Auditable Context-Aware HFMD Forecasting with Structured LLM Agents

Core Contribution: Decoupling Semantic Interpretation from Statistical Modeling

This work addresses the semantic gap in hand-foot-mouth disease (HFMD) forecasting—where external covariates (e.g., school schedules, temperature, public health policies) are poorly interpreted by conventional time-series models—by proposing a two-stage neuro-symbolic architecture that explicitly separates ‘contextual semantic parsing’ from ‘probabilistic time-series modeling’.

Technical Architecture: LLM Agent + Statistical Generator

• Event Interpreter: An LLM-driven structured agent ingesting unstructured/semi-structured inputs—including Ministry of Education academic calendars, China Meteorological Administration weather summaries, Chinese Center for Disease Control and Prevention (China CDC) surveillance bulletins, and clinical practice guidelines—and outputting a single normalized scalar: the transmission-impact signal, accompanied by traceable, verifiable reasoning traces;
• Forecast Generator: Integrates this signal with historical HFMD case counts (e.g., weekly national reports from China CDC) and produces probabilistic forecasts via Poisson or negative-binomial moment matching—enabling quantified uncertainty intervals and actionable risk thresholds.

Baseline Comparison: Highlighting the Semantic Gap

Classical models (ARIMA, Prophet) and foundation time-series models (Chronos, Moirai, TimesFM) treat external covariates as numeric embeddings, failing to explicitly encode epidemiological causal chains (e.g., ‘winter break ends → students return → contact rate rises → transmission accelerates’); this framework is the first to achieve end-to-end auditable coupling of symbolic contextual interpretation and probabilistic forecasting for HFMD.