Artificial Intelligence Benchmarking of Continuous Glucose Monitor Alarm Thresholds Under Unmeasured Skin-Tone-Dependent Sensor Lag: A Healthcare Equity-Constraint Evaluation

Continuous glucose monitor (CGM) alarm algorithms are benchmarked almost exclusively on accuracy and detection delay, with skin-tone-dependent sensor lag treated as an unmeasured nuisance rather than a design constraint. The framework is a before-after-control-impact design with eight, sixteen, twenty-four and thirty-two clusters, effect sizes of three, five and eight percentage points, an intraclass correlation of 0.640, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5, evaluated across 2000 replications per design cell. The analysis derives the bias of the naive threshold estimator under unmeasured lag, shows that the equity-constrained estimator is consistent when the lag is bounded, and establishes that fifty-four clusters are required to detect a five-percentage-point effect at eighty per cent power. The contribution is a benchmark specification that makes the equity constraint explicit and auditable rather than implicit in the choice of threshold.