Can You Forecast the Flu Season?

The America COMPETES Act authorizes US government agencies to host challenges to encourage innovation. Recently, the CDC used this authority to host the “Predict the Influenza Season Challenge.”

Infectious disease forecasting is an emerging area for flu season public health effort planning and response. Forecasting uses traditional surveillance mechanisms and predictive mathematical heuristics (commonsense rules) to predict the most likely influenza season course.

The CDC’s flu forecasting challenge looked at novel ways to determine the accuracy, capacity, and usefulness of forecasting techniques. New methods incorporating Twitter data (who’s tweeting about the flu), search engine queries (who’s Googling the flu), and Internet-based surveys seem to increase forecasting accuracy. A team of CDC researchers described the challenge’s results in an article published online in BMC Infectious Diseases.

The CDC challenged participants to predict the start, peak, and intensity of the 2013—2014 influenza season nationally and regionally. All participants submitted at least 9 biweekly reports at the national level. The CDC graded forecast accuracy based on the US Outpatient Influenza-like Illness Surveillance Network. Sixteen individuals or teams registered, and 9 teams produced forecasts.

Infectious disease forecasting using new methods was accurate in the short-term, but insufficient in its capacities for present use. Only 4 forecasts correctly predicted the start, peak, and intensity initially. However, the accuracy and precision of the forecasts improved as the season progressed.

The challenge stimulated interest in forecasting methodology development and cooperation among forecasters, subject matter experts, and public health decision-makers. Inaccuracy and imprecision was driven by unreliable data (eg, Google Flu Trends consistently overestimated influenza incidence) and the early state of study methodology development.

The CDC’s inaugural infectious disease forecasting competition shows that available forecasting models can predict the start, peak, and intensity of the 2013-2014 influenza season. The forecast developers have further developed their methodologies and will repeat the predictive process in the 2014-2015 and 2015-2016 seasons. In the meantime, this contest has brought the CDC and model developers into a closer relationship, shared forecasting methods, and standardized data.