Synthetic patient-intake conversations¶

generate_intake_conversations writes redistributable intake dialogues for conversational extraction and ambient-scribe evaluation. Each conversation has several speakers, caregiver or interpreter speech, hesitations, self-repairs, later corrections, explicit uncertainty and recording-consent events. Gold labels are attached per turn so a streaming evaluator can score turn t before it reads turn t + 1.

from openmed.training.synthetic.intake_dialogue import (
    build_intake_dialogue_manifest,
    generate_intake_conversations,
)

conversations = generate_intake_conversations(6, seed=7)
for record in conversations[0].iter_stream_records():
    gold = record["gold"]
    print(record["turn_index"], record["speaker_role"], len(gold["phi_spans"]))

manifest = build_intake_dialogue_manifest(conversations)

The generator uses repository-authored templates and fictional value pools. Phone numbers come from the reserved 555-01XX range, e-mail addresses use the reserved example.org domain, and ages are computed against a fixed reference date. It reads no dataset, makes no network call, and returns identical text, offsets and labels for the same seed, conversation index and scenario.

Conversation content¶

Conversation i uses scenarios[i % len(scenarios)], which defaults to all three scenarios:

Scenario	Speakers	Characteristic turns
`patient_only`	patient, clinician	patient self-repairs a birth date, then corrects an earlier phone number
`caregiver`	patient, caregiver, clinician	caregiver gives an uncertain birth date the patient corrects, gives the caregiver's own phone and e-mail, corrects the medication, and reports a symptom the patient denies
`interpreter`	patient, interpreter, clinician	interpreter speaks for the patient in the first person, makes an "interpreter correction", gives the interpreter's own name and ID, and corrects an earlier phone number

Every conversation contains a consent request and response, at least one hesitation, an explicit uncertainty cue, a within-turn self-repair and a later turn that corrects an earlier fact. Other phenomena rotate by conversation index: address answers are interrupted and completed on even indices, consent is declined when index % 5 == 3, and on odd indices the clinician mishears the birth date in a read-back and, unless consent was declined, the patient pauses the recording to share a private number.

Turn records¶

IntakeConversation.iter_stream_records() yields one JSON-compatible record per turn. Inputs sit at the top level and every label sits under gold. In this abridged patient turn, the speaker gives a birth date and then repairs it after "no, sorry"; the other spans and the utterance are elided:

{
  "schema_version": "openmed.training.synthetic_intake_dialogue.v1",
  "record_type": "intake_turn",
  "conversation_id": "synthetic-intake-7-0000",
  "turn_index": 4,
  "speaker_role": "patient",
  "content": "<synthetic utterance>",
  "stream_offset": 269,
  "gold": {
    "act": "identity_answer",
    "consent_state": "granted",
    "capture_permitted": true,
    "phi_spans": [
      {
        "span_id": "t04.s02",
        "start": 63,
        "end": 78,
        "stream_start": 332,
        "stream_end": 347,
        "label": "DATE_OF_BIRTH",
        "kind": "phi",
        "subject_role": "patient",
        "attribute": "date_of_birth",
        "slot": "patient.date_of_birth",
        "assertion": "affirmed",
        "revision": "self_repair",
        "supersedes": "t04.s01",
        "partial": false,
        "needs_adjudication": false,
        "conflicts_with": [],
        "text_sha256": "<SHA-256 of the span text>"
      }
    ],
    "clinical_spans": [],
    "events": [
      {
        "event_id": "t04.e00",
        "type": "self_repair",
        "start": 52,
        "end": 61,
        "stream_start": 321,
        "stream_end": 330,
        "attributes": {"repair": "t04.s02", "reparandum": "t04.s01"}
      }
    ],
    "fact_updates": {
      "patient.date_of_birth": "t04.s02",
      "patient.name": "t04.s00"
    }
  },
  "provenance": {"synthetic": true, "contains_real_phi": false, "seed": 7}
}

The content field follows the one-turn-per-line convention of openmed.multimodal.chatlog_jsonl, so the JSONL chat-log redactor can process these records and leave gold unchanged.

Each span carries:

start and end: half-open offsets into the turn's content.
stream_start and stream_end: offsets into IntakeConversation.stream_text, which is every turn followed by a newline (iter_stream_chunks()). Feeding those chunks in order to an incremental de-identifier such as openmed.core.streaming.StreamingDeidentifier produces spans in the same coordinates.
label: a canonical label. PHI spans use PERSON, DATE_OF_BIRTH, AGE, PHONE, EMAIL, STREET_ADDRESS, LOCATION, ID_NUM and DATE; clinical intake facts use CONDITION and MEDICATION.
speaker_role (on the turn) and subject_role (on the span): who spoke and whose fact it is. A caregiver can state the patient's birth date, and an interpreter's "my phone number" belongs to the patient.
slot: "<subject_role>.<attribute>", for example patient.phone or caregiver.phone.
assertion: affirmed, negated or uncertain, the same strings used by openmed.clinical.context.
revision and supersedes: self_repair, correction or completion, with the identifier of the earlier span this one replaces.
partial: the speaker was cut off before finishing the value.
needs_adjudication and conflicts_with: another speaker stated the same fact differently. The disagreement is flagged for review, never resolved by the generator.
text_sha256: a digest that binds the span to its characters. Gold labels never repeat surface text.

Events (hesitation, uncertainty, self_repair, correction, interruption, consent) are anchored to their cue words. Consent events carry action (request, grant, decline, pause, resume), scope and subject_role.

Causal gold¶

Labels in turn t only refer to turn t or earlier turns. supersedes, conflicts_with and event links always point backwards, and the constructors reject records that point forwards. Applying each turn's fact_updates in order reproduces fact_timeline(), the fact state after every turn. The end-of-conversation views are final_facts(), supersessions() and adjudication_items(), which to_dict() also exports.

consent_state is the state in effect when a turn starts. It is pending until consent is given, then granted, declined or paused; a consent event changes the state from the next turn on. capture_permitted is true only in the granted state. An identifier spoken before consent, after a decline, while recording is paused, or later superseded is still PHI: every PHI span remains a redaction target whatever its assertion, revision or consent state.

gold_openmed_spans(secret=...) returns the PHI gold as OpenMedSpan records in stream coordinates, with keyed hmac-sha256 text hashes instead of text.

Lineage and privacy¶

Each record's provenance states synthetic: true, contains_real_phi: false, restricted_data: false, the source openmed.synthetic.intake_dialogue, the Apache-2.0 license, the seed, the conversation index and the scenario. IntakeConversation.report() and build_intake_dialogue_manifest() contain only counts, schema versions, seeds and SHA-256 digests. Validation errors describe the broken rule without echoing any value.

Limits¶

The templates are short, English-only scripted dialogue acts. They are not transcribed speech, do not model recognition errors, and are not reviewed by clinicians, interpreters or linguists. Interpreted turns render the patient in English only. Symptoms, allergies and medications are fictional software stimuli, not clinical guidance. Using these conversations does not certify de-identification quality, consent compliance or any clinical decision.