Artificial Intelligence Benchmarking of Cervical Cancer Visual Inspection Classifiers Under Unmeasured Acetic Acid Application Interval Variability: A Healthcare Calibration-Transfer Evaluation

Visual inspection with acetic acid (VIA) is a widely used cervical cancer screening method in low-resource settings, and artificial intelligence classifiers trained on VIA images are increasingly proposed as decision support. However, the interval between acetic acid application and image capture is rarely recorded in routine practice, and classifiers may be evaluated on images acquired under heterogeneous, undocumented timing conditions. the present analysis formalises the problem as a calibration-transfer failure: a classifier calibrated on images from one distribution of application intervals will misestimate risk when deployed under another. We develop a theoretical framework that decomposes the expected calibration error into a component attributable to interval variability and a component attributable to classifier miscalibration proper. 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), we characterise the operating characteristics of the evaluation design across 2000 replications per cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points.52 clusters are required to detect a 5-percentage-point effect at 80% power. We derive a bound on the bias introduced by unmeasured interval variability and propose a sensitivity analysis that allows benchmark designers to quantify the robustness of classifier comparisons to this unmeasured confounder. The contribution is a formal treatment of a measurement problem that is pervasive in VIA-based AI benchmarking but rarely acknowledged.