Semantic Clarification of the Representation of Procedures and Diseases in SNOMED CT

Transcription

1 773 Semantic Clarification of the Representation of Procedures and Diseases in SNOMED CT Stefan Schulz a, Udo Hahn b, Jeremy Rogers c a Department of Medical Informatics, Freiburg University, Germany b Language And Information Engineering Lab, Jena University, Germany c School of Computer Science, Manchester University, United Kingdom Abstract SNOMED CT is emerging as a reference terminology for the entire health care process. It claims to be founded on logic-based modeling principles. In this work we analyze a special encoding scheme for SNOMED disease and procedure entities, the so-called relationship groups which had been devised in order to avoid ambiguities in entity definitions. We show that these artifacts may represent hidden mereological relations. We also report discrepancies encountered between the defined semantics of many SNOMED CT entity terms and their intuitive meaning, and inconsistencies detected between the definition of some complex composed entities and the definition of their top-level parents. As a result we formulate recommendations for improvements of SNOMED CT. Keywords: SNOMED, Knowledge Representation, Logic 1. Introduction SNOMED Clinical Terms (SNOMED CT) is a huge clinical terminology constructed by merging, expanding, and restructuring the previous SNOMED version RT and the Clinical Terms Version 3 (former Read Codes). SNOMED CT contains around 364,000 concepts, 984,000 terms and 1.45 million defined relationships between concepts 1. In the coming years it will be deployed for routine usage in several countries (U.S., U.K., Denmark), and it is intensively being analyzed by medical terminologists and decision makers in many other countries. SNOMED CT is a concept-oriented controlled vocabulary which has been designed according to previously suggested criteria for computer-based medical terminologies [1]. SNOMED CT is described as a clinical reference terminology, i.e. a set of concepts and relationships that provides a common reference point for comparison and aggregation of data about the entire health care process [2]. SNOMED CT concepts belong to multiple is-a hierarchies and are related with one another by various semantic relationships such as is-a, has-associated topography, has-action, has-associated morphology. Interestingly, SNOMED CT explicitly encodes part-of relationships only between 1

2 774 anatomic entities [3]. There is no relation with which to describe partonomy between diseases or procedure entities. 2. Relationship groups 2 in SNOMED SNOMED NT and CT follow a formal semantics based on KRSS, an early description logic [4], and have used a terminological classifier for terminology development [5,6]. In the UMLS distribution, however, the relational format of the MRREL table requires SNOMED CT content to be compressed into the common OAV (object attribute value triplets) format. For a logic-based model such a representation is ambiguous because it (i) obscures which attribute-value pairs are sufficient for the entity3 definitions, (ii) lacks role quantifications, and (iii) does not indicate whether sets of related attribute-value pairs should be interpreted as disjunction, conjunction or optional. Table 1 gives an example: Table 1. Example of OAV (object attribute-value) format of SNOMED CT entities such as distributed via the UMLS Metathesaurus SNOMED Concept 1 SNOMED Relationship SNOMED Concept 2 Renal glomerular disease has_finding_site Kidney Renal glomerular disease has_onset Gradual onset Renal glomerular disease has_onset Sudden onset Since 2004, the UMLS Rich Release Format (RRF) has encoded further information in the MRSAT table, including which entities are defined or primitive, which OAV triplets are optional (qualifiers) and which are restrictions, and which relationship group each belongs to. The purpose of relationship groups such as introduced by SNOMED is best explained by an example: Removal of foreign body from the stomach by incision involves Stomach structure and Digestive structure as values of the attribute has_procedure_site, and both Incision and Removal as values of the attribute has_method. However, such a simplistic representation is ambiguous: it could also be interpreted as Removal of the stomach and incision of a foreign body. Relationship groups declare associations between sets of OAV triplets (see Table 2). Although each group has an integer value, this does not imply any temporal or other ordering between groups. The SNOMED CT Technical Implementation and Technical Reference makes the following statement about relationship groups: Relationships, for a concept that are logically associated with each other. The Relationship group field in the Relationships Table is used to group these rows together for a concept. As relationship groups occur in about 17,000 disease entities and 13,000 procedure entities, according to [5], this phenomenon constitutes a major issue in SNOMED CT. In [5], Spackman et al. propose a description logics representation for relationship groups, in which they are expressed by an anonymous relation, named rg. From an ontological point of view, the proposed solution is, however, rather obscure. In the following we therefore explore the possible semantics of SNOMED CT relationship groups. We show that some basic assumptions of SNOMED CT are ontologically problematic, and we propose a 2 also called role groups [5] 3 Many ontological assumptions of SNOMED CT are still unclear. E.g., different things like Foot, Absent Foot, Football (qualifier value), Europe, Love, mmol, Yin excess, Kiel Classification, are concepts in SNOMED CT. Hence we use for the sake of neutrality the term entity for what SNOMED names concept.

