Presentation (CTI)

OBJECTIVES: Traditional health economic modeling requires substantial manual effort and multiple iterative validation steps. This study evaluated whether an agentic AI-assisted evidence-to-model workflow could reproduce a conventionally developed budget impact model while maintaining methodological transparency, reproducibility, and human oversight.
METHODS: A previously completed budget impact model for evaluating point-of-care rapid molecular testing, developed using conventional methods, served as the reference model. The same model was reconstructed using an agentic AI-assisted, human-in-the-loop workflow integrating protocol development, evidence synthesis, structured evidence, parameter extraction, model generation, sensitivity analyses, validation, and automated reporting. Outputs from the AI-assisted workflow were compared with the reference model for model structure, parameter agreement, model outputs, development time, and evidence traceability. Validation checkpoints followed ISPOR-SMDM modeling good practices and CHEERS 2022 reporting guidelines.
RESULTS: The agentic workflow successfully reproduced the reference model structure and generated equivalent budget impact estimates. Model inputs showed >90% agreement and development time was reduced by approximately 98%, while maintaining evidence traceability for all parameters. Human reviewers accepted much of AI-generated recommendations without modification. The AI-assisted workflow successfully integrated evidence from multiple sources into standardized model-ready data structures while maintaining parameter traceability. Human review remained integrated throughout model specification, parameter selection, structural validation, and interpretation, allowing iterative feedback and refinement at each stage. The workflow was successfully applied to decision tree, Markov, budget impact, and cost-effectiveness models while preserving transparent documentation of assumptions.
CONCLUSIONS: An agentic AI-assisted workflow successfully reproduced a conventional budget impact model while reducing development effort and maintaining transparency and human oversight. By combining standardized evidence structures with human-in-the-loop validation, agentic AI-assisted workflows may support consistent health economic model development for HTA, payer, regulatory, and clinical decision making. Future work will evaluate the workflow across additional HEOR model types and explore continuously updated ("living") models through real-time evidence integration.