Why are we doing this research?
One of the greatest problems for people in recovery is that it is a chronic, ongoing journey. People need access to different kinds of help and support throughout their lives, not just at their scheduled treatment visits.
What we need is a way for people who are in recovery to have that continuing support available whenever they need it. Our goal is to design a system that can be integrated into a smartphone app, which people can access 24/7, from anywhere. There are lots of self-help support apps, but so far none of them can provide personalized insights anywhere near as good as a therapist might be able to give. We’re hoping to change that.
What is this study about?
We already know from previous research that many people in recovery benefit from using “digital therapeutics” or health-related smartphone apps, to help support them in recovery. There are lots of health-related smartphone apps – you may be familiar with apps for tracking your sleep, your weight loss, or your glucose levels.
We think a health-related smartphone app for recovery would most helpful if it can detect when people are at a higher risk of drinking, and send them a personalized message to warn them. In this study, we’ll be testing a recovery support system, which can send people different kinds of personalized automated messages, to learn what types of messages best help people stick to their recovery goals when they are at higher or lower risk of drinking.
Learn more: Recovery warning signs
A lot of people in recovery have “slips,” where they drink alcohol, but are still committed to recovery. These slips are sometimes called “relapses,” “episodes,” or “lapses.” And while slips may seem to come out of nowhere and take people by surprise, research has found evidence of warning signs like really stressful events, big changes to life circumstances, changes in moods, or being around friends or places that remind people of their past drinking.
In our earlier studies, people in recovery shared information about their lives with us over many months, and this helped us to identify the feelings, events, and behavior patterns that can predict someone might be at risk of using alcohol and other drugs.
Some of these warning signs we found in what people told us about their own experiences. We asked people every day about their moods, experiences of stress, urges to use, and how much they wanted to avoid using. They were also asked every day if and when they had drank alcohol.
Another warning sign came from the location data they shared with us. People’s routines and patterns of where they go or spend their time, or changes to their routines and patterns, also contain warning signs.
How do the personalized predictions and personalized messages work?
Based on data from our previous studies, we’ve written a program called a “machine learning model” which can analyze everything people shared with us, and make a personalized prediction for every person in the study about whether they were likely to have a drinking lapse in the next day or week. It makes this prediction with about 90% accuracy, meaning sometimes it could be wrong, but about 90% of the time it was right about predicting whether people reported a the had a lapse.
In this study, we’ll use that same model to make daily personalized predictions for you. You’ll fill out a survey every day, so the predictions about your likelihood of lapsing are based on YOUR data. They’ll be different for every person in the study, and will change either a little or a lot over time, depending on how much your experiences are affecting your recovery.
That prediction as it comes out of the model is just a set of numbers, and doesn’t mean much to the average person. We’ll need to write a personalized message to tell you about your personal recovery prediction in a supportive way, and that means so many thousands of messages that we need to find an automated way to do it. So what we’ll do in this study is use the recovery support messaging system to send you a text message every day with an automated, but personalized, supportive message, based on what the machine learning program predicted about you.
You’ll be able to tell us every day how helpful and supportive you think that message is. That will help us improve the messaging system, and also the program that makes personalized predictions. Ultimately, we want to make this system available to anyone who wants it, so they can get a supportive message every day about their recovery, right on their smartphone.
Learn more: Personalized recovery predictions
We use a “machine learning model” to analyze your data and make predictions. You may have heard machine learning referred to as “AI” or Artificial Intelligence. However, our machine learning model is not AI, it’s just a computer program which can do very complicated math! We have collected data from hundreds of participants in studies similar to this one, and the computer program has “learned” what types of data predict that someone in recovery will have a slip where they drink. It’s more accurate to say that we have “trained” the data to find these patterns, because each data point is also marked with whether or not that person reported a slip where they used alcohol or other drugs.
Here are some other ways this machine learning program is not AI: it is a program that was written by our study staff. It runs on computers at the University of Wisconsin, owned by the University. Sometimes it takes a lot of computers to run, but we can even run some of it on our small laboratory workstations. Also, no data that goes into the model ever leaves the University.
When we feed your data into the program, it tries to predict whether or not you are likely to drink, by looking for the same patterns it found in its training data that were related to reports of people using alcohol. These predictions might be more or less accurate, and sometimes they might just be wrong. You’ll be able to tell us when a bad prediction is made, to help us improve the program
Learn more about: Automated personalized support messages
The recovery support messaging system, is how we tell you about the daily predictions made by the machine learning program, in plain, supportive language.
Since we need to test different types of messages, we need to write a LOT of them, and every message needs to be personalized to every person every day. It will require over 50 thousand messages! Our lab staff couldn’t write this many messages without them being repetitive and prone to human error.
In order to generate so many personalized support messages, we are using a Large Language Model, or LLM. It’s not “AI” even though some people call it that. It’s more like a fancy version of the predictive text that your phone uses to suggest the next word you want to type. But you have probably heard of this LLM: Chat-GPT.
In this study, we’ll be using a University-licensed instance of Chat-GPT which is run on University servers. It’s based on the same code as the public version of Chat-GPT, but the data we put into it never leaves the University servers and can’t be accessed, seen, or used by any other version of Chat-GPT outside or inside the University. When we’re done with the study, we’ll delete the LLM, and everything we’ve ever put into it will be gone, forever.
The data goes that goes into this LLM is very simple: It will be a random string of numbers assigned to your data, to tell it apart from other people in the study. It’s also a set of Yes/No values assigned to certain variables.
That helps the LLM write a message for a person identified by those numbers, who is at not yet at risk of drinking, but reported they are having strong cravings. There is NO information that can identify you will go into the LLM.
To tell the LLM the kind of message we need, we’ve worked for months with psychologists who run real-life clinics where they provide counseling and treatment for those in recovery. They’ve helped us figure out how to prompt the LLM how to write a helpful recovery support message. We think the automated messages look very similar to the ways a real clinical counselors might say things. But until we get feedback on which messages are helpful and which aren’t helpful, from people like you with the lived experience of being in recovery, the messages might not fit you perfectly. You’ll be able to tell us what kinds of messages you do and don’t like, and how helpful or supportive they are.