Semantic Data Is the Hidden Constraint in Radiopharmaceutical Networks

by Marina VasquezHead of Semantic Systems, HMND

GE HealthCare’s planned acquisition of SOFIE Biosciences would bring a United States network of 15 CMO sites operating 21 cyclotrons into its Pharmaceutical Diagnostics business, alongside a theranostics focused CDMO site. The visible asset is manufacturing capacity. The harder asset is shared meaning across every system that decides whether a radioactive product can reach the right patient before physics closes the door.

Radiopharmaceutical operations run on a cruelly precise clock. A product can be chemically correct, released according to procedure, and already unusable by the time a delivery exception reaches the scheduler. Short half lives make distance, production timing, transport, quality review, and patient readiness one coupled system.

A vague identifier can become a missed dose. The isotope decays while the systems argue about what they are looking at.

The noun that disappears between systems

Picture a morning handoff between a cyclotron site and a hospital. The manufacturing system calls the material “FDG.” The laboratory system records a lot number. The transport application knows a package identifier. The nuclear medicine department has a patient appointment tied to an accession number. The release record carries a different product code because quality assurance inherited a separate catalog.

Every system may be functioning exactly as designed. The network still lacks one answer to a basic question: which manufactured radioactive product, containing which isotope, from which batch, released under which decision, is traveling to which site for which patient window?

The entity that should exist is a stable radiopharmaceutical product identity. It needs relationships to its isotope, formulation, strength, presentation, manufacturing version, and regulatory context. “FDG” can remain a human friendly display name. It cannot carry the entire burden of identity.

The entity that should exist beside it is an isotope identity, with isotope species, activity units, calibration convention, reference time, decay model, and permitted use. Activity without reference time is numerically impressive and operationally incomplete. A dose of fluorine 18 is a moving quantity, which makes the timestamp part of the fact rather than decorative metadata.

A batch identity must connect the product to a specific manufacturing event, precursor material, equipment context, sampling record, test result, release decision, and expiry calculation. A batch is a collection of manufacturing-table rows that together describe one event.

Fifteen sites create a meaning problem

SOFIE’s network introduces geographic reach through 15 CMO sites and 21 cyclotrons. That density can support final mile access, yet every added site also adds local naming, local scheduling practices, local instrument identifiers, and local interpretations of “available.” A site may mean a licensed facility, a production room, a shipping origin, or a hospital facing address. Those are different entities wearing the same badge.

The semantic layer must represent site identity separately from the physical address and separately from the operating capability. A site can contain multiple cyclotrons. A cyclotron can have multiple target configurations. A manufacturing suite can produce several products under different validated processes. The graph should preserve those distinctions because routing depends on them.

A release identity also needs its own place. Release is a decision about a specific batch, at a specific time, under a specific procedure and authority. The release record should connect to test methods, results, deviations, electronic signatures, and the exact version of the specification that governed the decision. “Released” as a Boolean field is how audits acquire a sense of humor.

The delivery system then needs a delivery window identity, tied to origin, destination, route, handoff events, expected activity at administration, and contingency rules. A planned departure and a viable arrival are related facts. They describe different operational conditions.

Finally, the patient scheduling system needs a patient schedule identity that can connect an appointment slot to the required product, isotope, dose range, clinical protocol, site, and readiness state without exposing unnecessary patient data across every boundary. A semantic architecture should improve traceability while respecting the security boundary around protected health information. Identity must travel with the operational fact. Patient identity does not need to travel everywhere the fact travels.

A semantic layer has a job

A warehouse stores curated records for query and reporting. A dashboard displays selected measures. Both can be useful. Neither automatically gives a network shared meaning.

A semantic layer defines the entities, relationships, constraints, mappings, and vocabulary that let independent systems interpret a fact consistently. It can federate references to manufacturing execution systems, laboratory information systems, quality systems, transport platforms, enterprise resource planning systems, and scheduling applications without forcing every source into one giant migration project.

The practical implementation might combine an ontology, governed code lists, an identity registry, event contracts, and a knowledge graph that records provenance. The graph becomes useful when it can answer questions such as whether a released batch can satisfy a particular patient appointment after decay, transport duration, site capability, and current inventory are considered. A warehouse can answer a similar question only if someone has already encoded the meaning elsewhere. Calling the warehouse a knowledge graph does not make the missing identity appear. It merely gives the warehouse a more ambitious name.

