Inference engine

Compare a person with themselves. Then weigh the evidence.

Within-person change across the six RDoC domains with conformal guarantees, condition probabilities from validated instruments with stated priors, and a safety block that always reaches a clinician.

For research and clinician review. Not a diagnosis.

medera · /v1/inference
sample
Baseline · 24 sessions · todaySpeaking ratez −2.99Pause ratioz +3.21Pitch variabilityz −3.02RMSSDz −2.73Sleep efficiencyz −2.73Smile intensityz −1.91−4 sd−2 sd0+2 sd+4 sdRDoC domainsSensorimotorscore4.07p 0.04 · reviewPositive valencescore3.50p 0.04 · reviewArousal and regulatoryscore3.82p 0.04 · reviewSocial processesscore2.65p 0.04 · reviewgaze, pronounsConformal rank · SensorimotorToday ranks first of 25 sessionsp = 1/25 = 0.04, no distribution assumedTierstablewatchreviewMajor depression28.25%post-testprior 11.92%PHQ-9 13 (band 10–14)likelihood ratio 2.9091Safety: routine, no risk item
Explained

The engine, step by step.

1:10 with on-screen captions, on the published 24-session sample.

Read the transcript
  1. 01 · Within-person

    Compare a person with themselves. Then weigh the evidence.

  2. 02 · Baseline

    Every biomarker is compared with the same person's earlier sessions, never with other people.

    Here, 24 earlier sessions. Each gray dot is one of them.

  3. 03 · Deviation

    Today's values sit two to three standard deviations from this person's own baseline.

    Each is read in the direction published studies associate with burden.

  4. 04 · Domains

    Deviations pool into RDoC domains, weighted by the strength of the evidence.

    Correlated measures are discounted, so one signal is not counted twice.

  5. 05 · The guarantee

    Today's score is ranked among every session's: a conformal p-value.

    First of 25 gives p = 0.04 with no distributional assumption, and false discoveries are controlled.

  6. 06 · Tier

    Watch needs the conformal test. Review also needs the normal model at 0.01. Today: review.

  7. 07 · Probability

    A PHQ-9 of 13 multiplies the odds by 2.9091, the likelihood ratio from the DEPRESSD data.

    From a primary-care prior of 11.92% to 28.25%. A probability, not a diagnosis.

  8. 08 · Safety

    A positive item 9 or C-SSRS risk escalates the tier, whatever the probabilities say.

  9. 09 · One request

    One request: today's biomarkers, the baseline you hold, and the setting.

    Baselines travel with the request and are never stored.

  10. 10 · Recap

    Within-person change with stated error rates, evidence with stated priors, and a clinician decides.

How it reasons

Six steps, each one stated.

Nothing is learned from your data and nothing is hidden in a model: the engine is arithmetic you can check.

  1. 01Baseline

    Every biomarker is compared with the same person's earlier sessions, never with other people.

  2. 02Direction

    Each deviation is read in the direction published studies associate with burden, and pooled per RDoC domain.

  3. 03Correlation

    Related measures are down-weighted with an effective number of biomarkers, so one signal is not counted twice.

  4. 04Guarantee

    With enough sessions, the session is ranked among its own baseline: a conformal p-value valid without distributional assumptions.

  5. 05Probability

    Validated instruments move a setting prior by published likelihood ratios to a post-test probability.

  6. 06Safety

    Risk items route to a clinician, whatever the probabilities say.

Within-person change

The arithmetic, in full.

Default parameters are shown; each can be set per request within safe limits.

Deviation

z = (x − μ) / (σ · √(1 + 1/n))

Each biomarker against the same person's baseline mean μ and standard deviation σ over n sessions, widened for the uncertainty of a new observation.

Direction

c = +z or −z

Read in the direction published studies associate with burden: slower speech and lower RMSSD count as concern, faster or higher do not.

Domain score

S = Σ wᵢcᵢ / √(wᵀRw)

A weighted sum per RDoC domain. Weights follow the strength of the evidence; R holds the assumed correlation, 0.3 within a modality and 0.1 across.

Effective count

n_eff = (Σ wᵢ)² / (wᵀRw)

How many independent measures the domain really holds, so correlated biomarkers are not counted twice.

