2O24dpower2024/ESCALATE

---
language:
- en
license: cc-by-4.0
pretty_name: ESCALATE Dataset

task_categories:
- text-classification
- text-generation

task_ids:
- dialogue-modeling
- multi-class-classification

tags:
- healthcare
- medical
- llm
- multi-agent
- simulation
- dialogue
- evaluation

size_categories:
- 1K<n<10K
---

# ESCALATE: A Dataset for Safety-Critical Clinical Escalation Conversations

[](https://doi.org/10.5281/zenodo.19224182)
[](https://huggingface.co/datasets/2O24dpower2024/ESCALATE)
[](https://github.com/DavePower-cloud/ESCALATE-dataset)

# Dataset Card for ESCALATE

## Dataset Summary

ESCALATE is a synthetic dataset of clinical escalation-of-care conversations generated using multi-agent large language models.

It enables evaluation of communication quality and safety under structured (ISBAR) and unstructured conditions.

---

## Motivation

Clinical escalation is a safety-critical process where communication failures can lead to patient harm.

ESCALATE provides a structured way to evaluate how AI systems perform in this context.

---

## Composition

- 200 paired cases (control vs ISBAR)
- 10 deterioration archetypes
- 400 total transcripts

---

## Generation Process

- Case cards define ground truth
- Role-locked agents simulate:
- nurse
- registrar
- optional nurse-in-charge
- Models:
- OpenAI (GPT-4o-mini)
- Anthropic (Claude Haiku)
- Google (Gemini Flash)

---

## Evaluation Labels

Each transcript is scored using a structured rubric:

- omissions
- hallucinations
- actionability
- escalation appropriateness
- closed-loop communication

---

## Splits

- Train / Validation / Test
- Paired conversations kept together

---

## Intended Uses

- AI evaluation
- simulation research
- clinical communication analysis

---

## Out-of-Scope Uses

- real clinical decision support
- patient care

---

## Limitations

- synthetic data
- model-generated behaviours
- limited to defined archetypes

---

## Ethical Considerations

- no real patient data
- designed for research use only

---

## Citation

Power, D., & Power, T. (2026). ESCALATE: A Dataset for Safety-Critical Clinical Escalation Conversations (Version: 1.0.0) [Data set]. Zenodo.
https://doi.org/10.5281/zenodo.19224182
[](https://doi.org/10.5281/zenodo.19224182)

---

### BibTeX

@dataset{power2026escalate,\
author = {David Power},\
title = {ESCALATE: A Dataset for Safety-Critical Clinical Escalation Conversations},\
year = {2026},\
publisher = {Zenodo},\
doi = {10.5281/zenodo.19224182}\
}

---

## 🔗 Related Research

This dataset builds on prior work demonstrating the feasibility of multi-agent large language models for simulating clinical conversations:

*Power, D., & Power, T. (2026). Can Large Language Models Generate Role-Consistent Clinical Dialogue for Education? A Multi-agent Approach.* Under Review.

Preprint available at EdArViX: https://doi.org/10.35542/osf.io/etv6d_v

ESCALATE extends this work by introducing a structured dataset of escalation-of-care conversations with safety-focused evaluation labels

---

👤 Author

David Power

Healthcare Simulation Specialist | MSc Artificial Intelligence

💼 LinkedIn: https://www.linkedin.com/in/dave-power-47280a44/

💻 GitHub: https://github.com/DavePower-cloud

---

## 📜 Licensing

- Code is released under the MIT License
- Dataset is released under CC BY 4.0

Please cite appropriately when using the dataset.

---