On a Monday morning in early spring, Tomasz Kowalczyk settled into his office chair at a Warsaw addiction clinic, steam rising from his third coffee. Before his first patient arrived, he opened his laptop to a ninety-second briefing he had not written. Marek, thirty-eight, four months sober from alcohol, had been flagged. His heart rate variability had dropped twenty-two percent since Wednesday. He was waking nearly five times a night—more than double his baseline. His activity had cratered. Most telling: after six weeks of silence, he had reported three separate cravings over the weekend.
When Marek walked in, Tomasz skipped the ritual opener. Instead of "How was your week?" he said: "I'm seeing some patterns that concern me. Your sleep's been disrupted since Wednesday. You reported cravings Thursday and Friday for the first time in weeks. What's been happening?"
Marek's face shifted. "I didn't realize it was that obvious. I've been telling myself I'm fine, but…" He paused. "Work has been insane. And yeah, I've been thinking about drinking again. How did you know?"
Tomasz didn't know—not through intuition or clinical sixth sense. A system had noticed what Marek himself had not yet named, and had interrupted Tomasz's morning with a pattern worth exploring. That interrupt—not a diagnosis, not a directive, just a prompt to pay attention—is what continuous monitoring in addiction treatment can do when it works.
The case for such monitoring begins with a failure rate so consistent it reads almost like natural law. Across multiple longitudinal cohorts, a majority of patients with alcohol use disorder relapse within their first year; many studies report rates in the sixty-to-eighty-five percent range (Moos & Moos, 2006; Miller et al., 2001). For opioid use disorder without ongoing medication-assisted treatment, figures approach ninety percent within twelve months (NIH HEAL Initiative data, 2023). These numbers have remained stubbornly stable for decades.
Something in the model needs examination. A weekly fifty-minute session accounts for roughly half of one percent of a patient's week. The remaining one hundred sixty-seven hours—where recovery either solidifies or fragments—is also where relapse typically begins, often invisibly, days before the patient recognizes what is happening.
This is addiction's peculiar architecture. Unlike depression, which tends to descend gradually with symptoms patients can often articulate, relapse in substance use disorders unfolds rapidly. Multiple recent studies suggest a window of approximately three to seven days between detectable physiological dysregulation and conscious craving. Declining heart rate variability and increasing sleep fragmentation often reflect mounting stress before the patient can name it (Mahoney et al., 2023, Drug and Alcohol Dependence; Gustafson et al., 2014). Work from Rajita Sinha's laboratory at Yale has demonstrated that stress reactivity predicts craving in both laboratory and real-world settings, with physiological markers preceding subjective awareness (Sinha, 2008; Sinha et al., 2011).
The question continuous monitoring attempts to answer is not "How do we replace the therapist?" but "How do we help the therapist notice what is happening in those one hundred sixty-seven hours?" In this frame, measurement functions as an interrupt, not an authority: it suggests where to look, but the therapist and patient decide what it means and what to do.
Consider what the research actually shows, stripped of vendor enthusiasm. A 2023 study found that wearable sensors combined with machine learning could predict relapse within fourteen days with seventy-two percent sensitivity and eighty-one percent specificity (Carreiro et al., Drug and Alcohol Dependence). That sounds impressive until you flip the numbers: twenty-eight percent of relapses were still missed. A recent smartphone-based study combining ecological momentary assessment with machine learning found that recent substance use, stress levels, and environmental cues were the strongest predictors of next-day opioid craving—but accuracy varied considerably across individuals (Epstein et al., 2024).
What these studies suggest is not that algorithms prevent relapse, but that they can shift the timing of clinical attention. When heart rate variability, sleep quality, activity levels, and mood ratings all decline simultaneously from a patient's personal baseline—not population norms, but their own established pattern—the probability of imminent crisis rises. The therapist who receives this information can reach out mid-week, adjust contact intensity, have a different kind of conversation.
This is modest. It is also, potentially, lifesaving. The difference between catching deterioration at day three versus day ten can be the difference between a difficult conversation and an emergency department visit.