Conformal p-value

p = #{sessions with score ≥ today's} / (n + 1)

Every session, today's included, is scored against all the others. Valid under exchangeability, with no distributional assumption. Needs 19 or more sessions at α = 0.05.

Across domains

q = Benjamini–Hochberg(p)

Controls the false discovery rate across the domains that were tested.

stable

No domain beyond the baseline at α watch (0.05 by default).

watch

A domain is beyond the baseline at α watch, conformally when there are enough sessions.

review

Watch, and the domain is also beyond α review (0.01 by default) under the normal prediction model.

Safety overrides the tier: a positive PHQ-9 item 9 or any C-SSRS risk makes the tier at least watch, and an urgent safety priority makes it review.

RDoC domains

Read by domain, not by diagnosis.

Change is reported in the Research Domain Criteria of the National Institute of Mental Health: systems of function that cut across diagnoses. Only biomarkers with replicated or meta-analytic evidence carry full weight.

Negative valence systems

Responses to aversive situations: fear, anxiety, loss. Read from negative-affect language and brow lowering.

Positive valence systems

Responses to rewarding situations. Read from smiling, facial expressivity and vocal prosody (pitch and loudness variability).

Cognitive systems

Attention, working memory and processing speed. Read from response latency.

Social processes

Affiliation, social communication and self-perception. Read from gaze, non-affiliative expressions and self-referential language.

Arousal and regulatory systems

Arousal, circadian rhythms and sleep-wake regulation. Read from heart rate, vagal HRV, sleep and rest-activity rhythm.

Sensorimotor systems

Motor activity and psychomotor speed. Read from speech rate, pauses, head motion and physical activity.

½ marks half weight: a single-study association. Mixed or null findings never feed a domain.

Condition evidence

Post-test probability, with its working shown.

Post-test odds = pre-test odds × likelihood ratio. The pre-test probability comes from the setting you name, or from you; the likelihood ratio comes from the largest verified accuracy study for the instrument. With no prior that matches the setting, the API returns prior_required rather than a number.

Major depression

PHQ-9

Semistructured diagnostic interviews (SCID, SCAN, DISH); 29 studies, 6,725 participants, 924 with major depression.

ResultLikelihood ratio
Total 0–40.04
Total 5–90.33
Total 10–142.91
Total 15–2714.0

Setting priors

  • Default13.7%
  • Primary care11.9%
  • Specialty outpatient15.1%

Levis B, Benedetti A, Thombs BD; DEPRESSD Collaboration. Interval likelihood ratios are derived from pooled sensitivity and specificity at cutoffs 5, 10 and 15; they are point estimates without confidence intervals.

Generalized anxiety disorder

GAD-7

Structured or semi-structured clinical interviews; 35 studies, 15,274 participants (Cochrane 2025).

ResultLikelihood ratio
Positive: score 10 or more7.11
Negative0.40

Setting priors

  • Default12.0%
  • Primary care7.6%
  • General population8.0%

Aktürk Z, Hapfelmeier A, Fomenko A, et al. Pooled estimates are heterogeneous across settings; the review calls them rough averages.

Post-traumatic stress disorder

PCL-5

CAPS-5 in veterans receiving VA care (n = 104).

ResultLikelihood ratio
Positive: total of 33 or more2.84
Negative0.17

Setting priors

  • Veterans58.7%

Bovin MJ, Marx BP, Weathers FW, et al; National Center for PTSD, US Department of Veterans Affairs. Validated in a high-prevalence veteran sample; in other settings supply a pre-test probability.

Bipolar spectrum disorder

MDQ

Validated structured psychiatric interviews for DSM-IV (Carvalho 2015 meta-analysis).

ResultLikelihood ratio
Positive: positive screen3.14
Negative0.43

Setting priors

  • Default30.1%
  • Mental health30.1%

Carvalho AF, Takwoingi Y, Sales PMG, et al; Hirschfeld RMA, Williams JBW, Spitzer RL, et al. Sensitivity depends strongly on setting and falls when people with known bipolar disorder are excluded.

Your model, combined honestly

Add a biomarker model without double counting.

A deployed multimodal model can contribute a log-likelihood ratio per condition. The engine adds it to the instrument evidence and deflates the sum for their assumed correlation, so the same signal never counts twice.

