By Bio-IT World Staff
December 3, 2013 | This year marks the first real-time use of the flu forecast model developed by researchers at Columbia University's Mailman School of Public Health, which correctly predicted the seasonal influenza peak in 70% of the 108 cities for which forecasts were made. Based on the example of weather prediction, the flu forecast was built in 2012 to retroactively predict influenza trends in New York City using both CDC reports of individual flu cases and Google Flu Trends, which tracks the volume of search terms for flu and flu symptoms in different regions. Starting in November 2012, the system was extended to a much broader geographic scope and allowed to operate on real-time data, testing its predictive power. The model's success rate, and ability to call flu peaks up to nine weeks in advance, significantly outperformed traditional flu predictions that rely purely on historical trends.
If further validated, models like this could serve as early warning systems for cities to take preventive action against severe flu seasons, much like the National Weather Service guards against unanticipated weather disasters. Of course, flu prediction suffers from complicating factors that don't factor in weather prediction; in this first round of forecasts, the model picked up noise from the very attention it gathered, as media reports increased the incidence of flu-related search terms unrelated to actual cases of influenza. If the forecast becomes adopted as a tool for local health administrations, its own predictions could even prove disruptive, as measures like vaccination campaigns or school closings may forestall the same crises that inspired them.
Still, the promising first launch of the flu forecast suggests a future for this brand of big data disease prediction. This flu season, the model's forecasts will be made publicly available online through a forthcoming Columbia University-hosted website. Results from the first year of live forecasts appear today in Nature Communications.