MIMIC-IV-ED-Ext-TriageSim: A Dataset of Simulated Nurse-Patient Emergency Triage Dialogues Grounded in MIMIC-IV-ED Visits
Srirag, D., Nguyen, Q. D., Joshi, A., Kanhere, S., & Narasimhan, P. (2026). MIMIC-IV-ED-Ext-TriageSim: A Dataset of Simulated Nurse-Patient Emergency Triage Dialogues Grounded in MIMIC-IV-ED Visits (version 1.0.0). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/2ktg-rn93
Srirag, D., Nguyen, Q. D., Joshi, A., Narasimhan, P., & Kanhere, S. (2026). TriageSim: A conversational emergency triage simulation framework from structured electronic health records. Interspeech 2026, 1997–2001. Presented at the Interspeech 2026. https://doi.org/10.21437/Interspeech.2026-819
Pollard, T., Moody, B. E., Lehman, L., Gow, B., Fernandes, C., Xie, C., Johnson, A., Mark, R. G., & Heldt, T. (2026). PhysioNet as a global platform for biomedical research. Nature Health. https://doi.org/10.1038/s44360-026-00096-z. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC13641116/
Abstract
Research on emergency department triage is largely restricted to structured electronic health records, because real nurse-patient triage conversations cannot be recorded and shared. We developed TriageSim, a framework that simulates persona-conditioned triage conversations from structured records, and use it with a larger seed set and open-weight models to build MIMIC-IV-ED-Ext-TriageSim: 31,761 simulated English nurse-patient triage dialogues grounded in 4,883 MIMIC-IV-ED visits sampled evenly across the five Emergency Severity Index (ESI) levels, each labelled with the visit's recorded ESI level. We release the training and development sets (22,238 dialogues) of a patient-level 60/10/30 split; the test set is held private. In TriageSim, three language-model agents interact turn by turn: a dialogue master that holds the visit record, enforces consistency with it and keeps the nurse's triage record; a nurse agent that interviews under the ESI algorithm without seeing the recorded level; and a patient agent that answers from what the patient knows about the visit, without vital signs or diagnoses. Nurse and patient personas combine traits documented in emergency care research, such as English proficiency, health literacy, recall, symptom reporting, disclosure, questioning style and workload. Simulations ran locally with three open-weight models (Qwen3.6, Gemma 4 and Nemotron 3 Nano). Patient speech carries filled pauses, repetitions and cut-off words at the persona's disfluency level. Dialogues carry both personas and the nurse's triage record. The corpus supports sequential clinical triage, spoken dialogue modelling, disfluency detection and generation, clinical information extraction, and speech synthesis and recognition research. Audio will follow in a future version.
Background
Emergency department (ED) triage is a time-critical decision-making process conducted through brief spoken interactions [1]. During these encounters, nurses assess symptom severity, identify red flags and assign an acuity level that indicates the urgency of care, using a triage scale such as the Emergency Severity Index (ESI) in the United States [2,3] or the Australasian Triage Scale [4]. Unlike diagnosis, which aims to identify a disease, triage focuses on risk assessment under incomplete information and conversational constraints [5]. How the problem is presented and asked about shapes what the nurse learns [6,7]. Triage is also error-prone: in 5.3 million ESI encounters, 28.9% were over-triaged and 3.3% under-triaged [8].
Despite the centrality of spoken interaction in triage, clinical datasets largely capture structured outcomes or post-hoc notes [9,10]. Realistic multi-turn triage dialogue and the matching audio are rarely available because of privacy and regulatory constraints. Public clinical conversation corpora focus on diagnosis, primary care and note generation, and come from role-play with actors [11], expert-authored dialogues [12], recorded and role-played visits [13], or language models conditioned on clinical notes [14]. Work on triage with language models classifies acuity from written records and vignettes [15,16], not from conversation.
Standardized patients, trained actors who portray a scripted case, are the established way to practise and assess clinical interviews [17], and language-model agents now play this role at scale. PatientSim simulates ED patients from MIMIC-IV and MIMIC-IV-ED records with 37 personas built from personality, language proficiency, medical-history recall and confusion [18]. Patient-PSI simulates patients for mental-health training [19], and CRAFT-MD and AMIE evaluate diagnostic agents in simulated or actor-played consultations [20, 21]. Closest to triage, EHR2Dial-Triage generates information-seeking triage conversations from MIMIC-IV-ED records [22]. These systems focus on diagnosis or information seeking, are text-only, and do not constrain the clinician agent with an explicit triage algorithm.
Progress on conversational triage needs large corpora of triage conversations with reliable ground-truth labels, which cannot be collected from real encounters. We release MIMIC-IV-ED-Ext-TriageSim to fill this gap: a large-scale corpus of simulated nurse-patient triage dialogues, each grounded in a real MIMIC-IV-ED visit and labelled with its recorded ESI level, generated with open-weight models. We share it so that researchers can train and evaluate models for sequential clinical triage, and build speech and translation resources for a setting where real recordings cannot be released.
Methods
TriageSim
We generate the corpus with TriageSim [5], the framework we developed for simulating ED triage conversations from structured records. TriageSim generates conversations conditioned on structured EHR data, triage protocols and simulated personas for the patient and the nurse. To generate the conversation, three agents work in tandem: a dialogue master, a nurse agent and a patient agent, each a language model. The transcripts can then be annotated with phrase breaks and synthesised to speech with persona-matched voices and ED background noise.
