Clinical Plausibility

Validates that synthetic data respects hard biological boundaries and known correlations. This prevents the generation of impossible clinical scenarios.

• Lab Value Ranges: Serum potassium must fall within 3.5-5.0 mEq/L; a value of 15.0 is physiologically impossible.
• Demographic Consistency: A synthetic record cannot list a patient as both male and pregnant, or assign a pediatric weight to an adult.
• Temporal Logic: A diagnosis of metastatic cancer must chronologically follow a primary cancer diagnosis, not precede it.

Ensures all clinical concepts in the synthetic data are represented with valid, non-ghost codes from standardized medical ontologies like SNOMED CT.

• Hierarchical Integrity: A generated concept for 'Type 2 Diabetes Mellitus' must correctly roll up to 'Disorder of endocrine system' and not 'Injury of foot'.
• Post-Coordination Validation: When combining concepts (e.g., 'severe' + 'asthma'), the resulting expression must be a valid, logical combination within the ontology's compositional grammar.
• No Semantic Drift: Prevents the model from hallucinating a non-existent SNOMED code like 123456789.

Evaluates the temporal sequence of clinical events against known disease natural history. Synthetic patient trajectories must mirror realistic pathways.

• State Transition Validity: A patient cannot transition directly from 'Healthy' to 'Death' without an intervening pathological state.
• Temporal Density: The rate of clinical events (e.g., lab tests, procedures) should match real-world cadences for a given condition, avoiding both impossibly sparse and unrealistically dense timelines.
• Treatment-Response Logic: A synthetic patient with a bacterial infection who is generated with an antibiotic order should show a subsequent improvement in relevant lab markers (e.g., decreasing white blood cell count).

Checks for logical consistency across different data types within a single synthetic patient record, ensuring the structured data, narrative text, and imaging findings tell the same clinical story.

• Text-to-Data Alignment: A synthetic radiology report mentioning 'left lower lobe consolidation' must be reflected in the structured diagnosis codes for pneumonia.
• Lab-Imaging Correlation: A generated record with a troponin level of 50,000 ng/L (indicative of a massive heart attack) must not be paired with a synthetic chest X-ray reported as 'normal'.
• Medication-Reason Pairing: A synthetic prescription for insulin must be linked to a diagnosis of diabetes mellitus, not hypertension.

Assesses whether the synthetic cohort mirrors real-world prevalence, incidence, and comorbidity patterns without introducing spurious correlations.

• Prevalence Matching: The rate of a rare disease like Amyotrophic Lateral Sclerosis (ALS) in the synthetic dataset should approximate real-world prevalence (~5 per 100,000), not 10%.
• Comorbidity Preservation: Known associations, such as the high co-occurrence of hypertension and chronic kidney disease, must be maintained with realistic relative risk ratios.
• Simpson's Paradox Avoidance: The model must not create a trend in a synthetic subpopulation that reverses or disappears when the data is aggregated, unless that paradox is a documented feature of the real-world condition.

Validates synthetic medication data against drug knowledge bases for dosing, interactions, and contraindications.

• Dosage Form Logic: A synthetic order for 'insulin 100mg tablet' is invalid because insulin is not orally bioavailable and is dosed in units, not milligrams.
• Contraindication Checking: A synthetic record must not contain a prescription for a beta-blocker in a patient with a generated diagnosis of severe bradycardia.
• Age-Appropriate Dosing: A synthetic pediatric record should not contain an adult dose of medication; weight-based calculations must be logically consistent.

The process of mapping synthetic clinical concepts to Systematized Nomenclature of Medicine Clinical Terms, ensuring generated data uses standardized medical vocabulary. Plausibility checks verify that diagnoses, procedures, and findings adhere to the hierarchical is-a relationships and defining attributes within the ontology.

• Validates concept existence and semantic type consistency
• Prevents generation of nonsensical term combinations
• Enables interoperability with real-world EHR systems

A rule-based verification layer that enforces hard biological boundaries on synthetic data. This includes checking that laboratory values fall within survivable ranges, anatomical relationships are preserved, and temporal sequences respect biological causality.

• Example: Systolic blood pressure cannot exceed diastolic pressure
• Example: Hemoglobin A1c values must remain positive
• Example: Medication administration must precede its documented effect

Evaluation of whether synthetic patient trajectories follow clinically validated temporal patterns. This involves verifying that disease stages progress in the correct order, comorbidities cluster realistically, and treatment sequences align with established clinical guidelines.

• Uses Markov models and clinical pathways as ground truth
• Detects implausible reversals or skipped disease stages
• Validates time-to-event distributions against real-world registries

A generative model that incorporates causal structure and do-calculus to generate counterfactual data points. Unlike purely correlational generators, causal models preserve the underlying mechanisms of disease, enabling the simulation of interventions while maintaining clinical plausibility.

• Removes confounding biases from synthetic datasets
• Enables 'what-if' clinical scenario generation
• Validates that treatment effects follow causal pathways

An evaluation paradigm where a model is trained entirely on synthetic data and tested on real data. If the synthetic data lacks clinical plausibility, the model will fail to generalize to real patient populations, providing a practical utility benchmark.

• Measures whether synthetic data can substitute for real data
• Exposes gaps in physiological constraint preservation
• Preferred metric when original data cannot be shared