Electronic Health Record Generation

Synthetic EHRs must preserve longitudinal disease trajectories and clinically valid temporal sequences. This means generated records maintain realistic gaps between diagnoses, medication orders, and lab results.

• Medication ordering: A synthetic record should not prescribe a drug before the corresponding diagnosis appears in the timeline
• Lab value progression: Creatinine levels should change gradually, not spike randomly between encounters
• Age-appropriate conditions: Pediatric conditions should not appear in geriatric synthetic patients

Models like TimeGAN and EHR-M-GAN explicitly learn these temporal dependencies by combining adversarial training with supervised autoregressive objectives that capture transition probabilities between clinical states.

Real EHRs contain heterogeneous data types that must be synthesized jointly to preserve cross-modal correlations. A synthetic generator must simultaneously model:

• Continuous variables: Lab values, vital signs, BMI measurements with realistic non-Gaussian distributions
• Discrete categorical variables: ICD-10 diagnosis codes, CPT procedure codes, medication orders with severe class imbalance
• Unstructured clinical notes: Free-text physician narratives that correlate with structured billing codes
• Temporal event sequences: Admission-discharge-transfer patterns with variable-length encounter histories

CTGAN addresses tabular heterogeneity through mode-specific normalization, while EHR-Safe uses hierarchical encoder-decoder architectures to bind structured and unstructured data within the same synthetic patient.

Synthetic EHRs must provide provable privacy protections beyond simple de-identification. The Safe Harbor method of removing 18 HIPAA identifiers is insufficient when generative models can memorize and reproduce rare training records.

• Differential Privacy (DP): Injecting calibrated Gaussian noise into generator gradients during training, with privacy budgets typically set at ε ≤ 8 for clinical applications
• K-Anonymity enforcement: Ensuring each synthetic record is indistinguishable from at least k-1 other records in the generated dataset
• Membership inference resistance: Measured by Nearest Neighbor Adversarial Accuracy (NNAA), where values near 0.5 indicate attackers cannot distinguish real from synthetic records

DP-GAN and PATE-GAN architectures incorporate these guarantees directly into the training objective, providing mathematical bounds on re-identification risk.

Beyond statistical fidelity, synthetic EHRs must satisfy domain-specific medical knowledge constraints that purely data-driven models may violate. Clinical plausibility ensures generated records would be accepted as valid by a practicing clinician.

• Ontology adherence: Diagnoses must map to valid SNOMED CT or ICD-10 hierarchies without impossible parent-child relationships
• Physiological consistency: Lab values must fall within biologically possible ranges and respect known correlations (e.g., hemoglobin and hematocrit maintain a roughly 1:3 ratio)
• Drug-disease contraindication avoidance: Synthetic patients should not receive medications contraindicated for their documented conditions
• Demographic plausibility: Age, sex, and pregnancy status must align logically across encounters

Causal generative models enforce these constraints by incorporating directed acyclic graphs that encode clinical domain knowledge, preventing the generation of medically impossible records.

The ultimate test of synthetic EHR quality is whether models trained on synthetic data perform comparably on real clinical tasks. The Train-Synthetic-Test-Real (TSTR) paradigm quantifies this directly.

• Downstream task performance: A mortality prediction model trained on synthetic data should achieve AUROC within 5% of one trained on real data
• Propensity score matching: Distributions of propensity scores between real and synthetic cohorts should be indistinguishable
• Feature correlation preservation: Pairwise Pearson and Spearman correlations between clinical variables must be maintained
• Subgroup fidelity: Rare disease populations and demographic minorities must be represented with proportional accuracy

The Synthetic Data Quality Score aggregates these dimensions into a composite metric balancing fidelity, utility, and privacy for regulatory submissions.

Generative models trained on historically biased healthcare data risk amplifying existing disparities in synthetic EHRs. Fairness-aware generation techniques actively counteract this.

