Predicting risky drinking: It might be all in the words


August 15, 2017

Rachel Kornfield
PhD Candidate
School of Journalism and Mass Communication
Research Assistant 
University of Wisconsin-Madison 






The A-CHESS smartphone app provides addiction recovery services on-demand.  Analyzing the language used in A-CHESS discussion forums is helping researchers predict the likelihood of relapse. 
The words we say in daily conversation can provide a powerful window into our state of mind, including our moods, concerns, and priorities. General topics of discussion can be revealing (for example, if we’re talking about friends, the weather, or problems at work). But even more is often revealed by subtler styles of speech, including the pronouns we use, our emotional tone, and how we put our sentences together. These subtle linguistic differences are especially meaningful in an age when computers play an ever-increasing role in our lives. Technology and social media provide an array of new outlets through which to communicate. At the same time, computer science offers new tools to automatically measure subtle qualities of language. At the Center for Health Enhancement Systems Studies (CHESS), our research uses social media language not only to understand people better, but also to help people improve their health.

A-CHESS display 
For decades, our research center has been developing evidence-based computerized systems to guide and support individuals as they face different health challenges (e.g., cancer, aging). Recently, we have focused our attention on addiction by developing 
A-CHESS, a smartphone application that provides a range of on-demand recovery services, including self-help meeting directories, expert advice, games, and message forums that participants can use to instantly connect to their peers. Within these forums, participants can seek and provide support, or simply chat and build relationships. Having access to on-demand information and social support, even at 4am on a Tuesday, can make the challenges of recovery much more manageable. Our research has shown that providing individuals with A-CHESS reduced their heavy drinking, even a year later. However, until recently, we had not assessed the particular types of language produced by participants communicating with each other through A-CHESS.

...we found that those who went on to relapse show higher rates of swearing and negative emotion words in their messages, whereas those who did not relapse used more words related to achievements and information processing. 
As a doctoral candidate at CHESS, my research looks at how the language that individuals submit to online discussion boards can offer useful clues about future health and well-being. This research often involves very slowly coding messages by hand, but it can also involve computerized programs that automatically count dozens of subtle linguistic qualities in large bodies of text within seconds. Recently, we used a computer program called LIWC to identify a number of linguistic signals that precede episodes of risky drinking among A-CHESS participants.

For instance, we found that those who went on to relapse show higher rates of swearing and negative emotion words in their messages, whereas those who did not relapse used more words related to achievements. Taken together, individuals’ styles of writing provided a better indicator of future recovery success than survey-based measures. The surveys involved asking participants to report their level of confidence, social issues, and demographic characteristics, and took a lot of time and effort for participants to complete. In other words, automated linguistic analysis may improve our ability to predict outcomes while also reducing burden on patients – a win – win!
(This research will be reported in more detail in a forthcoming issue of Health Communication).

These linguistic insights also have practical applications for improving tools like A-CHESS. While A-CHESS provides a range of on-demand services, some individuals benefit from more personal attention. Knowing who is struggling will allow us to intervene to help people when they need help the most. We are already beginning to refine A-CHESS to do just this: In collaboration with computer scientists and engineers, we have programmed algorithms to run real-time scans of messages posted on the A-CHESS discussion board, looking for concerning words, so that we can direct help to those who need it, when they need it. For example, if a participant posts a message about stronger-than-usual cravings, our algorithm may recognize familiar patterns of words, which will alert a trained human moderator and allow her to respond by offering emotional support, advice, or even calling the participant on the phone.

Future research

A-CHESS word cloud
As we continue to improve A-CHESS, we expect to reduce the burden on our human moderators by incorporating automated conversational agents, which are computer programs that provide a human-like interaction. Regular conversations with a supportive human-like “chatbot” can engage and support participants throughout the day, while also collecting further information about their challenges, allowing trained human support to be pulled in only when necessary. Of course, there are many concerns we have about how people will respond to receiving help from “chatbots,” and how people respond will likely depend on our execution. As conversational agents play a greater role in more and more areas of our lives, understanding how people react to them will be an important and exciting research endeavor for years to come.

Individuals who use A-CHESS have often remarked that writing on the discussion forum has allowed them to get through their most difficult moments. Even when they don’t get an instantaneous response, individuals feel that they are sharing their thoughts, feelings, and concerns with an audience who cares. And they are correct. Not only will their peers read and consider their words, but the system itself is increasingly “listening” and responding, allowing us provide better care to more people. 

About our guest blogger: 

Rachel Kornfield is a doctoral candidate in Mass Communication at the University of Wisconsin-Madison. Her research is aimed at better understanding how emerging communication technologies facilitate therapeutic self-disclosure and social support exchange in order to improve individuals’ health and well being. Rachel has worked with the Center for Health Enhancement System Studies on several projects examining how individuals benefit from A-CHESS, a new mobile application for substance use disorders. 



3 comments:

  1. This comment has been removed by a blog administrator.

    ReplyDelete
  2. Thanks, Rachel, for your post. This is exactly the type of technology application that we in the recovery services provision system need to be utilizing. If Amazon can tell us what other products we might like, we ought to be able to use the chatbots to raise the red flag. I was fascinated with the mention of LIWC. What an important resource. I have an associate doing research on recovery stories and this will likely prove helpful to her.

    Roxanne Kibben

    ReplyDelete
  3. Thanks for sharing these insights into the work you're doing. One of the main challenges is getting people (especially those in early recovery) to communicate at all. If they don't post, there is nothing to analyze. Most alcoholics and addicts withdraw socially. Does your app measure lack of engagement as well? It would be interesting to track--on a user-by-user basis--levels of engagement and participation. Perhaps this would be another useful indicator that could help identify individuals at risk of relapse.

    - David S.

    ReplyDelete