Can the CDC use Twitter, AI and Google to forecast flu outbreaks?

It’s that time of year again, the beginning of flu season. The CDC and prevention are key components in helping hospitals and practitioners around the country prepare to combat the season, to provide the best care at the lowest cost.

Epidemiological forecasting, much like weather forecasting, may be able to bring together volumes of data – from retail sales of flu medications, to Google searches about flu symptoms, to tweets about symptoms or how someone might be feeling – to create a picture in near real-time that predicts the spread of the flu virus. If it’s successful, predicting the spread of disease will be as commonplace as predicting rain or snow.

For the past 4 years, the CDC has run a forecasting research initiative with the goals of building models and methods to predict what a particular flu season might bring. This year 28 organizations/participants submitted their own forecasting system and were judged based on their ability to accurately predict:

• When the season will start
  
• When it will peak
  
• How severe will the peak be
  
• How severe will it be in one, two, three and four weeks’ time
  

During the flu season participants submit a new forecast for each of the items above using new data that has been collected. At the end of the flu season the forecasts are compared with actual data. The results – 2 forecasts developed by Carnegie Mellon University’s Delphi research group claimed the number 1 and 2 spots. The Epicast system was able to predict with a level of accuracy at .451 and the Stat project’s accuracy at .0438. While their scores may seem low compared to 1.0, which would be 100% accurate, taking a simple average of previous data would have only netted an accuracy level of .237 and if you combined the accuracy of all participants into an ensemble forecast the accuracy was at .430 – a bit below the Delphi Stat and well below the Epicast system.

Is epidemiological forecasting the next big thing – read the full article.