Sixty-two adults being treated with buprenorphine for opioid use disorder at an outpatient clinic in California spent six months answering a short survey on their smartphones three times a day, describing their mood, cravings and surroundings in real time. Researchers fed those 14,322 individual survey responses into deep learning models, and the resulting system forecast which patients were at elevated risk of relapse or of dropping out of treatment with what the study’s authors called exceptional accuracy. The National Institutes of Health funded research, published in the Journal of Substance Use and Addiction Treatment in June, is one of the clearest demonstrations yet that AI can turn passive, moment-to-moment check-ins into an early warning system for one of medicine’s hardest problems to predict.
A crisis where timing is everything
The Food and Drug Administration estimates that between 65 and 70 percent of people treated for opioid use disorder relapse at some point, and the Substance Abuse and Mental Health Services Administration and other federal data sources put the number of Americans living with opioid use disorder at somewhere between five and seven million. Addiction counselors have long known that relapse rarely comes out of nowhere; it is typically preceded by a buildup of stress, boredom, exposure to triggering environments or emotional strain. The problem has always been detecting that buildup in time to intervene, since patients see counselors only periodically and often cannot recognize or report their own rising risk until it is too late.
What the smartphone data revealed
The study, conducted with participants recruited between June 2020 and January 2021, found that past-hour substance use was the single strongest predictor of continued use, but situational and emotional signals mattered too. Stress and physical pain tended to forecast relapse risk several days in advance, giving clinicians a meaningful window to intervene, while more immediate mood states such as boredom, exhaustion and low contentment, along with difficulty planning or following through on tasks, signaled elevated risk within hours rather than days, according to the research team, which included investigators from institutions across New Hampshire, New York, California, Massachusetts and Maryland.
Voices from the field
Caroline Easton, academic division chief of addiction psychiatry at the University of Rochester Medical Center, has said AI tools like this one could meaningfully reduce the burden on human therapists. “AI can help reduce therapists’ workloads, minimize compassion fatigue and burnout and provide additional resources for patient care,” Easton said. But Brenda Curtis, a researcher at the National Institute on Drug Abuse, has expressed more guarded enthusiasm, describing “trepidation” about how AI tools might be used in addiction treatment even as she said the field is taking a hard look at implementation questions before scaling them broadly.
The limits researchers are upfront about
The study’s authors were explicit about its constraints. The sample of 62 participants was modest and drawn from a population that was predominantly white, raising concerns that a model trained on this group could misread behavioral and linguistic cues from patients of different racial or ethnic backgrounds, a problem that has surfaced repeatedly in health AI more broadly. The models also produced false positives and false negatives, and performed less well predicting whether patients would simply stop complying with their medication regimen compared with predicting active relapse. None of this means the approach does not work, but it does mean the current results describe a promising pilot rather than a validated, ready-to-deploy clinical tool.
Two views of the same technology
Supporters see real-time behavioral prediction as a rare chance to intervene during the narrow window between rising risk and actual relapse, potentially allowing a counselor to reach out with a phone call or resource exactly when a patient needs it most rather than at the next scheduled appointment weeks later. Critics worry about the practical and ethical complications of monitoring vulnerable patients this closely: constant self-reporting can itself become burdensome, false alarms could erode trust in the technology, and there are open questions about who sees a patient’s risk score, how it is used, and whether it could ever be weaponized against a patient’s autonomy in treatment decisions rather than used purely to offer support.
What’s next
Researchers say larger, more demographically diverse studies are needed before predictive relapse models can be deployed at scale across addiction treatment programs nationally. If future trials replicate these findings across broader populations, the approach could become a standard complement to medication-assisted treatment, giving counselors a data-driven signal for when to check in rather than relying solely on scheduled visits. Given that the opioid crisis continues to claim tens of thousands of lives annually in the United States, even a modestly effective early-warning system, deployed responsibly, could translate into meaningful numbers of prevented relapses.
