FaXNet: a frequency-adaptive, explainable, and uncertainty-aware network for influenza forecasting.

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• Accurate influenza forecasting is crucial for public health preparedness.
• Existing models struggle with multi-scale temporal dynamics and uncertainty estimation, especially in China's diverse climate regions.
• Regional differences in influenza seasonality between northern and southern China pose forecasting challenges.

Purpose of the Study:

• To develop a deep learning framework (FaXNet) for accurate and interpretable influenza forecasting in China.
• To address challenges in capturing multi-scale temporal dynamics and providing reliable uncertainty estimates.
• To enable region-specific risk assessment and resource planning through actionable lead times.

Main Methods:

• Developed FaXNet, a frequency-adaptive, explainable, and uncertainty-aware deep learning framework.
• Integrated data-driven spectral representation with interpretable component selection and probabilistic forecasting.
• Utilized weekly influenza positivity rates from China and ERA5-Land meteorological data (temperature, dew point, precipitation) from 2011-2023 for northern and southern regions.

Main Results:

• FaXNet demonstrated superior performance in both northern and southern China across 1-4 week forecasting horizons.
• Achieved high R-squared values, e.g., 0.9319 (north) and 0.8665 (south) for 1-week-ahead forecasts.
• Model explanations identified precipitation as a key driver in the north and temperature in the south, validating frequency-adaptive modeling.

Conclusions:

• FaXNet delivers accurate, interpretable influenza forecasts with calibrated prediction intervals.
• The framework offers actionable lead times for region-specific public health planning and resource allocation.
• Future work may incorporate additional drivers like mobility and vaccination data for enhanced forecasting.