mdsajjadullah/synthetic-medical-ehr-dataset

🏥 Synthetic Privacy-Preserving Medical EHR Dataset

Dataset Description

A fully synthetic collection of 10,000 Electronic Health Records (EHRs) for binary classification research. The task is predicting adverse patient outcomes — deterioration or death — from clinical, demographic, and vital sign features. No real patient data was used at any stage. Privacy-safe under GDPR, CCPA, and HIPAA.

Supported Tasks

• tabular-classification: Predict adverse_outcome (0/1) from 13 clinical features
• Benchmark models: Random Forest, Gradient Boosting, XGBoost
• Class imbalance handling: SMOTE oversampling applied on training set only
• Interpretability: SHAP feature importance included in generation pipeline

Dataset Structure

Split	File	Rows	Strategy
train	train.csv	7,000	70% stratified by target
validation	val.csv	1,500	15% stratified by target
test	test.csv	1,500	15% stratified by target
full	full_dataset.csv	10,000	Complete unsplit dataset

Data Fields

Field	Type	Description
patient_id	string	UUID-based anonymous ID (e.g. P-3FA2C1B8)
age_range	string	Age bucket: 0-17 / 18-29 / 30-44 / 45-59 / 60-74 / 75-99
age_midpoint	int32	Midpoint of age bucket — used as numeric ML feature
gender	string	Male / Female / Non-binary
ethnicity	string	White / Black / Hispanic / Asian / Other
region	string	North / South / East / West / Central
icd10_code	string	ICD-10 primary diagnosis code
condition	string	Hypertension, Diabetes, Cancer, etc.
severity	string	Mild / Moderate / Severe
onset_year	int32	Year of diagnosis (2005–2025)
medication	string	Primary prescribed medication
num_visits	int32	Number of provider visits
adherence_pct	int32	Treatment adherence percentage 0–100
systolic_bp	int32	Systolic blood pressure in mmHg
diastolic_bp	int32	Diastolic blood pressure in mmHg
heart_rate	int32	Heart rate in beats per minute
cholesterol_mg_dl	int32	Total cholesterol in mg/dL
bmi	float32	Body Mass Index
outcome	string	Recovered / Stable / Deteriorated / Deceased
adverse_outcome	int32 TARGET	1 = Deteriorated or Deceased, 0 = otherwise

Privacy Methods

This dataset was synthetically generated in Python. Privacy techniques applied:

1. Anonymization — UUID patient IDs, no real names or SSNs
2. Generalization — age buckets, region labels instead of cities
3. Differential Privacy — Laplace mechanism noise (epsilon=2.0) on all vital signs
4. Data Perturbation — Gaussian noise on continuous features

No real patient records were used or referenced at any stage.

Benchmark Results

Model	Val ROC-AUC	Accuracy	F1 Adverse	CV 5-fold
Random Forest	0.9535	89%	0.82	0.9707 ± 0.013
Gradient Boosting	0.9630	90%	0.84	0.9760 ± 0.013
XGBoost (best)	0.9665	91%	0.85	0.9769 ± 0.012
XGBoost (test)	0.9542	90%	0.85	—

Use Cases

• Binary classification for healthcare outcome prediction
• Benchmarking tabular ML models without privacy risk
• Teaching SMOTE, scaling, hyperparameter tuning, SHAP
• Fairness analysis across demographic groups
• Differential privacy and privacy-preserving AI research

Citation

@dataset{mdsajjadullah_2025_synthetic_ehr,
  author    = {Md.Sajjad Ullah},
  title     = {Synthetic Privacy-Preserving Medical EHR Dataset},
  year      = {2025},
  publisher = {Zenodo},
  version   = {5.0},
  doi       = {10.5281/zenodo.18968480},
  url       = {https://doi.org/10.5281/zenodo.18968480}
}