Artificial Intelligence Benchmarking of Community Health Worker Triage Questionnaires Under Unmeasured Verbal Autopsy Recall Decay: A Healthcare Label-Noise Robustness Evaluation

Verbal autopsy (VA) remains the principal source of cause-of-death labels for community health worker (CHW) triage algorithms in low-resource settings, yet the recall interval between death and VA interview is rarely recorded and almost never modelled. the present analysis treats unmeasured recall decay as a structured label-noise process and characterises the operating characteristics of a before-after-control-impact (BACII) benchmarking design used to evaluate CHW triage classifiers under that noise. No observational data are used or claimed: the analysis is a replication-based evaluation of the estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.640), with power estimated across 2000 design cells spanning cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design attains 80% power at 54 clusters for a 5-percentage-point effect. The paper derives a bias bound showing that recall decay inflates the apparent between-cluster variance component and therefore the intraclass correlation, and it proposes a sensitivity-analysis protocol that reports benchmarking conclusions as a function of an assumed decay rate rather than as a single point estimate. The contribution is a design-level account of how label noise that is unmeasured but structurally bounded can be made visible in AI benchmarking for healthcare, together with the sample-size consequences of that visibility.