How Joe and Jane Tweet about Their Health: Mining for Personal Health Information on Twitter

Proceedings of Recent Advances in Natural Language Processing, pages 626–632, Hissar, Bulgaria, 7-13 September 2013. How Joe and Jane Tweet about Their Health: Mining for Personal Health Information on Twitter Marina Sokolova CHEO Research Institute University of Ottawa sokolova@uottawa.ca Stan Matwin Dalhousie University University of Ottawa stan@cs.dal.ca Yasser Jafer University of Ottawa yjafe089@uottawa.ca David Schramm CHEO and University of Ottawa dschramm@toh.on.ca Abstract With 19%–28% of Internet users partici- pating in online health discussions, it be- came imperative to be able to detect and analyze posted personal health informa- tion (PHI). In this work we introduce two semantic-based methods for mining PHI on social networks which will warn the users about potential privacy breaches. One method uses WordNet as a source of health-related knowledge, another - an ontology of personal relations. We use Twitter data to empirically evaluate our methods. We also apply Machine Learn- ing to demonstrate advantages of our ex- traction procedure when tweets contain- ing PHI have to be automatically identified among other tweets. Keywords: Text mining, Twitter, Personal Health Information, Machine Learning 1 Introduction Online networking websites Facebook, Twitter, PatientsLikeMe became popular communication hubs connecting millions of individuals. In casually written messages (posts, tweets, up- dates), people discuss life experience (i plan to stay home and watch christmas movies while I get better) and com- ment on various events (The discovery was made via CAT scans) 1 amongst others. While posting about personal health, a user re- veals details that in pre-social network era were usually discussed during visits to a health care provider or in a family setting. This detailed health description is called Personal Health Infor- mation (PHI) (Hersh, 2009). Posted online PHI 1All messages have authentical spelling and content. is used in several practical applications: formu- lating Web policies, including privacy and confi- dentiality concerns or information leak prevention (Ghazinour et al, 2013a), understanding popula- tion response on health care policies (vaccination, immunization)(Chew and Eysenbach, 2010), and an early detection of adverse health-related events (Lampos and Christianini, 2010). Recent studies of 11,000 posts on a social net- work showed deficiencies of traditional electronic sources of medical information in the task PHI de- tection (Ghazinour et al, 2013b). Our current work aims to show that it is pos- sible to considerably improve accuracy of PHI extraction from social networks. Our approach uses the PHI ontology presented in (Sokolova and Schramm, 2011). The ontology’s structure and terms reflect on patient communications in health care setting. In this paper, we present two semantic-based enhancements of the ontology and apply them to extract PHI in Twitter. One en- hancement uses WordNet as a source of health- related knowledge, another – an ontology of per- sonal relations. We use manual analysis to demonstrate that in- corporating semantic information significantly im- proves Precision and Fscore of the PHI text re- trieval. We also apply Machine Learning methods to show the advantage of our approach in auto- mated detection of PHI. Partial preliminary results of this work had been reported in …

Content truncated to 500 words