• FairGAN architectures: Enforce statistical parity constraints during training so synthetic data equalizes outcome distributions across protected groups
• Resampling strategies: Oversample underrepresented populations before training to prevent mode collapse on majority demographics
• Counterfactual fairness: Generate synthetic records that would remain consistent if sensitive attributes like race were altered, using causal structural equation models
• Bias auditing: Post-generation evaluation using fairness metrics like equalized odds difference and demographic parity ratio

Model Cards for synthetic EHR generators must transparently disclose these fairness assessments alongside privacy and utility evaluations to meet emerging regulatory requirements.

A GAN architecture that conditions generation on specific patient attributes, enabling controlled synthesis of EHRs for underrepresented cohorts. Key mechanisms:

• Auxiliary conditioning: Feeds labels like ICD-10 codes or demographics into both generator and discriminator
• Class-conditional batch normalization: Modulates feature maps based on conditional vectors
• Temporal conditioning: Incorporates time-series metadata to preserve disease progression trajectories

Example: Generating synthetic records for diabetic patients aged 65+ with specific comorbidity profiles to augment rare disease datasets.

A mathematical framework providing provable privacy guarantees by injecting calibrated noise into the training process, ensuring synthetic EHRs do not leak individual patient information. Core components:

• Epsilon (ε): Privacy budget parameter controlling the privacy-utility tradeoff; lower values (ε < 1) provide stronger guarantees
• Gaussian mechanism: Adds noise proportional to the sensitivity of the query
• DP-SGD: Modifies stochastic gradient descent by clipping gradients and adding noise during GAN training

Clinical context: Enables compliance with HIPAA Safe Harbor and GDPR when sharing synthetic datasets across institutions.

A privacy audit technique where an adversary determines whether a specific patient's record was used to train the generative model, exposing potential data leakage. Attack methodology:

• Shadow model training: Trains attack models on synthetic datasets with known membership labels
• Loss-based inference: Exploits differences in model confidence between training and non-training samples
• Nearest neighbor analysis: Compares distances between synthetic records and real patient vectors

Mitigation: Combine with differential privacy and nearest neighbor adversarial accuracy (NNAA) evaluation to quantify and reduce re-identification risk.

The degree to which synthetic EHRs adhere to established physiological constraints, medical ontologies, and realistic disease progression pathways. Validation dimensions:

• Ontological consistency: Alignment with SNOMED CT, LOINC, and RxNorm coding systems
• Temporal coherence: Lab values follow clinically valid trajectories (e.g., HbA1c changes over months, not hours)
• Comorbidity realism: Disease co-occurrence patterns match epidemiological prevalence rates
• Drug interaction logic: Prescriptions respect contraindication rules and dosage guidelines

Example metric: Measuring the proportion of synthetic records with impossible combinations like pregnancy diagnosis codes assigned to male patients.

An evaluation paradigm measuring synthetic data utility by training predictive models entirely on generated EHRs and testing on real patient records. Workflow:

• Training phase: Train a mortality prediction or readmission model exclusively on synthetic data
• Testing phase: Evaluate performance on a held-out real patient cohort
• Utility ratio: Compare TSTR performance to Train-Real-Test-Real (TRTR) baseline

Interpretation: A TSTR/TRTR ratio > 0.9 indicates high utility; the synthetic data effectively substitutes for real data in downstream clinical ML tasks without exposing protected health information.

A generative model combining unsupervised GAN training with supervised autoregressive objectives to capture temporal dynamics in longitudinal patient records. Architecture components:

• Embedding network: Maps temporal features to latent representations preserving conditional distributions
• Recovery network: Reconstructs original sequences from latent embeddings
• Sequence generator/supervisor: Produces realistic temporal trajectories with stepwise transitions
• Sequence discriminator: Distinguishes real from synthetic time-series sequences

EHR application: Generating multi-year patient histories with realistic lab value trends, medication adherence patterns, and disease progression timelines.