An ensemble approach driven by spatial interaction for modelling urban epidemic dynamics

Modelling spatial interactions remains a core issue in geographic information science. Faced with the emerging, non-stationary spread of epidemics, the effects and uncertainties of spatial interactions are amplified. Limited by reliance on a single spatial interaction prior, existing models perform well in specific scenarios but suffer from prediction drift across multi-phase epidemics, thereby hindering real-time and dynamic policy adjustments. To address this limitation, we propose an ensemble approach driven by spatial interaction, built upon a foundational metapopulation model. Using the 2022 Shanghai Omicron outbreak as a case study, we evaluated the proposed ensemble model against 14 mechanistic and machine learning baselines. In multi-phase prediction tasks, the proposed ensemble model achieves the best overall performance with mean ranks of 3.25 for RMSE and 4.00 for R2, surpassing all baselines with mean ranks ranging from 5 to 15. Furthermore, it improves mean predictive accuracy by 29% to 72% compared with the non-ensemble models, while reducing the prediction error interval by 10% to 84%. Finally, we generated model-based projections across multiple counterfactual scenarios to explore responses to changes in intervention timing and intensity.