The question is meaning

by Marina VasquezHead of Semantic Systems, HMND

A federated scientific data layer earns its keep only when it changes a choice. A query that crosses silos and comes back with enough semantic fidelity to move a trial, a cohort, or a batch release is decision infrastructure.

The part everyone keeps skipping

The useful pattern is a governed federation with a query layer at the center, identity managed at the source, and access rules enforced before the question reaches the data. The recent literature keeps landing on the same point from different angles. A federated semantic metadata registry, local RDF and triple store generation, and federated querying all depend on shared semantics and local governance, because the answer has to be trustworthy enough for clinical research and care, or for multi site clinical record retrieval.

That is the real intersection between pharma and IT. The software layer has to preserve provenance, keep the consent or access boundary intact, and map local terms to a shared model such as FHIR, OMOP, SNOMED CT, or LOINC so the result means the same thing in every site. If a study team asks for anemia across hospitals and one site maps lab values while another maps diagnosis codes, the federation returns noise dressed as coverage. Meaning has not been shared. Identity has not been earned.

A recent Swiss federated clinical dataset shows why this matters in practice. Participating hospitals keep semantic annotation local, store SPHN compliant RDF in dedicated triple stores, and rely on a common semantic model spanning millions of patients, billions of triples, and thousands of code mappings. That is a serious operational answer, because it lets teams ask one question across sites without pretending the local systems have magically become identical. The data stays where the law and the workflow require it to stay. The decision does not.

Where the effort stalls

The failure mode is usually a meaning problem wearing an integration badge. Teams build a connector, declare victory, then discover that sample, subject, encounter, and specimen were never aligned tightly enough to support the actual decision they wanted to make. The federation looks alive. The query result cannot be trusted.

The mapping layer is where the work lives. FHIR to OMOP guidance asks teams to identify the information needed for the study first, then align FHIR elements, test gaps against the ontology, preserve provenance, and only then populate the common model. SPHN’s connector story says the same thing in a cleaner accent, with local transformation, deidentification, validation, and globally unique identifiers holding the semantic line at the source. That is the kind of engineering that makes federation operational instead of decorative. It is not dashboard glitter. It is the machinery that decides whether a cohort definition survives contact with reality.

The choice that should drive the design

Discovery asks whether a target subgroup exists and whether the evidence is consistent enough to spend the next quarter on it. Clinical development asks whether eligibility, safety signals, or endpoint history can be assembled across institutions without breaking the consent boundary. Manufacturing asks whether batch, assay, and quality data can be traced across systems well enough to spot drift before release becomes a guess. Those are different questions, but they all depend on one thing. The federated answer has to carry identity, provenance, and shared meaning all the way through the query path.

A semantic MDR or a federated SPARQL layer sounds abstract until a team needs semantically equivalent elements from two EHR systems and needs to know they are equivalent for real, not just in a slide deck. That is where ontology and access policy stop being plumbing and start being the reason the scientist can act. Call it federation if you want. The useful version is closer to a negotiated truth service.

Curious thing, really. So many programs call this interoperability when what they want is decision confidence. If the federation cannot change a choice, it is just another place to admire the data from a safe distance.

If this kind of boundary work is on your desk, hello@example.com is where the conversation starts.