Ontology governance supplies the discipline. Each important entity needs an owner, a definition, an allowed lifecycle, and rules for reconciliation when two sources disagree. Product identity belongs with product governance. Batch identity needs manufacturing and quality participation. Delivery windows require operations and logistics. Patient schedule semantics require clinical operations and privacy expertise. A committee can approve the terms, but an accountable owner must be able to change them and explain why.

The failure arrives quietly

The dangerous failure does not always look like a crashed interface. It can appear as a successful order with the wrong isotope mapping, a release event associated with a superseded batch code, or a transport plan that treats the appointment time as a preference instead of a hard clinical constraint.

A scheduler may see supply as available because the inventory system reports activity at calibration time. The delivery planner may calculate arrival using a different reference time. Quality may have released the batch, while the receiving site still lacks the release document associated with the identifier in its local system. Everyone sees green. The patient sees a cancellation.

Radiopharmaceutical logistics already have a narrow margin because decay makes time a material property of the shipment. Short lived isotopes may require production and distribution close to the patient, especially when the isotope’s half life limits the useful delivery window. The architecture therefore needs event time, reference time, location, and identity treated as first class data, with validation rules that fail early when those facts cannot be reconciled.

A useful rule would reject an order when the requested product has no resolvable isotope identity. Another would prevent release status from propagating when the batch identifier maps to more than one manufacturing event. A third would flag a delivery plan whose calculated arrival leaves insufficient activity for the clinical protocol. These are ontology rules expressed as safety controls. The database can refuse a dangerous ambiguity without a dramatic speech.

Integration needs a boundary, not another label

Acquisition integration often begins with interface maps. Teams connect the manufacturing system to the enterprise resource planning system, then connect quality to manufacturing, then add transport, then discover that each connection carries a different definition of product, batch, and status. The result is a polite mesh of contradictions.

An architect should place the semantic boundary before the interface boundary. Source systems keep their operational authority. The semantic layer records canonical identities, mappings, provenance, and relationships. APIs and event streams publish governed business facts such as batch manufactured, sample tested, batch released, shipment dispatched, shipment received, and appointment confirmed. Each event carries identifiers that can be resolved across the network, along with timestamps whose meaning is explicit.

Security belongs in the design rather than in the apology after deployment. Manufacturing and quality users may need batch and release relationships. Logistics may need origin, destination, package, activity, and timing. Scheduling may need product requirements and appointment status. Only authorized services should resolve the patient identity behind a schedule. The semantic model can separate a clinical appointment from the personally identifiable subject while preserving the operational link required to deliver the dose.

This is why adoption stalls. Teams fear that a canonical model will erase local reality, and they have seen enough “enterprise integration” projects to distrust any noun followed by a steering committee. Local terms can survive. Local identity can survive. What must disappear is the assumption that a local label can cross a safety boundary without a mapping, an owner, and a declared meaning.

GE and SOFIE have a concrete test

The planned SOFIE acquisition gives GE HealthCare a useful test of architectural seriousness. The network’s 15 CMO sites and 21 cyclotrons create a real federated operating environment, with manufacturing capacity distributed across locations and patient access dependent on time sensitive coordination. The first integration question should not be which dashboard shows network output.

Ask whether every product can be identified across catalog, manufacturing, quality, transport, and scheduling. Ask whether every isotope has a governed identity with a declared time basis. Ask whether a batch can be followed from precursor and production run through test, release, shipment, receipt, and administration. Ask whether a site capability can be queried without confusing a facility with an address. Ask whether the delivery window can be recomputed when a cyclotron run slips. Ask whether the patient schedule can state its product requirement without exposing more patient information than the receiving service needs.

Those questions sound almost embarrassingly basic. The expensive truth often lives there.

Semantic data becomes visible only when it fails. A missed dose, an unusable batch, and a shipment whose activity has decayed below protocol are all operational consequences of nouns that never received stable identity. If two teams cannot agree what a sample, batch, or product means, the network has a meaning problem wearing infrastructure as a disguise.

GE HealthCare can treat the acquisition as a collection of sites and machines, or it can treat it as a chance to establish a federated semantic backbone for radiopharmaceutical operations. The second choice requires fewer heroic reconciliations at 4:30 in the morning, which is a modest but underappreciated form of patient safety.

If your systems keep losing the identity of a product between release and delivery, hello@example.com is a reasonable place to start the conversation.