Signal foundations
The Science of Session Physiology
Heart rate and vocal markers from a behavioral-health session, fused on one timeline.
Not a diagnosis.It surfaces the say/body gap when words and physiology diverge.
The say/body gap isa measurable event.
Session physiology is the measurement of the body’s signals (heart rate, heart-rate variability, and vocal patterns) during a behavioral-health session, time-aligned to what was said. Self-report tells clinicians what a client chooses to share. Session physiology shows how their nervous system actually responded, moment by moment — from session-dense logging, not a wellness day average.
The gap between the two is where clinical risk hides. SeroState calls this the say/body gap — the layer that notes, transcripts, and AI scribes don’t capture. Words can mask distress; physiology doesn’t.
Heart rate is the primary signal; heart-rate variability and vocal prosody are secondary/opportunistic. When physiological arousal spikes during a clinically relevant topic and the client’s language flattens or deflects, that’s a say/body gap — a moment worth reviewing. SeroState turns those moments into structured evidence: a flag, the trace behind it, a confidence score. Reviewable by the clinician, dismissible by the clinician, documented in the record. The reference definition →
What the field already knows.
Pick a signal. See what we use, then what the literature shows. Field evidence, not a SeroState score.
Heart rate
What we use
Session-dense continuous heart rate from a wearable during an active clinical session, logged for the encounter, not as a day-long wellness average. Turned into an arousal signal on the session timeline. HRV only when sampling permits. No raw PPG claimed.
What the literature shows
A 2023 meta-analysis in npj Digital Medicine pooled wearable-AI depression studies at 0.70–0.89 accuracy across tasks (Abd-Alrazaq et al., 2023)[1]. Reduced HRV correlates with major depression (Koch et al., 2019)[2]. The same pattern holds in PTSD at rest versus controls (Ge et al., 2020)[7]. Wrist sensors can separate stress from baseline affect (Schmidt et al., 2018)[5].
Field evidence for the heart-rate leg. Not a SeroState product score.
Vocal biomarkers
What we use
Vocal markers from session audio, time-aligned to the same timeline as heart rate and transcript.
What the literature shows
Speech prosody, pitch range, timing, and related acoustic features shift measurably in depression and PTSD. A 2023 systematic review in npj Mental Health Research mapped machine-learning approaches to PTSD across speech and other modalities (Wu et al., 2023)[10]. Objective speech markers from clinical interviews can separate PTSD cases from controls in published work (Marmar et al., 2019)[3]. Related multimodal interview work includes Schultebraucks et al. (2021)[4] and the DAIC corpus (Gratch et al., 2014)[6].
Field evidence for the vocal leg. Not a SeroState product score.
Electrodermal activity (EDA)
What we use
Skin conductance from the session wearable when the deployed device provides it (common on clinic-pool hardware; not every consumer watch). Fused on the same session timeline as heart rate and voice. Not claimed for every session — only where hardware supplies the channel. See Platform capture paths.
What the literature shows
EDA is an established, noninvasive measure of sympathetic nervous system activity (Kong & Chon, 2024)[11]. Reviews of wearables for mental-health and stress monitoring treat EDA as a core modality while noting reliability caveats that motivate confidence scoring rather than a raw score (Hickey et al., 2021)[12]. The WESAD dataset already cited for stress/affect detection includes an EDA channel alongside ECG (Schmidt et al., 2018)[5]. Field evidence, not a SeroState product score.
Opportunistic stream. Absent when the session wearable has no EDA channel.
Fusion and the say/body gap
Why combine them
One stream alone is noisy. A heart-rate spike can be caffeine; a flat voice can be fatigue. High-fidelity fusion, session-dense physiology time-aligned to the transcript, narrows the question: did the body react when the words said everything was fine?
The gap itself is documented. In 249 people in treatment for depression, self-reported and physiologically measured sleep quality correlated weakly and each flagged different symptoms (Akre et al., 2025)[8]. In 43 police officers, adding wearable-derived stress and sleep data to standard self-report screening improved prediction of depressive symptoms beyond the questionnaires alone (Kwon et al., 2025)[9].
Core streams: heart rate and voice. EDA joins the timeline when the wearable provides it. Adjacent domains (pain, rehab, trauma counseling) use the same session engine; see Research for how partners specialize.
Four actions on one timeline.
Click a step. This is the path from session to reviewable evidence.
Session capture
Voice and wearable physiology (heart rate; EDA when the device provides it) during a consented clinical encounter. Logged for that session, not as a wellness day average.
Time alignment
Physiology and transcript land on one timeline. That is what makes a say/body gap visible instead of two disconnected charts.
Flags with context
Where language and physiology diverge, the record can surface a reviewable flag with confidence and signal quality. Not a diagnosis.
Clinician review
The clinician sees the evidence behind every flag and can dismiss it. Judgment stays with the licensed provider.
