Presentation (CTI)

OBJECTIVES: Health economic modelers spend considerable time on evidence gathering, iterating model structures, and writing documentation before a model is even running. These early-phase tasks are manual and repetitive. We developed an agentic AI platform that accelerates early-stage modeling while keeping simulation in the hands of community-supported open-source packages. The platform is designed with clear trust boundaries and an extensible architecture, allowing new model types and computational engines to be added without changing the core workflow.
METHODS: AI agents with deep research capabilities synthesize clinical and economic evidence from published literature, reducing manual search time. Modelers can quickly test alternative model structures through a YAML-based model specification—chosen for being human-readable, easy to review, and token-efficient for AI workflows. This specification serves as both documentation and executable input. Simulation is handled by open-source packages—heemod and hesim in R, pyDICE in Python—routed automatically based on model taxonomy. The platform can export executable code (R/Python) or Excel workbooks. All AI-extracted parameters remain traceable to source documents, and results are reproducible through the underlying packages.
RESULTS: In a proof-of-concept, the platform demonstrated the feasibility of shortening the path from published evidence to a running simulation from weeks to hours. For supported model classes, changing model structure no longer requires recoding; modifying the YAML specification and re-running the pipeline is sufficient. Model specifications were routed to suitable computational engines based on model taxonomy, demonstrating interoperability through a shared specification framework and engine-specific translators. Generated documentation maintained traceability to source evidence.
CONCLUSIONS: We found AI most useful when focused on evidence synthesis, structuring, and documentation, while validated packages handle computation. This separation of concerns avoids additional review burden and builds on tools the HTA community already trusts, with output available as executable code or Excel for downstream review.