3 775 solution for clarification which will be mostly compatible with the current SNOMED CT architecture. Table 2. Entries in the SNOMED CT core relationships table for the entity : Removal of foreign body from the stomach by incision, using three relationship groups SNOMED Concept 1 SNOMED Relationship SNOMED Concept 2 RG Removal of Foreign Access Open Approach 0 Body from the Stomach Is A Removal of foreign body from 0 by Incision digestive system Is A Removal of foreign body from 0 stomach Is A Incision of stomach 0 Method Removal - action 1 Direct Morphology Foreign body 1 Procedure site-indirect Digestive structure 1 Method Incision - action 2 Procedure site Stomach Structure 2 3. Ontological Analysis of Relationship Groups We refer to the same parsimonious variant of description logics as used by [5]. Entity names are characterized by initial capital letters. They can be joined by the AND operator. As an example, the expression AcuteDigestiveSystemDisorder AND AcuteInflammatoryDisease denotes inflammatory diseases of the digestive system, i.e. the intersection of entities subsumed by AcuteDigestiveSystemDisorder with all those subsumed by the entity AcuteInflammatoryDisease (or the set of all entities subsumed by both). Relation symbols begin with lower case, e.g. hasassociatedmorphology. Roles are formed by a quantifier (here only the existential quantifier,, is used), a relation symbol, followed by a dot and an entity symbol. For example, hasassociatedmorphology.inflammation denotes the entity whose instantiation is the set of all individuals related to an instance of Inflammation by the relation hasassociatedmorphology. We can therefore rewrite the role group 1 and 2 entries in Table 2: (1) RemovalOfForeignBodyFromTheStomachByIncision IMPLIES rg.( hasproceduresite.stomachstructure AND hasmethod.incisionaction) AND rg.( hasproceduresite. DigestiveStructure AND hasdirectmorphology.foreignbody AND hasmethod.removalaction) Let us now look at the parent entities, RemovalOfForeignBodyFromDigestiveSystem and IncisionOfStomach 4. The first is an is-a descendent of RemovalProcedure, and the latter an is-a descendent of IncisionProcedure. Consequently, all instances of the entity RemovalOfForeignBodyFromTheStomachByIncision are instances of both IncisionProcedure and RemovalProcedure and as a result must therefore inherit the properties of both Incision and Removal. This is hardly imaginable: In this case, objects 4 As contained in the SNOMED CT sources from the UMLS, or visualized by the SNOMED CT Browser at

4 776 would be equally incised and removed. In a strict upper level ontology Incision and Removal are expected to be mutually exclusive. In reality, the surgeon first performs the incision and then the removal: Incision and Removal are two separate sub-procedures and so are properly not parents but parts 5 of the entity RemovalOfForeignBodyFromTheStomachByIncision. Fig. 1 gives a graphic outline of this procedure which begins with the incision of the wall of the stomach, followed by the removal action and the closure of the wound (the latter is not mentioned in the procedure definition). These time-dependent sub-procedures stand to the main procedure in a part-of relationship. This is concordant with the commonly accepted mereological (part-whole) view of actions and processes, which are, according to [8] characterized by time-dependent parts. Having this in mind it seems straightforward to re-interpret the relationship group attribute rg in (1) as the mereological primitive has-part: RemovalOfForeignBodyFromTheStomachByIncision IMPLIES has-part.( hasproceduresite.stomachstructure AND hasmethod.incisionaction) AND has-part.( hasproceduresite. DigestiveStructure AND hasdirectmorphology.foreignbody AND hasmethod.removalaction) (2) Similarly, for the parent entities we obtain: IncisionOfStomach IMPLIES has-part.( hasproceduresite.stomachstructure AND hasmethod.incisionaction) (3) RemovalOfForeignBodyFromDigestiveSystem IMPLIES has-part.( hasproceduresite. DigestiveStructure AND hasdirectmorphology.foreignbody AND hasmethod.removalaction) (4) Looking up the SNOMED CT hierarchy, we obtain exactly these definitions after replacing rg by has-part. At a first glance this seems strange, since the main rationale for relationship groups, viz. the avoidance of ambiguities, makes no sense, here. Entity names such as Incision of Stomach, suggest definitions without the has-part role: IncisingAStomach IMPLIES hasproceduresite.stomachstructure AND hasmethod.incisionaction (5) Fig. 1. Graphical Representation of the Process Removal of Foreign Body from the Stomach by Incision 5 Entity A has B as part is equivalent to the DL expression A IMPLIES has-part.b,

