Paper

  • Title: Detecting influenza epidemics using search engine query data
  • Authors: Jeremy Ginsberg, Matthew H. Mohebbi, Rajan S. Patel, Lynnette Brammer, Mark S. Smolinski & Larry Brilliant
  • DOI: 10.1038/nature07634
  • Tags:
  • Year: 2009

Summary

  • What

    • Propose a method of analyzing large numbers of Google search queries to track influenza-like illness in a population which can shorten reporting lag for 1-2 weeks.

  • Why analyze Google queries possible

    • Innovative surveillance systems become available, e.g. telephone traige advice lines and over-the-counter drug sales.
    • A large number of American adults (about 90 million) tend to search online for information about specific disease.
    • Data of required queries can be accessed through several websites which collect such information.
    • Correlation between a set of Yahoo search queries containing keywords about influenza-like illness and virological and mortality surveillance data were found.

  • How

    • Obtain data
      • Query fraction
        • Create a time series of weekly counts for 50 million of the most common search queries of in the US from search queries logs between 2003 and 2008.
        • The time series data was separately calculated for nine states.
        • Query fraction was computed to remove the effect of totla numbers of queries.
      • ILI data
        • Publicly available historical data from the CDC’s US Influenza Sentinel Provider Surveillance Network.
        • Only record data of weeks inside annual influenza seasons.

    • Method

      • Linear model
        • $\mathit{logit}(I(t)) =\alpha \mathit{logit}(Q(t)) + \varepsilon$
        • where $I(t)$ is the percentage of ILI physician visits, $Q(t)$ is the ILI-related query fraction at time $t$, $\alpha$ is the multiplicative coefficient, and $\varepsilon$ is the error term. $\mathit{logit}(p)$ is simply $\ln (p/(1 - p))$.
      • Select ILI-relatied queries
        • A list of highest scoring search queries
          • Use one query at a time to measure the efficiency of fitting in each region.
          • Reward query which shows regional variations similar to the regional ILI data.
          • Obtain a list of highest scoring queries, sorted by mean Z-transformed correlation across nine regions.
        • A set of search queries
          • Consider different sets of top $n$ queries in the list. $n=45$ was found to be the best result. -w547
          • The 45 queries appear to be consistently related to ILIs. Other 55 queries tend to coincide with influenza season in the US. -w534

    • Evaluation

      • Correlation between $\mathit{logit}(I(t))$ and CDC-reported weekly ILI percentages.

  • Results

    • For nine region together
      • Train set: A mean correlation of 0.90 (min=0.80, max=0.96, n=9 regions; Fig.2). -w535
      • Validation set: A mean correlation of 0.97 (min= 0.92, max = 0.99, n = 9 regions) across 42 validation points.
      • Estimate consistently the current ILI percentage 1-2 weeks ahead of the publication of reports by the CDC’s US Influenza Sentinel Provider Surveillance.

    • For one state (Utah) only

      • A correlation of 0.90 across 42 validation points. -w537

    • Conclusions

      • Pros
        • Resulting ILI estimates were consistently 1-2 weeks ahead of CDC’s reports.
        • Up-to-date report may enable public health officials and health professionals to respond better to seasonal epdemics.
      • Cons
        • Can’t replace traditional surveillance networks or supplant the need for laboratory-based diagnoses and surveillance.
        • The correlations are only meaningful across large populations since the queries are not only submitted by users who are experiencing influenza-like symptoms.
        • The system remains susceptible to false alerts caused by a sudden increase in ILI-related queries, such as drug recall event.