richardyoung/synthea-575k-patients

Synthea Synthetic Patient Records (575K Patients)

A comprehensive synthetic healthcare dataset containing 575,415 patients with complete medical histories, generated using Synthea - an open-source synthetic patient generator.

Dataset Curator: Richard Young | DeepNeuro.AI

Dataset Description

This dataset provides realistic but entirely synthetic patient records suitable for:

• Machine learning model training and evaluation
• Healthcare analytics research
• Clinical NLP development
• Medical AI/ML experimentation
• Educational purposes

All data is 100% synthetic - no real patient information is included.

Dataset Statistics

Table	Records	Description
patients	575,415	Core patient demographics
encounters	~93M	Healthcare visits/encounters
observations	~470M	Clinical observations (vitals, labs)
conditions	~56M	Diagnosed conditions
medications	~54M	Prescribed medications
procedures	~140M	Medical procedures
claims	~30M	Insurance claims
claims_transactions	~1.9B	Detailed claim transactions
immunizations	~17M	Vaccination records
careplans	~6M	Care plan records
allergies	~230K	Patient allergies
devices	~16M	Medical devices
imaging_studies	~32M	Imaging study records
supplies	~11M	Medical supplies
payers	10	Insurance payers
payer_transitions	~12M	Insurance changes
providers	~24K	Healthcare providers
organizations	~24K	Healthcare organizations

Total Size: 134GB (Parquet), 977GB (CSV source)

Data Processing Pipeline

Generation Process

The dataset was generated using a batched approach to handle the scale:

• 20 batches of 25,000 patients each
• 40GB Java heap allocation per batch
• 32 CPU cores for parallel generation
• Custom batch merging to preserve CSV headers (avoiding Synthea's append_mode bug)

Compression & Optimization

The raw CSV output (977GB) was converted to Parquet format achieving 86% compression:

Format	Size	Compression
CSV (raw)	977 GB	-
Parquet	134 GB	86% reduction

Conversion method: Memory-efficient streaming using Polars scan_csv() + sink_parquet():

• Processes files in chunks without loading entire files into memory
• Handles 30GB+ CSV files without OOM errors
• Schema overrides for mixed-type columns (claims, procedures, observations)
• 8 parallel workers for optimal throughput

Data Verification

Data integrity was verified through:

1. Header Validation: All 19 CSV files confirmed to have correct headers
2. Row Count Verification: Patient counts validated at each batch merge
3. Parquet Integrity: All 18 Parquet files successfully written and readable
4. Foreign Key Validation: Patient IDs verified across related tables
5. Schema Consistency: Column types verified during Parquet conversion

Data Schema

patients.parquet

• Id - Unique patient identifier (UUID)
• BIRTHDATE - Date of birth
• DEATHDATE - Date of death (if applicable)
• SSN - Social Security Number (synthetic)
• DRIVERS - Driver's license number (synthetic)
• PASSPORT - Passport number (synthetic)
• PREFIX - Name prefix (Mr., Mrs., etc.)
• FIRST - First name
• MIDDLE - Middle name
• LAST - Last name
• SUFFIX - Name suffix
• MAIDEN - Maiden name
• MARITAL - Marital status
• RACE - Race
• ETHNICITY - Ethnicity
• GENDER - Gender (M/F)
• BIRTHPLACE - Place of birth
• ADDRESS - Street address
• CITY - City
• STATE - State
• COUNTY - County
• FIPS - FIPS code
• ZIP - ZIP code
• LAT - Latitude
• LON - Longitude
• HEALTHCARE_EXPENSES - Total healthcare expenses
• HEALTHCARE_COVERAGE - Insurance coverage amount
• INCOME - Annual income

encounters.parquet

• Id - Encounter ID
• START - Start datetime
• STOP - End datetime
• PATIENT - Patient ID (FK)
• ORGANIZATION - Organization ID (FK)
• PROVIDER - Provider ID (FK)
• PAYER - Payer ID (FK)
• ENCOUNTERCLASS - Type (ambulatory, emergency, inpatient, etc.)
• CODE - SNOMED-CT code
• DESCRIPTION - Encounter description
• BASE_ENCOUNTER_COST - Base cost
• TOTAL_CLAIM_COST - Total claim cost
• PAYER_COVERAGE - Amount covered by payer
• REASONCODE - Reason code (SNOMED-CT)
• REASONDESCRIPTION - Reason description

conditions.parquet

• START - Condition onset date
• STOP - Condition resolution date
• PATIENT - Patient ID (FK)
• ENCOUNTER - Encounter ID (FK)
• SYSTEM - Coding system (SNOMED-CT)
• CODE - Condition code
• DESCRIPTION - Condition description

