Time series-based forecasting of infectious disease outbreak using information systems in public health

The escalating frequency of infectious disease outbreaks underscores the urgent need for reliable forecasting systems to support timely public health interventions. Existing approaches-ranging from statistical heuristics to black-box deep learning models-often lack domain awareness, adaptability, and interpretability, limiting their utility in dynamic outbreak scenarios.

This study proposes EpiCastNet, a forecasting framework that integrates spatiotemporal attention mechanisms with a hybrid encoding architecture to jointly model empirical patterns and rule-based constraints. A key component is the Causal Regularization with Semantic Anchors (CRSA) module, which incorporates epidemiological principles, such as intervention efficacy and seasonal transmission dynamics, into the model's differentiable training process. This enhances both semantic alignment and robustness under real-world uncertainties.

Empirical evaluations on public health time-series datasets, including COVIDcast and JHU COVID-19, demonstrate that EpiCastNet consistently outperforms state-of-the-art methods in terms of RMSE, MAE, R², and MAPE, while maintaining high stability under noisy and incomplete data conditions. These findings highlight the framework's effectiveness and interpretability in epidemic forecasting, offering a practical tool for data-driven decision-making in public health surveillance.