Researchers at Columbia University’s Mailman School of Public Health have devised a flu predictor that could pinpoint peak cases as much as nine weeks early.
That could help to contain the yearly flood of influenza cases, which stretch from December to April and can surge unpredictably. Knowing when specific cities or regions might be hit especially hard could help people to be more vigilant about washing their hands, as well as allow doctors and public health officials to stockpile vaccines and other flu remedies.
The model, which is described in the journal Nature Communications, incorporates flu-related search data from Google Flu Trends and flu cases from the Centers for Disease Control. The Google data relies on search terms as indicators of flu activity, but the forecaster improves on the data’s predictive capabilities. While more cases of flu may certainly lead to more searches for influenza-related issues, these search queries can be thrown off by media coverage of the flu, which aren’t always connected to the intensity of infections but could simply reflect, for example, the start of flu season. The model also borrows strategies used in weather prediction, which make it one of the most accurate predictors of peak flu cases so far. The weather techniques include taking advantage of physical characteristics — in the case of influenza, things such as the dynamics of how influenza particles travel through the air and transfer from person to person — as well as historical information on previous infection trends.
“We have a model of influenza that describes the propagation of the influenza virus through a population,” says the study’s lead author Jeffrey Shaman, an assistant professor of Environmental Health Sciences. “The idea is that you take the real-time observations from the last 10 weeks up to the present and you use those to inform the model. You train the model, and by doing that, the model doesn’t fly off and make some predication or represent reality in a way that is not correct,” says Shaman. Every week, the tool incorporates the actual number of cases of flu reported and compares it to its own forecast; if it’s predicating too many or too few cases, the model adjusts.
When the scientists applied their predictor to the 2012-2013 flu season, it accurately identified peak cases in the southeast in December, while forecasting that most of the rest of the country would see surges in January.
The Columbia model takes such potential bias into account, and has so far outperformed existing predictors. And if you’re curious about whether you’re surrounded by flu, the predictor tool will be available to the public on Columbia’s Mailman School of Public Health website in the a few weeks.