Abstract
Global escalation of infectious disease outbreak risks necessitates advanced predictive models. Despite methodological advances, errors in initial states and parameters of epidemiological dynamic models remain a key limitation to prediction reliability. To address this limitation, we propose an optimized data assimilation framework for combined state-parameter optimization based on Ensemble Kalman Filter. We design space transformations and adaptive covariance inflation driven by epidemic development and prediction errors, achieving a more stable update process and rapid response to epidemic changes. Through synthetic experiments and real-world case studies, the proposed scheme significantly reduces initial state and parameter errors, leading to a substantial improvement in prediction accuracy during the early stages of an epidemic. Compared with predictions without data assimilation, the average prediction error rate decreased by more than 50 % for 1-day predictions and by approximately 15 % for 7-day predictions. The prediction accuracy rate for the peak day of the epidemic and the peak number of infected cases reached more than 70 % in advance by 3 days. Critically, simple dynamical model integrated with our data assimilation framework outperform complex models without data assimilation. This study establishes data assimilation as an essential tool for epidemic forecasting and provides an extensible framework adaptable to multiple infectious diseases, offering critical support for public health decision making.