Results and Methodological Implications of the Digital Epidemiology of Prescription Drug References Among Twitter Users: Latent Dirichlet Allocation (LDA) Analyses

Abstract

Background: Social media is an important information source for a growing subset of the population and can likely be leveraged to provide insight into the evolving drug overdose epidemic. Twitter can provide valuable insight into trends, colloquial information available to potential users, and how networks and interactivity might influence what people are exposed to and how they engage in communication around drug use.

Objective: This exploratory study was designed to investigate the ways in which unsupervised machine learning analyses using natural language processing could identify coherent themes for tweets containing substance names.

Methods: This study involved harnessing data from Twitter, including large-scale collection of brand name (N=262,607) and street name (N=204,068) prescription drug-related tweets and use of unsupervised machine learning analyses (ie, natural language processing) of collected data with data visualization to identify pertinent tweet themes. Latent Dirichlet allocation (LDA) with coherence score calculations was performed to compare brand (eg, OxyContin) and street (eg, oxys) name tweets.

Results: We found people discussed drug use differently depending on whether a brand name or street name was used. Brand name categories often contained political talking points (eg, border, crime, and political handling of ongoing drug mitigation strategies). In contrast, categories containing street names occasionally referenced drug misuse, though multiple social uses for a term (eg, Sonata) muddled topic clarity.

Conclusions: Content in the brand name corpus reflected discussion about the drug itself and less often reflected personal use. However, content in the street name corpus was notably more diverse and resisted simple LDA categorization. We speculate this may reflect effective use of slang terminology to clandestinely discuss drug-related activity. If so, straightforward analyses of digital drug-related communication may be more difficult than previously assumed. This work has the potential to be used for surveillance and detection of harmful drug use information. It also might be used for appropriate education and dissemination of information to persons engaged in drug use content on Twitter.

Keywords: LDA; NLP; Twitter; data mining; digital epidemiology; drug; drug use; epidemiological; epidemiology; machine learning; natural language processing; pharmaceutic; pharmaceutical; pharmaceuticals; pharmacy; prescription; social media; text mining; tweet; tweets; unsupervised analysis.