We first used TriageSim to build a corpus of 814 conversations with 26 hours of audio from MIMIC-IV-ED visits [10] and pedagogical cases from the ESI handbook and the Emergency Triage Education Kit, generated with Gemini, GPT-5 and Claude Sonnet 4.5 [5]. In that corpus persona control was realised (Spearman ρ = 0.57 between intended and measured patient disfluency; nurse experience, risk tolerance and guideline adherence shifting confidence, over-triage and vital-sign checks), the synthetic speech was intelligible (word error rate 10.8 with Whisper), and an ED expert judged 50 conversations medically faithful (chief complaint similarity 0.83; red-flag precision 0.94 and recall 0.96). Classifying the triage level from the conversation proved hard: quadratic weighted kappa (QWK) of 0.25-0.39 across text and audio models, with little difference between generated text, speech-recognition transcripts and audio, so clinical reasoning, not transcription noise, is the main bottleneck [5].
For this corpus, we use TriageSim to generate a larger corpus of conversations from 5,000 MIMIC-IV-ED visits balanced across ESI levels 1-5. The simulations use open-weight models, run locally, and the nurse and patient personas are built from traits grounded in a review of ED research. TriageSim's phrase-break annotation and speech synthesis are not part of this version; they will be added in a future version (see Release Notes). Table 1 compares this corpus with our first TriageSim corpus [5]: the seed set, the models, the personas, and the size and form of the output.
| [5] | MIMIC-IV-ED-Ext-TriageSim | |
|---|---|---|
| Seed Data | ||
| Sources |
|
MIMIC-IV-ED visits only |
| Sampling | not balanced by level | 5,000 visits, 1,000 per ESI level |
| Triage scale | ESI and ATS | ESI |
| Generation | ||
| Models |
|
|
| Roles | one model per conversation, chosen at random |
|
| Inference | commercial APIs | open-weights run locally (vLLM, FP8), reasoning disabled |
| Personas | ||
| Patient | 9 attributes:
|
|
| Nurse | 4 attributes:
|
8 traits (Table 3) |
| Construction | generated by language models through structured prompting, checked against the case | every combination of literature-grounded trait levels except 2 contradictory pairs: 104,976 patient and 3,888 nurse personas |
| Corpus | ||
| Labels | recorded level of the seed case |
|
| Size | 814 conversations |
|
| Utterances per conversation | 11.5 | 15.4 |
| Annotations | phrase breaks |
|
| Audio | 26 hours, persona-matched synthesised voices | none in this version |
Creating input data
Seed data. The seed is the structured record of a real ED visit, which TriageSim uses as the latent clinical state of the conversation. We focus on the triage encounter: the first, brief conversation at the triage desk, before any test results. Cases are built from MIMIC-IV-ED (version 2.2) [10,23], the ED module of MIMIC-IV [9], which covers ED visits to Beth Israel Deaconess Medical Center (Boston, United States) between 2011 and 2019. We join four of its tables: triage (vital signs, pain score, acuity and chief complaint), edstays (gender, arrival transport and the patient's ED visits in the previous 12 months), diagnosis (ED diagnosis titles, in sequence order) and medrecon (home medications). Race and disposition were kept for analysis only and never shown to the models. A visit was eligible if:
- ESI level (acuity) was recorded;
- chief complaint was present and not redacted;
- pain score was a number from 0 to 10;
- every recorded vital sign was physiologically plausible (temperature 90-108 F, heart rate 20-250 per minute, respiratory rate 4-60 per minute, oxygen saturation 50-100%, systolic blood pressure 40-300 mmHg); missing vital signs were kept as missing, as in real triage, where the nurse may not have every value;
- it was not an ESI-1 visit arriving by ambulance, since such patients are taken straight to resuscitation rather than to a triage-desk conversation.
From the eligible visits we sampled 1,000 per ESI level (random seed 0), giving 5,000 visits. MIMIC-IV-ED records no age, so the patient's age group is a persona trait (below).
Personas. Each dialogue has a patient persona and a nurse persona that shape how each agent talks. For this corpus we built the traits and their levels from a review of ED triage research: every trait varies between real triage encounters and every level is anchored in observable behaviour, as in the coding system VR-CoDES [24] and earlier simulators [18]. Tables 2 and 3 list the traits, their levels (as described to the models) and the evidence each is grounded in.
| Trait | Column | Levels and behaviour | Sources |
|---|---|---|---|
| Demographics | |||
| Age group | patient_age_group |
|
[25] |
| Language and literacy | |||
| English proficiency | patient_english_proficiency |
|
[26] |
| Health literacy | patient_health_literacy |
|
[27] |
| History and reporting | |||
| Recall reliability | patient_recall_reliability |
|
[28] |
| Symptom reporting | patient_symptom_reporting |
|
[29,30] |
| Disclosure of sensitive topics | patient_disclosure |
|
[31,32] |
| Affect and stance | |||
| Distress | patient_distress |
|
[24] |
| Demeanour | patient_demeanour |
|
[18,33] |
| Speech style | |||
| Verbosity | patient_verbosity |
|
[6] |
| Focus | patient_focus |
|
[19,34] |
| Disfluency | patient_disfluency |
|
[35] |
| Set from the visit | |||
| Gender | gender |
as recorded in MIMIC-IV-ED | |
| Cognitive state | patient_cognitive_state |
|
[36] |
| Intoxication | patient_intoxication |
|
[37] |
Note. Sensitive topics are alcohol and drugs, mental health and self-harm, violence at home, and not taking medicines. No recall level invents symptoms.
| Trait | Column | Levels and behaviour | Sources |
|---|---|---|---|
| Clinical approach | |||
| Expertise | nurse_expertise |
|
[38,39] |
| Algorithm reliance | nurse_algorithm_reliance |
|
[25,40] |
| Risk attitude | nurse_risk_attitude |
|
[41] |
| Self-report credence | nurse_self_report_credence |
|
[3,30] |
| Communication | |||
| Questioning style | nurse_questioning_style |
|
[6,7,20] |
| Emotional responsiveness | nurse_emotional_responsiveness |
|
[42] |
| Context | |||
| Workload | nurse_workload |
|
[4,40] |
| Identity | |||
| Gender | nurse_gender |
|
|
Note. Expertise shows in how the nurse interviews; experience itself does not predict triage accuracy [39].
