The primary goal for interface terminologies is to support the interaction between clinical users and structured representations of medical data, often by serving to support structured documentation into electronic health record systems. 5
Terminological attributes that maximize the efficiency of data entry may make interface terminologies more usable. Such attributes include a rich synonymy, a level of detail that matches the natural language common to relevant biomedical discourse and a balance between pre-coordination and post-coordination which facilitates searching for or composing terms when needed. However, these attributes may make the interface less able to provide formal knowledge representation required for clinical data exchange, aggregation, and analysis. 27
In contrast, reference terminologies are typically designed to provide a formal representation of medical knowledge and may be used to provide ontologic rigor and standardization to interface terminologies. A reference terminology would be expected to be able to represent a given interface terminology only if the two covered the same knowledge domain.
The current study characterized how well the concepts and semantic structures available in the reference terminology SNOMED CT could represent the concepts underlying a random selection of terms from two interface terminologies that have been used in clinical practice. The reviewers in this study found that SNOMED CT had excellent coverage for the concepts underlying the sample interface terms evaluated, with better coverage for those from CHISL than from MEDCIN. In the sample, the interface terms from MEDCIN were more complex in terms both of average degrees of freedom and implied semantics. That SNOMED CT concepts covered the study interface terms generally supports the NCVHS' call to map commonly used interface terminologies to standard reference terminologies such as SNOMED CT.
While coverage by SNOMED CT for the two interface terminologies CHISL and MEDCIN was high in the current study, it may actually be higher than measured by the reviewers. As in the example above, some interface terms did not match SNOMED CT terms, even though the underlying concept could be represented. Interface terminology users may benefit from having access to a rich synonymy 5
that allows them to represent clinical entities using the words or phrases that they prefer. A rich synonymy can improve the efficiency of searching a terminology for a needed term and may enhance the expressivity and accuracy of a document coded using the terminology. In the above example, reviewers did not find a concept to represent “hysterical dysbasia” even though it could be composed in SNOMED CT. If SNOMED CT is used to represent commonly used interface terminologies for data storage and aggregation rather than at the human-terminology interface, missing synonyms may not be a problem.
Overall concept coverage statistics in the current study were calculated without requiring that all necessary semantic linkages be present when complex interface terms required compositional expressions in SNOMED CT. When mapping terminologies, requiring that terms be mapped using both correct concepts and complete semantics may improve how completely and correctly the knowledge implied by interface terms is formally represented by the mapped reference terminology. However, requiring that semantics be used for all compositional expressions may reduce the number of interface terms that can be completely represented by an interface terminology and would increase the complexity of the mapping task. In the current study, such a requirement would have decreased coverage by up to 41% for MEDCIN and 33% for CHISL. For example, the study sample included the complex CHISL term above, “femoral artery systolic pressure over 40 mmHg above brachial systolic pressure,” which is synonymous with “Hill's sign,” and is not represented directly in SNOMED CT. Reviewers represented this concept using four concepts and two qualifiers from SNOMED CT, “Entire femoral artery,” “Systolic blood pressure,” “Over,” “mmHG,” “Entire brachial artery,” and “Systolic blood pressure.” The review software and human reviewers did not specify any semantic linkages for structuring the complex CHISL interface term, although it may be reasonable to construct the following semantic structure from SNOMED CT: [(271649006, Systolic blood pressure) HAS_FINDING_SITE (244332003, Entire femoral artery)] and [(271649006, Systolic blood pressure) HAS_FINDING_SITE (181322008, Entire brachial artery)]. An additional semantic linkage would need to be added to describe the relationship between the femoral and brachial artery blood pressures implied by this interface term. Without the added semantic linkage, the reference terminology representation for this interface term would be incomplete.
It is unclear the degree to which a complex interface term containing numerous atomic concepts needs complete semantics to be unambiguously modeled by a reference terminology. For example, reviewers could not fully model the semantic linkages implied by the interface term, “right thumb adduction”, but they could find the underlying concepts. It is likely that modeling the interface term to SNOMED concepts and semantic linkages, [(302540006, Entire thumb) HAS_LATERALITY (24028007, Right (qualifier value)] and (11554009, Adduction) is correct and unambiguous, even though not all the necessary semantic linkages were found. The presence of a semantic linkage between [(302540006, Entire thumb) HAS_LATERALITY (24028007, Right (qualifier value)] and (11554009, Adduction) would reduce ambiguity further, but it is not clear to what degree and with what incremental utility. Terminology evaluators should measure the incremental utility in real-world systems of having complete semantic representations of compositional expressions implied by complex interface terms. Likewise, reference terminology developers should enrich the set of available semantic linkages to ensure that interface terms and other natural language phrases can be completely and unambiguously covered.
The percent positive agreement among reviewers differed for the two interface terminologies, with agreement for both concept mapping and semantic categorization tasks lower for MEDCIN than for CHISL. The authors speculate that the lower rates of agreement for MEDCIN relate to two factors. First, in the current study, complexity measures for MEDCIN were higher than they were for CHISL (as above, MEDCIN had 3.8 degrees of freedom, on average, versus 1.8 for CHISL). Second, MEDCIN is much a much larger terminology than CHISL and as a result likely contains more specialized terms. These two factors would have increased the difficulty of the reviewers' task of finding the most appropriate concepts for covering the study terms.
The authors have previously speculated that developers may be able to enrich interface terminologies by using reference terminology concepts as a starting point. 5
With this approach, interface terms could be created either by identifying them or composing them from those contained in standard reference terminologies such as SNOMED CT. This method would require that developers and knowledge domain experts assemble clinically meaningful compositions, appropriate synonyms and linkages between concepts and related concepts or modifiers. While this approach may be somewhat labor-intensive, it may permit the underlying formal structure provided by the source reference terminology to remain, while simultaneously presenting clinicians with complex and meaningful interface terms. Starting with a single standard reference terminology such as SNOMED CT would permit a uniform back-end representation regardless of the diverse terms that users require.
The current study did not explore how well MEDCIN and CHISL function in their role as interface terminologies when used directly at the human-terminology interface, or the relative frequency with which specific terms in these vocabularies are actually used. Studies of interface terminologies, as the investigators have previously speculated, 5
should directly test synonymy, the degree of pre-coordination and ability to support post-coordination, and how well assertional medical knowledge is used to link together related concepts and modifiers. Other attributes, such as a consistent syntactic approach to constructing concepts' preferred terms, including attributes to support natural language generation, independence from the application that uses it and a formal semantic structure, may also enhance interface terminology usability. The investigators are not aware of any studies evaluating the fitness of MEDCIN or CHISL as interface terminologies according to these attributes.