A mobility-aware and patch-based transformer for predicting spatiotemporal transmission patterns of epidemics

Accurately forecasting the spatiotemporal spread of infectious diseases is critical for enabling timely and data-driven public health responses. This paper presents GCN-PEFormer, a mobility-aware deep learning framework that integrates spatial dependencies through graph convolutional networks (GCNs) and temporal dynamics via a patch-based Transformer. To model spatial interactions, we construct mobility-informed dynamic graphs based on inter-state air travel data, further combined with geographic proximity. These graphs provide a realistic representation of transmission pathways across regions, enabling the model to capture both static and dynamic spatial influences. Besides, to better capture the non-stationary patterns of epidemic progression, we introduce a novel phase-aware patching strategy that aligns temporal segmentation with epidemiological stages such as acceleration, peak, and stabilization. The framework is trained and evaluated on a multivariate dataset encompassing daily COVID-19 cases, hospitalizations, public health interventions, and mobility records across 48 U.S. states from August 2020 to October 2022. Experimental results demonstrate that GCN-PEFormer consistently outperforms existing baselines, with up to 34.5 % and 28.9 % reductions in RMSE and MAE, respectively, for 60-day forecasts. The model’s forecasts also exhibit strong alignment with national-level patterns (R2 = 0.966), validating its robustness. Beyond predictive accuracy, the framework offers valuable support for regional public health decision-making by enabling targeted interventions and resource allocation. This work contributes a scalable and domain-informed approach to epidemic forecasting, with practical implications for regional health preparedness and response.