Artificial Intelligence Constrained Multi-Objective Optimisation of Community Clinic Snakebite Antivenom Stocking Under Variable Species-Imprecise Envenomation Diagnosis: A computed Healthcare Mortality Equity Evaluation

Community clinics in rural sub-Saharan Africa stock polyvalent antivenom under diagnostic uncertainty: the biting species is rarely identified, yet the antivenom's species coverage and the patient's mortality risk depend on that unidentified species. the present analysis formulates antivenom stocking as a constrained multi-objective optimisation problem in which the decision variable is a vector of vial quantities across antivenom products, the objectives are expected mortality and mortality equity across presentation profiles, and the constraints are budget, cold-chain capacity and product-specific species coverage. Because species-imprecise diagnosis induces a partially observable state, the formulation is a constrained partially observable Markov decision process; a Lagrangian relaxation yields a scalarised objective whose subgradient is computable from the belief state, and a projected subgradient algorithm is proposed. The framework is evaluated in a computed environment parameterised from published snakebite epidemiology and antivenom cost data; no field data were collected. Across 500 computed clinic-years, the constrained optimiser reduced expected mortality by 23.4% relative to a coverage-maximising heuristic and by 11.7% relative to an unconstrained mortality-minimising allocation, while narrowing the inter-profile mortality gap from 4.1 to 1.9 deaths per 1,000 presentations. The equity constraint was binding in 62% of computed clinic-years, indicating that mortality-minimising allocations systematically neglect low-presentation-frequency profiles. The contribution is a formal optimisation framework that makes the equity cost of species-imprecise diagnosis explicit and computable, together with a proof that the Lagrangian dual bound is tight under product-independence of species coverage.