Monitoring Influenza Activity in the United States: A Comparison of Traditional Surveillance Systems with Google Flu Trends

Abstract

BackgroundGoogle Flu Trends was developed to estimate US influenza-like illness (ILI) rates from internet searches; however ILI does not necessarily correlate with actual influenza virus infections.Methods and FindingsInfluenza activity data from 2003–04 through 2007–08 were obtained from three US surveillance systems: Google Flu Trends, CDC Outpatient ILI Surveillance Network (CDC ILI Surveillance), and US Influenza Virologic Surveillance System (CDC Virus Surveillance). Pearson's correlation coefficients with 95% confidence intervals (95% CI) were calculated to compare surveillance data. An analysis was performed to investigate outlier observations and determine the extent to which they affected the correlations between surveillance data. Pearson's correlation coefficient describing Google Flu Trends and CDC Virus Surveillance over the study period was 0.72 (95% CI: 0.64, 0.79). The correlation between CDC ILI Surveillance and CDC Virus Surveillance over the same period was 0.85 (95% CI: 0.81, 0.89). Most of the outlier observations in both comparisons were from the 2003–04 influenza season. Exclusion of the outlier observations did not substantially improve the correlation between Google Flu Trends and CDC Virus Surveillance (0.82; 95% CI: 0.76, 0.87) or CDC ILI Surveillance and CDC Virus Surveillance (0.86; 95%CI: 0.82, 0.90).ConclusionsThis analysis demonstrates that while Google Flu Trends is highly correlated with rates of ILI, it has a lower correlation with surveillance for laboratory-confirmed influenza. Most of the outlier observations occurred during the 2003–04 influenza season that was characterized by early and intense influenza activity, which potentially altered health care seeking behavior, physician testing practices, and internet search behavior.

The Parable of Google Flu: Traps in Big Data Analysis

“…No, greater value can be obtained by combining GFT with other near-real time health data (2,20). For example, by combining GFT and lagged CDC data, as well as dynamically recalibrating GFT, we can substantially improve on the performance of GFT or the CDC alone (see the chart).…”

“…Because a simple lagged model for flu prevalence will perform so well, there is little room for improvement on the CDC data for model projections [this does not apply to other methods to directly measure flu prevalence, e.g. (20,27,28)]. If you are 90% of the way there, at most you can gain that last 10%.…”

Epidemiological and statistical basis for detection and prediction of influenza epidemics

“…A possible reason for the underperformance of telenursing data during the winter influenza season in 2007-08 may be that this season comprised only a small proportion of young adults (Figure 2), which may have led to comparatively fewer individuals contacting the telenursing call service centers. For GFT, the outcomes are consistent with the findings of previous studies conducted at national, state, and local levels, which also reported strong associations between GFT and influenza-diagnosis cases (Ortiz et al 2011, Malik et al 2011, Dugas et al 2012, Kang et al 2013). Similar to telenursing data, in this study, GFT showed lower correlations with influenza-diagnosis cases for the 2007-08 and 2008-09 winter seasons when comparatively fewer young adults were diagnosed with influenza.…”

“…The latter data source was restricted to a limited period and only used for validation purposes. On the other hand, the effect size of the correlation between the microbiological and clinical diagnosis rates observed was large during the validation period, and corresponded to the findings reported from other settings (Ortiz et al 2011). …”

Web GIS-Based Public Health Surveillance Systems: A Systematic Review

“…In addition to the six WGPHSSs analyzed in this paper, previous studies also demonstrated the widespread use of this informal online information. For example, Twitter has been used to monitor dental pain [50], H1N1 [51], cholera [52], and Enterohemorrhagic Escherichia coli (EHEC)/ hemolytic uremic syndrome (HUS) [53]; Web-user search data has been proven to have the possibility to effectively monitor H1N1 [54] and toxicological outbreaks [55]; and Google Flu Trends has been employed for flu surveillance in various countries worldwide [56][57][58][59][60][61]. Unlike official data, these informal data are much less costly to obtain, more timely (often 1-2 weeks earlier [52,56,62]) thus allowing for early outbreak detection and responses, more frequently posted, and are generated worldwide making large spatial scale PHS a reality [50].…”