Neural network models for influenza forecasting with associated uncertainty using Web search activity trends

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Influenza poses a significant global health burden, causing millions of illnesses, hospitalizations, and deaths annually.
• Accurate forecasting of influenza prevalence is crucial for timely public health interventions against seasonal and novel strains.
• Current influenza forecasting models face challenges in achieving high accuracy, necessitating improved methodologies.

Purpose of the Study:

• To develop and validate a novel methodological framework for enhancing the accuracy of influenza-like illness (ILI) rate forecasting in the United States.
• To integrate web search activity time series with historical ILI data for improved predictive modeling.
• To incorporate uncertainty quantification into forecasting models for more reliable estimates.

Main Methods:

• Utilized neural network (NN) architectures, including Bayesian layers, for time series analysis and prediction.
• Employed web search activity data alongside historical ILI rates as input features for NN training.
• Developed and evaluated an iterative recurrent neural network (IRNN) architecture as the best-performing model.

Main Results:

• The proposed framework demonstrated improved state-of-the-art forecasting accuracy for ILI rates.
• The best performing model, IRNN, reduced mean absolute error by 10.3% compared to existing methods.
• The IRNN model improved forecasting skill by an average of 17.1% across four consecutive flu seasons.

Conclusions:

• The proposed methodological framework significantly enhances the accuracy of influenza-like illness forecasting.
• The integration of web search data and advanced NN architectures, like IRNN, offers a powerful tool for public health surveillance.
• The developed models provide reliable forecasts with uncertainty intervals, complementing traditional epidemiological approaches.