Robust two-stage influenza prediction model considering regular and irregular trends

Area of Science:

• Epidemiology
• Computational epidemiology
• Public health informatics

Background:

• Influenza causes significant global mortality annually, necessitating accurate prediction for medical preparedness.
• Traditional flu prediction models rely on historical data, which struggle with irregular outbreaks.
• User-generated content (UGC) offers advantages for detecting irregular phenomena but lacks historical trend analysis.

Purpose of the Study:

• To develop a novel flu prediction model integrating both historical and UGC data.
• To address the limitations of existing models in capturing both regular and irregular influenza trends.
• To create a robust and adaptable model for seasonal flu prediction.

Main Methods:

• A two-stage model was proposed, first estimating regular trends using historical data.
• Irregular trends were then predicted using a separate model based on UGC data (e.g., search queries, tweets).
• Models were trained separately for practical utility and robustness against changes in UGC providers.

Main Results:

• Experiments on US and Japan datasets demonstrated the feasibility of the proposed two-stage approach.
• The model showed robustness against outliers in dropout tests, simulating changes in UGC services.
• The combined approach effectively leverages the strengths of both historical and UGC data for flu prediction.

Conclusions:

• The proposed two-stage model offers an effective solution for predicting seasonal influenza.
• This novel approach enhances prediction accuracy by integrating diverse data sources.
• The model's modular design ensures adaptability and stability in dynamic data environments.