Comparing Twitter data to routine data sources in public health surveillance for the 2015 Pan/Parapan American Games: an ecological study

Yasmin Khan, Garvin J. Leung, Paul Belanger, Effie Gournis, David L. Buckeridge, Li Liu, Ye Li, Ian L. Johnson

This study examined Twitter for public health surveillance during a mass gathering in Canada with two objectives: to explore the feasibility of acquiring, categorizing and using geolocated Twitter data and to compare Twitter data against other data sources used for Pan Parapan American Games (P/PAG) surveillance. Syndrome definitions were created using keyword categorization to extract posts from Twitter. Categories were developed iteratively for four relevant syndromes: respiratory, gastrointestinal, heat-related illness, and influenza-like illness (ILI). All data sources corresponded to the location of Toronto, Canada. Twitter data were acquired from a publicly available stream representing a 1% random sample of tweets from June 26 to September 10, 2015. Cross-correlation analyses of time series data were conducted between Twitter and comparator surveillance data sources: emergency department visits, telephone helpline calls, laboratory testing positivity rate, reportable disease data, and temperature. The frequency of daily tweets that were classified into syndromes was low, with the highest mean number of daily tweets being for ILI and respiratory syndromes (22.0 and 21.6, respectively) and the lowest, for the heat syndrome (4.1). Cross-correlation analyses of Twitter data demonstrated significant correlations for heat syndrome with two data sources: telephone helpline calls (r = 0.4) and temperature data (r = 0.5). Using simple syndromes based on keyword classification of geolocated tweets, we found a correlation between tweets and two routine data sources for heat alerts, the only public health event detected during P/PAG. Further research is needed to understand the role for Twitter in surveillance.