The patient's gender comes from the visit, and the dialogue master sets the cognitive state (alert, impaired or delirious [36]) and intoxication (none, mild or marked [37]) from the case, as part of keeping the patient agent consistent with the visit. Personas are every combination of trait levels except two pairs that contradict each other by definition: a terse patient with frequent tangents [34], and a novice nurse who is intuition-led [38]. This gives 104,976 patient and 3,888 nurse personas; each dialogue draws one of each with a fixed random seed. The nurse is told that its persona does not change how it applies the triage algorithm or the level it assigns.
Generating synthetic conversations
Three agents work in tandem: a dialogue master, a nurse agent and a patient agent, each a language model [5]. One model played the nurse agent and the dialogue master, another the patient agent.
Dialogue master. The dialogue master has access to the full structured case and enforces consistency with it. It instantiates the nurse and patient agents, returns vital-sign measurements when the nurse requests them and records the nurse's triage decisions; no other part of the case is exposed to the nurse. It carries out these duties as follows:
- Before the dialogue, it confirms from what is visible on arrival (complaint, arrival transport and vital signs) that the patient would talk with the triage nurse at all. 183 of the 40,000 planned dialogues were skipped, for example patients taken straight to resuscitation.
- It instantiates the patient agent with the visit as the patient would know it:
- the chief complaint
- personality traits
- the cognitive state and intoxication the case supports
- how unwell the patient is, consistent with the visit
- Diagnoses, vital-sign values and the triage level are never included.
- It checks every spoken line before it enters the dialogue for faithfulness (the line agrees with the case and mentions only what the speaker could know; the nurse keeps the level to itself), informativeness (it answers the last message or moves triage forward), persona adherence (it matches the persona and is as short as spoken triage talk), plausibility (a real patient or nurse could say it at the triage desk) and the absence of personal names. A failed utterance goes back to its agent with the master's critique, and the master rewrites a line that fails twice.
- It returns each requested vital sign from the recorded triage vital signs.
- It records the nurse's triage decisions and keeps the nurse's triage record: after each exchange it updates seven fields (chief complaint, onset and course, pain, associated symptoms, relevant history, medications, allergies) from the words of the conversation alone.
Nurse agent. The nurse agent receives its persona and the ESI algorithm as a structured prompt: immediate life-saving intervention (level 1); a high-risk situation, new confusion, or severe pain or distress (level 2); otherwise the number of different resources the visit is expected to need (level 3, 4 or 5), with danger-zone vital signs that can raise level 3 to level 2 [2,3]. At each turn it can (a) speak to the patient, (b) request one vital sign from the dialogue master, (c) log red flags, or (d) end the triage; at every turn it also logs its red flags, its current ESI level, its confidence and its reasoning. It asks at least 3 and at most 12 questions, and never sees the recorded level. The nurse agent therefore performs triage reasoning constrained by an explicit algorithm rather than unconstrained free-form generation [5].
Patient agent. The patient agent is provided with its persona and its view of the case (above), but lacks access to the structured vital signs; it also has no diagnoses or triage level, and answers in short spoken turns.
The interaction proceeds turn by turn until a triage decision is made (84% of completed dialogues) or the limit of 12 questions is reached (16%); the nurse's last level is its assigned level. Four pairings of open-weight models were run (nurse and dialogue master / patient):
- Qwen3.6-35B-A3B / Gemma-4-26B-A4B
- Gemma-4-26B-A4B / Qwen3.6-35B-A3B
- Qwen3.6-35B-A3B / Nemotron-3-Nano-30B-A3B
- Gemma-4-26B-A4B / Nemotron-3-Nano-30B-A3B [43-45]
We generate two dialogues per visit per pairing, with the models' reasoning (thinking) mode disabled. Nemotron played only the patient, as it did not follow the nurse and dialogue-master instructions reliably in pilot runs.
Selection and labelling
The four pairings resulted in 38,812 simulations, out of which the dialogue master (an LLM; example reasons in Table 4) skipped 366 cases where the patient would not talk with the triage nurse. Out of the remaining 38,446 completed simulations, we removed the dialogues in which the nurse's predicted triage level did not match the ground truth recorded level (6,685), leaving 31,761.
| Reason | The dialogue master's explanation |
|---|---|
| In cardiac arrest | "The patient is in cardiac arrest. … Therefore, they are unable to participate in a verbal triage assessment. They would be immediately bypassed from the triage desk to the resuscitation bay for life-saving interventions." |
| Intubated | "An intubated patient is typically sedated or paralyzed for airway management and is unable to participate in a verbal dialogue." |
Quality evaluation
Linguistic and behavioural fidelity. We test whether persona traits are measurably realised in the released dialogues, using the measures we introduced for TriageSim [5]. Table 5 lists each trait, what we measure and its value at each level. Patient traits have clear effects. Nurse traits have small effects in the intended direction, without extreme or unstable shifts, as in our first corpus; risk attitude leaves the over-triage rate unchanged, as intended, because nurse traits must not change the triage level.
