Artificial Intelligence Benchmarking of Tuberculosis Contact-Tracing Prioritisation Under Unmeasured Household Screening Delay Reporting: A Healthcare Transmission-Model Evaluation

Contact-tracing prioritisation algorithms for tuberculosis are benchmarked against transmission outcomes that depend on household screening delay, yet the delay itself is rarely measured and never reported in the routine data on which those algorithms are trained. the present analysis formalises the resulting identification problem and derives its consequences for benchmark validity. A latent-delay transmission model is specified in which the true household screening delay is an unobserved confounder of the relationship between algorithmic priority scores and onward transmission; the model yields a closed-form bias expression for the benchmarked effectiveness of any prioritisation rule. First, the ranking of competing prioritisation algorithms is not identified from observable benchmarking data whenever delay is correlated with the covariates the algorithm uses. Second, a partial-identification bound is available from the observable distribution of notification intervals, and this bound is tight when delay is independent of household size. Third, the bias is signed: algorithms that preferentially prioritise large households are systematically over-credited, because large households experience longer unmeasured delays and therefore contribute more secondary cases per index case regardless of algorithmic quality. The framework is illustrated with the distribution of household computer possession across 27 African countries (56 country-survey observations, mean 7.893, standard deviation 5.144), which proxies the household-level data infrastructure that would be required to measure delay directly. The contribution is a benchmark-validity criterion and a partial-identification estimator that can be applied to existing contact-tracing evaluations without new delay measurement.