Presentation (CTI)

OBJECTIVES: Health technology assessment (HTA) submissions require robust economic models to demonstrate value and inform reimbursement decisions. Traditional model development is resource-intensive, requiring 12-16 weeks and specialized expertise. This study presents a novel multi-agentic artificial intelligence (AI) system utilizing retrieval-augmented generation (RAG) to automate HTA-ready health economic model development in Excel, significantly reducing development time while maintaining methodological rigor and human oversight.
METHODS: A multi-agentic AI framework was developed comprising specialized agents (clinicians, statisticians, mathematicians, economists, health economic modelers, Excel experts, and VBA programmers) to generate decision tree, Markov, and hybrid models. The system accepts model development protocols as primary input, extracting model structure, transition probabilities, cycle length, time horizon, discounting rates, clinical inputs, and cost parameters through RAG architecture. Users can provide additional details or specifications through iterative feedback loops. The algorithm is designed to generate fully functional Excel models with standardized sheets: Home, General Settings, Clinical Inputs, Cost Inputs, Quality-of-Life Inputs, Parameters (with appropriate statistical distributions per best practices), One-Way and Probabilistic Sensitivity Analysis, Calculations, and centralized data storage. Advanced features include automated dropdown menus and data validation cells. Human oversight validates outputs through multi-perspective verification including formula accuracy, cell validation, cross-sheet linkages, and structural assessment.
RESULTS: AI-generated models demonstrated 65-70% alignment with subject matter expert specifications across 3 validation cases. Development time decreased by 75% compared to manual workflows (from 15-16 weeks to 3-4 weeks). Formula linkage accuracy reached 82%, with cross-sheet references correctly established in 583/710 of instances. Parameter sheet completeness achieved 90%, with statistical distributions correctly assigned in 96% [126/130] of cases. User acceptance testing by 5 health economists showed 80% satisfaction with model usability.
CONCLUSIONS: Multi-agentic AI systems with human oversight substantially accelerate HTA-ready model development while maintaining quality standards. VBA automation limitations require resolution. Future developments will incorporate advanced validation algorithms and complex model architectures.