DEVELOPMENT OF A MULTI-AGENT "DEEP" AI SYSTEM FOR HEOR EVIDENCE AND INTELLIGENCE GENERATION
Author(s)
Jag Chhatwal, PhD1, Mine Tekman, PhD2, Ismail F. Yildirim, MSc2, Sumeyye Samur, PhD2, Turgay Ayer, PhD2;
1Massachusetts General Hospital/Harvard Medical School, Boston, MA, USA, 2Value Analytics Labs, Boston, MA, USA
1Massachusetts General Hospital/Harvard Medical School, Boston, MA, USA, 2Value Analytics Labs, Boston, MA, USA
OBJECTIVES: While Generative AI offers promise, standard Large Language Models (LLMs) often lack the reasoning depth and traceability required for rigorous HEOR tasks. We introduce ValueGen.AI, a novel multi-agentic system designed to revolutionize evidence gathering by transitioning from simple prompting to autonomous, deep agent workflows.
METHODS: We developed a hierarchical, multi-agent system utilizing the LangGraph Python library to execute complex research workflows. The architecture is composed of three functional layers: 1) A Main Orchestrator Layer, where primary agents decompose queries, manage workflow logic, and evaluate information sufficiency; 2) A Deep Agent Layer, comprising specialized sub-agents tasked with domain-specific retrieval; and 3) A Tool Layer utilizing the Model Context Protocol (MCP) via FastMCP to interface with heterogeneous data sources (databases, APIs, and web repositories). Inter-agent communication is managed via an asynchronous message queue, enabling parallel processing. Unlike linear query systems, this graph-based structure allows agents to iteratively pull data, critique relevance, and verify sources before synthesis.
RESULTS: The system successfully deploys 1000+ concurrent sub-agents to produce verifiable and traceable HEOR outputs. In testing, the platform generated complex deliverables, including health economic conceptual models, therapeutic analogs, and comparative HTA insight reports across multiple international jurisdictions. The multi-layer verification process significantly mitigated hallucination risks by enforcing strict citation constraints, ensuring that all generated insights are directly linked to retrieved primary sources.
CONCLUSIONS: Deep agentic AI workflows represent a paradigm shift from passive information retrieval to autonomous research assistance. By structuring AI as a hierarchy of specialized agents utilizing standardized protocols like MCP, HEOR professionals can significantly accelerate evidence synthesis and strategy development without compromising the scientific rigor required for value demonstration.
METHODS: We developed a hierarchical, multi-agent system utilizing the LangGraph Python library to execute complex research workflows. The architecture is composed of three functional layers: 1) A Main Orchestrator Layer, where primary agents decompose queries, manage workflow logic, and evaluate information sufficiency; 2) A Deep Agent Layer, comprising specialized sub-agents tasked with domain-specific retrieval; and 3) A Tool Layer utilizing the Model Context Protocol (MCP) via FastMCP to interface with heterogeneous data sources (databases, APIs, and web repositories). Inter-agent communication is managed via an asynchronous message queue, enabling parallel processing. Unlike linear query systems, this graph-based structure allows agents to iteratively pull data, critique relevance, and verify sources before synthesis.
RESULTS: The system successfully deploys 1000+ concurrent sub-agents to produce verifiable and traceable HEOR outputs. In testing, the platform generated complex deliverables, including health economic conceptual models, therapeutic analogs, and comparative HTA insight reports across multiple international jurisdictions. The multi-layer verification process significantly mitigated hallucination risks by enforcing strict citation constraints, ensuring that all generated insights are directly linked to retrieved primary sources.
CONCLUSIONS: Deep agentic AI workflows represent a paradigm shift from passive information retrieval to autonomous research assistance. By structuring AI as a hierarchy of specialized agents utilizing standardized protocols like MCP, HEOR professionals can significantly accelerate evidence synthesis and strategy development without compromising the scientific rigor required for value demonstration.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR236
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas