Synthetic OMOP cohorts¶

openmed.training.synthetic.omop_cohorts maps explicitly synthetic-marked existing ClinicalFact graphs into coherent OMOP-shaped person, visit, condition, drug, measurement, procedure and observation tables. It also renders synthetic source notes with actual content hashes and stable note/event links, using the existing OMOP cohort checker for relational/concept/provenance validation.

generate_omop_cohort(facts, seed=..., cohort_condition_codes=...) derives positive BIGINT identities from graph identity and seed. Missing visits, cross-person encounters, duplicate facts, invalid dates and future causal parents reject. Visits span the declared synthetic effective event dates. Concept records use explicit OPENMED_SYNTHETIC placeholders with no standard-concept claim; no Athena, UMLS, SNOMED or CPT content is distributed. Person demographic values are fictional fixture defaults, with birth year 30 years before the earliest visit (clipped at year 1) so date-shifted early visits cannot precede birth. Numeric measurement truth remains None when the graph contains only a code, rather than being invented.

The artifact contains tables and gold cohort membership. Membership uses the intersection of all requested controlled condition codes, calculated from declared synthetic condition facts. validate_synthetic_omop independently checks existing runtime relational rules, visit/event chronology, synthetic markers, actual note hashes and cohort truth. Orphan people, mismatched visits, reversed dates and altered membership are separate negative controls.

Only 1,000 graph facts and 128 criterion codes are accepted; validation is capped at 10,000 rows. No database write, network, real-record import or restricted terminology lookup occurs. These are structural/utility test fixtures, not clinical plausibility, full CDM conformance, disease progression or treatment benefit assurances. Cohort flags are synthetic gold labels, never clinical eligibility recommendations.