HG-DCM: History Guided Deep Compartmental Model for Early Stage Pandemic Forecasting

Ziming Wei, Michael Lingzhi Li

Keywords: Pandemic Forecasting, Time-Series Forecasting, Deep Compartmental Modeling, Explainable AI, Epidemiological Modeling

TL;DR: We introduce HG-DCM, a history-guided deep compartmental model that integrates neural networks with epidemiological modeling to deliver interpretable and more accurate early-stage pandemic forecasts.

Abstract: Early-stage pandemic forecasting is fundamentally constrained by a lack of data. When a new pathogen emerges, there is insufficient historical context to calibrate standard epidemiological models. We introduce the History-Guided Deep Compartmental Model (HG-DCM), a framework designed to overcome this scarcity by systematically transferring knowledge from historical pandemics to the current outbreak. Rather than relying solely on the sparse data of an unfolding crisis, HG-DCM leverages a deep learning backbone to extract universal temporal patterns and parameter dynamics from a comprehensive dataset of past global outbreaks. By integrating these historical insights with epidemiological and demographic metadata, our approach infers robust, interpretable parameters for compartmental forecasting even when current data is minimal. Experimental results on early-stage COVID-19 tasks demonstrate that leveraging historical guidance significantly reduces overfitting and improves stability compared to standard compartmental models and data-isolated deep learning approaches. HG-DCM establishes a new paradigm for pandemic modeling that moves beyond the limitations of single-outbreak data by integrating the collective history of global epidemiology.

Primary Area: learning on time series and dynamical systems

Submission Number: 14697