A Unified Framework for Epidemic Forecasting and Contact Reconstruction via Bidirectional Enhancement

Yuanyuan Zeng, Kai Xiang

Abstract: Epidemic forecasting and contact network reconstruction are inherently coupled, yet most existing methods address them independently, limiting performance and interpretability. In this work, we propose EpiLoop, a unified closed-loop framework that enables bidirectional enhancement between case forecasting and latent contact network inference. EpiLoop integrates two synergistic components: DualView-GraphLSTM, a spatiotemporal graph neural network that models infection dynamics and mobility patterns, and VGAE-CR, a variational graph autoencoder that reconstructs time-varying subpopulation-level contact networks. Crucially, EpiLoop establishes a feedback mechanism: reconstructed networks are fed back to refine future forecasts, while latent infection states from forecasting guide network reconstruction—enabling mutual co-refinement through iterative optimization. We evaluate EpiLoop on real-world COVID-19 trajectory data from four Chinese provinces, using a comprehensive metapopulation network. Results show that EpiLoop achieves state-of-the-art performance, outperforming decoupled baselines in both forecasting accuracy (lower MAE and RMSE) and contact reconstruction (AUC up to 88.8%) under partial observability and behavioral changes. By unifying prediction and mechanism discovery, EpiLoop advances epidemic modeling toward transparent, self-improving systems and generalizes to behavior-modulated dynamics beyond COVID-19.

External IDs:doi:10.1109/tetci.2025.3637844