Synthetic EHR Generation

The synthetic dataset must preserve the joint multivariate distribution of the original data. This means maintaining not just univariate column statistics (mean, variance) but also complex bivariate and higher-order correlations between features like lab results, diagnoses, and medications.

• Column-wise metrics: Kolmogorov-Smirnov tests for continuous variables, Chi-squared tests for categorical codes.
• Pairwise correlation preservation: Pearson and Spearman correlation matrices must show minimal divergence.
• Propensity Mean Squared Error (pMSE): A discriminative model should not be able to easily distinguish synthetic from real records.

The gold standard for utility is the Train-Synthetic-Test-Real (TSTR) paradigm. A predictive model trained exclusively on synthetic data must perform comparably to one trained on real data when evaluated on a held-out real test set.

• Downstream task validation: Measure AUC-ROC for mortality prediction or readmission risk using synthetic training data.
• Feature importance alignment: SHAP values from models trained on synthetic data should rank clinical features similarly to models trained on real data.
• Subgroup analysis: Performance parity must hold across sensitive demographic strata to avoid introducing bias.

EHRs are inherently longitudinal. Synthetic records must respect clinical chronology—a diagnosis cannot precede a patient's birth date, and a medication refill must follow the initial prescription.

• Event sequence logic: Procedures like 'Cesarean Section' must only occur after a 'Pregnancy' diagnosis.
• Time-series plausibility: Lab values like HbA1c must change gradually, not spike randomly between visits.
• State transition validity: Markov transition probabilities between disease states must mirror real clinical progression patterns.

High fidelity must not come at the cost of memorization. The model must generalize the training distribution without reproducing verbatim patient trajectories. Formal privacy guarantees are essential.

• Differential Privacy (DP): Training with DP-SGD provides a provable ε-bound on information leakage.
• Nearest Neighbor Adversarial Accuracy (NNAA): Measures the risk of identity disclosure by comparing distances between synthetic and real records.
• Membership Inference Resistance: An attacker should not be able to determine if a specific individual was in the training set with confidence significantly above random chance.

Synthetic data must respect the relational database schema of the source EHR system. Foreign key constraints and code set validity are non-negotiable.

• Code ontology adherence: All ICD-10, SNOMED CT, and LOINC codes must be valid and active in the target vocabulary.
• Referential integrity: Every synthetic 'visit' record must link to a valid synthetic 'patient' record.
• Cross-table consistency: A synthetic lab result for 'Serum Potassium' must be consistent with a synthetic diagnosis of 'Hyperkalemia' in the conditions table.

The generator must cover the long tail of clinical presentations, including rare diseases, while also producing novel combinations not explicitly copied from the training set.

• Rare disease representation: Conditions with prevalence < 0.1% must still appear in the synthetic cohort.
• Novelty score: The proportion of synthetic records that are exact duplicates of real records must be near zero.
• Boundary pushing: The model should generate plausible but unseen comorbidity clusters, demonstrating generalization rather than rote memorization.