Abstract
Epidemic response planning is a critical yet labor-intensive task in public health, often hindered by manual processes and static protocols. We present EpiPlanAgent, a multi-agent AI system designed to automate the generation of actionable response plans. By grounding Large Language Models (LLMs) in a structured epidemiological knowledge base and employing an iterative agentic workflow, EpiPlanAgent reduces plan generation time by approximately 94% while improving plan completeness and alignment with official guidelines.
Introduction
Traditional epidemic response planning relies heavily on human experts manually synthesizing information from static, paper-based protocols. This process is inherently slow, prone to inconsistency, and difficult to scale during rapid outbreaks. There is a critical need for digital tools that can transform these static guidelines into standardized, actionable protocols in real-time. EpiPlanAgent addresses this by leveraging the reasoning capabilities of LLMs within a structured, multi-agent framework to produce high-quality contingency plans in minutes.
Methodology
EpiPlanAgent is built upon the SigmaFlow framework, which orchestrates task dependencies and parallel execution. The system utilizes DeepSeek-V3 as its core reasoning engine and follows a multi-stage agentic workflow:
- Epidemic Identification: Extracts disease types from unstructured user reports.
- Structured Retrieval-Augmented Generation (RAG): Performs exact-match retrieval from a pre-classified JSON knowledge base of official guidelines, preventing the semantic drift common in vector-based RAG.
- Trigger Analysis & Case Structuring: Decomposes complex trigger conditions into atomic judgment points to convert unstructured inputs into structured facts.
- Iterative Refinement: Employs a feedback loop where the agent refines task lists based on field data or expert input, simulating a human-in-the-loop process.
The system's knowledge base was constructed from over 40 official textual resources, covering 8 major infectious diseases across respiratory, enteric, and vector-borne categories.
Results and Discussion
EpiPlanAgent was evaluated against manual planning by 25 public health professionals across 16 real-world-inspired outbreak scenarios.
Efficiency and Quality
The system achieved a 93.9% reduction in planning time, decreasing the average generation time from 24.5 minutes to 1.5 minutes. Simultaneously, the Plan Completeness Score improved from 68.7% (manual) to 82.4% (assisted). The high Pearson correlation (r = 0.92) with human-authored expert content demonstrates strong alignment with professional standards.
Key Insights
A significant technical finding is the superiority of structured JSON retrieval over standard vector-based RAG for high-stakes domains. Exact-match retrieval ensures 100% adherence to official policy, eliminating hallucinations. Furthermore, the iterative workflow (Round 2 vs. Round 1) showed measurable gains in completeness, validating the agentic approach for complex reasoning tasks.
Conclusion
EpiPlanAgent demonstrates the potential for agentic AI to "level up" public health response by standardizing expertise and drastically reducing response latency. Future work will focus on integrating dynamic web search for emerging pathogens and expanding the multi-agent architecture to handle long-term resource allocation and scenario simulation.
Summary updated