Print 10 comment(s) - last by ven1ger.. on Feb 19 at 5:49 PM

Typically accurate algorithm may need tweaking to deal with new trends

Launched in 2008Google Flu Trends is one of Google Inc.'s (GOOG) most intriguing internet services.  Leveraging Google's mastery of data mining and industry-leading search engine position, the system injects data on health-related searches into a complex model which spits out an estimation of the number of people in each region that are infected with the flu virus.  

I. Flu Trends' Bad Forecast

The system performs extraordinarily well, typically matching up closely with data from the Centers for Disease Control (CDC).  The system typically is several days ahead of CDC predictions in 29 countries and today also tracks a second disease -- dengue.

But this flu season was a rough one for the Google service.  Google badly over predicted the Christmas peak of flu season, estimating that over 10 percent of the U.S. population had flu, versus the CDC's data, which showed only around 6 percent had it.

Google Flu Trends miss
Google Flu overpredicted, while Flu Near You underpredicted. [Image Source: Nature]

So what happened?

Experts believe part of the problem was public fear influencing search behavior in an unusual way.  The flu season got off to its earliest seasonal start in nearly a decade, starting in November and peaking in December.  Of the three major flu virus strains, the most virulent strain -- H3N2n -- dominated, as in the severe 2003 outbreak.

There were multiple deaths, particularly among the elderly from the outbreak.  As a result, the mass media latched onto the story and focus a great deal of coverage on the flu -- more so than usual.  Experts say that may have led people to increase searches for flu terms, even if they didn't have flu.

It was also a bad season for norovirus outbreaks, which may have led to people searching for flu terms.  Norovirus is a separate pathogen that manifests itself via intestinal symptoms.  However, many in the public mistake norovirus infections for flu.

II. Self-Tracking Efforts Also Miss

Professor John Brownstein, a Harvard Medical School epidemiologist, told Nature Magazine that the miss by Google shows the difficulty in adapting algorithms to new variables.  He comments, "You need to be constantly adapting these models, they don’t work in a vacuum.  You need to recalibrate them every year."

Professor Brownstein helps maintain "Flu Near You" a Boston Children's Hospital/HealthMap project which tracks flu in the U.S. via volunteer self-reporting.  The 2011 program tracks a sample set of 70,000 people and has 46,000 participants (who report on their own health and health of family members).

Flu Virus
Tracking the flu-virus in real time is a daunting challenge. [Image Source: Navco]

To Google's credit, Flu Near You missed nearly as badly, underestimating the peak infection rate.  It peaked around a little over 4 percent.  It's unclear why that underestimation happened.

Digital tracking of influenza-like illness (ILI) -- symptoms such as fever and sinus issues -- first began in 1985 in France with the creation of the Sentinelles network.  Today France's continues that tradition.  Similar to Flu Near You, GrippeNet uses self-reporting from volunteers and has 5,500 participants.

The CDC's tracking comes directly from a network of 2,700 health institutions, which cover 30 million patients.

III. What About Twitter Tracking?

The misses of Flu Near You and Google Flu Trends open the door to new analysis projects.  A pair of new efforts -- MappyHealth and Sickweather -- look to track Twitter posts to determine flu rates.  Johns Hopkins University Professor Michael Paul argues that Twitter monitoring may offer less noise than search term monitoring, while offering big sample sets than self-monitoring.

He comments, "I suspect that passive monitoring of social media will always yield more data than systems that rely on people to actively respond to surveys, like Flu Near You."

Some are turning to Twitter for flu tracking.

But some are skeptical of Twitter-mining.  Lyn Finelli, head of the CDC’s Influenza Surveillance and Outbreak Response Team, argues that Twitter crowdsourcing is less effect that search-term efforts, because Twitter's userbase tends to be younger, less reliable users.  She comments, "The Twitter analyses have much less promise."

The 2012 miss was the worst since 2009 for Google, which led to some major algorithmic adjustments.  In 2009 Google under predicted the H1N1 swine flu outbreak.  That mistake was the subject of a PLOS One paper entitled "Assessing Google Flu Trends Performance in the United States during the 2009 Influenza Virus A (H1N1) Pandemic".

Source: Nature Magazine

Comments     Threshold

This article is over a month old, voting and posting comments is disabled

By chromal on 2/18/2013 12:13:27 PM , Rating: 1
I always thought that, while interesting, zeitgeist-like data is at best 'soft' data, saying more about what people are thinking about than any actual cold hard reality. Further distorting any meaning of the data would be flu-related searches inspired by the new media flu coverage, probably driven (ironically) in part by the same Google predictions.

It's all sillyness. Who cares what people *think* or *believe*, what is actual and real is what matters. I realize that the flu is tricky to track. Because some folks are asymptomatic or don't seek medical attention, it's a tip-of-the-iceberg sort of situation.

"The whole principle [of censorship] is wrong. It's like demanding that grown men live on skim milk because the baby can't have steak." -- Robert Heinlein
Related Articles
Google Predicts the Flu
November 12, 2008, 10:10 AM

Copyright 2016 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki