1 of 1

Mining Digital Media to Understand the

Individual and Societal Impact of Substance Use Disorders

Daniel Habib (Preceptors: Brenda Curtis, Salvatore Giorgi, Douglas Bellew, TingTing Liu)

Technology and Translational Research Unit, National Institute on Drug Abuse Intramural Research Program

Introduction

Stigma, privacy concerns, and legal repercussions drive many people with a substance use disorder (SUD) to seek support online (Bergman and Kelly, 2021).
Unlike other platforms such as Twitter, Reddit can be filtered by subforum (i.e., subreddit) and does not have a character limit, which proves more fruitful for language processing (Cohan et al., 2018).
Behavioral health datasets rarely include substance use (SU) data, and when they do, they only include the most common ones.

Methods

Aims

Create SU-related and self-disclosed SUD/recovery datasets of Reddit posts
Improve reproducibility and uniformity for assessing substance use via social media

Future Directions

Discussion

All subreddits and posts from 01/2006-12/2020 were extracted from the Pushshift Reddit Dataset and uploaded to MySQL (Baumgartner et al., 2020).
SU keywords were used to find SU subreddits and compile posts within them into the SU-related Dataset.
The Self-disclosed SUD and Recovery Dataset was compiled from Reddit users within the SU subreddits who self-disclosed an SUD diagnosis. All of these users’ posts both within and outside of SU subreddits were included in the dataset. Additionally, posts from users with similar traits within health and general subreddits were added as matched controls.

Compare how individuals with an SUD have been dehumanized on various media platforms
Optimize drug term discovery methods
Geolocate Reddit users to yield location-specific data (Balsamo et al., 2019)
Study the intersectionality of an SUD with gender or other SUDs
Generalize findings to different contexts (Huang et al., 2012)
Analyze users’ language to study/predict users with an SUD

SU-related Dataset: corpus for studying the language around SU
Self-disclosed SUD and Recovery Dataset: centralized source for testing algorithms that automatically detect/predict users with an SUD
These datasets fill a gap in behavioral health data that is not addressed by mental health datasets.

Figure 1. Schema of the SU-related Dataset and the Self-disclosed SUD and Recovery Dataset. Posts are within quotes and subreddits begin with “r/”. The SU-related Dataset is public facing while the Self-disclosed SUD and Recovery Dataset must requested.

Limitations

Real users with bot-like names
Capturing slang terms
False positives in diagnosis patterns
Matched controls
Representativeness of users who self-disclose

Ethical Considerations

Consent and anonymity
Responsibilities of social media companies

Stigma, privacy measures, and legal concerns surrounding substance use has led to fewer people seeking traditional treatment and a relative lack of behavioral health data compared to other health domains. While other researchers have leverage social media to study depression and suicide prevention algorithms as you can see at the bottom right, our lab taps into the wealth of online data to study the language around substance use, which has been relatively overlooked in online behavioral health research.

�

Specifically, I sought to compile two datasets. The first was the substance use related dataset that includes all social media posts regardless of whether the poster was the one using the substance or not. This would facilitate the study of dehumanizing drug language, evolving drug slang terms, and patterns in substance use.

�

The second was the self-disclosed substance use disorder and recovery dataset that included all posts by users who divulged their diagnoses with an SUD. From these data, we could train computer programs to automatically detect people who need help just by analyzing their language.

�

For these purposes, Reddit posts seemed to fit the bill nicely. Reddit is divided into subforums called subreddits within which users submit posts that have no character limit and are relevant to the subreddit title. First, I read in the billions of Reddit posts from January 2006 to December 2020. For the SU-related dataset, I searched for subreddit titles containing substance use keywords like cocaine, weed, et cetera, and compiled the millions of posts within these subreddits. Hence, posts were filtered at the subreddit level.

�

In contrast, the self-disclosed substance use disorder and recovery dataset was filtered at the user level. Users that posted about their SUD diagnosis within a substance use subreddit were added to the self-disclosed SUD and recovery dataset. So, not just their SU-related posts but all of their posts across all subreddits were included. As you can see in the figure, there is necessarily some overlap between the datasets. All posts about self-disclosed SUD diagnoses are in both datasets. Notice, though, that if a user posts in a substance use subreddit, their unrelated posts are not included in the SU-related dataset. On the other hand, posts irrelevant to substance use are included in the self-disclosed SUD and recovery dataset if the user self-disclosed their diagnosis in a substance use subreddit. This is because prediction algorithms must detect language patterns for people known to have an SUD without being trained on posts that explicitly contain the words “diagnosed with an SUD”. That would not be very helpful. The key is to detect language patterns indicative of an SUD in everyday speech. For instance, it doesn’t achieve much to predict depression in a user who states “i have depression”. Instead, detecting subtle clues like a high proportion of words with negative connotation et cetera provides novel insight into earlier detection.

�

3.5 mins

�

While all of the data is public, care must be taken to avoid associating the anonymous usernames to people. Also, the sensitive nature of compiling diagnosed users led us to not publicly release the self-disclosed SUD and recovery dataset though it can be requested. This research has profound ethical implications for social media companies such as censoring dangerous language or detecting users’ risk of overdose, but these must be balanced with the rights of free speech and being able to post without the fear of being given a behavioral health assessment.

�

Some difficulties we’ve run into are capturing drug terms with double meanings and checking for false positives. Also, we’ve had to compile control posts from people who are similar to people with an SUD in terms of social media activity (like post frequency or number of likes) but do not post about substance use. It is also important to note that the type of people divulging their SUD diagnoses might have different traits than people with an SUD but who do not post about it.

�

Aside from what I’ve already mentioned, there are numerous future directions for these Reddit datasets, such as providing location specific data based on local subreddits like Trees Maryland. We will also track the usage of stigmatizing words such as ‘junkie’ over time to see how the portrayal of people with an SUD has evolved. Thank you everyone for your guidance and support on this project and thank you all for listening.

Introduction

- Tapping into these data captures a more complete picture of each patient and allows for changes in behavioral health status to be identified in real time (Gkotsis et al., 2017).

Ethical Considerations

- Social media users do not expect their online discussions to be used for behavioral health assessments. Thus, utilizing the Self-disclosed SUD and Recovery Dataset requires additional tact.

- Social media companies might have additional responsibilities for SU prevention and might be held culpable when the failure of automatic detection leads to overdose.

Potential censorship of harmful stigmatizing language online calls into question the balance between free speech and users’ well-being.

Limitations

Double meanings for drug terms (e.g., coke)
Unknown and internet-specific slang terms (e.g., doja and ents)
Matching protocols relying on evidence that users’ personality, psychological, and demographic characteristics can be estimated by social media activity (Schwartz et al., 2016)
The type of user who chooses to self-disclose not being representative of the general target population (Lippincott & Carrell, 2018)

Future Directions

using Natural Language Processing and Machine Learning
Establish Linguistic equivalency between platforms to capture the various ways words are semantically represented in

- by subreddit, post content, and user description