STAN: spatio-temporal attention network for pandemic prediction using real-world evidence

Area of Science:

• Epidemiology
• Computational Biology
• Public Health Informatics

Background:

• Accurate prediction of infectious disease spread is crucial for effective pandemic response and resource allocation.
• Traditional epidemiological models often struggle to incorporate real-world, localized data for timely predictions.
• The need for advanced computational models that integrate diverse data sources for enhanced pandemic forecasting is evident.

Purpose of the Study:

• To develop a hybrid model for earlier and more accurate prediction of pandemic-infected cases.
• To leverage patient claims data, demographic similarities, geographical proximity, and transmission dynamics.
• To improve the prediction of local disease status and medical resource utilization during pandemics.

Main Methods:

• Proposed a spatio-temporal attention network (STAN) model for pandemic prediction.
• Utilized a graph attention network to capture spatio-temporal disease dynamics.
• Integrated a dynamics-based loss term to enhance long-term prediction accuracy.
• Tested STAN using real-world patient claims data and COVID-19 statistics across US counties.

Main Results:

• STAN demonstrated superior performance compared to traditional epidemiological models (SIR, SEIR) and existing deep learning models.
• Achieved up to an 87% reduction in mean squared error for both long-term and short-term predictions.
• Validated the model's effectiveness using diverse datasets including patient claims and public health statistics.

Conclusions:

• The hybrid STAN model effectively combines claims data and case counts for improved disease status prediction.
• STAN offers enhanced accuracy in predicting medical resource utilization during public health crises.
• This approach provides a more robust tool for pandemic forecasting and management.