Options for Drugs Product Role

Background

The modeling of dispositions and roles has been under discussion for ten years now, mainly as part of the substance hierarchy redesign effort. It is true that the involved patterns are not limited to substances and products, and therefore consistent adoption of an Ontology Design Pattern (ODP) would be desirable.

Early work on this area has been influenced by the approach to ontology normalization proposed by Alan Rector to "disentangle" ontologies:

• Modularisation of Domain Ontologies Implemented in Description Logics and related formalisms including OWL

• Normalisation of ontology implementations: Towards modularity, re-use, and maintainability

Some of the normalization principles discussed in those papers were difficult to implement because of the complex nature of clinical data in SNOMED CT, but other principles were applicable. Among them:

• Reduce/avoid Is A relationship overloading, using a primary partition criteria for standalone concepts and maintain poly-hierarchies by modeling (partitioning) properties and

• Leverage the assistance of a reasoner to maintain inferred supertypes (instead of entangling the hierarchy with manual assertions that fix them and limit reasoning assistance).

Later work on the Substance Hierarchy Redesign (SHR) project suggested that a pattern for modeling roles should be considered. Initially this was influenced by the concepts described in the OntoClean methodology in the early 2000's. Attention was given at that time to the anti-pattern "roles subsuming types" (using OntoClean categories).

• Examining SNOMED from the Perspective of Formal Ontological Principles: Some Preliminary Analysis and Observations

• An overview of OntoClean

The distinction between dispositions and roles was introduced in the redesign discussions after influential work from the Basic Formal Ontology thread:

• Function, Role, and Disposition in Basic Formal Ontology

• Realizable Entities in Basic Formal Ontology

• Representing dispositions

In the last few years, there has been growing agreement that dispositions could be represented in the substance hierarchy, as a secondary "partitioning" criteria implemented with sufficiently defined groupers, a disposition hierarchy under Qualifier values, and a "has disposition" property. An initial iteration of this substance content remodeling approach has been delivered as part of the January 2018 release. Further cleanup of the substance groupers will be delivered in future iterations.

A major problem in the substance hierarchy has been the presence of therapeutic groupers that represent therapeutic roles or applications of certain substances. Some of those therapeutic role groupers are heavily used as targets of defining relationships in other hierarchies. Therefore, while there is consensus that (eventual) therapeutic roles would be better modeled in the products hierarchy rather than in substances, the transition would require at least three steps:

1. Modeling of therapeutic roles as groupers in the Products hierarchy

2. Changes to the concept model to use products instead of substances as targets of defining relationships, when appropriate

3. Inactivation of the therapeutic roles in the Substances hierarchy (a few exceptions might remain if absolutely necessary)

Note that "product disposition groupers" are already handled in the product model by relating them to the corresponding substance disposition groupers. So groupers like "Beta-blocker product" would be sufficiently defined as "Product with beta-adrenergic receptor antagonist mechanism of action" or similar depending on the final naming conventions. In the substances hierarchy, they are modeled using properties (has disposition) and sufficiently defined groupers. Given that most of the substance content will not be sufficiently defined, the side effect of modeling a realizable as part of the default set of definitions has limited impact.

Therefore, the modeling of dispositions in Substances and Products is not the main focus of this thread given a model is already being implemented. We recognize though that they are also a kind of realizables, and further discussion might be required to apply these concepts to other semantic areas.

It has been decided that the therapeutic roles (like oral anticoagulant, for example) would be handled at an appropriate product hierarchy level (MP, MPF, or CD), based on the following requirements and constraints:

NOTE: the diagrams were produced based on existing diagrams and customized with Powerpoint templates. They include some inconsistencies, like not representing attributes that are outside of the scope of this discussion. In addition, the proximal primitive for all these concepts would be "Medicinal product" rather than "Pharmaceutical / biologic product". A new set of diagrams would be needed for the final documentation of the selected alternative. Native diagrams would be extracted from the authoring tool once the definitive pattern is represented or tested in the development environment.

Option 1 - Add "IS A" in a new axiom

Therapeutic Role parents in the Core Module for those concepts required to model other hierarchies, but comphrensive coverage is out of scope. National extensions could add additional roles per local requirements. 

For example, a product can be defined by its active ingredient of Timolol and manufactured dose form of eye drops. The additional definition is the product role as an anti-glaucoma preparation.

Open

Advantages and disadvantages:

The product therapeutic groupers are modeled as primitive concepts without properties, eventually forming a hierarchy under "Product categorized by therapeutic role".

• The advantage would be that there would not be a need to model a parallel hierarchy of role qualifiers to sufficiently define them.

• This might also be a disadvantage if a uniform pattern is to be adopted for modeling roles in SNOMED CT, as those roles would be represented in different hierarchies and be specific to them.

The therapeutic grouper is represented as the target of an asserted Is A relationship modeled in a separate subclass axiom

• A disadvantage could be that lower levels of the hierarchy might not represent the implication in the concept inferred view, as there would be more specific subsumers that would "hide" the inheritance of that parent as it is already implied in the reduced transitive closure. To find the grouper the user would need to navigate up in the hierarchy.

This alternative supports the requirements described in the background section.

