Presentation (CTI)

OBJECTIVES: Healthcare decision-makers must navigate large volumes of clinical, health technology assessment (HTA), and regulatory evidence, making traditional evidence synthesis slow and resource-intensive. This study developed and validated an artificial intelligence (AI)-assisted visualization platform and chatbot to present clinical effectiveness and real-world treatment patterns in KRASG12C-mutant metastatic colorectal cancer (mCRC)
METHODS: The platform used a Retrieval-Augmented Generation (RAG) framework + multi-agentic architecture to enable transparent and traceable evidence retrieval. Source documents (PDF, Word, PPT, and TXT) were processed via Optical Character Recognition when needed and standardized into structured markdown. Content was embedded in a vector database for semantic search. User queries retrieved the most relevant evidence, which was passed to a large language model to generate grounded, source-linked responses. All outputs were reviewed by subject matter experts (SMEs) for accuracy, relevance, and traceability.
RESULTS: The platform generated customized visual outputs (i.e., pie chart, bar graph, etc.) showcasing distribution of baseline characteristics, treatment patterns and clinical effectiveness for 17 studies in the systematic literature review (SLR) of KRASG12C-mutant mCRC. An SME validated all outputs for accuracy, traceability, and relevance. Performance metrics revealed 92% accuracy in AI-generated responses versus manual synthesis, with response generation time averaging 10 seconds versus 30-60 minutes for traditional approaches. The system effectively integrated and balanced evidence from multiple sources, significantly enhancing data interpretation efficiency while maintaining high accuracy.
CONCLUSIONS: This AI-assisted visualization platform and chatbot advances evidence generation by enabling quick access to contextualized evidence with multiple visualization options. It bridges the evidence-to-utilization gap by supporting efficient transformation of generated outputs into a range of health economic and outcomes research deliverables. The platform can ingest any SLR to generate visual outputs and/or summary text, to support effective communication of SLR findings. Additionally, the integrated chatbot functionality enables users to generate tailored insights by querying any data type within the included publications.