olaflaitinen/fedmml-ed-triage

Dataset Card for FedMML Emergency Department Triage Dataset

Dataset Summary

The FedMML Emergency Department Triage Dataset contains 87,234 synthetic emergency department encounters from 6 hospitals across 3 countries (Denmark, Turkey, Latvia). The dataset integrates three critical data modalities (clinical notes, vital signs, laboratory data) for predicting Emergency Severity Index (ESI) levels 1-5.

This dataset is specifically designed to support:

• Federated Learning: Multi-institutional collaboration without data sharing
• Multimodal Fusion: Integration of text, structured clinical data, and labs
• Privacy-Preserving ML: Research in differential privacy and secure aggregation
• Clinical AI: Emergency department triage automation

Supported Tasks and Leaderboards

• Primary Task: Multi-class classification of Emergency Severity Index (ESI) levels
• Benchmark Performance (from paper):
  • FedMML: AUROC 0.923, Accuracy 89.2%
  • Centralized Multimodal: AUROC 0.931
  • Traditional Triage: AUROC 0.762

Languages

• English (clinical notes and chief complaints)

Dataset Structure

Data Instances

{
  "encounter_id": "ENC3005432",
  "patient_id": "PAT045123",
  "site_id": 3,
  "country": "Turkey",
  "age": 67,
  "sex": "M",
  "arrival_timestamp": "2022-08-15 14:23:00",
  "chief_complaint": "Chest pain",
  "clinical_notes": "67yo M c/o Chest pain. Patient in moderate distress...",
  "systolic_bp": 165.0,
  "diastolic_bp": 95.0,
  "heart_rate": 102.0,
  "respiratory_rate": 22.0,
  "temperature": 37.8,
  "spo2": 94.0,
  "pain_score": 7.0,
  "wbc": 11.2,
  "hemoglobin": 13.8,
  "platelet_count": 235.0,
  "sodium": 138.5,
  "potassium": 4.1,
  "creatinine": 1.2,
  "glucose": 145.0,
  "troponin": 0.12,
  "bnp": 280.0,
  "lactate": 1.8,
  "inr": 1.1,
  "esi_level": 2
}

Data Fields

See detailed field descriptions in the README.md file.

Key Fields:

• esi_level (int): TARGET VARIABLE - Emergency Severity Index (1=Immediate, 5=Non-Urgent)
• clinical_notes (str): Triage nurse assessment (98.2% complete)
• Vital signs: systolic_bp, diastolic_bp, heart_rate, etc. (93.7% complete)
• Laboratory values: wbc, hemoglobin, troponin, etc. (69.5% complete)

Data Splits

The dataset is provided as a single file and can be split by:

1. Site-based (for federated learning):

   • 6 sites with 12,415-18,234 encounters each

2. Temporal (for time-based validation):

   • 4 years (2020-2023) with ~25% data each year

3. Random stratified (traditional ML):

   • Recommended: 70/15/15 train/val/test split

Dataset Creation

Curation Rationale

This dataset was created to support reproducible research in federated learning for healthcare while addressing privacy concerns. It enables:

1. Multi-institutional collaboration research without real patient data sharing
2. Development of multimodal clinical AI systems
3. Benchmarking of privacy-preserving machine learning techniques
4. Study of non-IID data distributions across healthcare institutions

Source Data

Initial Data Collection and Normalization

The dataset is entirely synthetic, generated to statistically match real emergency department patterns while protecting patient privacy. Generation process:

1. Demographic sampling: Age, sex distributions matched to published ED statistics
2. ESI distribution: Based on typical ED acuity patterns (ESI-3 most common at ~47%)
3. Vital signs: Generated from ESI-conditional distributions (critical vitals for ESI-1, normal for ESI-4/5)
4. Laboratory values: Sampled from clinical reference ranges with ESI-dependent abnormalities
5. Clinical notes: Template-based generation with realistic medical terminology
6. Missing data: Simulated based on typical ED workflows (labs ordered selectively)

Who are the source language producers?

Clinical terminology and notes written in medical English, following standard emergency medicine documentation practices.

Annotations

Annotation process

ESI levels (target variable) are assigned based on:

• Simulated clinical severity
• Vital sign stability
• Chief complaint acuity
• Resource utilization needs

This matches the real-world Emergency Severity Index triage system used in emergency departments.

Who are the annotators?

The dataset is synthetic; no human annotators were involved. Generation logic implements clinical triage guidelines.

Personal and Sensitive Information

This dataset contains NO real patient data. All encounters, patients, and clinical information are entirely synthetic.

Considerations for Using the Data

Social Impact of Dataset

Positive Impacts:

• Enables privacy-preserving healthcare AI research
• Democratizes access to emergency medicine datasets
• Supports development of clinical decision support systems
• Facilitates federated learning algorithm development

Potential Risks:

• Models trained on synthetic data require validation before clinical deployment
• May not capture all real-world clinical complexities
• Should not be used directly for patient care without proper validation

Discussion of Biases

Addressed Biases:

• Balanced sex distribution (~51% female)
• Age distribution representative of adult ED population
• Geographic diversity (3 countries)

Potential Limitations:

• European healthcare settings only (Denmark, Turkey, Latvia)
• May not represent populations from other continents
• Synthetic generation may miss rare clinical presentations
• COVID-19 pandemic period may affect generalizability

Other Known Limitations

1. Synthetic Nature: While statistically calibrated, this is not real clinical data
2. Missing Rare Conditions: Focus on common ED presentations
3. Language: English-only clinical notes
4. Temporal Scope: 2020-2023 (includes pandemic period)
5. ESI Interobserver Variability: Not modeled in synthetic generation

Additional Information

Dataset Curators

• Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov (Technical University of Denmark)
• Sila Burde Dulger (Gaziantep University, Turkey)

Licensing Information

Creative Commons Attribution 4.0 International (CC BY 4.0)

You are free to share and adapt this dataset for any purpose, provided you give appropriate credit.

Citation Information

@article{lundstrom2024fedmml,
  title={Federated Multimodal Learning for Real-Time Emergency Department Triage Optimization: A Privacy-Preserving Framework Integrating Clinical Notes, Vital Signs, and Laboratory Data},
  author={Laitinen-Fredriksson Lundström-Imanov, Gustav Olaf Yunus and Dulger, Sila Burde},
  journal={Preprint submitted to Elsevier},
  year={2024},
  url={https://huggingface.co/datasets/olaflaitinen/fedmml-ed-triage}
}

Contributions

This dataset was created to support reproducible research in federated learning and multimodal clinical AI. We welcome feedback and contributions from the research community.

Contact: olaf.laitinen@gmail.com | silaburdedlgr@outlook.com