The inferred form for this concept would look like: (simplified representation not considering intermediate concepts)

Open

The additional axiom is lost and the "Is a" relationship becomes part of the only one definition set represented in the relationship table inferred view. However, computing equivalence or subsumption is expected to be performed against the stated form represented in the OWL refset, that has both axioms explicitly represented.

An important consideration is that clinical drugs below this level would not inherit the "anti-glaucoma preparation "Is a" because if would be implied and removed as redundant during the Distribution Normal Form computation. While it is possible to eventually consider changes to the DNF computation, those would need to be discussed in a separate thread as several technical options would need to be considered beyond the content-centric focus of this thread.

Clinical drugs that would classify under the above concept would have an inferred form like:

Open

Option 2 - Add new role attribute in a new axiom

This pattern could be implemented with two variations (diagrams represent asserted/stated forms):

2a. Adding only the new role attribute in a new axiom:

Open

This modeling pattern (just specifying an existential restriction as the only component of the second axiom) is supported and natural in OWL.

However, there is an implied supertype (Thing, and in SNOMED CT, the root concept). It would mean any SNOMED CT concept that plays this role.

Therefore, it might be good practice to restrict the supertype to the proximal primitive.

2b. Adding a proximal primitive parent and the new role attribute in a new axiom

Open

Both 2a and 2b assume that the therapeutic grouper would be modeled as a sufficiently defined concept like:

Open

And also both assume that the hierarchy of roles are modeled in the Qualifier value hierarchy (this pattern is very similar to the one used for Dispositions in substances, the only difference is that the assertion is placed in a separate axiom.

Open

For both 2a and 2b, the inferred form would look like: (simplified representation not considering intermediate concepts)

Open

However, descendants of this product would inherit the attribute from the additional axiom, flattened into a single set in the inferred form represented in the relationships table:

Open

Advantages and disadvantages:

The product therapeutic groupers are modeled as sufficiently defined concepts with a "plays potential role" property. The arrangement of this kind of product groupers would depend on the hierarchy of target values represented in the qualifier value hierarchy. This pattern is very similar to the one already in use for substance dispositions. The grouper "Product categorized by therapeutic role" could be fully defined in this case.

• The disadvantage would be that there would be a need to model a parallel hierarchy of role qualifiers to sufficiently define the the product role groupers.

• An advantage is that an uniform pattern is adopted for modeling roles and dispositions in SNOMED CT

The therapeutic grouper would be inferred as a parent after classification. As in all other alternatives, the inferred form is flattened into a single set of definitions.

• An advantage is that lower levels of the hierarchy will inherit the attribute in the concept inferred view, making the identification of assigned roles easier in an authoring environment. To find the grouper the user would still need to navigate up in the hierarchy if the inferred parent is more than one level above.

This alternative supports the requirements described in the background section.

Option 3 - New attribute is non-defining (eg "additional" rather than stated or inferred)

This alternative is not supporting the requirements described above as these assertions would not affect classification and would require manual maintenance. In addition, non-defining relationships should be deprecated once the "Part of" relationships become defining ones again. Any non-defining assertions could be represented in association refsets.

Option 4 - Specify role using a map Refset

This alternative ignores that we need some of the product role groupers represented as concepts in SNOMED CT that could serve to define other concepts. It is not an option from the modeling point of view. However, this alternative should be considered and recommended for extensions willing to adopt other therapeutic classifications that would not affect the core content maintenance (e.g. would not be referenced in SNOMED CT concept definitions). For example, we might need to model the concept "Poisoning due to anti-depressants" using a SNOMED CT concept as a target. While it would be possible to map some of the therapeutic role product groupers to other domain terminologies, their representation might be different enough to be represented with mappings. For example, in ATC:

N06A ANTIDEPRESSANTS

• N06AA Non-selective monoamine reuptake inhibitors

• N06AB Selective serotonin reuptake inhibitors

• N06AF Monoamine oxidase inhibitors, non-selective

• N06AG Monoamine oxidase A inhibitors

• N06AX Other antidepressants

Some of the ATC categories would be therapeutic roles, while others would be disposition groupers. Residual categories are an aggregation pattern not aligned with description logics-based terminologies.

Option 5 - Relationships from different modules into separate axioms.

This workaround was discarded after the decision to migrate the stated definitions to a versioned OWL representation. Therefore, at the core level it is no longer necessary, as starting with the July 2018 release it would be possible to add additional axioms and support multiple definition sets. Following practices common to OWL imports of dependencies on other terminologies, assertions created in extensions would be considered part of extension-specific axioms rather than modifying the core axioms. Therefore, this mechanism is also potentially useful for extensions that would like to extend the International edition groupers with greater granularity. In that case, they can use those new concepts as targets of relationships defining other concepts in the hierarchy.

This option has been objected to by ADHA (Australia) who rely on the current flattening of stated relationships into a single axiom to allow them to make required changes to core concepts.  If any relationship they added to a core concept was put into a new separate axiom, it would not be possibly to augment existing axioms. However, later discussions on SNOMED CT extensibility accomodate both options (flattening is not recommended but is allowed for countries that have adopted that alternative in the past). The reason for not recommending flattening is that change requests should be submitted to the International Edition, so it is expected that future evolution to continuous release and collaborative editing would not requiring the modification of core content in an extension with a potential impact on interoperability.

Comparison table