Research highlight

Virology: Predicting seasonal influenza in the US

Nature Communications

December 4, 2013

Accurate forecasting of seasonal influenza peaks in the US is reported in Nature Communications. The work illustrates that accurate forecasts of influenza timing, with lead times of up to nine weeks, can be generated with a simple model that uses real-time observations of virus incidence.

While the general wintertime peak of influenza incidence in temperate regions is well described and vaccination programmes are in place, the specific timing, magnitude and duration of individual local outbreaks in any given year are highly variable. If these characteristics were to be reliably forecast, public health response efforts may be better coordinated.

Jeffrey Shaman and colleagues report weekly forecasts of seasonal influenza in real time for 108 cities in the US during the 2012-2013 season. They base their system on a method used previously to generate retrospective forecasts in New York City that incorporates numbers of newly affected people, numbers of susceptible people, regional peaks and total outbreak cases. In doing this, reliable forecasts of influenza outbreak peak timing, with lead times of up to nine weeks in some states, were produced. The authors report that the accuracy of weekly predictions did differ between cities. Some areas, e.g. Birmingham, Kansas City, Buffalo, were accurately forecast throughout the influenza season, both before and after the observed local peak had passed whereas outbreak peaks in other cities, such as Chicago and New Orleans, were never well predicted. Many cities, including San Diego, Atlanta, and Boston showed increasing accuracy of prediction as the season progressed. Overall, forecast accuracy across cities increased from 19% to 74% as the season progressed and by week 52 63% of all ensemble forecasts of peak week of the virus were accurate within 1 week.

The authors report that forecasts significantly outperformed alternate prediction methods including those derived from the resampling of historical outcomes.

doi: 10.1038/ncomms3837

Return to research highlights

PrivacyMark System