Beginning in late November of 2012, the researchers used the flu forecasting system to perform weekly estimates for 108 cities. They shared the results with the CDC and posted them online in an academic archive. Near the end of 2012, four weeks into the flu season, the system had predicted 63% of cities accurately. As the season progressed, the accuracy increased. By week four, it successfully predicted the seasonal peak in 70% of the country. It was able to give accurate lead-times up to nine weeks in advance of the peak; most lead-times were two to four weeks.
The flu forecasts were also much more reliable than those made using alternate, approaches that rely on historical data. "Our method greatly outperformed these alternate schemes," says Dr. Shaman.
The researchers saw regional differences in the accuracy of the system, but they were likely within normal variation. "As an example, retrospectively, we've been able to predict the flu in Chicago very well; this year we did a terrible job in that city. For other cities, the opposite held. It averages out. On the whole the system performed very well," Dr. Shaman says. However, there were hints of geographical differences. "We were able make better predictions in smaller cities. Population density may also be important. It suggests that in a city like New York, we may need to predict at a finer granularity, perhaps at the borough level. In a big sprawling city like Los Angeles, we may need to predict influenza at the level of individual neighborhoods."
Google Flu Trends Goes "Off the Rails"
The researchers designed the flu forecasting system to use combined data from 1) Google Flu Trends, which makes estimates of outbreaks based on the number of flu-related search queries, and 2) region-specific reports from the Centers for Disease Control on verified cases of flu. The system app
|Contact: Timothy S. Paul|
Columbia University's Mailman School of Public Health