Pre-training Epidemic Time Series Forecasters with Compartmental Prototypes

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• Accurate epidemic forecasting is vital for public health preparedness but challenged by data scarcity and distribution shifts in current models.
• Existing models often fail when encountering new pathogens or evolving viral dynamics due to limited training data and lack of generalizability.
• Decades of disease surveillance data and compartmental modeling principles offer a rich, yet underutilized, source of transferable knowledge for improving forecasting.

Purpose of the Study:

• To introduce CAPE (Compartmental Attention Prototype Embeddings), the first open-source pre-trained model for epidemic forecasting.
• To develop a model that leverages transferable knowledge from diverse historical disease data to overcome limitations of current forecasting methods.
• To enable robust and accurate epidemic forecasting, even in scenarios with limited data or significant epidemiological shifts.

Main Methods:

• CAPE models epidemic dynamics using mixtures of latent compartmental population states, termed 'compartmental prototypes'.
• A flexible dictionary of compartment prototypes is learned from extensive simulation data, linking observed infections to latent population states.
• The model employs a next-token-prediction paradigm with epidemic-aware regularization for pre-training, ensuring epidemiological relevance and robust generalization.

Main Results:

• CAPE demonstrates significant zero-shot forecasting performance across a benchmark of 17 diverse diseases.
• The model substantially outperforms existing strong baseline methods in epidemic forecasting tasks.
• The learned compartmental prototypes effectively capture and represent epidemiological dynamics from varied disease contexts.

Conclusions:

• CAPE represents a significant advancement in epidemic forecasting, offering a transferable and epidemiologically grounded approach.
• The pre-trained nature of CAPE allows for rapid and accurate forecasting of new outbreaks with minimal specific training data.
• This work paves the way for more robust and reliable epidemic preparedness through advanced AI and epidemiological modeling.