Synthetic Privacy-Preserving Medical EHR Dataset: 10,000 Synthetic Electronic Health Records for Machine Learning Classification Research

This dataset contains 10,000 fully synthetic Electronic Health Records (EHRs) generated for binary classification research in healthcare AI. The primary task is predicting adverse patient outcomes — defined as clinical deterioration or death — from patient demographics, medical conditions, treatment information, and vital signs. No real patient data was used at any stage of dataset creation. The dataset was generated using Python with clinically realistic signal embedded into the generation process. Adverse outcomes are genuinely correlated with risk factors including older age, high-risk diagnoses (lung cancer, coronary artery disease, chronic kidney disease, type 2 diabetes), elevated blood pressure, high BMI, and low treatment adherence. This makes the dataset a meaningful benchmark for clinical prediction models. Privacy-preserving techniques applied:(1) Anonymization — real identifiers replaced with UUID patient IDs(2) Generalization — exact ages replaced with age buckets, cities replaced with regional labels: North, South, East, West, Central(3) Differential Privacy — Laplace mechanism noise (epsilon=2.0) added to all vital sign columns during generation(4) Data Perturbation — Gaussian noise applied to all continuous features to prevent exact record replication The dataset is fully compliant with GDPR, CCPA, and HIPAA. No personally identifiable information (PII) or protected health information (PHI) is present anywhere in the dataset. Three machine learning models were trained and benchmarked:- Random Forest: Validation ROC-AUC 0.9535, Accuracy 89%, Adverse class F1 0.82, CV 0.9707 ± 0.013- Gradient Boosting: Validation ROC-AUC 0.9630, Accuracy 90%, Adverse class F1 0.84, CV 0.9760 ± 0.013- XGBoost (best model): Validation ROC-AUC 0.9665, Accuracy 91%, Adverse class F1 0.85, CV 0.9769 ± 0.012- XGBoost on test set: ROC-AUC 0.9542, Accuracy 90%, Adverse class F1 0.85 The complete ML pipeline includes StandardScaler preprocessing, SMOTE oversampling for class imbalance on training set only, RandomizedSearchCV hyperparameter optimisation with 15 iterations and 3-fold CV, 5-fold cross-validation on all final models, and SHAP-based model interpretability. The dataset is pre-split into training 70%, validation 15%, and test 15% with stratification on the target variable. Files included:- full_dataset.csv: All 10,000 records with 20 columns- train.csv: 7,000 training records, 70% stratified split- val.csv: 1,500 validation records, 15% stratified split- test.csv: 1,500 test records, 15% stratified split- metadata.json: Complete pipeline metadata, best hyperparameters, and evaluation scores Target column: adverse_outcome (binary: 1 = Deteriorated or Deceased, 0 = Recovered or Stable)Class distribution: approximately 68% Stable, 32% AdverseTotal features: 20 columns, 13 used as ML features