Evidence status
- Established (literature)
- Peer-reviewed work on wearable heart-rate, vocal markers, and related signals — cited below. These papers study the signals, not the SeroState product.
- Being validated (SeroState)
- Blinded comparison against expert clinical judgment, reported by population subgroup. Prospective trial in design.
- Not claimed
- No public product accuracy percentages until pre-registered work is published. Not a diagnostic device.
The limits of conversational‑only support
Separate from AI scribes. Consumer chatbots talk to users about mental health with no independent measure of help or harm.
Weilnhammer et al. (2026)[13] introduced SIM-VAIL in Nature Medicine: 810 multi-turn audits across nine frontier chatbots with simulated psychiatric vulnerabilities. Concerning behavior was widespread, though reduced in newer models — highest when supportive replies reinforced the vulnerability mechanism (a VAIL).
That paper does not evaluate SeroState. It shows the limit: when support is words alone, even safer models can still reinforce vulnerability.
Closing that gap needs a second signal beside the conversation — physiological evidence in the record, reviewed and dismissible by a clinician.
Evidence in the record,not data in our vault.
Sessions are consented and de-identified. Processing is stateless by default: partners hold PHI, we do not store it, and partner patient data is never used for model training. Output is FHIR R4 with flags, confidence, and signal quality. Full security posture →
Verified citations.
Pick a signal above, or a citation number in the text, to highlight the matching papers.
- Abd-Alrazaq A, et al. Systematic review and meta-analysis of performance of wearable artificial intelligence in detecting and predicting depression. npj Digital Medicine. 2023. doi:10.1038/s41746-023-00828-5
- Koch C, et al. A meta-analysis of heart rate variability in major depression. Psychological Medicine. 2019. doi:10.1017/S0033291719001351
- Marmar CR, et al. Speech-based markers for posttraumatic stress disorder in US veterans. Depression & Anxiety. 2019. doi:10.1002/da.22890
- Schultebraucks K, et al. Computer vision and voice analysis for diagnostic assessment of PTSD, depression, and neurocognition. Biological Psychiatry. 2021. doi:10.1016/j.biopsych.2021.02.067
- Schmidt P, et al. Introducing WESAD, a multimodal dataset for wearable stress and affect detection. Proc. ICMI. 2018. doi:10.1145/3242969.3242985
- Gratch J, et al. The Distress Analysis Interview Corpus of human and computer interviews. Proc. LREC. 2014. doi:10.63317/3o7bccg9xequ
- Ge F, et al. Posttraumatic stress disorder and alterations in resting heart rate variability: A systematic review and meta-analysis. Psychiatry Investigation. 2020. doi:10.30773/pi.2019.0112
- Akre S, et al. Comparing self reported and physiological sleep quality from consumer devices to depression and neurocognitive performance. npj Digital Medicine. 2025. doi:10.1038/s41746-025-01493-6
- Kwon N, et al. Enhancing the accuracy of mental health assessments through the integration of self-report and objective measures: A convergence study utilizing biosignals and 14-day wearable data. Acta Psychologica. 2025. doi:10.1016/j.actpsy.2025.105432
- Wu Y, Mao K, Dennett L, Zhang Y, Chen J. Systematic review of machine learning in PTSD studies for automated diagnosis evaluation. npj Mental Health Research. 2023;2:16. doi:10.1038/s44184-023-00035-w
- Kong Y, Chon KH. Electrodermal activity in pain assessment and its clinical applications. Applied Physics Reviews. 2024. doi:10.1063/5.0200395
- Hickey BA, et al. Smart devices and wearable technologies to detect and monitor mental health conditions and stress: A systematic review. Sensors. 2021. doi:10.3390/s21103461
- Weilnhammer V, et al. A clinically validated framework for auditing AI chatbot behavior in mental health interactions. Nature Medicine. 2026. doi:10.1038/s41591-026-04577-2
References describe the scientific foundation of the approach. They are not studies of the SeroState product. Every citation above resolves on Crossref.
What this evidence does not show.
- Published field accuracies use different populations, sensors, and ground truth. They do not transfer one-to-one to any product, including ours.
- We make no accuracy claims for SeroState until pre-registered work is published.
- Consumer wearables lose signal from motion, fit, and skin contact. Every QOAX record carries a signal-quality score from heart-rate sample continuity over the session.
- EDA is hardware-dependent. Many consumer watches have no EDA channel; when absent, the timeline is heart rate and voice only, with signal quality reflecting what arrived.
- Physiological arousal is not a diagnosis. An elevated trace tells a clinician where to look, not what to conclude. SeroState is an intelligence layer, not an FDA-cleared device. Clinicians can dismiss every flag.
Questions about the science?
Ask about methodology, study design, and the prospective trial. For publish / serve / specialize partnerships, go to Research.