Health economic model conceptualisation moves to agentic AI: two cases match the core staging logic of NICE appraisals, with automated runs of about an hour

The conceptualisation stage of a health economic model can be carried out by an agentic generative-AI framework, which returns two outputs: a disease concept diagram and a model structure diagram. Conceptualisation has been a manual process requiring health economics, data analysis and clinical expertise across a multidisciplinary team, and it is time-consuming, resource-intensive and often underweighted. Two proof-of-concept cases covered data-rich advanced HR-positive, HER2-negative breast cancer and data-sparse hATTR-PN; outputs were reviewed against industry standards by two modellers with 10–12 and 5–6 years of experience and two clinicians, and externally compared with published economic evaluations.

The framework chains reasoning, retrieval and generation into one auditable multi-step pipeline in which the agent decides for itself when more information is needed. A single one-shot large language model query does not capture the nuances of a real case study, so the task is decomposed, retrieved for on demand and iterated on partial solutions. Reasoning used chain-of-thought with self-consistency, re-issuing the same prompt as 5 independent runs and keeping the states and transitions that recur most often by majority voting judged by another large language model; retrieval used GraphRAG, which organises guidelines and trial material into a graph of nodes and edges whose structure human experts reviewed and approved before retrieval.

In the data-sparse rare-disease case the framework reproduced the PND-staging structure but did not add stratification of each stage by NT-proBNP threshold. The company model in the NICE Highly Specialised Technology appraisal HST10 stratified each of the six PND stages by NT-proBNP threshold, giving 12 alive health states rather than 6. The gap is attributed to a retrieval corpus built from published literature and clinical guidelines rather than patient-level data, and expert reviewers judged that splitting a small ultrarare-disease population across 12 states would reduce the observations available for each stratum-specific transition probability.

The report is aimed at researchers and engineers in health economics and health technology assessment, and at readers tracking the role of generative AI in evidence generation. The demonstrated scope is two proof-of-concept cases: data-rich advanced HR-positive, HER2-negative breast cancer and data-sparse hereditary transthyretin amyloid polyneuropathy, with outputs consisting of a disease concept diagram and a model structure diagram compared structurally against published economic evaluations. The authors propose future blinded or double-blinded expert review with structured protocols, and further research to test generalisability and decision impact.

The authors state that the outputs are satisfactory for proof-of-concept purposes but should not be considered final, since a thorough clinical review using a Delphi panel and structured expert elicitation has not yet been conducted. External validation was not blinded: reviewers knew the outputs came from generative AI and knew the comparator published models, so reviewer expectations may have influenced judgements. Although the retrieval corpus was kept free of existing models in the disease area, the authors note this does not preclude the possibility that the large language model was already pre-trained on such data. They also state that reproducing the cardiac-biomarker stratification would require the agent to have direct access to patient-level data or to be explicitly prompted with a pre-derived stratification scheme.