The clinically interesting part of AI in addiction treatment and recovery is not that a model can label relapse risk after the visit is missed or after non-prescribed opioid use has already occurred. The useful question is whether it can notice deterioration early enough for a care team to do something safe and proportionate.
That is why the most important recent evidence is not a broad claim about artificial intelligence, but a narrow one: in a 2025 study summarized by the Recovery Research Institute, Heinz and colleagues analyzed 62 people with opioid use disorder from one California clinic, using 14,322 daily smartphone observations, and recurrent deep learning models predicted next-day non-prescribed opioid use with AUC as high as 0.97; the same work found signals for treatment dropout emerging 5 to 8 days before dropout occurred.[1]
For an opioid treatment program, the timing matters. A next-day warning could change when outreach happens. A 5- to 8-day dropout signal could change whether a counselor waits for the next appointment or checks in sooner. Prediction becomes clinically meaningful only when it creates a window in which the response can still be low-burden, respectful, and plausible.
What the smartphone model actually measured
The Heinz study used smartphone-based ecological momentary assessments: repeated, real-time or near-real-time check-ins that ask people about recent experiences, rather than relying only on clinic visits or retrospective recall. The repeated structure is the point. A single report of craving, irritability, boredom, or recent use can be noisy; a sequence of reports can show whether someone is moving toward instability.
Recurrent deep learning models are built for that kind of sequence. They do not just look at today’s answer in isolation; they can incorporate patterns across recent observations. For clinical readers, the important distinction is not the architecture itself, but the input-output relationship: daily smartphone check-ins went in, and near-term outcomes such as next-day non-prescribed opioid use or later treatment dropout came out.
The performance varied by feature group, which is more informative than the headline AUC alone. Past-hour substance use data produced the highest reported next-day relapse prediction, with AUC up to 0.97. Social and environmental context alone reached AUC 0.90, and self-regulation difficulties reached AUC 0.92.[1]
| Prediction task or feature group | Reported finding | Clinical reading |
|---|---|---|
| Next-day non-prescribed opioid use using past-hour substance use data | AUC up to 0.97 | Strongest signal, but also closest to the outcome being predicted |
| Next-day non-prescribed opioid use using social and environmental context alone | AUC 0.90 | Suggests risk was not captured only by direct substance-use reporting |
| Next-day non-prescribed opioid use using self-regulation difficulties | AUC 0.92 | Clinically plausible because loss of control often precedes use |
| Treatment dropout prediction | Risk signals appeared 5 to 8 days in advance | Potentially useful lead time for outreach before disengagement |
The past-hour substance use result deserves a careful reading. It is impressive, but it is not the same as predicting relapse from vague digital exhaust. A model that knows recent substance use is working with information that is already clinically close to the outcome. That does not make it unhelpful. It may still identify a short-term trajectory that a weekly visit would miss. But it should not be interpreted as proof that passive smartphone data, by itself, can reliably detect impending opioid use.
The social and environmental context and self-regulation findings are arguably more interesting for care planning. They imply that daily instability around setting, relationships, impulse control, or coping may carry a near-term signal even when the model is not relying only on recent use. That is closer to the clinical intuition many addiction teams already use: relapse risk often appears first as a change in context, routine, affect, or self-management.
Dropout prediction may be the more actionable warning
Relapse prediction gets attention because the outcome is urgent. Dropout prediction may be easier to translate into a safe workflow. In the Heinz study, dropout risk appeared 5 to 8 days ahead through features that included mood states such as boredom and exhaustion, cravings, and self-regulation lapses.[1]
That kind of warning does not require a clinic to make a high-stakes determination that a patient is about to use opioids. It can justify a lower-intensity response: a supportive message, appointment troubleshooting, medication access check, transportation discussion, or brief counselor outreach. The ethical and operational burden is different when the alert says “this patient may disengage soon” rather than “this patient will relapse tomorrow.”
The lead time also fits the reality of addiction care staffing. Same-day rescue is difficult. A week of warning is still narrow, but it gives a team more options than discovering disengagement after a no-show, a gap in medication, or a return to non-prescribed opioid use.
AUC is not a deployment plan
An AUC of 0.97 should make clinicians pay attention. It should not make them skip the methods section. AUC measures how well a model ranks cases across thresholds; it does not tell a clinic which threshold to use, how many alerts staff will receive, how many will be false positives, or what intervention should follow.
The sample is the first constraint. Heinz and colleagues’ result came from 62 participants at a single California clinic.[1] That is a reasonable early study design for detecting signal in intensive longitudinal data. It is not enough to assume the same performance in rural programs, correctional-transition settings, office-based buprenorphine practices, methadone clinics with different visit patterns, or populations with different smartphone access and reporting behavior.
The observation count can also be misleading if read casually. Fourteen thousand three hundred twenty-two daily smartphone observations sounds large, and for modeling repeated within-person states it is valuable.[1] But those observations are clustered within 62 people. The model is learning from many time points, not from thousands of independent patients.
