A semi-mechanistic modeling strategy for infectious diseases forecasting: Error correction and probabilistic prediction

Global climate change and technological advancements have intensified the threats of pandemics, while complex transmission dynamics challenge infectious disease forecasting. Traditional compartmental models struggle to fully capture both the dynamic transmission processes and their associated uncertainties. Here, we develop a novel hybrid methodology that integrates dynamic modeling with statistical approaches, establishing a semi-mechanistic model for error correction and probabilistic prediction. Our error analysis of the dynamic model reveals that frequent population mobility compromises the accuracy of dynamic predictions and that meteorological conditions further modulate forecast performance by regulating human movement patterns. To capture these effects, we implement a quantile regression long short-term memory (QRLSTM) network to estimate prediction errors of the epidemic dynamic model based on mobility and environmental data. This hybrid approach corrects dynamic prediction errors and generates probabilistic forecasts. Validation using multi-state the United States (U.S.) coronavirus disease 2019 (COVID-19) outbreak data shows that our framework reduces dynamic prediction errors by over 50 %. Compared with pure deep learning approaches, the semi-mechanistic model significantly enhances long-term prediction performance and interpretability. By integrating mechanistic modeling with data-driven learning, the proposed approach improves the predictive accuracy and reliability of models in real-world outbreaks, thereby delivering more effective decision support for public health interventions.