Passive smartphone sensing as a between-session clinical instrument
- 82 adults with body dysmorphic disorder (BDD) completed 28 days of EMA reporting on suicidal ideation, BDD-related avoidance, and time spent on BDD concerns. In parallel, their phones passively collected GPS, accelerometer, and demographic data for three months.
- Random forest models predicted same-day clinical state: Pearson r = .74-.75 for time spent on BDD behaviours; r = .70-.73 for suicidal ideation (max and mean); r = .56-.62 for avoidance.
- The most predictive features were step count and demographic context (living situation, education) — not exotic signals, but the basics.
- Trial: ClinicalTrials.gov NCT04254575. Massachusetts General Hospital / Harvard Medical School collaboration with the Fenway Institute digital phenotyping group.
Scales measured in-session are a snapshot of a patient's self-report on one day. Between sessions, the patient's state drifts — sometimes toward crisis. This study does not invent a new scale. It demonstrates that the smartphone already on the patient's hip is enough to detect when BDD is worsening on a given day, without asking the patient a single question.
What the tool does
Participants ran passive sensing on their phones — GPS (where you are, how far you move), accelerometer (activity, step count). A random forest trained on these signals plus a handful of demographic variables matched daily self-reports of BDD severity with r up to .75. That is not a clinical-grade diagnostic, but it is roughly what you would get from a weekly self-report questionnaire — with zero patient burden.
The finding that step count and living situation carry most of the signal matters. It means this is reproducible without proprietary pipelines; any competent digital health team can build it. The signal is behavioural (withdrawal, reduced movement, housing instability), not introspective.
For your practice
This is not a scale you administer. It is a framework for what the next generation of clinical tools will look like: passive, continuous, integrated into the device the patient already uses. The immediate clinical use case is just-in-time intervention — send a push notification when the model flags worsening, route the patient back to a distress-tolerance skill, or prompt the therapist to schedule earlier contact.
For now, treat this as reason to ask a fuller set of behavioural questions at intake: current step count, current living stability, time spent in one location. These free, coarse signals explain more of the variance than most therapists assume. For your digital health committee: the reporting quality across digital phenotyping studies is poor (see the parallel JMIR scoping review — PMID 41877492). Demand external validation before deploying anything like this at a population level.
The clinical scale of the next decade is not a scale. It is the phone already in the patient's pocket.
Single-condition (BDD), single-site, majority-female sample. No causal claim — the model predicts a same-day self-report, not a downstream outcome. No real-world deployment yet.