samwell/synthea-ncd-instructions

Synthea NCD Instructions

Synthetic EHR-based instruction-tuning dataset for training LLMs to predict non-communicable disease (NCD) risk, specifically Type 2 Diabetes and Hypertension.

Quick Start

from datasets import load_dataset

dataset = load_dataset("samwell/synthea-ncd-instructions")

# View a sample
print(dataset["train"][0])

Dataset Description

This dataset contains instruction-tuning examples derived from synthetic patient records generated using Synthea. Each example presents a patient's clinical data and asks the model to assess their NCD risk.

Why This Dataset?

• —NCD burden: Diabetes and hypertension affect billions globally, especially in LMICs
• —Clinical decision support: LLMs can help with risk stratification at scale
• —Open & reproducible: Fully synthetic, no privacy concerns, Apache 2.0 licensed
• —Instruction-tuned format: Ready for fine-tuning Gemma, Llama, Mistral, etc.

Intended Use

• —Fine-tuning LLMs for clinical decision support
• —Training models for NCD risk stratification
• —Research in medical AI and clinical NLP
• —Educational purposes
• —Prototyping healthcare AI applications

Not Intended For

• —Direct clinical use without validation
• —Replacing clinical judgment
• —Diagnosis or treatment decisions
• —Production deployment without clinician review

Dataset Structure

Each example contains:

Example

{
  "instruction": "Based on the following patient record, assess the risk of Type 2 diabetes and hypertension. Provide risk levels (LOW, MODERATE, HIGH, or DIAGNOSED) with supporting factors, and clinical recommendations.",
  "input": "Patient: 52yo Female\nVitals: BP 142/88 mmHg, BMI 29.4, Weight 78.2 kg\nLabs: Glucose 118 mg/dL, HbA1c 6.0%, Total Cholesterol 210 mg/dL\nActive conditions: Prediabetes\nMedications: None",
  "output": "## Risk Assessment\n\n**Type 2 Diabetes: MODERATE**\n- Fasting glucose 118 mg/dL (prediabetes range 100-125)\n- HbA1c 6.0% (prediabetes range 5.7-6.4%)\n- BMI 29.4 (overweight)\n- History of prediabetes\n\n**Hypertension: HIGH**\n- BP 142/88 mmHg (Stage 1 hypertension)\n\n## Recommendations\n1. Lifestyle counseling: diet modification, increase physical activity\n2. Recheck glucose/HbA1c in 3-6 months\n3. Confirm elevated BP on 2 separate occasions\n4. Consider initiating antihypertensive therapy"
}

Risk Levels

Data Splits

Clinical Parameters

Observations Used

Vitals:

• —Blood Pressure (systolic/diastolic)
• —BMI
• —Body Weight
• —Heart Rate

Labs:

• —Fasting Glucose
• —HbA1c
• —Total Cholesterol, HDL, LDL, Triglycerides
• —Creatinine, eGFR

Conditions Tracked

Generation Process

1. 1.Synthea generated synthetic patient populations with realistic disease progression
2. 2.Patient records were filtered for those with relevant NCD observations
3. 3.Risk assessments were generated using clinical guidelines:
4. 4.ADA criteria for diabetes/prediabetes
5. 5.ACC/AHA guidelines for hypertension
6. 6.Data was formatted for instruction-tuning (Alpaca-style)

Reproducibility

The generation scripts are available at: github.com/HopeOS/training

# Generate synthetic patients
./run_synthea -p 50000 --exporter.csv.export=true

# Transform to instruction format
python synthea_to_instructions.py --input ./synthea/output/csv --output ./data

Limitations

• —Synthetic data: Does not capture all real-world clinical complexity
• —US-based demographics: Synthea defaults to US population characteristics
• —Simplified risk model: Does not include family history, lifestyle factors, or genetic risk
• —English only: All text is in English
• —No longitudinal reasoning: Each example is a snapshot, not a time-series

Changelog

v1.0.0 (April 2026)

• —Initial release
• —~50,000 synthetic patients
• —Diabetes and hypertension risk assessment
• —Train/val/test splits (80/10/10)

Roadmap

Planned improvements (contributions welcome!):

• —[ ] Ghana/African demographics: Custom Synthea config for African population characteristics
• —[ ] Additional NCDs: Chronic kidney disease, cardiovascular disease, obesity
• —[ ] Multilingual: French, Twi, Hausa translations for West African context
• —[ ] Longitudinal examples: Multi-visit patient trajectories
• —[ ] Family history: Incorporate genetic risk factors
• —[ ] Lifestyle factors: Diet, exercise, smoking, alcohol
• —[ ] Validated models: Release fine-tuned Gemma/Llama checkpoints

Contributing

We welcome contributions! Here's how you can help:

1. 1.Report issues: Found an error in the data? Open an issue
2. 2.Improve generation: Submit PRs to the generation scripts
3. 3.Add demographics: Help create Synthea configs for other regions
4. 4.Validate clinically: Are you a clinician? Help us review the risk assessments
5. 5.Translate: Help translate to other languages

Citation

@dataset{synthea_ncd_instructions_2026,
  title={Synthea NCD Instructions: A Synthetic Dataset for Clinical Risk Assessment},
  author={samwell},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/samwell/synthea-ncd-instructions},
  note={Living dataset - check for updates}
}

License

Apache 2.0 - free to use, modify, and distribute with attribution.

Acknowledgments

• —Synthea - Synthetic patient generation
• —Unsloth - Efficient fine-tuning
• —Clinical guidelines: ADA, ACC/AHA, WHO
• —HopeOS team for the initial implementation

Contact

• —Maintainer: @samwell
• —Discussions: Dataset discussions
• —Issues: Report data quality issues in discussions

Last updated: April 2026