to detect a 5-percentage-point stability effect at 80% power. The implication is that IG-based audit reports for retinopathy graders should report dilation-stratified stability rather than a single faithfulness score.