Flusion: Integrating multiple data sources for accurate influenza predictions.

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• The Centers for Disease Control and Prevention (CDC) conducts an annual influenza forecasting challenge to enhance public health preparedness.
• Recent shifts in forecasting targets to National Healthcare Safety Network (NHSN) hospital admissions present a challenge due to limited historical data.

Purpose of the Study:

• To develop and evaluate an influenza forecasting model, Flusion, capable of producing accurate predictions with limited target surveillance data.
• To identify key factors contributing to the success of the Flusion model in the CDC's influenza prediction challenge.

Main Methods:

• Flusion, an ensemble model, combines gradient boosting and Bayesian autoregressive models.
• Data augmentation strategies included using historical Influenza-Like Illness (ILI+) proportions and laboratory-confirmed hospitalization rates.
• Models were trained jointly on data from multiple surveillance signals and locations to leverage shared information.

Main Results:

• Flusion was the top-performing model in the 2023/24 CDC influenza prediction challenge.
• The model's success was attributed to a gradient boosting component trained on combined data from multiple surveillance signals and locations.
• Joint training across signals and locations significantly improved predictive performance.

Conclusions:

• Integrating diverse data streams and locations enhances influenza forecasting accuracy, particularly when target data is scarce.
• Flusion's performance highlights the value of information sharing across surveillance systems for improved public health situational awareness.
• The study demonstrates the effectiveness of machine learning approaches in addressing public health surveillance challenges.