Google dot org blog - News from Googles Philanthropic Arm

Just published: Assessing Google Flu Trends performance

We’re pleased to announce that “Assessing Google Flu Trends Performance in the United States during the 2009 Influenza Virus A (H1N1) Pandemic” is now published in PLoS One. If you’ve wondered what changed between our pre and post swine flu Google Flu Trends models in the US, or if you’re just a flu data nerd, then you’ll definitely want to read it.

Each year we review our Google Flu Trends models and may update them, if an update would make them more accurate when compared with the official influenza data we benchmark against. In this paper, we, together with a member of the US CDC influenza division, provide details of how the original pre-H1N1 Google Flu Trends model for the US performed in comparison with the updated Google Flu Trends model when compared against the CDC’s ILINet (Influenza-Like-Illness) data.

Since the original Google Flu Trends model for the US was built using only seasonal influenza data (pandemic flu data didn’t exist in the five years prior to 2009), we did not know how Google Flu Trends would perform during an outbreak of pandemic flu. If the symptoms and complications of pandemic flu were similar to seasonal flu, we expected Google Flu Trends would be able to detect it.


Time series plots of ILINet data and original and updated Google Flu Trends estimates
A. ILINet data and Google Flu Trends estimates from 2009
B. ILINet data and Google Flu Trends estimates for the entire time period where GFT estimates are available: 2003-2009.
(Click on image to enlarge)

We found that while both the original and updated Google Flu Trends models performed well prior to H1N1, the updated model performed better during H1N1--particularly during the first wave of H1N1. Though generally getting the trend correct, the original model underestimated the magnitude of ILI activity during the H1N1. Why was this the case? In short, search behavior changed during H1N1. This was especially true for the categories “influenza complications” and “term for influenza." This is not unexpected, since differing complications and the fact that H1N1 began spreading during the northern hemisphere summer rather than winter likely played a role. It is interesting to see how the categories and examples of specific queries changed, though. For some specific examples, read the paper.

We’re soon headed into flu season for the half of the world in the Northern hemisphere, so let’s stay healthy!

Permalink | Links to this post |

The comments you read here belong only to the person who posted them. We do, however, reserve the right to remove off-topic comments.