Improving Google Flu Trends estimates for the United States through transformation

Area of Science:

• Epidemiology
• Computational epidemiology
• Public Health Surveillance

Background:

• Google Flu Trends (GFT) utilizes internet search queries for early influenza-like illness (ILI) detection.
• GFT estimates for the US Centers for Disease Control and Prevention's (CDC) ILINet data showed significant overestimation and delayed peak detection during the 2012-13 season.

Purpose of the Study:

• To investigate the relationship between GFT estimates (%GFT) and CDC's ILINet data (%ILINet) from 2010-14.
• To develop a transformation method for %GFT to improve its correlation with %ILINet.

Main Methods:

• Analysis of relative changes between %GFT and %ILINet data from 2010-14.
• Development and application of a transformation equation to adjust %GFT estimates.
• Comparison of transformed %GFT accuracy against original %GFT and %ILINet values.

Main Results:

• Transformed %GFT estimates were within ±10% of %ILINet for 17/29 weeks above baseline (2010-13), compared to only 2/29 weeks for original %GFT.
• The 2012-13 peak estimate using transformed %GFT was 2% lower and one week later than %ILINet, versus 74% higher and three weeks later for original %GFT.
• The transformation method also improved estimates for the recalibrated 2013 GFT model in early 2013-14.

Conclusions:

• A novel transformation method significantly enhances the accuracy and timeliness of Google Flu Trends estimates for influenza surveillance.
• Transformed GFT data can be available approximately one week earlier than official CDC ILINet reports.
• The stable transformation equation developed over 2010-13 facilitates improved utilization of GFT for public health decision-making.