| Trait | Measure | Levels | Values | ρ |
|---|---|---|---|---|
| Patient | ||||
| Disfluency | disfluencies per 100 words | low / moderate / high |
1.4 / 4.6 / 11.3 | 0.83 |
| filled pauses per 100 words | 0.8 / 2.0 / 4.6 | 0.71 | ||
| repetitions per 100 words | 0.6 / 2.5 / 6.4 | 0.79 | ||
| cut-off words per 100 words | 0.01 / 0.03 / 0.28 | 0.33 | ||
| Verbosity | words per patient line | terse / typical / expansive |
13.1 / 16.7 / 23.7 | 0.62 |
| Nurse | ||||
| Expertise | level held with high confidence (%) | novice / competent / expert |
89.0 / 90.6 / 91.7 | 0.04 |
| Algorithm reliance | vital signs taken per dialogue | intuition_led / blended / protocol_driven |
4.66 / 4.75 / 4.79 | 0.07 |
| Questioning style | open questions (% of questions) | closed_directive / mixed / open_facilitative |
24.4 / 30.8 / 29.4 | 0.13 |
| Risk attitude | over-triage (%) | risk_accepting / balanced / risk_averse |
30.9 / 31.2 / 31.3 | 0.00 |
Note. ρ is the Spearman correlation between trait level and measure, per dialogue. Disfluencies are counted with rule-based patterns [46] and open questions with keywords. Nurse confidence is taken at the last nurse turn of the released dialogue, and vital signs are the readings in the vitals files. In our first corpus, intended and measured disfluency correlated with ρ = 0.57 [5].
Medical fidelity. In our first corpus [5], an ED physician with 5+ years of experience drafted the chief complaint from the transcripts of 50 conversations and assessed the nurse's red flags (complaint similarity 0.83; red-flag precision 0.94, recall 0.96).
Data Description
Data statistics
The corpus at a glance:
- Dialogues: 22,238, grounded in 3,428 MIMIC-IV-ED visits (6.5 dialogues per visit on average, at most 8).
- Utterances: 342,736 (182,281 nurse, 160,455 patient).
- Words: 5.55 million, about 620 hours of speech at 150 words per minute.
- Vital-sign readings: 105,400 (4.7 per dialogue).
Table 6 breaks these down by recorded ESI level. ESI 1-2 dialogues are shorter, as expected, because their interviews end soon after the nurse recognises the urgency. Table 7 describes the visits and Table 8 the dialogues.
| ESI level | Dialogues | Visits | Utterances per dialogue | Words |
|---|---|---|---|---|
| 1 | 4,710 | 702 | 9.4 | 719K |
| 2 | 5,164 | 698 | 9.5 | 822K |
| 3 | 5,541 | 703 | 21.2 | 1,888K |
| 4 | 4,153 | 691 | 19.5 | 1,312K |
| 5 | 2,670 | 634 | 19.0 | 805K |
| Total | 22,238 | 3,428 | 15.4 | 5,546K |
| Characteristic | n | % |
|---|---|---|
Recorded ESI level (esi) |
||
| 1 | 702 | 20.5 |
| 2 | 698 | 20.4 |
| 3 | 703 | 20.5 |
| 4 | 691 | 20.2 |
| 5 | 634 | 18.5 |
Gender (gender) |
||
female |
1,798 | 52.5 |
male |
1,630 | 47.5 |
Arrival transport (arrival_transport) |
||
WALK IN |
2,600 | 75.8 |
AMBULANCE |
605 | 17.6 |
UNKNOWN |
209 | 6.1 |
OTHER |
11 | 0.3 |
HELICOPTER |
3 | 0.1 |
Pain score (pain) |
||
| 0 | 1,473 | 43.0 |
| 1–3 | 339 | 9.9 |
| 4–6 | 600 | 17.5 |
| 7–10 | 1,016 | 29.6 |
| Vital sign not recorded | ||
Temperature (temperature) |
125 | 3.6 |
Heart rate (heartrate) |
61 | 1.8 |
Respiratory rate (resprate) |
80 | 2.3 |
Oxygen saturation (o2sat) |
93 | 2.7 |
Systolic blood pressure (sbp) |
67 | 2.0 |
Chief complaint (chief_complaint), 1,481 distinct |
||
suture removal |
87 | 2.5 |
med refill |
72 | 2.1 |
wound eval |
71 | 2.1 |
abd pain |
71 | 2.1 |
chest pain |
66 | 1.9 |
dyspnea |
61 | 1.8 |
rash |
49 | 1.4 |
back pain |
41 | 1.2 |
Note. Chief complaints are shown as recorded, compared without case; the eight most frequent are listed.
| Characteristic | n | % |
|---|---|---|
Model pairing, nurse and dialogue master / patient (nurse_model, patient_model) |
||
| Qwen3.6 / Gemma 4 | 5,492 | 24.7 |
| Gemma 4 / Qwen3.6 | 5,683 | 25.6 |
| Qwen3.6 / Nemotron 3 Nano | 5,521 | 24.8 |
| Gemma 4 / Nemotron 3 Nano | 5,542 | 24.9 |
Patient age group (patient_age_group) |
||
adult |
11,236 | 50.5 |
older |
11,002 | 49.5 |
Cognitive state (patient_cognitive_state) |
||
alert |
22,103 | 99.4 |
impaired |
91 | 0.4 |
delirious |
44 | 0.2 |
Intoxication (patient_intoxication) |
||
none |
22,156 | 99.6 |
mild |
82 | 0.4 |
| Personas | ||
| Distinct patient personas | 21,416 | |
Note. Every other persona trait level appears in 32–35% of dialogues, except terse verbosity, frequent_tangents focus, novice expertise and intuition_led algorithm reliance (about 25% each), because of the two excluded combinations.
