Skip to content

EpiPlanAgent: Agentic Automated Epidemic Response Planning

🕒 Published (v1): 2025-12-11 06:03 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EpiPlanAgent is a multi-agent system for automated epidemic contingency plan generation, built on SigmaFlow—an open-source Python graph-based orchestration framework—combined with DeepSeek-V3 and a structured RAG knowledge base. The system converts unstructured outbreak reports into actionable, guideline-aligned JSON task lists through a sequential node pipeline with iterative feedback refinement. In a controlled study with 25 public health professionals, it raised plan completeness from 68.7% to 82.4% while cutting generation time by 93.9%.

Problem

Traditional epidemic response planning is expert-dependent, slow (≈24.5 min/plan), inconsistent across jurisdictions, and fails to scale during rapidly evolving outbreaks. Existing LLM applications in public health are limited to single-turn tasks (triage, information extraction); no prior system automated end-to-end structured contingency plan generation with iterative refinement and domain grounding.

Method

EpiPlanAgent orchestrates a directed-graph workflow using SigmaFlow, which models tasks as typed nodes: Model Nodes (LLM calls via DeepSeek-V3), Tool Nodes (RAG, Code Execution, Web Search), and Logic Nodes (Branch, For Loop). The pipeline executes sequentially: 1. Epidemic identification: a Code Node retrieves candidate disease types from the knowledge base; a Model Node extracts the epidemic type; an RAG Node performs exact-match retrieval of structured JSON "Candidate Plans" keyed by disease type (bypassing vector search to avoid semantic drift). 2. Trigger condition decomposition: a Code Node extracts trigger conditions; a Model Node decomposes them into atomic boolean judgment points; another Model Node maps the free-text report to a structured case object. 3. Task list generation + iteration: a Model Node synthesizes the structured case + candidate plans into a JSON task list (action, responsible party, deadline). A Logic Node checks for prior feedback; if present, the agent re-enters a refinement branch for multi-turn iterative optimization (Round 1 → Round 2).

The knowledge base covers 8 diseases across 40+ official CDC/municipal documents, stored as structured JSON with fields: Action, Trigger Condition, Work Requirement, Responsible Party, Time Limit.

Key Contributions

  • SigmaFlow-orchestrated agentic pipeline: first application of a graph-based LLM orchestration framework to epidemic response planning, with explicit dependency management and parallel execution.
  • Exact-match structured RAG: replaces vector similarity search with disease-type-keyed JSON retrieval, ensuring high recall/precision against authoritative guidelines without semantic drift.
  • Multi-round iterative refinement: Logic Node-driven feedback loop demonstrably improves completeness (R1: 78.0% → R2: 82.4%) without significant time cost (1.8 → 1.5 min).
  • Experience-level standardization: junior professionals achieve plan completeness (81.7%) statistically comparable to senior professionals (83.5%), demonstrating quality floor-lifting.
  • Controlled human-subject evaluation: 25 public health professionals, 16 real-world scenarios across 8 diseases, pre-post design with expert panel scoring.

Results

  • Completeness Score: 82.4 ± 6.3% (EpiPlanAgent R2) vs. 68.7 ± 7.9% (manual), p < 0.001
  • R1 already significantly exceeds manual: 78.0 ± 6.0%, p < 0.01
  • Generation time: 1.5 ± 0.4 min vs. 24.5 ± 5.1 min manual — 93.9% reduction, p < 0.001
  • Expert content alignment: Pearson r = 0.92 (95% CI: 0.87–0.96, p < 0.001) vs. gold-standard human-authored plans
  • Per-disease completeness range: 79.8% (Hand-Foot-and-Mouth) to 85.1% (COVID-19)
  • By disease category: Respiratory 82.9%, Foodborne/Contact 82.2%, Vector-borne 81.2%
  • User satisfaction: 91.5% rated system "very helpful" or "extremely helpful"

Limitations

  • Static knowledge base: covers only 8 pre-compiled diseases; cannot handle novel or emerging pathogens lacking structured guidelines.
  • Single-stage scenario model: workflow processes one report snapshot; no real-time dynamic re-planning as case counts or resources change.
  • LLM dependency: epidemic type extraction and task list generation quality are bounded by the backbone model's instruction-following capability.
  • Small evaluation cohort: N=25 participants, 16 scenarios — limited generalizability across health system contexts and jurisdictions.
  • No cross-jurisdictional validation: knowledge base sourced from Shanghai/national Chinese CDC standards only.

Relevance to Harnesses / Meta-Harnesses

EpiPlanAgent is a domain-specific harness built on SigmaFlow, a reusable meta-harness for LLM/MLLM/multi-agent task-flow orchestration. SigmaFlow is the structural contribution of broader interest: it formalizes workflows as directed graphs of typed nodes (Model, Tool, Logic), manages inter-node dependencies, and enables parallel execution — exactly the architecture pattern that generic agentic meta-harnesses (LangGraph, SmoLAgents, AutoGen) provide. The paper is a concrete instantiation and evaluation demonstrating that a graph-based orchestration layer with explicit node typing, exact-match retrieval integration, and Logic Nodes for conditional branching/iteration produces measurably better outputs than manual baselines, validating the harness design pattern in a high-stakes domain. The iterative refinement loop (Logic Node checking feedback state → branch re-execution) is a clean demonstration of stateful multi-turn harness control flow.