Presentation (CTI)

OBJECTIVES: Generative AI has attracted growing attention in HEOR, but foundational AI methods designed explicitly for reasoning and decision-making under uncertainty remain underexplored. These methods include decision-theoretic planning, reinforcement learning, game theory, probabilistic graphical models, and symbolic reasoning. This study mapped how far these paradigms have been applied to explicit health-economic decisions, as distinct from estimation-oriented uses.
METHODS: A targeted scoping review informed by PRISMA-ScR was conducted with a taxonomy of foundational decision-AI methods developed a priori. PubMed was searched using method-specific and HEOR terms, yielding 888 records across 13 algorithm streams. PMIDs were verified against PubMed metadata. Records were classified using LLM-assisted implementation of a predefined framework distinguishing Decision-AI uses, defined as methods linked to quantified economic objectives such as costs, QALYs, ICERs, and resource allocation, from estimation-oriented uses such as prediction, disease progression modelling, and evidence synthesis. LLM re-screening of 180 stratified records assessed rule-application consistency (κ=0.63); 66 priority records underwent blind human adjudication.
RESULTS: Among 888 screened records, 16 unique studies (under 2%) met Decision-AI criteria, concentrating in decision-theoretic planning applied to screening, treatment-timing, and resource-allocation problems with operationalized cost-effectiveness objectives. Reinforcement learning appeared in healthcare applications but rarely with economic outcomes in the reward function. Game-theoretic and symbolic reasoning were minimally represented. Strikingly, none of 179 Bayesian-network records met Decision-AI criteria, despite probabilistic graphical models being widely used in HEOR for estimation tasks including network meta-analysis, disease progression, and risk stratification.
CONCLUSIONS: Foundational AI methods for reasoning and sequential decision-making are present in HEOR but rarely implemented within economic decision frameworks. Limited uptake likely reflects structural factors such as unspecified economic objectives and disciplinary boundaries, rather than fundamental methodological limitations. The most tractable near-term opportunity is reinforcement learning, where the healthcare infrastructure already exists and the missing piece is specifying economic outcomes in the reward function.