  • Connected per deployment through a private endpoint; the response reports applied, unavailable or not_configured.
  • Combined evidence is deflated with an effective count at an assumed correlation of 0.5 between sources for the same condition.
  • Without a model, probabilities use validated instruments only, and the response says so.
  • A model can never lower a safety priority.
Limits

What it will not claim.

Baselines travel with the request and are never stored. You hold the history.

  • A biomarker needs at least 3 baseline sessions.
  • Conformal p-values need 19 or more sessions at alpha 0.05; with fewer, p-values are parametric and labeled.
  • No probability is returned without a prior that matches the setting.
POST /v1/inference

One request. Try it now.

The sample holds 24 synthetic weekly sessions, today's biomarkers, a PHQ-9 and the primary-care setting.

# The published sample request (synthetic, no real person's data)
curl -s https://www.medera.info/api/v1/samples/inference -o inference.json

curl -s https://www.medera.info/api/v1/inference \
  -H "Authorization: Bearer mk_sandbox_medera_public" \
  -H "Content-Type: application/json" \
  --data @inference.json
POST/v1/inference
Sample request

Sends the synthetic sample to the live sandbox with the public key and shows the full response. Nothing personal is sent and nothing is stored.

Response for the sample · 200
{
  "object": "inference",
  "api_version": "v1",
  "mode": "sandbox",
  "engine": {
    "name": "medera-inference",
    "version": "1.0.0",
    "vocal": "BH-ASR voice layer 1.2.0 (TypeScript port, number-for-number with the Python reference)"
  },
  "intended_use": "Measured biomarkers and within-person change for research and clinician review. Not a diagnosis and not a medical device.",
  "tier": "review",
  "change": {
    "status": "computed",
    "method": "conformal",
    "baseline_sessions": 24,
    "alpha": {
      "watch": 0.05,
      "review": 0.01
    },
    "guarantee": "Watch: under exchangeability of the baseline sessions and this session (no real change), a domain is flagged with probability at most alpha_watch, with no distributional assumption when p-values are conformal; Benjamini–Hochberg across domains controls the false discovery rate. Review: the domain must also exceed alpha_review under the normal prediction model.",
    "notes": [],
    "biomarkers": [
      {
        "key": "speakingRateWpm",
        "name": "Speaking rate",
        "modality": "vocal",
        "value": 128,
        "baseline": {
          "mean": 153.8028,
          "sd": 8.463,
          "n": 24
        },
        "z": -2.987,
        "domain": "sensorimotor",
        "concern": "down",
        "concern_z": 2.987
      },
      {
        "key": "pauseRatio",
        "name": "Pause ratio",
        "modality": "vocal",
        "value": 0.29,
        "baseline": {
          "mean": 0.2029,
          "sd": 0.0266,
          "n": 24
        },
        "z": 3.213,
        "domain": "sensorimotor",
        "concern": "up",
        "concern_z": 3.213
      },
      {
        "key": "meanPauseS",
        "name": "Mean pause",
        "modality": "vocal",
        "value": 0.84,
        "baseline": {
          "mean": 0.6336,
          "sd": 0.0692,
          "n": 24
        },
        "z": 2.92,
        "domain": "sensorimotor",
        "concern": "up",
        "concern_z": 2.92
      },
      {
        "key": "f0SdSemitones",
        "name": "Pitch variability",
        "modality": "vocal",
        "value": 2.2,
        "baseline": {
          "mean": 3.1118,
          "sd": 0.2955,
          "n": 24
        },
        "z": -3.023,
        "domain": "positive_valence",
        "concern": "down",
        "concern_z": 3.023
      },
      {
        "key": "firstPersonSingularPer100",
        "name": "First-person singular",
        "modality": "clinical",
        "value": 9.1,
        "baseline": {
          "mean": 6.6013,
          "sd": 1.3289,
          "n": 24
        },
        "z": 1.842,
        "domain": "social",
        "concern": "up",
        "concern_z": 1.842
      },
      {
        "key": "rmssdMs",
        "name": "RMSSD",
        "modality": "physiological",
        "value": 31,
        "baseline": {
          "mean": 43.8116,
          "sd": 4.5905,
          "n": 24
        },
        "z": -2.734,
        "domain": "arousal_regulatory",
        "concern": "down",
        "concern_z": 2.734
      },
      {
