Styxis
Company
Knowledge

Must medicines AI data wait for fully harmonised metadata?

Compare reuse paths by the definitions required for the intended purpose and the differences that must remain visible.

What to harmonise and what to interpret

One route to reusing medicines-regulatory information for AI is to wait for a common metadata model. Another is to align only the definitions needed for a particular purpose. The better route depends on that purpose. Truthound's role should likewise be a defined input boundary, not a declaration that all regulatory data has become interoperable.

The October discussion

EMA's published notes of the October 5, 2026 AI stakeholder meeting discuss machine readability, accessibility and interoperability. Participants distinguished evidentiary expectations by intended use and noted that some AI advances may reduce the need for fully harmonised metadata in particular cases. This was a discussion, not a demonstrated outcome or final guidance.

Two preparation paths

The first preparation path normalises sources into a common model. It can simplify later work but may erase a source distinction too early. The second preserves original definitions and maps only what one question requires. It starts smaller, while accepting continuing responsibility for differences behind apparently identical names.

Work backwards from intended use

Narrow the purpose before choosing. A document-discovery index and an adverse-reaction analysis input require different meanings. List source, document version, product identifier and context of use, then decide which differences are acceptable and which stop review. An identical product label should not silently establish equivalent clinical meaning.

The table Truthound can check

Truthound Core Python SDK 3.3.4, operating on a prepared Polars table, provides InSetValidator for supplied category lists and NotNullValidator for required fields. These can check an agreed context-of-use category and its evidence reference. The business owner supplies the allowed set. Populated fields establish one input condition, not the adequacy of evidence for a medicines decision.

Choosing the contract boundary

Soda Core 4.30.0 offers another path through YAML schema, missing-value and allowed-value contracts. A team already maintaining those on a supported source may improve its existing definition rather than migrate. Compare a prepared-table Python boundary with the established connection and YAML ownership. Soda v4's Elastic License 2.0 also needs consideration for the actual deployment form.

Less harmonisation still needs accountability

Using less harmonised metadata does not remove accountability. Ambiguity accepted during document discovery should not travel unchanged into a different analytical purpose. When the purpose changes, revisit definitions and review ownership. A successful small experiment is not an approval for every later use.

A failure that would reverse the choice

Prepare cases that could reverse the choice: the same product label with different versions, an unapproved use category, or a row without provenance. Examine where each path holds them. Accepting more rows is not inherently better; it may mean a distinction required by the decision was quietly removed.

Decide with a small reuse contract

The useful output can be one maintainable reuse contract rather than a universal model. Explain what meaning it preserves and what it leaves unresolved before adding another purpose. That grows the defensible scope of reuse instead of merely increasing the amount of metadata made to look alike.

Conditions

State intended use, provenance and version for lawfully accessible regulatory data.

Alternatives

An existing Soda Core 4.30.0 YAML contract on a supported data source can remain a candidate validation path.

Limitations

  • An interoperability discussion is not final regulatory guidance or evidence of clinical AI validity.

Evidence

Event checked
Subject checked
Modified