Artificial Intelligence Benchmarking of Under-Five Mortality Prediction Ensembles Under Unmeasured Verbal Autopsy Cause-of-Death Misclassification: A Healthcare Decision-Curve Evaluation

Verbal autopsy (VA) remains the primary source of cause-of-death information for under-five mortality in settings without complete civil registration, yet the cause labels it produces are known to be misclassified in ways that are rarely quantified and almost never propagated into the machine-learning pipelines that consume them. We define a misclassification operator on the label simplex, derive the bias it induces in ensemble risk scores and in decision-curve net benefit, and show that the ranking of ensembles by area under the receiver operating characteristic curve is invariant to symmetric label noise only under restrictive conditions that fail for the asymmetric confusion structures typical of VA. The framework is illustrated with a decision-curve evaluation protocol that separates ensemble discrimination, calibration and clinical utility, and that reports the range of misclassification rates over which a given ensemble ranking is preserved. We argue that benchmarking practice in this domain should report a misclassification-robustness envelope alongside conventional performance metrics, and we specify the sensitivity analysis that such reporting requires. The contribution is methodological: it does not report new empirical results but provides the formal apparatus and reporting standard needed before ensemble comparisons on VA-derived labels can be interpreted as evidence about clinical decision quality.