medications.parquet

• START - Prescription start date
• STOP - Prescription end date
• PATIENT - Patient ID (FK)
• PAYER - Payer ID (FK)
• ENCOUNTER - Encounter ID (FK)
• CODE - RxNorm code
• DESCRIPTION - Medication description
• BASE_COST - Base cost
• PAYER_COVERAGE - Payer coverage amount
• DISPENSES - Number of dispenses
• TOTALCOST - Total cost
• REASONCODE - Reason code
• REASONDESCRIPTION - Reason description

observations.parquet

• DATE - Observation date
• PATIENT - Patient ID (FK)
• ENCOUNTER - Encounter ID (FK)
• CATEGORY - Category (vital-signs, laboratory, etc.)
• CODE - LOINC code
• DESCRIPTION - Observation description
• VALUE - Observation value
• UNITS - Units of measurement
• TYPE - Data type

procedures.parquet

• START - Procedure start datetime
• STOP - Procedure end datetime
• PATIENT - Patient ID (FK)
• ENCOUNTER - Encounter ID (FK)
• SYSTEM - Coding system (SNOMED-CT)
• CODE - Procedure code
• DESCRIPTION - Procedure description
• BASE_COST - Base cost
• REASONCODE - Reason code
• REASONDESCRIPTION - Reason description

Usage

import polars as pl

# Load a single table
patients = pl.read_parquet("hf://datasets/richardyoung/synthea-575k-patients/data/patients.parquet")

# Load and join tables
encounters = pl.read_parquet("hf://datasets/richardyoung/synthea-575k-patients/data/encounters.parquet")
conditions = pl.read_parquet("hf://datasets/richardyoung/synthea-575k-patients/data/conditions.parquet")

# Example: Get all conditions for a patient
patient_conditions = conditions.filter(pl.col("PATIENT") == "some-uuid")

# Using pandas
import pandas as pd

patients = pd.read_parquet("hf://datasets/richardyoung/synthea-575k-patients/data/patients.parquet")

# Using datasets library
from datasets import load_dataset

dataset = load_dataset("richardyoung/synthea-575k-patients", data_files="data/patients.parquet")

Demographics Summary

Based on 575,415 synthetic patients:

• Gender Distribution: ~50% Female, ~50% Male
• Race Distribution: ~83% White, ~8% Black, ~6% Asian, ~3% Other
• Ethnicity: ~90% Non-Hispanic, ~10% Hispanic
• Age Range: 0-111 years
• Geographic Coverage: Massachusetts, USA

Generation Details

• Generator: Synthea v3.x
• Generation Date: November 2025
• Configuration: Full clinical documentation enabled
• Modules: 89 disease modules, 157 submodules
• Seed Range: 3000-22000 (20 batches of 25K patients)
• Hardware: 62GB RAM, 32 CPU cores, 40GB Java heap

Limitations

1. Synthetic Data: All records are computationally generated and may not perfectly reflect real-world distributions
2. Geographic Bias: Patients are generated for Massachusetts only
3. Temporal Coverage: Records span multiple decades but follow Synthea's simulation model
4. Clinical Realism: While based on medical literature, some edge cases may not be realistic

Citation

If you use this dataset, please cite both the dataset and Synthea:

@misc{young2025synthea575k,
  title={Synthea Synthetic Patient Records (575K Patients)},
  author={Young, Richard},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/richardyoung/synthea-575k-patients}},
  note={Generated using Synthea. Curated by DeepNeuro.AI}
}

@article{walonoski2018synthea,
  title={Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record},
  author={Walonoski, Jason and Kramer, Mark and Nichols, Joseph and Quina, Andre and Moesel, Chris and Hall, Dylan and Duffett, Carlton and Dube, Kudakwashe and Gallagher, Thomas and McLachlan, Scott},
  journal={Journal of the American Medical Informatics Association},
  volume={25},
  number={3},
  pages={230--238},
  year={2018},
  publisher={Oxford University Press},
  doi={10.1093/jamia/ocx079}
}

License

This dataset is released under the MIT License, consistent with Synthea's licensing.

Acknowledgments

• Dataset Curation: Richard Young - DeepNeuro.AI
• Data Generation: Synthea - The MITRE Corporation
• Processing Tools: Polars for memory-efficient data processing

Contact

For questions or issues with this dataset, please contact:

• Richard Young: deepneuro.ai/richard
• Open an issue on this dataset's community tab