Improving probabilistic infectious disease forecasting through coherence

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Influenza causes significant economic and health burdens in the U.S.
• The Centers for Disease Control and Prevention (CDC) forecasts weighted influenza-like illness (wILI) to predict flu spread.
• Current forecasting models often generate independent predictions for different regions, neglecting national-regional data consistency.

Purpose of the Study:

• To develop a novel algorithm for generating probabilistically coherent influenza-like illness forecasts.
• To ensure national wILI forecasts are a consistent weighted sum of regional wILI forecasts.
• To improve the accuracy and reliability of influenza spread predictions.

Main Methods:

• Proposed a new algorithm to transform independent forecast distributions into probabilistically coherent ones.
• Applied the algorithm to existing influenza forecasting models.
• Evaluated the impact of probabilistic coherence on forecast skill across multiple flu seasons.

Main Results:

• The novel algorithm successfully enforced probabilistic coherence in influenza forecasts.
• Forecast skill improved for 79% of tested models after enforcing probabilistic coherence.
• Demonstrated the importance of respecting geographical hierarchies in forecasting systems.

Conclusions:

• Enforcing probabilistic coherence is a valuable method for enhancing influenza forecast accuracy.
• The proposed algorithm offers a way to improve national and regional wILI predictions.
• Respecting the hierarchical structure of geographical data is critical for robust epidemiological forecasting.