Leveraging Twitter Data to Explore the Feasibility of Detecting Negative Health Outcomes Related to Vaping

Adverse health outcomes (e.g., respiratory infections, lung injury, death) related to vaping were reported at significantly higher rates in healthcare systems starting in the fall of 2019. This study seeks to leverage artificial intelligence (AI) techniques, such as latent dirichlet allocation (LDA) methods, to determine whether a signal of these negative health outcomes could have been detected by the frequency of Twitter content posted about vaping and these health outcomes prior to this increase. We utilized a random sample of 3,523 tweets from 2019 and performed LDA methods on this sample to cluster the tweets and identify latent topics. We then utilized keywords from within the health-related cluster (topic) to manually verify the frequency of these tweets across previous years to approximate topic trends. LDA methods resulted in 4 distinct topics of tweets, including a health-related topic. Keywords from this topic were found to increase slightly in 2017 and 2018, with a dramatic increase in 2019. Further, the highest performing keyword combination was found to increase most significantly beginning in August 2019. The results of this study support the feasibility of leveraging artificial intelligence techniques for surveillance of public health concerns such as vaping and adverse health outcomes reported in Twitter. Further research is needed into the development of such models, which could promote earlier detection of public health issues and timely outreach to those groups most at risk.