Frequency-aligned loss and spectral filtering improve long-range influenza forecasting

Area of Science:

• Epidemiology
• Time Series Analysis
• Machine Learning

Background:

• Long-horizon seasonal influenza forecasting is challenged by rapid error growth.
• Forecasting is complicated by the interplay of seasonal cycles and outbreaks.
• Existing models often overlook autocorrelation in incidence data, impacting accuracy.

Purpose of the Study:

• To develop an advanced forecasting model for improved long-range seasonal influenza prediction.
• To address limitations in current forecasting methods, specifically error growth and autocorrelation.
• To enhance the interpretability and stability of influenza forecasts.

Main Methods:

• Introduced a frequency-aware pipeline combining a Spectral Adaptive Filtering Network and a Frequency-Aligned Direct Loss.
• The model isolates spectral bands and uses window-specific filters for transient events.
• Employed a convex loss function for simultaneous time and frequency domain supervision to reduce bias.

Main Results:

• Achieved 6-15% lower Mean Squared Error (MSE) and 2-20% lower Mean Absolute Error (MAE) at 24-week horizons across various US regions.
• Demonstrated interpretable band-pass responses aligning with known epidemiological periodicities.
• Confirmed the necessity of joint time-frequency supervision and dual filtering for optimal performance.

Conclusions:

• Explicit spectral decomposition and autocorrelation-aware training provide a robust method for stable, interpretable long-range influenza forecasting.
• The proposed modular objective can be integrated into other architectures for similar error reductions.
• This approach offers a principled way to improve the accuracy and reliability of influenza predictions.