Artificial Intelligence Graph Neural Network Reconstruction of Facility Catchment Populations Under Variable Health Information System Boundary Reassignment: A Spatio-Temporal Evaluation of Healthcare Planning Denominators

Health information systems in low- and middle-income countries routinely reassign facility catchment boundaries, so the denominators used for planning are unstable even when the underlying population is not. The framework treats facilities as nodes in a spatial graph, propagates population signals along edges weighted by travel impedance, and learns a boundary-invariant representation that can be transferred across administrative reconfigurations. The empirical component uses the Demographic and Health Surveys indicator 'Households possessing a computer' (HC_HEFF_H_CMP) across 27 African countries and 56 country-survey observations from 2015 to 2025, with a mean of 7.893 and a standard deviation of 5.144. The paper formalises catchment reconstruction as a constrained graph signal recovery problem, proves a stability bound on the reconstructed denominator under boundary perturbation, and specifies an evaluation protocol that separates boundary-induced variance from true population change. The contribution is a formal account of when graph-based reconstruction is identifiable and a design for evaluating it against the boundary reassignments that health systems actually perform.