BACKGROUND
Opioid misuse is a major health problem in the United States, and can lead to addiction and fatal overdose. The United States is in the midst of an opioid epidemic; in 2018, an average of approximately 130 Americans died daily from an opioid overdose and 2.1 million have an opioid use disorder (OUD). In addition to electronic health records (EHRs), social media have also been harnessed for studying and predicting physical and behavioral outcomes of OUD. Specifically, it has been shown that on Twitter the use of certain language patterns and their frequencies in subjects’ tweets are indicative of significant healthcare outcomes such as opioid misuse/use and suicide ideation. We sought to understand personal traits and behaviors of Twitter chatters relative to the motive of opioid misuse; pain or recreational.
OBJECTIVE
.
METHODS
We collected tweets using the Twitter public developer application programming interface (API) between April 13, 2018 – and May 21, 2018. A list of opioid-related keywords were searched for such as methadone, codeine, fentanyl, hydrocodone, vicodin, heroin and oxycodone. We manually annotated tweets into three classes: no-opioid misuse, pain-misuse and recreational-misuse, the latter two representing misuse for pain or recreation/addiction. We computed the coding agreement between the two annotators using the Cohen’s Kappa statistic. We applied the Linguistic Inquiry and Word Count (LIWC) tool on historical tweets, with at least 500 words, of users in the dataset to analyze their language use and learn about their personality raits and behaviors. LIWC is a text processing software that analyzes text narratives and produces approximately 90 variables scored based on word use that pertain to phsycological, emotional, behavioral, and linguistic processes. A multiclass logistic regression model with backward selection based on the BIC criterion was used to identify variables associated with pain and recreational opioid misuse compared to the base class; no-opioid misuse.. The goal was to understand whether personal traits or behaviors differ across different classes. We reported the odd ratios of different variables in both pain and recreational related opioid misuse classes with respect to the no-opioid misuse class.
RESULTS
The manual annotation resulted in a total of 1,164 opioid related tweets. 229 tweets were assigned to the pain-related class, 769 were in the recreational class, and 166 tweets were tagged with no opioid misuse class. The overall inter-annotator agreement (IAA) was 0.79. Running LIWC on the tweets resulted in 55 variables. We selected the best model based on BIC. We examined the variables with the highest odd ratios to determine those associated with both pain and recreational opioid misuse as compared to the base class. Certain traits such as depression, stress, and melancholy are established in the literature as commonplace amongst opiod abuse indiviuals. In our analysis, these same characteristics, amongst others, were identified as significantly positively associated with both the Pain and Recreational groups compared to the no-opioid misuse group. Despite the different motivaions for opiod abuse, both groups present the same core personality traits. Interestingly, individuals who misuse opioids as a pain management tool exhibited higher odds ratios for psychological processees and personal traits based on their tweet language. These include a strong focus on discipline, as demonstrated by the variables “disciplined”, “cautious” and “work_oriented”. Their tweet language is also indicative of cheerfulness, a variable absent in the recreational misuse group. Variables associated with the reacreational misuse group revolve around external factors. They are generous and motivated by reward, while maintaining a religious orientation. Based on their tweet language, this group is also characterized as “active”; we understand that these individuals are more social and community focused .
CONCLUSIONS
To our best knowledge, this is the first study to investigate motivations of opioid abuse as it relates to tweet language. Previous studies utilizing Twitter data were limited to simply detecting opiod abuse likelihood through tweets. By delving deeper into the classes of opioid abuse and its motivation, we offer greater insight into opioid abuse behavior. This insight extends beyond simple identification, and explores patterns in motivation. We conclude that user language on Twitter is indicative of significant differences in personal traits and behaviors depending on abuse motivation: pain management or recreation.