Investigating Twitter as a Source for Studying Behavioral Responses to Epidemics

Introduction Public health policies combatting the spread of infectious disease rely on bio-surveillance systems for coordinating responses such as vaccination programs. Classical epidemiology assumes that population behavior remains constant during an epidemic. However, the public response to an epidemic can have a serious impact on an epidemic’s course. As a result, modeling the influence of population behavior has become a recent trend in epidemiology research (Funk, Salathé, and Jansen 2010; Bisset et al. 2009), and there have been a few proposed mathematical models to explain individual responses to an outbreak, including game theoretic models (Reluga 2010) and agent based models (Epstein et al. 2008). Model validation has relied on simulated epidemic data because of a lack of empirical data. Consider influenza (seasonal flu) which, as the season progresses, leads to an increased awareness of infection, which may cause individuals to respond by obtaining a flu vaccination or limiting social contact. Fear of an illness can itself be a contagion (Epstein 2009), and disease propagation models could be applied. However, research will be limited without high quality empirical data reflecting population behavioral responses. One promising source of such data are social media, such as Twitter. Recent studies have shown an ability to track influenza rates from Twitter (Paul and Dredze 2011; Aramaki, Maskawa, and Morita 2011; Signorini, Segre, and Polgreen 2011; Lampos, De Bie, and Cristianini 2010; Culotta 2010) since Twitter users tweet illnesses (“i am home sick with the flu”). However, users may also tweet concerned awareness of illness (“don’t want to get sick, need a flu shot”). Identifying these messages can support proposed computational epidemic response models. We present preliminary results for mining concerned awareness of influenza tweets. We describe our data set construction and experiments with binary classification of data into influenza versus general messages and classification into concerned awareness and existing infection.