How Weather Models and Google Could Help Forecast Flu Season

Last month, despite the tragic consequences of Hurricane Sandy, one thing became apparent—the powerful weather models now available have become better and better at helping forecasters predict where storms like Sandy are going next.

That technology is more useful than just storm prediction. In a study published yesterday in the Proceedings of the National Academy of Sciences, a pair of researchers have harnessed this tech to predict the spread of influenza. With real-time data from Google Flu Trends, their models can forecast where, when and how severely seasonal flu outbreaks will occur across the country.

“ findings indicate that real-time skillful predictions of peak timing can be made more than seven weeks in advance of the actual peak,” writes Jeffrey Shaman, an environmental scientist from Columbia University, and Alicia Karspeck of the National Center for Atmospheric Research, in their paper. “This work represents an initial step in the development of a statistically rigorous system for real-time forecast of seasonal influenza.” If such hopes come to fruition, there could be something like an advance flu warning system (“flu rates are projected to peak in your area next week”) similar to those for hurricanes and other severe weather events.

Both weather and flu transmission are examples of non-linear systems: ones in which a small change in starting conditions can bring about an enormous change in outcomes. In building weather models, scientists look at historical data about how these sorts of small changes (slightly warmer water in the Caribbean, say) have affected outcomes (a hurricane with much more strength when it makes landfall on the East Coast). By assimilating years of data and running countless simulations, they can generate a reasonably accurate prediction for the odds of hypothetical weather events occurring within a period of about a week.

In the new study, the researchers used principles derived from these models and applied them to the spread of the flu. For inputs, in addition to atmospheric measurements of temperature, pressure and wind, they used Google Flu Trends, a service that provides real-time data on flu transmission around the world by closely examining search terms entered into Google. While not every person searching for “flu” necessarily has influenza, Google researchers have shown that flu-related search terms can be an accurate proxy for flu transmission rates around the globe—if many people in a particular area are suddenly googling for “flu,” it’s a good bet that the infection has arrived en masse.

Influenza seems to behave according to probabilistic principles involving atmospheric conditions similar to the weather. Other factors to consider include an area’s population density. In combining factors like humidity and temperature with data from Google and actual flu rate information kept by hospitals, the researchers were able to develop models that approximate how flu has been transmitted in the years since officials have been keeping track.

To test their model, the researchers assessed New York City flu data from 2003 to 2008. By entering data on flu transmission up to a given time and asking the model to provide a weekly forecast for how the flu would behave, they were able to produce accurate forecasts of when the infection would peak, sometimes up to seven weeks ahead of time. Additionally, as with weather models, the system can distinguish between several different scenarios and provide estimates of how likely each one is to occur.

With continued development and real-time data like Google Flu Trends available, this type of technology could theoretically be put to use to generate a flu forecast for local areas, even down to the state or city level.

Joseph Stromberg | | READ MORE

Joseph Stromberg was previously a digital reporter for Smithsonian.