Forecasting influenza outbreak dynamics in Melbourne from Internet search query surveillance data

Area of Science:

• Epidemiology
• Computational modeling
• Public health surveillance

Background:

• Seasonal influenza epidemics pose challenges for healthcare due to yearly variations in timing, size, and duration.
• Bayesian estimation techniques show promise for predicting influenza peaks using surveillance data.
• Existing methods require tailoring to specific populations and surveillance systems.

Purpose of the Study:

• To assess the accuracy of influenza epidemic forecasting in metropolitan Melbourne, Australia.
• To evaluate if similar predictive accuracy to previous studies can be achieved.
• To adapt and apply advanced forecasting methods to a new geographical context.

Main Methods:

• Utilized a bootstrap particle filter combined with a mechanistic infection model.
• Employed weekly Internet search query data from Google Flu Trends (2006-2014).
• Developed a novel method for selecting optimal observation models, treating forecasts like receiver operating characteristic (ROC) curves.

Main Results:

• Accurate prediction of influenza epidemic peak timing 4-6 weeks in advance was achieved.
• Forecasting the magnitude of the epidemic peak and overall disease burden proved more challenging.
• The chosen methods demonstrated the feasibility of advanced forecasting in metropolitan Melbourne.

Conclusions:

• Influenza peak timing is predictable using internet search data, aiding public health preparedness.
• Forecasting epidemic magnitude requires further refinement of infection and observation models.
• Improved forecast robustness can enhance decision-making for healthcare providers during influenza seasons.