COVID-19 prediction with doubly multi-task Gaussian Process.

Area of Science:

• Computational epidemiology
• Machine learning for public health
• Time-series analysis

Background:

• Accurate prediction of COVID-19 cases and deaths is crucial for public health response.
• Existing models often struggle with the complex, correlated nature of multi-country, multi-task time-series data.
• Understanding inter-country dynamics is vital for effective pandemic management.

Purpose of the Study:

• To propose a novel Doubly Multi-Task Gaussian Process (DMTGP) model for simultaneous prediction of COVID-19 confirmed cases and deaths.
• To incorporate task-wise correlations, leveraging both individual (task-specific) and shared (cross-task) information.
• To model and analyze dynamic relationships between multiple countries using attention mechanisms.

Main Methods:

• Development of the Doubly Multi-Task Gaussian Process (DMTGP) model.
• Utilizing a Transformer encoder layer for cross-attention to model inter-country interactions.
• Construction of a database for Japan, South Korea, and Taiwan, focusing on confirmed cases and deaths.
• Qualitative analysis of attention score maps to interpret model behavior.

Main Results:

• The DMTGP model demonstrated superior performance compared to baseline models in handling doubly multiple tasks.
• The model successfully predicted the number of confirmed cases and deaths across the selected East Asian countries.
• Attention score analysis confirmed the model's ability to capture dynamic, time-varying relationships between countries.

Conclusions:

• The proposed DMTGP model is effective for multi-task, time-series prediction problems with correlated data.
• Incorporating cross-task correlations and attention mechanisms enhances prediction accuracy in epidemiological modeling.
• The framework provides a robust approach for understanding and forecasting disease spread across different regions.