Expanding optimization ensemble model methods for forecasting seasonal influenza in the U.S

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Seasonal influenza epidemics exhibit significant year-to-year variability, necessitating accurate forecasting for effective public health response.
• The Centers for Disease Control and Prevention (CDC) FluSight competition gathers influenza forecasts to inform public health strategies.
• Existing ensemble methods, like the CDC's, provide a baseline for influenza prediction accuracy.

Purpose of the Study:

• To develop and evaluate novel weight-based ensemble forecasting methods for predicting laboratory-confirmed influenza hospital admissions.
• To compare the performance of the proposed methods against the established CDC ensemble forecast.
• To assess the impact of different forecasting horizons and epidemic phases on method performance.

Main Methods:

• Introduction of Expanding Window Optimization (EWO), a weight-based ensemble method with week-by-week updated optimal weights to minimize mean squared error (MSE).
• Development of Adjusted Weighted EWO (Adw-EWO), which enhances EWO with a correction term based on horizon-0 forecast errors.
• Utilizing an expanding time window approach for generating forecasts throughout the influenza season.

Main Results:

• EWO demonstrated superior average performance over the CDC ensemble, achieving lower mean absolute error (MAE) and weighted interval score (WIS) across all forecast horizons.
• Adw-EWO further improved upon EWO, particularly at horizon 0, by incorporating a correction term.
• Phase-based analysis indicated Adw-EWO's optimal performance during epidemic growth and peak phases, while EWO excelled in early growth, and the CDC ensemble was more effective during the decay phase.

Conclusions:

• The proposed EWO and Adw-EWO methods offer improved accuracy for influenza hospitalization forecasting compared to the current CDC ensemble.
• Adw-EWO provides enhanced predictive capabilities, especially for short-term forecasts (horizon 0) and during epidemic growth and peak periods.
• These advanced forecasting techniques can significantly aid public health preparedness and resource allocation for seasonal influenza.