Presentation (CTI)

OBJECTIVES: When head-to-head clinical trial evidence is unavailable, indirect treatment comparisons (ITCs) are essential for health technology assessments, but their validity depends on complex, multi-domain feasibility evaluations. This study evaluates the performance of a Generative AI (GenAI)-based framework designed to conduct transparent, traceable, and HTA-aligned feasibility assessments for ITCs and to validate its outputs against expert human assessments.
METHODS: A multi-agentic, retrieval-augmented generation (RAG)-based architecture was developed to support feasibility analysis for indirect treatment comparisons. Clinical trial data from multiple sources were uploaded, standardized, and indexed using a RAG pipeline to enable the retrieval of trial characteristics, baseline variables, and outcomes. Multiple GenAI agents were configured with domain-specific prompts to independently evaluate four feasibility domains: population similarity, evidence network structure, trial design heterogeneity, and outcome definition alignment. All outputs were generated in accordance with HTA, ISPOR, Cochrane, and JCA methodological guidance.
RESULTS: The GenAI framework produced a feasibility report with sections covering network diagrams, comparisons of study and patient characteristics, statistical tests, and analysis recommendations, including sensitivity analyses. The AI-assisted feasibility flagged the outlier studies and proposed data-driven sensitivity and scenario analyses. The AI-assisted feasibility assessment required human editing in terms of managing/deleting repetitive information and formatting changes. Human ITC expert marked readiness of AI-assisted feasibility at 85%, requiring only10-15% human intervention to finalize it.
CONCLUSIONS: The GenAI framework produced ITC feasibility assessments that closely matched expert evaluations, requiring only minimal human refinement, and shows strong potential to improve the efficiency, consistency, and rigor of HTA-aligned evidence synthesis.