PAN-cODE: COVID-19 forecasting using conditional latent ODEs.

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• The COVID-19 pandemic highlighted the critical need for accurate disease spread modeling.
• Data-driven forecasting is essential for informing policy decisions on non-pharmaceutical interventions (NPIs).

Purpose of the Study:

• To introduce PAN-cODE, a novel deep learning method for forecasting pandemic infections and deaths.
• To enable the generation of alternative caseload trajectories based on varying NPI adoption scenarios.
• To allow caseload estimation for previously unseen geographical regions.

Main Methods:

• Development of PAN-cODE, a deep conditional latent variable model.
• Utilizing a deep learning approach for pandemic caseload forecasting.
• Automated model training with potentially less detailed data.

Main Results:

• PAN-cODE demonstrates performance comparable to state-of-the-art methods for 4- and 6-week-ahead forecasting.
• The model successfully generates realistic alternative outcome trajectories for selected US regions.
• Capability to estimate caseloads in regions not included in the training data.

Conclusions:

• PAN-cODE offers a robust and adaptable tool for pandemic forecasting and policy simulation.
• The model's ability to handle unseen regions and NPI scenarios enhances its practical utility.
• This data-driven approach supports informed decision-making in managing public health crises.