Unweighted Percentages of Weekly Outpatient Visits for ILI and Google Flu Trends dataWe retrieved the weekly unweighted percentages of patient visits due to influenza-like illness (ILI), reported through the U.S. Outpatient Influenza-like Illness Surveillance Network (ILINet), from http://www.cdc.gov/flu/weekly/ on 10th December 2013. Here, ILI is defined as fever with a temperature of 100°F or greater, accompanied by a cough or a sore throat. Note that the data recorded for a given week can be updated in subsequent weeks, if the CDC have reason to believe that an updated figure would be more accurate. Here, we focus our analysis on the latest data available on the date of retrieval. We obtained the weekly time series of query volume for searches relating to ILI symptoms from Google Flu Trends (http://www.google.org/flutrends) on 18th December 2013. This time series is restricted to searches made in the United States, and has been shown by Ginsberg et al. to be correlated with the percentage of physician visits in which a patient presents with influenza-like symptoms. The creators of Google Flu Trends state that their algorithm for identifying influenza related searches is constantly evaluated against figures reported by the CDC and is occasionally updated to reflect changes in human online search behaviour. Since publication of the work carried out by Ginsberg et al., the algorithm underwent updates in 2009 and 2013. Data analysed here is therefore an amalgamation of two different Google Flu Trends algorithms, with the transition occurring in August 2013. In both the patient visit and search engine query time series, weeks start on Sundays and end on Saturdays.PreisMoat2014.csv

Seasonal influenza outbreaks and pandemics of new strains of the influenza virus affect humans around the globe. However, traditional systems for measuring the spread of flu infections deliver results with one or two weeks delay. Recent research suggests that data on queries made to the search engine Google can be used to address this problem, providing real-time estimates of levels of influenza-like illness in a population. Others have however argued that equally good estimates of current flu levels can be forecast using historic flu measurements. Here, we build dynamic ‘nowcasting’ models; in other words, forecasting models that estimate current levels of influenza, before the release of official data one week later. We find that when using Google Flu Trends data in combination with historic flu levels, the mean absolute error (MAE) of in-sample ‘nowcasts’ can be significantly reduced by 14.4%, compared with a baseline model that uses historic data on flu levels only. We further demonstrate that the MAE of out-of-sample nowcasts can also be significantly reduced by between 16.0% and 52.7%, depending on the length of the sliding training interval. We conclude that, using adaptive models, Google Flu Trends data can indeed be used to improve real-time influenza monitoring, even when official reports of flu infections are available with only one week's delay.