Presentation (CTI)

OBJECTIVES: Health economics and outcomes research (HEOR) publications often contain key evidence in tables, figures, graphs, and complex document layouts that are challenging for conventional text-based AI workflows to interpret accurately. This study evaluated a multimodal generative artificial intelligence (GenAI) framework to automatically extract, interpret, and synthesize textual and visual evidence from HEOR publications to support AMCP dossier development within a human-in-the-loop workflow
METHODS: A Python-based multimodal AI framework was developed using a Retrieval-Augmented Generation (RAG) architecture. A multimodal preprocessing pipeline extracted textual content and native tables using document parsing and table extraction workflows, while specialized multimodal AI agents interpreted figures, graphs, and tables embedded as images. Outputs were independently evaluated by subject matter experts (SMEs) using a structured scoring framework across six dimensions: relevance, completeness, accuracy, language quality, traceability, and overall quality
RESULTS: Approximately 50 publications systematically selected and analyzed to generate different sections of an AMCP dossier. SME evaluation confirmed that the framework accurately extracted and interpreted evidence from narrative text, structured tables, figures, graphs, Kaplan-Meier curves, and tables embedded as images, enabling the generation of relevant, complete, and traceable AMCP dossier sections. Compared with a conventional manual workflow, the framework reduced development time by approximately 80%, representing a substantial efficiency gain without compromising evidence traceability. Minor refinements were limited to poor-quality or blurred images and graphs with poorly defined scales, highlighting the continued importance of expert review
CONCLUSIONS: Evidence data are presented in three primary forms: text, tables, and visual elements such as figures and graphs. Appropriate preprocessing of each form is essential for accurate interpretation and evidence synthesis. Multimodal GenAI integrated within a RAG architecture demonstrated feasibility for scalable, traceable AMCP dossier development, substantially reducing manual effort while preserving evidence integrity. Future work should improve interpretation of poor-quality visual content while retaining expert oversight