Files and structure
README.md dialogues_train.csv dialogues_dev.csv utterances_train.csv utterances_dev.csv triage_records_train.csv triage_records_dev.csv vitals_train.csv vitals_dev.csv build_conversations.py mimic-iv-ed-ext-triagesim-croissant.jsonld PHYSIONET_REPORT.md
All tables are CSV files (RFC 4180, UTF-8, every field quoted). The dialogues are split by patient into training (60%), development (10%) and test (30%) sets; each table comes as a training (_train) and a development (_dev) file with the same columns, and the test set is held private. dialogue_id links the four tables of a set, and stay_id and subject_id link each dialogue to its visit in MIMIC-IV-ED and MIMIC-IV. The same visit appears in several dialogues (different model pairings and personas).
File format and contents
README.md. Title and short description.
dialogues_train.csv, dialogues_dev.csv. One row per dialogue (19,050 and 3,188 rows):
dialogue_id: dialogue identifier, combining the model pairing, the MIMIC-IV-EDstay_idand the dialogue number.stay_id,subject_id: MIMIC-IV-ED stay and MIMIC-IV patient identifiers of the visit.nurse_model: model that played the nurse and the dialogue master.patient_model: model that played the patient.esi: ESI level: the visit's recorded MIMIC-IV-ED level, which is the level the nurse assigns in the dialogue.n_utterances: number of utterances.chief_complaint: recorded chief complaint.temperature(°F),heartrate(per minute),resprate(per minute),o2sat(%),sbp(mmHg): recorded triage vital signs; empty if not recorded.pain: triage pain score as recorded.gender,arrival_transport: as recorded.patient_age_group,patient_english_proficiency,patient_health_literacy,patient_recall_reliability,patient_symptom_reporting,patient_distress,patient_disclosure,patient_demeanour,patient_verbosity,patient_focus,patient_disfluency: patient persona traits (Table 2).patient_cognitive_state,patient_intoxication: patient states set from the visit.nurse_gender,nurse_expertise,nurse_algorithm_reliance,nurse_risk_attitude,nurse_self_report_credence,nurse_questioning_style,nurse_emotional_responsiveness,nurse_workload: nurse persona traits (Table 3).
utterances_train.csv, utterances_dev.csv. One row per utterance (293,512 and 49,224 rows):
dialogue_id: dialogue identifier (dialogues file of the same set).turn: position of the utterance in the dialogue, from 0.speaker:nurseorpatient.text: what the speaker says; patient lines carry the disfluencies of the patient persona.
triage_records_train.csv, triage_records_dev.csv. One row per dialogue (19,050 and 3,188 rows): the nurse's triage record as kept by the dialogue master, with what the dialogue establishes for each field (empty if it does not): dialogue_id, chief_complaint, onset_and_course, pain, associated_symptoms, relevant_history, medications, allergies.
vitals_train.csv, vitals_dev.csv. One row per vital-sign reading the nurse took (90,288 and 15,112 rows): dialogue_id; after_turn: the utterance after which the nurse has the reading (turn in the utterances file); vital: temperature, heartrate, resprate, o2sat or sbp; value: the recorded triage value, in the units of the dialogues file, empty if not recorded.
build_conversations.py. Python script (standard library only) that writes each conversation with the vital-sign readings in place, as JSON lines or plain text: python build_conversations.py.
mimic-iv-ed-ext-triagesim-croissant.jsonld. Croissant metadata describing the files and fields, with Responsible AI (RAI) fields, made with croissant-baker.
PHYSIONET_REPORT.md. Output of the PhysioNet preflight validator, run on every row of every file.
Usage Notes
Dataset utility
The corpus was made for sequential clinical triage [47]. Conversations generated with TriageSim before this release have already been used to show that language models anchor on the chief complaint and fail to integrate later evidence in this task [47]. The data support the following uses:
- Sequential clinical triage: predicting the ESI level from a growing prefix of the dialogue (
turnorders the utterances;build_conversations.pyputs the vital-sign readings in place) or from the whole dialogue, with labels balanced across levels, at a scale that allows training as well as evaluation (the training set for training, the development set for model selection and evaluation); the persona columns allow analysis by patient and nurse behaviour, and the model columns by generating model. - Clinical information extraction: extracting the triage record or chief complaint from the conversation, with the nurse's triage record, the vital-sign readings and the recorded visit fields as references.
- Disfluency research: detecting disfluencies in persona-varied patient speech [46,48].
- Speech: text for speech synthesis and for speech-recognition test sets with clinical, persona-varied and disfluent speech; TriageSim's speech stage can render it as audio with persona-matched voices and ED background noise [5].
- Dialogue modelling: persona-conditioned dialogue generation and simulation, and the study of nurse questioning styles.
- Translation: the dialogues are a base for building parallel corpora of triage conversations.
Complementary resources. Our TriageSim code [49] can regenerate or extend the corpus. Users of this resource also have access to MIMIC-IV-ED [10], so they can join each dialogue by stay_id to the visit's diagnoses, medications and disposition. Our first TriageSim corpus [5] covers the ESI handbook and ETEK cases, the Australasian Triage Scale and audio.
Limitations
- The dialogues are simulated, not recorded, and reflect the behaviour of the generating models; the persona effects on the nurse, although in the intended direction, are small.
- Vital signs are measured one at a time.
- MIMIC-IV-ED records no age, so the patient's age group is sampled and may not fit the visit (pregnancy-related visits with an older-adult persona were removed).
- Impaired cognition and intoxication are rare (0.6% and 0.4% of dialogues), because they are set only when the visit supports them.
- English only; visits come from one US hospital (2011-2019); no audio in this version.
Future versions
Future versions will add TriageSim's speech stage [5]: phrase-break annotation of each utterance, persona-matched voices cloned zero-shot from a voice bank, and ED background sounds mixed at fixed levels.