        "key": "meanHrBpm",
        "name": "Mean heart rate",
        "modality": "physiological",
        "value": 74,
        "baseline": {
          "mean": 66.122,
          "sd": 2.6602,
          "n": 24
        },
        "z": 2.902,
        "domain": "arousal_regulatory",
        "concern": "up",
        "concern_z": 2.902
      },
      {
        "key": "sleepEfficiencyPct",
        "name": "Sleep efficiency",
        "modality": "physiological",
        "value": 82,
        "baseline": {
          "mean": 89.1211,
          "sd": 2.5576,
          "n": 24
        },
        "z": -2.728,
        "domain": "arousal_regulatory",
        "concern": "down",
        "concern_z": 2.728
      },
      {
        "key": "smileIntensityMean",
        "name": "Smile intensity",
        "modality": "facial",
        "value": 0.19,
        "baseline": {
          "mean": 0.2787,
          "sd": 0.0454,
          "n": 24
        },
        "z": -1.912,
        "domain": "positive_valence",
        "concern": "down",
        "concern_z": 1.912
      },
      {
        "key": "expressivityMean",
        "name": "Facial expressivity",
        "modality": "facial",
        "value": 0.07,
        "baseline": {
          "mean": 0.0888,
          "sd": 0.0089,
          "n": 24
        },
        "z": -2.069,
        "domain": "positive_valence",
        "concern": "down",
        "concern_z": 2.069
      },
      {
        "key": "gazeAversionFraction",
        "name": "Gaze aversion",
        "modality": "facial",
        "value": 0.27,
        "baseline": {
          "mean": 0.1824,
          "sd": 0.0412,
          "n": 24
        },
        "z": 2.083,
        "domain": "social",
        "concern": "up",
        "concern_z": 2.083
      },
      {
        "key": "headMotionDegPerSec",
        "name": "Head motion",
        "modality": "facial",
        "value": 4.4,
        "baseline": {
          "mean": 6.0537,
          "sd": 1.3778,
          "n": 24
        },
        "z": -1.176,
        "domain": "sensorimotor",
        "concern": "down",
        "concern_z": 1.176
      }
    ],
    "domains": [
      {
        "domain": "sensorimotor",
        "label": "Sensorimotor systems",
        "biomarkers": 4,
        "effective_biomarkers": 2.5,
        "score": 4.07,
        "p_value": 0.04,
        "p_conformal": 0.04,
        "p_parametric": 0,
        "q_value": 0.04,
        "tier": "review",
        "contributors": [
          {
            "key": "pauseRatio",
            "concern_z": 3.213,
            "weight": 1
          },
          {
            "key": "speakingRateWpm",
            "concern_z": 2.987,
            "weight": 1
          },
          {
            "key": "meanPauseS",
            "concern_z": 2.92,
            "weight": 1
          },
          {
            "key": "headMotionDegPerSec",
            "concern_z": 1.176,
            "weight": 1
          }
        ]
      },
      {
        "domain": "arousal_regulatory",
        "label": "Arousal and regulatory systems",
        "biomarkers": 3,
        "effective_biomarkers": 1.88,
        "score": 3.818,
        "p_value": 0.04,
        "p_conformal": 0.04,
        "p_parametric": 0.0001,
        "q_value": 0.04,
        "tier": "review",
        "contributors": [
          {
            "key": "meanHrBpm",
            "concern_z": 2.902,
            "weight": 1
          },
          {
            "key": "rmssdMs",
            "concern_z": 2.734,
            "weight": 1
          },
          {
            "key": "sleepEfficiencyPct",
            "concern_z": 2.728,
            "weight": 1
          }
        ]
      },
      {
        "domain": "positive_valence",
        "label": "Positive valence systems",
        "biomarkers": 3,
        "effective_biomarkers": 2.25,
        "score": 3.502,
        "p_value": 0.04,
        "p_conformal": 0.04,
        "p_parametric": 0.0002,
        "q_value": 0.04,
        "tier": "review",
        "contributors": [
          {
            "key": "f0SdSemitones",
            "concern_z": 3.023,
            "weight": 1
          },
          {
            "key": "expressivityMean",
            "concern_z": 2.069,
            "weight": 1
          },
          {
            "key": "smileIntensityMean",
            "concern_z": 1.912,
            "weight": 1
          }
        ]
      },
      {
        "domain": "social",
        "label": "Social processes",
        "biomarkers": 2,
        "effective_biomarkers": 1.82,