But measurement's value depends entirely on who controls it and how it is used. Here the history of Central Europe casts a particular shadow. In Poland, where bureaucratic surveillance was a tool of Soviet-era control, any system that tracks and reports carries echoes that younger clinicians may not feel viscerally but that older patients—and their families—remember in their bones. In Germany, post-war sensitivity to institutional misuse of medical data runs deep. When a Berlin therapist introduces continuous monitoring to a patient whose grandparents lived under two totalitarian regimes, they are not merely explaining technology. They are negotiating with historical trauma.
For that reason, continuous monitoring in these contexts must be voluntary, transparent, and reversible: patients should know exactly what is collected, where it is stored, who can see it, and how to stop or delete it. Data sovereignty matters—servers within EU jurisdiction, GDPR compliance, patient-controlled sharing permissions, audit logs accessible to both parties. Without these guarantees, the historical anxieties are not only understandable; they are justified.
What we oppose, explicitly: Outcome and biometric data must never be used to rank therapists or produce performance league tables. Data must not be used by payers to deny care based on "insufficient progress." Continuous monitoring must not be mandated where it conflicts with clinical judgment or harms the therapeutic relationship. Algorithms do not override clinician responsibility; final decisions rest with the treating professional. If measurement becomes coercive, we oppose it—even when it threatens platforms we might otherwise support.
Even when appropriately used, continuous monitoring fails in ways deserving honest acknowledgment. Technology access creates exclusions: wearables require smartphones, stable internet, digital literacy that elderly patients, rural populations, and refugees often lack. The Syrian mother in a German resettlement program may not find a fitness-tracker-based stability index relevant to her circumstances. Most validation studies come from Anglo-American samples whose generalizability to Polish farmers or German refugees remains uncertain.
There is a subtler failure too, one that matters in traditions where the unconscious matters. A French psychoanalyst working in the Lacanian tradition might reasonably argue that reducing distress to questionnaire items fundamentally misunderstands what therapy addresses. The classical psychodynamic position—that deeper structural change resists quantification, that symptoms are communications rather than targets—is not conservative resistance to innovation. It is a coherent intellectual position deserving engagement rather than dismissal.
When not to measure: In acute crisis, when the priority is immediate containment and relational presence. In early sessions where trust is fragile and patients experience measurement as scrutiny. In depth-oriented work where symptom scores would derail the process. These are not failures of measurement; they are appropriate clinical priorities.
Tomasz's session with Marek did not end in triumph. There was no dramatic breakthrough, no Hollywood moment where data saved the day. Instead, there was a difficult conversation about work stress and isolation, an adjustment to the treatment plan, an agreement to check in mid-week. Marek left uncertain whether he would make it through the month.
But here is what happened: a session that would have started with "Fine" instead started with specificity. A pattern that might have gone unnoticed for another week was named. A therapist arrived prepared. And a patient, surprised that his struggle had been seen before he could articulate it, felt perhaps slightly less alone in the project of staying sober.
Our research agenda, honestly stated, is not to prove that measurement "works" in some global sense, but to discover where it helps, where it harms, and where it simply does not matter. For this journal, that means a standing commitment: we will publish negative findings; we will invite and platform critics of measurement-based care; we will advocate retiring metrics that mislead, even popular ones; we will prioritize studies comparing low-burden monitoring with fully digital systems, especially in under-resourced settings; we will be explicit about conflicts of interest when presenting data from commercial platforms.
That is the modest promise of continuous monitoring in addiction treatment. Not revolution, not algorithmic replacement of clinical wisdom. Just the interrupt—the tap on the shoulder that says, look here, something may be happening. What the therapist does with that interrupt remains entirely theirs to decide. The data flags. You interpret. You act. Always.
For patients whose recovery depends on catching fragility in that narrow window before conscious craving emerges, those interrupts may matter. For clinicians navigating the impossible task of staying present to dozens of patients between sessions, they may offer a way to notice what would otherwise slip past.
If that sounds useful, perhaps it is worth trying—carefully, with appropriate skepticism, with the ethical lines firmly drawn. If not, that is reasonable too. The numbers are weak narrators. But they are, sometimes, decent interrupters.