From Human-Readable Mappings to Machine-Verifiable SDTM Transformation Specifications with AI-Assisted Authoring

About this session

Part of Session 7, Track A: Foundational Standards - SDTM (Chair: Alana St. Clair, CDISC).

SDTM implementation relies on human-readable mapping specifications that are valuable for collaboration but difficult to verify consistently before execution. Downstream standards such as Define-XML and CORE validate final deliverables, yet many organizations lack a formal, machine-verifiable layer for the transformation logic itself.
We describe a specification-driven approach grounded in production experience at Merck processing approximately 4,000 SDTM updates annually. Transformation intent is captured as structured (machine-readable) specifications declaring source fields, target fields, approved versioned capabilities, and dependencies. A validation layer checks technical executability, governance compliance, and domain correctness before any pipeline runs. AI assists with authoring candidate specifications but remains subject to the same validation boundary as human-authored work.
The approach complements Define-XML and CORE by formalizing upstream transformation intent. We discuss the validation model, governed function registries, similarity-based specification retrieval achieving 60-80% baseline cross-study mapping reuse for standardized studies, controlled AI integration, and implications for CDISC standards alignment.

Speakers

• RM

  Rostislav Markov

  Amazon Web Services

  Speaker

• AC

  Alana St. Clair

  CDISC

  Chair