        "score": 2.646,
        "p_value": 0.04,
        "p_conformal": 0.04,
        "p_parametric": 0.0041,
        "q_value": 0.04,
        "tier": "review",
        "contributors": [
          {
            "key": "gazeAversionFraction",
            "concern_z": 2.083,
            "weight": 1
          },
          {
            "key": "firstPersonSingularPer100",
            "concern_z": 1.842,
            "weight": 1
          }
        ]
      }
    ],
    "overall": {
      "score": 19.83,
      "p_value": 0.04,
      "p_conformal": 0.04,
      "p_parametric": 0,
      "tier": "review",
      "biomarkers": 12
    }
  },
  "conditions": [
    {
      "condition": "mdd",
      "label": "Major depression",
      "status": "computed",
      "prior": {
        "value": 0.1192,
        "source": "Levis 2019, primary care: 377 of 3,163 (11.9%)."
      },
      "evidence": [
        {
          "source": "instrument",
          "name": "PHQ-9",
          "observation": "total 13 (band 10–14)",
          "likelihood_ratio": 2.9091,
          "derivation": "(0.88 − 0.56) / (0.96 − 0.85)",
          "imprecise": false
        }
      ],
      "combined_likelihood_ratio": 2.9091,
      "probability": 0.2825,
      "references": [
        "Levis2019"
      ],
      "caveats": [
        "Interval likelihood ratios are derived from pooled sensitivity and specificity at cutoffs 5, 10 and 15; they are point estimates without confidence intervals.",
        "Levis 2019 excluded youth and psychiatric settings; supply your own pre-test probability there."
      ]
    },
    {
      "condition": "gad",
      "label": "Generalized anxiety disorder",
      "status": "no_evidence",
      "prior": {
        "value": 0.0756,
        "source": "Spitzer 2006, primary care: 73 of 965 interviewed (7.6%)."
      },
      "evidence": [],
      "combined_likelihood_ratio": null,
      "probability": null,
      "references": [
        "Akturk2025"
      ],
      "caveats": [
        "Pooled estimates are heterogeneous across settings; the review calls them rough averages.",
        "The GAD-7 is used here as a screen for GAD, not for any anxiety disorder."
      ]
    },
    {
      "condition": "ptsd",
      "label": "Post-traumatic stress disorder",
      "status": "no_evidence",
      "prior": null,
      "evidence": [],
      "combined_likelihood_ratio": null,
      "probability": null,
      "references": [
        "Bovin2016",
        "NCPTSD_PCL5"
      ],
      "caveats": [
        "Validated in a high-prevalence veteran sample; in other settings supply a pre-test probability.",
        "No peer-reviewed diagnostic-accuracy meta-analysis of the PCL-5 was found; cutoffs vary by population."
      ]
    },
    {
      "condition": "bipolar",
      "label": "Bipolar spectrum disorder",
      "status": "no_evidence",
      "prior": null,
      "evidence": [],
      "combined_likelihood_ratio": null,
      "probability": null,
      "references": [
        "Carvalho2015",
        "Hirschfeld2000"
      ],
      "caveats": [
        "Sensitivity depends strongly on setting and falls when people with known bipolar disorder are excluded.",
        "Primary-care and general-population estimates are very imprecise."
      ]
    }
  ],
  "safety": {
    "review_required": false,
    "priority": "routine",
    "reasons": []
  },
  "biomarker_model": {
    "status": "not_configured",
    "detail": "No multimodal condition model is deployed on this endpoint; condition probabilities use validated instruments only."
  }
}
Computed by the engine for the published sample. Each call also carries an id and a created timestamp.
FAQ

The engine, answered.

Still wondering about something?

A real person will answer, usually the same day.

Voice, physiology and expression vary widely between people and setups. Change within one person, measured the same way, is far more informative than a population cutoff.

From validated instruments and the largest verified accuracy studies: the DEPRESSD meta-analysis for the PHQ-9, the 2025 Cochrane review for the GAD-7, Bovin 2016 for the PCL-5 and Carvalho 2015 for the MDQ. Each result shows its likelihood ratio, prior and source.

Yes. A deployed multimodal model can contribute log-likelihood ratios, which the engine combines with the instrument evidence and deflates for correlation.

Start measuring.

Sandbox today. Production under agreement.

Or make a first request with the sandbox key.