Dynamic feature selection improves influenza forecasting accuracy and generalization across countries.

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Influenza forecasting faces challenges due to varying epidemic patterns and influencing factors.
• Existing methods struggle with predictive accuracy and cross-regional generalizability.
• Novel approaches are needed to enhance influenza prediction capabilities.

Purpose of the Study:

• To develop AdaFluDR, an adaptive feature selection model for influenza forecasting.
• To address the time-varying nature of influenza transmission drivers.
• To improve the accuracy and generalizability of influenza predictions across different regions and time periods.

Main Methods:

• AdaFluDR integrates SpaceTime and Crossformer models with a correlation-driven mechanism.
• A comprehensive score integrating temporal, frequency, and time domain information dynamically adjusts feature pathways.
• A multilayer perceptron (MLP) models nonlinear relationships for prediction.

Main Results:

• AdaFluDR demonstrated superior predictive performance compared to traditional and other machine learning methods.
• The model showed robust performance across multiple forecasting horizons (1-4 weeks).
• AdaFluDR exhibited strong generalization ability across the United States, Canada, and Portugal.

Conclusions:

• AdaFluDR offers a novel and practical framework for accurate influenza forecasting.
• The model possesses cross-national applicability, enhancing global epidemic preparedness.
• This framework provides a valuable tool for improving public health response strategies.