Clinical Computer Vision for Cervical Precancer Detection Under Variable Acetic Acid Application Timing: A Test-Time Adaptation Evaluation of Artificial Intelligence Colposcopy Referral Sensitivity

Colposcopic artificial intelligence (AI) systems for cervical precancer detection are typically trained and validated on images acquired at a fixed interval after acetic acid application, yet in routine screening the interval varies with operator workload, patient tolerance and clinic throughput. the present analysis formalises that mismatch as a covariate-shift problem and evaluates, in a replication-based design study, whether test-time adaptation can preserve referral sensitivity when the acetic acid interval is not controlled. The analysis characterises the operating characteristics of a before-after-control-impact estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) across 2000 replications per design cell, with cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Power to detect a 5-percentage-point sensitivity difference reached 80% at 52 clusters. The results indicate that unadjusted comparisons of adapted and non-adapted AI systems are biased toward the null when the intraclass correlation is high, and that test-time adaptation should be evaluated with cluster-randomised designs sized to the expected shift magnitude rather than with fixed validation sets.