Towards Reliable Spatiotemporal Epidemic Forecasting via Steering Diffusion Inference

Guanghui Min, Tianhao Huang, Ke Wan, Qi R. Wang, Chen Chen

Keywords: Computational Epidemiology, Spatiotemporal Forecasting, Physics-informed models

Abstract:

Reliable epidemic prediction is vital for public health response and resource allocation, especially in rapidly evolving outbreaks.
Despite the recent attempts to integrate the epidemic mechanistic model into data-driven forecasting models, existing approaches still lack explainability and robustness.
To bridge this gap, we propose **EpiDiff**, an epidemiology-aware diffusion framework that incorporates mechanistic estimations and their posterior uncertainties into the forecasting process. **EpiDiff** features a flexible and high-capacity diffusion backbone specifically designed for spatiotemporal epidemic data, enabling accurate and robust sequence prediction. By quantifying the uncertainty of mechanistic forecasts and using it to steer the diffusion model at inference, **EpiDiff** dynamically adjust the data-driven prediction with the guidance from epidemic model.
Extensive experiments on real-world epidemic datasets demonstrate that **EpiDiff** consistently outperforms state-of-the-art baselines in both accuracy and robustness, while offering improved explainability for epidemic forecasting.
Our code and datasets are available at https://anonymous.4open.science/r/epidiff-4782.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 9683