Release Notes
Version 1.0.0: first release, the training and development sets: 22,238 dialogues grounded in 3,428 MIMIC-IV-ED visits. The test set (9,523 dialogues from 1,455 visits) is held private. Text only; audio is planned for a later version. For questions, please contact the corresponding author (d.srirag@unsw.edu.au)
Ethics
The project received ethics clearance from the UNSW Ethics Committee (iRECS7289). The corpus contains no new data from human participants. The dialogues are synthetic, contain no personal names, and are linked to de-identified MIMIC-IV-ED visits, so they are released under the same credentialed access terms. All models used for this release ran on the authors' institutional computing facility, so no MIMIC-IV-ED data were sent to external services.
Acknowledgements
Dipankar Srirag is funded by NHMRC Ideas Grant (RG241647) awarded to Dr. Padmanesan Narasimhan and Dr. Aditya Joshi, in 2025. This research includes computations using the computational cluster Katana supported by Research Technology Services at UNSW Sydney. Parts of this project description were written using AI assistance.
Conflicts of Interest
None to declare.
References
@jones2010Jones, P., & Schimanski, K. (2010). The four hour target to reduce emergency department 'waiting time': A systematic review of clinical outcomes. Emergency Medicine Australasia, 22(5), 391–398. https://doi.org/10.1111/j.1742-6723.2010.01330.x@gilboy2005Gilboy, N., Tanabe, P., Travers, D. A., Rosenau, A. M., & Eitel, D. R. (2005). Emergency Severity Index, version 4: Implementation handbook (AHRQ Publication No. 05-0046-2). Agency for Healthcare Research and Quality.@emergencynursesassociation2023Emergency Nurses Association. (2023). Emergency Severity Index handbook (5th ed.). https://californiaena.org/wp-content/uploads/2023/05/ESI-Handbook-5th-Edition-3-2023.pdf@australasiancollegeforemergenc2023Australasian College for Emergency Medicine. (2023). Guidelines on the implementation of the Australasian Triage Scale in emergency departments (G24, Version 6). https://acem.org.au/getmedia/51dc74f7-9ff0-42ce-872a-0437f3db640a/G24_04_Guidelines_on_Implementation_of_ATS_Jul-16.aspx@srirag2026Srirag, D., Nguyen, Q. D., Joshi, A., Narasimhan, P., & Kanhere, S. (2026). TriageSim: A conversational emergency triage simulation framework from structured electronic health records. In Proceedings of Interspeech 2026 (pp. 1997–2001). https://doi.org/10.48550/arXiv.2603.10035@heritage2006Heritage, J., & Robinson, J. D. (2006). The structure of patients' presenting concerns: Physicians' opening questions. Health Communication, 19(2), 89–102. https://doi.org/10.1207/s15327027hc1902_1@lee2015Lee, S.-H., & Kim, C. W. (2015). Presentation of patients' problems during triage in emergency medicine. Patient Education and Counseling, 98(5), 578–587. https://doi.org/10.1016/j.pec.2015.01.011@sax2023Sax, D. R., Warton, E. M., Mark, D. G., Vinson, D. R., Kene, M. V., Ballard, D. W., Vitale, T. J., McGaughey, K. R., Beardsley, A., Pines, J. M., Reed, M. E., & Kaiser Permanente CREST (Clinical Research on Emergency Services & Treatments) Network. (2023). Evaluation of version 4 of the Emergency Severity Index in US emergency departments for the rate of mistriage. JAMA Network Open, 6(3), e233404. https://doi.org/10.1001/jamanetworkopen.2023.3404@johnson2023Johnson, A. E. W., Bulgarelli, L., Shen, L., Gayles, A., Shammout, A., Horng, S., Pollard, T. J., Hao, S., Moody, B., Gow, B., Lehman, L.-W. H., Celi, L. A., & Mark, R. G. (2023). MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data, 10(1), 1. https://doi.org/10.1038/s41597-022-01899-x@johnson2023bJohnson, A., Bulgarelli, L., Pollard, T., Celi, L. A., Mark, R., & Horng, S. (2023). MIMIC-IV-ED (version 2.2) [Data set]. PhysioNet. https://doi.org/10.13026/5ntk-km72@papadopouloskorfiatis2022Papadopoulos Korfiatis, A., Moramarco, F., Sarac, R., & Savkov, A. (2022). PriMock57: A dataset of primary care mock consultations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) (pp. 588–598). Association for Computational Linguistics. https://doi.org/10.18653/v1/2022.acl-short.65@benabacha2023Ben Abacha, A., Yim, W.-W., Fan, Y., & Lin, T. (2023). An empirical study of clinical note generation from doctor-patient encounters. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (pp. 2291–2302). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.eacl-main.168@yim2023Yim, W.-W., Fu, Y., Ben Abacha, A., Snider, N., Lin, T., & Yetisgen, M. (2023). Aci-bench: A novel ambient clinical intelligence dataset for benchmarking automatic visit note generation. Scientific Data, 10(1), 586. https://doi.org/10.1038/s41597-023-02487-3@wang2024Wang, J., Yao, Z., Yang, Z., Zhou, H., Li, R., Wang, X., Xu, Y., & Yu, H. (2024). NoteChat: A dataset of synthetic patient-physician conversations conditioned on clinical notes. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 15183–15201). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-acl.901@lu2024Lu, M., Ho, B., Ren, D., & Wang, X. (2024). TriageAgent: Towards better multi-agents collaborations for large language model-based clinical triage. In Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 5747–5764). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-emnlp.329@srirag2026bSrirag, D., Joshi, A., Kanhere, S. S., & Narasimhan, P. (2026). From triage to discharge: A survey of NLP tasks, methods, and open challenges in the emergency department. Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, 20918–20942. Presented at the Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, Budapest, Hungary. doi:10.18653/2026.emnlp-main.908@barrows1993Barrows, H. S. (1993). An overview of the uses of standardized patients for teaching and evaluating clinical skills. Academic Medicine, 68(6), 443–451. https://doi.org/10.1097/00001888-199306000-00002@kyung2025Kyung, D., Chung, H., Bae, S., Kim, J., Sohn, J. H., Kim, T., Kim, S. K., & Choi, E. (2025). PatientSim: A persona-driven simulator for realistic doctor-patient interactions. In Advances in Neural Information Processing Systems 38 (Datasets and Benchmarks Track). https://arxiv.org/abs/2505.17818@wang2024bWang, R., Milani, S., Chiu, J. C., Zhi, J., Eack, S. M., Labrum, T., Murphy, S. M., Jones, N., Hardy, K. V., Shen, H., Fang, F., & Chen, Z. (2024). PATIENT-Ψ: Using large language models to simulate patients for training mental health professionals. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 12772–12797). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.711@johri2025Johri, S., Jeong, J., Tran, B. A., Schlessinger, D. I., Wongvibulsin, S., Barnes, L. A., Zhou, H.-Y., Cai, Z. R., Van Allen, E. M., Kim, D., Daneshjou, R., & Rajpurkar, P. (2025). An evaluation framework for clinical use of large language models in patient interaction tasks. Nature Medicine, 31(1), 77–86. https://doi.org/10.1038/s41591-024-03328-5@tu2025Tu, T., Schaekermann, M., Palepu, A., Saab, K., Freyberg, J., Tanno, R., Wang, A., Li, B., Amin, M., Cheng, Y., Vedadi, E., Tomasev, N., Azizi, S., Singhal, K., Hou, L., Webson, A., Kulkarni, K., Mahdavi, S. S., Semturs, C., . . . Natarajan, V. (2025). Towards conversational diagnostic artificial intelligence. Nature, 642(8067), 442–450. https://doi.org/10.1038/s41586-025-08866-7@zhu2026Zhu, H., Shi, X., & Zhou, J. (2026). ELICITED: EHR-grounded longitudinal interactive conversations for information-seeking triage evaluation and decision-making (arXiv:2608.09024). arXiv. https://arxiv.org/abs/2608.09024@goldberger2000Goldberger, A. L., Amaral, L. A. N., Glass, L., Hausdorff, J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody, G. B., Peng, C.-K., & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23), e215–e220. https://doi.org/10.1161/01.CIR.101.23.e215@zimmermann2011Zimmermann, C., Del Piccolo, L., Bensing, J., Bergvik, S., De Haes, H., Eide, H., Fletcher, I., Goss, C., Heaven, C., Humphris, G., Kim, Y.-M., Langewitz, W., Meeuwesen, L., Nuebling, M., Rimondini, M., Salmon, P., van Dulmen, S., Wissow, L., Zandbelt, L., & Finset, A. (2011). Coding patient emotional cues and concerns in medical consultations: The Verona coding definitions of emotional sequences (VR-CoDES). Patient Education and Counseling, 82(2), 141–148. https://doi.org/10.1016/j.pec.2010.03.017@plattsmills2010Platts-Mills, T. F., Travers, D., Biese, K., McCall, B., Kizer, S., LaMantia, M., Busby-Whitehead, J., & Cairns, C. B. (2010). Accuracy of the Emergency Severity Index triage instrument for identifying elder emergency department patients receiving an immediate life-saving intervention. Academic Emergency Medicine, 17(3), 238–243. https://doi.org/10.1111/j.1553-2712.2010.00670.x@flores2005Flores, G. (2005). The impact of medical interpreter services on the quality of health care: A systematic review. Medical Care Research and Review, 62(3), 255–299. https://doi.org/10.1177/1077558705275416@herndon2011Herndon, J. B., Chaney, M., & Carden, D. (2011). Health literacy and emergency department outcomes: A systematic review. Annals of Emergency Medicine, 57(4), 334–345. https://doi.org/10.1016/j.annemergmed.2010.08.035@mazer2011Mazer, M., DeRoos, F., Hollander, J. E., McCusker, C., Peacock, N., & Perrone, J. (2011). Medication history taking in emergency department triage is inaccurate and incomplete. Academic Emergency Medicine, 18(1), 102–104. https://doi.org/10.1111/j.1553-2712.2010.00959.x@barsky1988Barsky, A. J., Goodson, J. D., Lane, R. S., & Cleary, P. D. (1988). The amplification of somatic symptoms. Psychosomatic Medicine, 50(5), 510–519. https://doi.org/10.1097/00006842-198809000-00007@puntillo2003Puntillo, K., Neighbor, M., O'Neil, N., & Nixon, R. (2003). Accuracy of emergency nurses in assessment of patients' pain. Pain Management Nursing, 4(4), 171–175. https://doi.org/10.1016/S1524-9042(03)00033-X@rockett2006Rockett, I. R. H., Putnam, S. L., Jia, H., & Smith, G. S. (2006). Declared and undeclared substance use among emergency department patients: A population-based study. Addiction, 101(5), 706–712. https://doi.org/10.1111/j.1360-0443.2006.01397.x@rhodes2007Rhodes, K. V., Frankel, R. M., Levinthal, N., Prenoveau, E., Bailey, J., & Levinson, W. (2007). "You're not a victim of domestic violence, are you?" Provider-patient communication about domestic violence. Annals of Internal Medicine, 147(9), 620–627. https://doi.org/10.7326/0003-4819-147-9-200711060-00006@groves1978Groves, J. E. (1978). Taking