Presentation (CTI)

OBJECTIVES: Parameterisation of health economic models requires systematic extraction, synthesis, and validation of inputs from randomized controlled trials (RCTs) and published literature, a process that is resource-intensive and prone to inconsistency. This study evaluated a generative AI retrieval-augmented generation (RAG) pipeline designed to support automated parameterisation of health economic models, benchmarking its performance against manual extraction and validation.
METHODS: A Gen-AI RAG pipeline was developed to retrieve evidence from curated sources including peer-reviewed publications and RCT reports, extract candidate model parameters (e.g., clinical effectiveness, utilities, adverse event rates, and resource use), and generate structured parameter tables with supporting citations.Validation was conducted across multiple case studies representing different therapeutic areas and model types. Extracted parameters were compared with manually curated inputs across three evaluation domains: (1) data accuracy and completeness, (2) traceability to source evidence, and (3) suitability for direct use in cost-effectiveness model parameterisation. Outputs were independently reviewed by senior health economists.
RESULTS: Across case studies, the RAG pipeline successfully retrieved relevant evidence and generated structured parameter tables aligned with manual reference standards.The majority of core model inputs were correctly identified and accurately extracted, with full citation traceability to source documents. Discrepancies primarily related to contextual interpretation (e.g., subgroup definitions or outcome timing), requiring expert adjudication. Exact-match rates were highest for clinical efficacy data (>88%) but showed moderate variance in complex utility derivations from SLRs. Use of the RAG pipeline reduced parameterisation time by approximately 60-70% compared with manual workflows.
CONCLUSIONS: In this validation study, a Gen-AI RAG pipeline demonstrated the potential to substantially accelerate model parameterisation while maintaining traceability to published evidence. When used as a decision-support tool alongside expert review, such approaches may improve efficiency and consistency in health economic model development without compromising methodological rigor.