Interpretability Failure Modes of Computer Vision Triage Classifiers Under computed Camera Occlusion in Low-Resource Clinics: A Healthcare Robustness Evaluation

Computer vision triage classifiers deployed in low-resource clinics are routinely evaluated on clean images, yet their interpretability outputs are rarely stress-tested against the partial occlusions that arise from dust, condensation, poor lighting and improvised camera mounting. the present analysis characterises the interpretability failure modes that emerge when saliency-based explanations for such classifiers are evaluated under computed camera occlusion. Because no field data are available, the study is framed as a design-based robustness evaluation: a before-after-control-impact (BACI) estimator is characterised under a declared variance structure with intraclass correlation 0.64, between-cluster SD 6.0 and residual SD 4.5, across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per design cell.54 clusters are required to detect a 5 percentage point effect at 80% power, and it derives a formal decomposition of interpretability failure into attribution displacement, attribution collapse and attribution inversion. The contribution is a taxonomy of these failure modes together with a power-calibrated evaluation protocol that treats explanation stability, not only classification accuracy, as the primary endpoint for triage deployment in low-resource settings.