Artificial Intelligence replication of District Hospital Maternal Referral Ambulance Dispatch Under Unmeasured Road-Traffic Delay Censoring: A Healthcare Robustness-Constrained Reinforcement Learning Evaluation

Maternal referral ambulance dispatch in district hospitals is routinely planned with travel-time estimates that are right-censored by unmeasured road-traffic delay, so that the reinforcement learning policies trained on such data are evaluated against a distorted notion of what a dispatch action actually achieves. the present analysis formalises the censoring mechanism as a latent delay process that truncates observed arrival times, derives the bias it induces in the action-value function, and proposes a robustness-constrained policy evaluation that bounds the worst-case value loss over an ambiguity set consistent with the censoring pattern. The methodological argument is developed analytically and is accompanied by a replication-based evaluation of the ordinary least squares estimator that would underlie any empirical calibration of the delay model: 600 data sets of N = 500 were drawn from a declared data-generating process with intercept 12.0, exposure 0.42, group -3.1, severity 1.35 and error SD 4.2, using seed 445059128. Across the replications the intercept bias was 0.05317 with RMSE 0.86519, the exposure bias was -0.00104 with RMSE 0.01556, and empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. Robustness-constrained dispatch evaluation is feasible without imputing the censored delays, provided the ambiguity set is calibrated to the observable censoring rate, and it identifies the conditions under which the constrained policy coincides with the nominal one.