Artificial Intelligence Benchmarking of Referral-Network Graph Neural Triage Under Unmeasured Inter-Facility Transport Time Censoring: A Healthcare Safety-Robustness Evaluation

Graph neural networks (GNNs) are increasingly proposed for triage across inter-facility referral networks, yet their safety evaluation typically assumes that transport times between facilities are observed. In many low- and middle-income referral systems, transport time is unmeasured and right-censored: only the fact that a patient arrived, or did not, is recorded. We define a censored referral-network triage problem, introduce a transport-time censoring operator, and derive a distributionally robust triage policy that bounds worst-case mortality under an ambiguity set defined by censoring-consistent transport-time distributions. We prove that the robust policy admits a tractable convex reformulation and that its safety gap relative to the oracle policy is bounded by the Wasserstein radius of the ambiguity set. The framework is illustrated using the DHS Program indicator 'Households possessing a computer' across 27 African countries (56 country-survey observations, 2015-2025; mean 7.893, SD 5.144), which serves as a proxy for digital readiness of referral nodes. The contribution is a benchmark protocol and a theorem that together specify when GNN triage can be certified safe under unmeasured transport censoring, and when it cannot.