There is also a workflow problem hiding behind every predictive score. If an alert fires, who receives it? During what hours? Is the response documented? Does the patient know that daily answers can trigger outreach? What happens when the patient denies risk, ignores the message, or is already frustrated by surveillance? These are not objections to prediction; they are the conditions that determine whether prediction becomes care or simply a new liability.
Related evidence points in the same direction, but not the same workflow
A useful comparison comes from Curtis and colleagues, who used a BERT language model on pre-treatment social media language to predict 90-day treatment dropout. That model reached AUC 0.81 and outperformed standard psychometric assessments, with the comparison reported as statistically significant at p<0.001.[1]
That finding supports a broader point: clinically relevant dropout risk may be detectable in patient-generated language and behavior before it becomes visible in attendance records. But it is a different kind of tool. Pre-treatment social media language is not the same input as daily smartphone check-ins, and 90-day dropout is not the same outcome as next-day non-prescribed opioid use. The comparison strengthens the case that AI can identify disengagement signals; it does not validate any one relapse model for routine clinical use.
The field still has a reproducibility problem
The broader machine learning literature in substance use disorder treatment is promising but uneven. In a 2024 systematic review, de Mattos and colleagues examined 28 machine learning studies on SUD treatment outcomes and found promising accuracy for adherence and relapse prediction, while also identifying major limitations: most studies used US samples, external validation was absent, code transparency was limited, and reproducibility remained weak.[2]
The review’s sample geography matters. Eighty-two percent of the reviewed machine learning studies used US samples, which limits confidence that models will generalize across health systems, treatment cultures, drug markets, and patient populations.[2] That is not a complete equity analysis, but it is enough to keep performance claims from being treated as portable clinical truth.
External validation is the dividing line between an interesting model and a candidate clinical tool. A single-site model can learn real patterns and still fail elsewhere because intake practices, medication access, visit cadence, patient mix, or reporting norms differ. Without independent validation, the honest conclusion is that the model worked in the study setting, not that it is ready for broad deployment.
Code transparency is not an academic nicety here. Addiction treatment programs are being asked to trust tools that could influence outreach, documentation, staffing, and patient experience. If outside teams cannot inspect, reproduce, or stress-test a model, then a high AUC remains a study result rather than a dependable clinical asset.
What would make this clinically usable
A clinic considering smartphone-based relapse or dropout prediction would need more than a performance statistic. It would need a clear statement of the input data, outcome definition, population, validation setting, alert threshold, expected alert volume, and recommended response. Those details determine whether the tool supports care or creates noise.
- Input data: whether the model uses active daily check-ins, recent substance use reports, mood, craving, self-regulation items, social context, passive sensor data, or some combination.
- Outcome definition: whether “relapse” means any next-day non-prescribed opioid use, confirmed toxicology, self-report, treatment interruption, overdose, or another endpoint.
- Validation: whether the model has been tested outside the clinic, region, and patient population where it was trained.
- Workflow: who receives alerts, how quickly they act, what intervention is allowed, and how patient consent and expectations are handled.
- Monitoring: whether the clinic tracks false positives, missed events, alert fatigue, differential performance, and unintended consequences after implementation.
The safest early use is probably not autonomous decision support. It is clinician-supervised risk monitoring, especially for dropout risk, where the response can be supportive rather than punitive. A model that prompts a counselor to check whether a patient is overwhelmed, running out of medication, losing transportation, or drifting from care has a different risk profile than a model used to intensify surveillance or restrict treatment.
Patient-facing design also matters. Daily ecological momentary assessment creates burden. People may stop responding when they feel better, worse, watched, or simply tired of the phone. Missingness is not just a statistical nuisance; in addiction care it may itself be part of the clinical signal, and it may also reflect ordinary life constraints. A usable system has to distinguish enough of those possibilities to avoid turning nonresponse into automatic suspicion.
Where the evidence stands in Q3 2026
Smartphone-based AI relapse and dropout prediction is one of the more concrete applications of AI in opioid use disorder care because it is tied to a real clinical timing problem. The strongest available evidence shows that daily smartphone assessments can generate high-performing next-day relapse predictions and dropout warnings days in advance in a small, single-site sample.[1]
The appropriate conclusion is neither dismissal nor adoption. As of Q3 2026, this belongs in an evidence-grounded emerging-tool category. The signal is clinically plausible, the lead time is meaningful, and the workflow use case is easy to imagine. But routine standalone clinical adoption would require larger and more diverse samples, external validation, transparent methods, reproducible code, and intervention protocols that specify what clinicians should do when the model says risk is rising.
References
- Artificial intelligence and smartphones: predicting opioid use outcomes. Recovery Research Institute.
- Machine learning in the prediction of treatment outcomes for substance use disorders: A systematic review. International Journal of Mental Health and Addiction, 2024.
Comments
Join the discussion with an anonymous comment.