5 777 The semantic difference is the following: Whereas IncisingAStomach denotes the atomic procedure of performing an incision onto a stomach, SNOMED s IncisionOfStomach subsumes any complex procedure during which an incision of stomach is being performed. Analogously, RemovalOfForeignBodyFromDigestiveSystem subsumes any complex procedure during which a foreign body is extracted from the digestive system. Looking still higher up the SNOMED CT hierarchy, IncisionOfStomach is a child entity of IncisionProcedure, which is itself related to an Incision action by the relation has_method: IncisionProcedure IMPLIES rg.( hasmethod.incisionaction) (6) As indicated above, the semantics of rg may be improved to derive: IncisionProcedure IMPLIES has-part.( hasmethod.incisionaction) (7) An IncisionProcedure is, therefore, any procedure which has a part characterized by the enactment of an Incision. Only this broader definition justifies IncisionProcedure being the ancestor of nearly one thousand entities: so many distinct flavours of incision do not exist, but more than one thousand surgical procedures have an incision as part of their description. We have taken our example from the procedure branch of SNOMED CT. We could have used, as well, numerous examples from the disease / disorders branch, e.g. Acute Perforated Appendicitis, which is subsumed by both Inflammation and Perforation (in the above sense). However, there may be situations in SNOMED where the translation of rg as has-part would not be correct. For example, within the current SNOMED content, it is possible to construct the post-coordinated composition of an urgent swab of the left eye, by adding the urgent and left qualifiers as appropriate. The flattened (not role grouped) representation that would result would be: Entity 1 Relationship Entity2 isa Specimen from Conjunctiva specimensourcetopography Conjunctival Structure Urgent Swab of the Left Eye specimenprocedure Taking of Swab priority Urgent laterality Left Here it would then be impossible to tell whether the attribute Left should be applied to the swab or the eye. So, we might want to formalize: UrgentSwabOfLeftEye IMPLIES Specimen from Conjunctiva AND rg.( specimensourcetopography.conjunctivalstructure AND Laterality.Left) AND rg.( specimenprocedure.takingoffswab AND.Priority.Urgent) (8) 4. Conclusion Our analysis of relationship groups in SNOMED CT revealed weaknesses which motivated us to make some recommendations which would improve SNOMED CT in clarity and which would remove inconsistencies from the terminology. These suggestions

6 778 would encompass only minor modifications of the SNOMED CT architecture: Rename the relationship group attribute rg by has-part or has-subprocess where it appears between a complex process and its subprocesses (i.e. especially in the disease and procedure chapters of SNOMED CT). Make a clearer distinction between atomic entities (such as IncisionAction) and those entities which have atomic entities as parts (such as IncisionProcess). The present entity names are misleading. Finally, one has to take into account, there are scenarios in which the use of relationship groups seem adequate, without, however, corresponding to a mereological relation. A more detailed ontological inquiry of these cases is still due. 5. Acknowledgments This work was supported by the EU Network of Excellence Semantic Interoperability and Data Mining in Biomedicine (NoE ), cf. We also thank Ulrike Sattler (Manchester, UK) and Kent Spackman (Portland, OR, U.S.), for their helpful discussions. 6. References [1] Cimino JJ. Desiderata for controlled medical vocabularies in the twenty-first century. Methods of Information in Medicine, 1998: 37(4/5): [2] Spackman KA, Campbell K, and Cote RA. SNOMED RT: A reference terminology for health care. In Daniel R. Masys, editor, AMIA'97 - Proceedings of the 1997 AMIA Annual Fall Symposium, pp Philadelphia, PA: Hanley & Belfus, [3] Spackman KA and Reynoso G. Examining SNOMED from the perspective of formal ontological principles: Some preliminary analysis and observations. In Hahn U, Schulz S, and Cornet S, editors, KR-MED Proceedings of the 1st International Workshop on Formal Biomedical Knowledge Representation, Collocated with the 9th International Conference on the Principles of Knowledge Representation and Reasoning (KR 2004), pp Whistler, B.C., Canada, June 1, Bethesda, MD: American Medical Informatics Association (AMIA), Published via [4] Baader F, Calvanese D, McGuinness D, Nardi D, Patel-Schneider P. The Description Logic Handbook. Theory, Implementation and Applications. Cambridge, U.K. Cambridge University Press. [5] Spackman KA, Dionne R, Mays E, and Weis J. Role grouping as an extension to the description logic of ONTYLOG, motivated by entity modeling in SNOMED. In Kohane IS, editor, AMIA Proceedings of the Annual Symposium of the American Medical Informatics Association, pp Philadelphia, PA: Hanley & Belfus, [6] Spackman KA and Campbell KE. Compositional entity representation using SNOMED: Towards further convergence of clinical terminologies. In Chute CG, editor, AMIA'98 - Proceedings of the 1998 AMIA Annual Fall Symposium, pp Philadelphia, PA: Hanley & Belfus, [7] Schulz S and Hahn U. Medical knowledge reengineering: Converting major portions of the UMLS into a terminological knowledge base. International Journal of Medical Informatics 2001: 64(2/3): [8] Simons P. Parts: A Study in Ontology. Oxford: Clarendon Press, Address for correspondence PD Dr. med. Stefan Schulz, Abteilung Medizinische Informatik, Universtätsklinikum Freiburg Stefan-Meier-Str. 26, D Freiburg (Germany), stschulz@uni-freiburg.de,