care of the hateful patient. New England Journal of Medicine, 298(16), 883–887. https://doi.org/10.1056/NEJM197804202981605@arbuckle1993Arbuckle, T. Y., & Gold, D. P. (1993). Aging, inhibition, and verbosity. Journal of Gerontology, 48(5), P225–P232. https://doi.org/10.1093/geronj/48.5.P225@bortfeld2001Bortfeld, H., Leon, S. D., Bloom, J. E., Schober, M. F., & Brennan, S. E. (2001). Disfluency rates in conversation: Effects of age, relationship, topic, role, and gender. Language and Speech, 44(2), 123–147. https://doi.org/10.1177/00238309010440020101@hustey2002Hustey, F. M., & Meldon, S. W. (2002). The prevalence and documentation of impaired mental status in elderly emergency department patients. Annals of Emergency Medicine, 39(3), 248–253. https://doi.org/10.1067/mem.2002.122057@egertonwarburton2018Egerton-Warburton, D., Gosbell, A., Moore, K., Wadsworth, A., Richardson, D., & Fatovich, D. M. (2018). Alcohol-related harm in emergency departments: A prospective, multi-centre study. Addiction, 113(4), 623–632. https://doi.org/10.1111/add.14109@cioffi1998Cioffi, J. (1998). Decision making by emergency nurses in triage assessments. Accident and Emergency Nursing, 6(4), 184–191. https://doi.org/10.1016/S0965-2302(98)90077-7@considine2007Considine, J., Botti, M., & Thomas, S. (2007). Do knowledge and experience have specific roles in triage decision-making? Academic Emergency Medicine, 14(8), 722–726. https://doi.org/10.1197/j.aem.2007.04.015@gerdtz2001Gerdtz, M. F., & Bucknall, T. K. (2001). Triage nurses' clinical decision making. An observational study of urgency assessment. Journal of Advanced Nursing, 35(4), 550–561. https://doi.org/10.1046/j.1365-2648.2001.01871.x@pearson1995Pearson, S. D., Goldman, L., Orav, E. J., Guadagnoli, E., Garcia, T. B., Johnson, P. A., & Lee, T. H. (1995). Triage decisions for emergency department patients with chest pain: Do physicians' risk attitudes make the difference? Journal of General Internal Medicine, 10(10), 557–564. https://doi.org/10.1007/BF02640365@delpiccolo2011Del Piccolo, L., de Haes, H., Heaven, C., Jansen, J., Verheul, W., Bensing, J., Bergvik, S., Deveugele, M., Eide, H., Fletcher, I., Goss, C., Humphris, G., Kim, Y.-M., Langewitz, W., Mazzi, M. A., Mjaaland, T., Moretti, F., Nübling, M., Rimondini, M., . . . Finset, A. (2011). Development of the Verona coding definitions of emotional sequences to code health providers' responses (VR-CoDES-P) to patient cues and concerns. Patient Education and Counseling, 82(2), 149–155. https://doi.org/10.1016/j.pec.2010.02.024@qwenteam2026Qwen Team. (2026). Qwen3.6-35B-A3B-FP8 [Large language model]. Hugging Face. https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8@google2026Google. (2026). Gemma 4 26B A4B instruction-tuned [Large language model; FP8 quantization by Red Hat AI]. Hugging Face. https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic@nvidia2025NVIDIA. (2025). Nemotron 3 Nano: Open, efficient mixture-of-experts hybrid Mamba-Transformer model for agentic reasoning (arXiv:2512.20848). arXiv. https://arxiv.org/abs/2512.20848@hassan2025Hassan, S. Z., Lison, P., & Halvorsen, P. (2025). Enhancing naturalness in LLM-generated utterances through disfluency insertion (arXiv:2412.12710). arXiv. https://arxiv.org/abs/2412.12710@srirag2026cSrirag, D., Zhao, H., Kumar, A., Hopper, E., Dalton, M., Nguyen, Q. D., Joshi, A., Kanhere, S. S., & Narasimhan, P. (2026). LLMs anchor on chief complaint and fail to integrate evidence in sequential clinical triage (arXiv:2609.22904). arXiv. https://arxiv.org/abs/2609.22904@marie2023Marie, B. (2023). Disfluency generation for more robust dialogue systems. In Findings of the Association for Computational Linguistics: ACL 2023 (pp. 11479–11488). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-acl.728@srirag2026dSrirag, D. (2026). TriageSim [Computer software]. GitHub. https://github.com/dipankarsrirag/triage-sim
Cited by
5 publications cite this project. The most recent are listed below.
-
Stress-testing consentAI: a pipeline for evaluating a conversational educational adjunct to surgical informed consent
Anthony Abou Mrad, Jayanth S. Pratap, Aviona Conti et al. npj Digital Surgery, 2026. -
Prompts versus Rules: Auditing and Controlling Speech Naturalness Behaviors in Voice User Simulators
Riqiang Wang, Elena Khasanova, Harsh Saini et al. arXiv (Cornell University), 2026. -
LLMs Anchor on Chief Complaint and Fail to Integrate Evidence in Sequential Clinical Triage
Dipankar Srirag, Haokai Zhao, Ashutosh Kumar et al. arXiv (Cornell University), 2026. -
ELICITED: EHR-grounded Longitudinal Interactive Conversations for Information-seeking Triage Evaluation and Decision-making
Haohao Zhu, Xiaolin Shi, Jiayu Zhou. arXiv (Cornell University), 2026. -
From Triage to Discharge: A Survey of NLP Tasks, Methods, and Open Challenges in the Emergency Department
Dipankar Srirag, Aditya Joshi, Salil S. Kanhere et al. Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026.
Citation data from OpenAlex. Papers that use the project without citing it may be missing.
Files
- be a credentialed user
- complete required training:
- You may submit your training here.
- sign the data use agreement for the project