Random Forest of epidemiological models for Influenza forecasting

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• Accurate forecasting of influenza hospitalizations is essential for effective public health preparedness and resource allocation.
• Existing forecasting methods are submitted to the CDC for real-time public health communication during influenza seasons.

Purpose of the Study:

• To enhance influenza hospitalization forecasting by developing a novel machine learning approach.
• To combine predictions from multiple mechanistic models into an improved, automated forecast.

Main Methods:

• A Tree Ensemble model was designed to leverage individual predictors from a baseline model (SIkJalpha).
• Each predictor was generated by varying a set of hyperparameters, creating diverse potential trajectories.
• The approach was fully automated, requiring no manual tuning, and was tested in the FluSight challenge.

Main Results:

• Submissions consistently ranked in the top 33% of models across multiple influenza seasons (2022-2024).
• The Random Forest-based approach improved forecast accuracy, coverage, and weighted interval score compared to individual predictors.
• Retrospective analysis showed superior performance in mean absolute error and weighted interval score for the 2021-22 season.

Conclusions:

• The proposed machine learning method effectively combines mechanistic model outputs for improved influenza hospitalization forecasting.
• Automated, ensemble-based forecasting offers a robust and reliable strategy for public health planning.
• This approach demonstrates the potential of machine learning to enhance real-time disease surveillance and response.