Can a Fitbit predict craving or substance use for one person?
Not reliably in this exploratory study. Personalised models using heart rate and step count from 16 participants produced modest discrimination across 47 person–substance–task scenarios, with mean AUROC values of 0.569 to 0.617.
The short version
- What happened
- Sixteen participants generated 47 person–substance–task prediction scenarios: 25 for substance use and 22 for craving across six substances.
- Who it affects
- People seeking support should not be told that a consumer wristband can reliably know when they are craving or using a substance on the basis of this evidence.
- What remains unproven
- Only 16 participants contributed data, despite 47 person–substance–task scenarios.
Keep in mind
Only 16 participants contributed data, despite 47 person–substance–task scenarios.
Heart rate and step count are affected by many causes unrelated to substance use or craving.
Self-reported ecological momentary assessments may contain timing and reporting error.
Average AUROC can hide large variation between people, substances and tasks.
No clinical intervention, relapse-prevention benefit or patient outcome was tested.
Source analysis · 1 linked source. Primary analysed source date: 14 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: personalisation did not make the signal dependable
Researchers tested whether machine-learning models could recognise moments of substance use or craving from Fitbit heart rate and step count in everyday life. Even when models were tailored to individuals, average discrimination remained modest: mean area under the receiver operating characteristic curve ranged from 0.569 to 0.617 across the approaches and tasks.
An AUROC of 0.5 represents chance ordering and 1.0 perfect discrimination. Values around 0.57 to 0.62 may contain some statistical signal, but they are far from a safe basis for deciding that a particular person is craving or using a substance. The result is therefore a useful negative finding: a consumer wristband's limited streams did not become a reliable detector simply because the model was personalised.
The study does not show that wearables have no role in substance-use care. It shows that heart rate and step count, used largely in isolation, are not specific enough for the task tested. Stress, exercise, sleep, illness, caffeine and many ordinary experiences can change the same physiological measures.[1]
Sixteen people produced 47 prediction scenarios
The exploratory study included 16 participants and constructed 47 participant–substance–task scenarios spanning six substances. Twenty-five scenarios concerned substance use and 22 concerned craving. Each scenario represented a particular person, substance and prediction target rather than an independent new participant.
Participants wore Fitbits in free-living conditions and supplied ecological momentary assessments — short reports recorded during daily life — that served as the target labels. For each scenario, the team compared global supervised learning, global self-supervised learning and personalised self-supervised approaches.
This structure respects the fact that physiological responses and patterns of use differ between people. It also means the headline count of 47 must not be mistaken for 47 independent patients. Repeated scenarios from a small participant pool can help explore feasibility, but they provide limited evidence about how a model would perform across a diverse population or in a new clinic.[1]
Heart rate reflects arousal, not one cause
The biological rationale is understandable. Several substances can acutely alter heart rate, and craving may involve autonomic arousal. Conditioning heart-rate changes on activity level through step count could, in principle, help distinguish drug-related changes from exercise.
In practice, the same signals have many competing explanations. A faster pulse might reflect anxiety, climbing stairs, pain, infection, poor sleep or an emotionally charged conversation. Step counts cannot fully describe exertion, posture, temperature or context. A model can learn an individual's recurring pattern and still confuse causes that produce similar sensor traces.
The researchers examined whether differences in data coverage and adherence to the collection protocol explained poor performance and found that they did not account for the gaps. That makes the negative result more informative: collecting more of the same limited signals may not solve a problem of physiological specificity.[1]
False alarms and missed moments would affect real people
A digital intervention might use a prediction to offer support at a high-risk moment. If the detector fires during exercise or ordinary stress, repeated false alarms could feel intrusive, stigmatising or exhausting. If it misses a genuine high-risk moment, a user may wrongly believe the system is providing protection that it cannot deliver.
Those risks make discrimination only one part of evaluation. Developers need positive predictive value at the prevalence seen in the intended population, calibration, false alarms per day and performance across substances and individuals. They also need to know whether people want an intervention at that moment and whether it improves outcomes rather than merely generating notifications.
Privacy is unusually sensitive here. Heart-rate streams, location-adjacent routines and self-reports about substance use can expose health and behavioural information. Any future service would need clear consent, minimal collection, secure storage and a rule against using uncertain predictions for punishment, insurance, employment or policing.[1]
A better target may be a well-characterised antecedent
The authors suggest digital-intervention research may gain more by predicting better-characterised antecedents such as stress than by trying to infer use or craving directly. That is a narrower and potentially more actionable goal because wearables have a clearer relationship with some forms of arousal, even though stress detection also remains imperfect.
A support tool could ask rather than assume: a sensor change might prompt a private check-in that lets the person describe stress, craving or another cause. User confirmation can reduce the harm of an opaque label and create a more useful intervention point. The system should still be evaluated for burden and benefit.
Researchers could also test richer signals, including sleep, electrodermal activity, context supplied voluntarily and patterns over longer windows. More data are not automatically better: every additional stream increases privacy and implementation costs, and a more complex model must still prove that it helps people.[1]
Evidence limits and what would change the assessment
The paper is peer reviewed, open access and explicit about limited utility. It compares several modelling strategies and examines data coverage and adherence. Funding came from the US National Science Foundation, the National Institutes of Health, the Hawaiʻi Community Foundation, the Weill Neurohub and AWS cloud credits; the authors declare no competing interests.
The participant denominator is very small, the six substances may differ substantially, and ecological momentary reports can be delayed, incomplete or inaccurate. The abstract reports averages across scenarios, which can conceal people or substances with better or worse performance. The study evaluates prediction, not diagnosis, treatment, relapse prevention or emergency response.
Confidence would change with larger prospective cohorts, enough participants for substance-specific estimates, external validation and transparent reporting of calibration and daily false alarms. The decisive test is a consented trial showing that an intervention triggered by the model improves a meaningful outcome without unacceptable intrusion or unequal error. The present evidence does not support routine deployment.[1]