Recovery leaves a measurable trace.
Addiction is a chronic, relapsing condition, yet it is monitored with some of the sparsest data in medicine. Our work is to turn everyday signal — how a person speaks, sleeps, moves and feels — into an objective, longitudinal record of how recovery is actually going, and to test that work in the open with NHS and academic partners.
Speech timing, prosody, pause structure, psychomotor slowing and facial affect — markers with established links to craving, withdrawal state and cognitive load in substance use disorders.
Autonomic load expressed through heart-rate variability, resting heart rate, sleep architecture and circadian drift. Slow-moving, individual, and difficult to consciously mask.
Validated craving, mood and functioning instruments delivered at a cadence services cannot staff — anchoring the passive signals to measures clinicians already trust.
A thirty-second check-in yields more than a monthly questionnaire.
Patients record a short video in the app. Frames and audio are reduced to numerical features — speech rate, pause distribution, vocal tremor, gaze stability, facial affect — and the underlying recording is never required for the prediction itself.
Those features join passive phone phenotyping and phone-derived autonomic data to form a daily feature vector unique to each patient, which is what the model actually reads.
A phone camera, six steps, and a measurable nervous system.
Step through the pipeline, then scrub along the recording to watch a single beat resolve. The waveform is generated for explanation — it is not a recording of a person.
A fingertip rests on the rear camera and flash. Each frame is a picture of light passing through tissue. On its own it looks like nothing at all.
Camera photoplethysmography is a well-described technique, but signal quality depends on lighting, movement and skin tone. We treat every derived measure as within-person and validate it against reference devices before it informs any clinical view.
Subreal-General is a multimodal foundation model for addiction recovery, pretrained on longitudinal patient journeys: EHR timelines, clinical notes, prescriptions, psychometrics and digital biomarkers.
It is then adapted to specific clinical tasks — relapse, overdose, treatment disengagement and pharmacological response — rather than trained separately for each from scratch.
- Longitudinal, not cross-sectionalEach patient is their own baseline. Deviation from a personal norm carries more information than any population cut-off.
- Multimodal fusionNo single channel is sufficient. Signal that is ambiguous alone becomes informative when speech, routine, autonomic state and self-report move together.
- Pre-specified thresholdsEscalation rules are agreed with the treating service before deployment, not tuned after the fact.
- Clinician in the loopThe system surfaces risk and context. Every clinical decision remains with the treating team.
Every claim we make has a sensor, a reason and a status.
This working framework maps each proposed capability from sensing method to scientific rationale and current validation status. It is presented as a research roadmap, with current evidence in the field.
Physiological stress can be tracked daily without a wearable.
Thirty seconds of fingertip video at 30 fps, reduced on-device to inter-beat intervals and heart-rate variability indices.
Beat-to-beat variation indexes parasympathetic withdrawal, which is repeatedly associated with craving states, withdrawal and stress reactivity in substance use disorders.
Camera PPG is validated against reference ECG in controlled settings. Our own work is establishing within-person reliability in everyday, unsupervised use.
- PaperCHEATA · Nottingham University Hospitals NHS TrustJMIR · 2025 · doi.org/10.2196/87186
- ProtocolSheffield Hallam UniversityOSF protocol · doi.org/10.17605/OSF.IO/9WXGV
- QualitativeDigital health from the patients' perspective: a thematic analysis of semi-structured interviewsSheffield Hallam University
- PolicyInnovation pathways in addiction: funding, R&D and commercialisation — a global study of longitudinal pathways of innovative companies in addictionUniversity of St Andrews
Subreal-CARE is investigational and not yet CE-marked or UKCA-marked. It does not diagnose, does not replace clinical judgement, and is not a crisis service. Predictive performance is not yet clinically established; future validation will require appropriately approved studies with academic and NHS partners.