EpiNarrate: Agentic AI turns messy outbreak data into trustworthy health reports

EpiNarrate, developed by Rituparna Datta and eight co-authors, tackles a critical pain point in public health communication: translating high-dimensional ensemble projections from models like the COVID-19 Scenario Modeling Hub into clear, accurate, and context-rich narratives. Instead of feeding raw numbers directly into a large language model (LLM)―which often leads to inconsistencies, omissions, or fragile outputs―EpiNarrate uses an agentic framework that first extracts scenario axes and organizes them into a partial-order schema. This enables systematic traversal of the multidimensional space (covering intervention assumptions, geography, demographics, outcomes, time horizons, and uncertainty quantiles). It then builds an augmented dataset and applies a comparison grammar that enforces both semantic and arithmetic consistency when generating quantitative statements.

To avoid redundancy while maximizing coverage, the framework employs an interestingness-driven selection mechanism grounded in maximum-entropy principles. In experiments, EpiNarrate generated narratives that were not only factually more accurate than naive LLM summaries but also covered a broader range of salient epidemiological patterns. The output style closely resembles that of expert-written reports, making it suitable for policymakers and the general public. The paper is available on arXiv (ID: 2607.15544) and marks a notable step in applying agentic AI to structured data-to-text problems in high-stakes domains like public health.