The impressive performance of state of the art weather forecasting systems led to strikingly accurate forecasts for Superstorm Sandy at long lead times. A new study finds the same technology that has led to such weather forecasting feats may help us determine when we’re most likely to catch the flu, which kills 35,000 Americans each year.
Thanks to a collaboration between scientists at Columbia University’s Mailman School of Public Health and the National Center for Atmospheric Research, a model for predicting flu outbreaks has been developed that operates like a modern day weather modeling system.
It processes gobs of data and conveys not only the most likely forecast but also the range of possibilities to provide an idea of the uncertainty.
“Analogous to weather prediction, this system can potentially be used to estimate the probability of regional outbreaks of the flu several weeks in advance,” said Alicia Karspeck, NCAR scientist and study co-author.
Karspeck and Columbia University colleague Jeffrey Shaman ran their model for the 2003-2008 flu seasons in New York City to determine how well it could predict the timing of outbreaks. The results were published in the Proceedings of the National Academy of Sciences.
“The findings indicate that real-time skillful predictions of peak timing [of a flu outbreak] can be made more than 7 weeks in advance of the actual peak,” the study said.
The modeling system Shaman and Karspeck developed optimizes itself by continually bringing into the model near-real-time data from Google Flu Trends, a tracker of flu activity.
The data are critical, because they make up for incomplete understanding about the relationship between weather and flu and non-weather factors.
“It has been difficult to develop an influenza/weather approach that has any degree of accuracy because of the many intervening variables that are not related to meteorology,” said Larry Kalkstein, a weather and health researcher at the University of Miami, not involved in the study.
Without the Google data, Shaman says, small errors in the model’s calculation of flu transmission - which must make simplistic assumptions - would amplify over time.
“A flu model left to its own devices will fly off and have errors,” Shaman said. “By having the observations we can recursively nudge the model towards those observations.”
Shaman stressed “entraining” the data into the model not only keeps its forecasts on track, but also makes the model work better.
“[By assimilating the data] we optimize the model so its internal dynamics are primed better to evolve in a manner that’s consistent with the dynamics of the evolving [flu] epidemic,” Shaman said.
Much in the same way, weather modeling systems have been improved by bringing in more data from satellites, weather balloons, aircraft, and ground observations to help overcome the complexity of weather processes and their chaotic behavior.
While study authors caution their study represents merely a “first step” in influenza forecasting, they see tremendous potential for its application.
“The forecasts developed here indicate that we will soon reach an era when reliable forecasts of some infectious pathogens are as commonplace as weather predictions,” the study concludes.