When you’re addicted, almost everything can be a trigger. A location. A song on the radio. A memory. The smallest thing can unleash an irresistible craving that leads to relapse.
It’s what makes overcoming addiction to alcohol and substances a lifelong daily battle for the millions who suffer from it. And it’s one of the things that has driven John Curtin, a professor in the Department of Psychology, to spend the last decade developing technology-based tools that can assist users in avoiding relapse. Today, Curtin and his colleagues are deploying machine-learning algorithms to hone and give feedback on the effectiveness — and fairness — of the tools they have developed.
“Broadly, we’re interested in supporting people through their recovery from alcohol and other substance-use disorders,” says Curtin. “The research community is very clear that alcohol and substance-use disorders are chronic disorders that require lifelong management. We don’t seem to treat them that way very well. We often provide initial treatments to help people stop or reduce their use, but long-term continuing care is rarely available. It’s very difficult to pay attention throughout the rest of your life to the integrity of your recovery without additional support.”
People in recovery from addiction need to pay attention to a laundry list of concerns to maintain healthy balance: stress levels, positive events, the social network they’ll encounter when they leave the house, their daily lifestyles. The resources each individual needs to support their recovery vary wildly, especially over time.
Back in 2014, Dave Gustafson (’62, MS’63, PhD’66), an engineering professor who heads the Center for Health Enhancement Systems Studies (CHESS), created an app-based intervention program to help manage these concerns. In many ways, the app was extremely effective: It cut heavy drinking days in half, and it almost doubled the odds of abstinence over the first year of recovery for individuals following inpatient treatment.
But Gustafson noticed that many of the people who lapsed or relapsed during the first year hadn’t used the app in the days leading up to that lapse, and others hadn’t used the supports in the app that he thought would have been most effective for them. Gustafson approached Curtin and his lab with a simple question: Could you predict precisely when someone might lapse back to alcohol or substance abuse, and could you identify why and how best to support them?
It’s the first time we’re actually going to be able to give a tool to the participant that can potentially help them with their recovery.
The answer was yes. Backed by the first of several grants from the National Institutes of Health, Curtin was able to use the signals they were monitoring to begin predicting relapses for alcohol abuse. These signals included ecological momentary assessments (EMA). At various intervals, a set of questions was pushed out to users’ smartphones to gauge things like stressors, cravings and emotional state. The app also tracked a patient’s location, calls and texts.
“Those signals, by themselves, are useful for understanding how broadly healthy the individual is and how healthy their recovery is,” says Curtin, who got into addiction research in part because of what he calls a “dense family history” of alcohol and substance abuse. “We can see over time the shifts in their patterns and their movements.”
All that data was fed into machine-learning algorithms that developed predictions as to precisely when a patient was at risk of relapse. Over time and over grant cycles, Curtin’s lab has been able to use the algorithms in combination with interpretable AI techniques that allow them to understand why the models are making the predictions that they’re making — and how they might be improved.
“For one individual, it may be predicting that in the next day, they’re at high probability of lapsing, because they’ve had a number of stressful events recently,” explains Curtin. “For someone else, they may be at high probability of lapsing because they’ve been craving a lot.”
The interpretive models also suggest factors that could be contributing to the probability of relapse and recommend personalized interventions to address them based on the data being collected. Under the latest grant, the algorithms will collect data and provide messages to participants daily over a six-month period. At the end, Curtin will be able to measure how useful they find it and how willing they are to engage.
“That’s what we’re most excited about,” says Curtin. “It’s the first time we’re actually going to be able to give a tool to the participant that can potentially help them with their recovery. Everything up until now has been building the system, and now we can start to release and try to implement the system.”
The team is now working on addressing issues of algorithmic fairness. For example, the difference between predictors in an urban versus a rural setting can be stark, and algorithms are historically bad at making predictions for people of color, because the data used to train them aren’t always broadly representative. Curtin has intentionally addressed this problem in his research because he wants to make sure that this resource is a tool that anyone who is struggling with addiction can benefit from, regardless of their background.
“We need to be able to serve everyone with the tools that people can use to help with this, and that’s really what we’re hoping to be able to do,” he says.
