Clinical Data Mappings

An illustrative end-to-end pipeline

How CDASH-like collection, SDTM-like tabulation, ADaM-like analysis, FHIR exchange, and OMOP analysis relate — and where they do not.

The problem

People talk past each other with acronyms. A warehouse engineer is asked to “make it CDISC” and also “put it in OMOP” and also “accept FHIR from the hospital.” Those are different pipelines that can share identifiers, not one automatic conversion.

A working example

A public, widely used research-shaped path (also sketched in pharmaverse and official CDISC materials):

  1. Collection — forms designed with CDASH in mind (what is asked at the visit).
  2. Tabulation (SDTM) — collected data reorganized into submission/standard domains (DM, AE, LB, VS, …). Official IGs govern this.
  3. Analysis (ADaM) — ADSL and analysis datasets derived from SDTM, then tables, listings, and figures. Official ADaM materials govern this.
  4. Traceability — each step can point at the previous one. See traceability.

A public real-world / EHR-shaped path:

  1. Exchange (FHIR) — Patient, Encounter, Observation, AdverseEvent, and related resources leave the EHR.
  2. Research tabulation (FHIR → SDTM) — official FHIR to CDISC IG for the domains it covers; then the usual SDTM discipline.
  3. Observational analysis (FHIR → OMOP) — official FHIR to OMOP IG plus CDM v5.4.

The synthetic subjects on this site (SYN-001, SYN-002) were walked a short way down both research-shaped and EHR-shaped paths: source rows → SDTM-shaped DM/AE, and source rows → OMOP-shaped PERSON / CONDITION_OCCURRENCE with blank concept ids. That is a demo of grain and precision, not a full pipeline.

The gotcha

SDTM and OMOP are not interchangeable warehouses. One is a clinical-trial tabulation model; the other is an observational CDM with vocabulary-standardized events. You can map between them only with an explicit, official-or-specified project. There is no honest SELECT … INTO omop FROM sdtm.

FHIR is not a warehouse model. It is an exchange model with nested, repeated elements. Persistent analytic tables (SDTM-like or OMOP-like) require flattening decisions. See FHIR to SDTM.

Licensing cuts across both paths. MedDRA, some drug dictionaries, and some code systems cannot be redistributed just because your pipeline needs them. Leave them in licensed stores; see MedDRA levels and vocabulary process.

When not to use this

Verification

The two-path sketch is a reading of public CDISC, HL7, OHDSI, and pharmaverse materials linked from this site. The synthetic rows were produced in DuckDB 2.0.0. Nothing here is a complete or validated pipeline.

Official / public docs