-Objectives. To identify missing hierarchical relations in SNOMED CT from logical definitions based on the lexical features of concept names. Methods. We first create logical definitions from the lexical features of concept names, which we represent in OWL EL. We infer hierarchical (subClassOf) relations among these concepts using the ELK reasoner. Finally, we compare the hierarchy obtained from lexical features to the original SNOMED CT hierarchy. We review the differences manually for evaluation purposes. Results. Applied to 15,833 disorder and procedure concepts, our approach identified 559 potentially missing hierarchical relations, of which 78% were deemed valid. Conclusions. This lexical approach to quality assurance is easy to implement, efficient and scalable.
SNOMED CT; quality
Quality assurance of large biomedical terminologies
remains an active area of research [
Both lexical features and logical definitions have been used
for quality assurance purposes. Approaches based on lexical
features generally exploit the presence of specific words in
SNOMED CT terms or contrast sets of words for terms across
concepts to suggest relations among concepts (e.g., [
The objective of this investigation is to identify missing hierarchical relations in SNOMED CT from logical definitions based on the lexical features of concept names. More specifically, we propose to leverage description logics for representing the lexical features of concept names and infer hierarchical relations based on these lexical features with a reasoner. The hierarchical relations inferred from lexical features but not present in SNOMED CT are candidates for missing relations.
Developed by the International Health Terminology Standard Development Organization (IHTSDO), SNOMED CT is the world’s largest clinical terminology. With 320,000 active concepts, it provides broad coverage of clinical medicine, including findings, diseases, and procedures for use in electronic medical records [9].
SNOMED CT provides a preferred name and synonyms for each concept (“descriptions” in SNOMED CT parlance). The “fully specified name” is guaranteed to be unique for each concept and consists of the preferred term followed by a semantic tag (e.g., Blepharorrhaphy (procedure) (388008)). In addition to names, all concept have a logical definition, based on definitional characteristics of the concept (not on the lexical features of the concept names). For example,
EquivalentTo:
and (Method some Closure - action) and (Procedure site - Direct some Structure of palpebral fissure)
and (Using device some Surgical suture, device)
In SNOMED CT, the logical definitions are processed with a description logic reasoner for consistency validation and to generate the hierarchical structure by inferring subClassOf relations among the concepts.
The version of SNOMED CT used in this work is the U.S. edition dated March 2016.
Description logics (DL) are a family of knowledge
representation languages often used as ontology languages, and
defined as a trade-off between expressivity and tractability
[
Among the various flavors of DL languages available, the
EL family offers sufficient expressivity for the simple
definitions resulting from lexical features, as well as scalability
to a large number of classes [
As illustrated above, SNOMED CT relies on DL for representing the logical definitions it provides for its concepts. It also makes use of a reasoner for testing the consistency of these definitions across the whole ontology, as well as for inferring the hierarchy of concepts. In this work, we apply the reasoner not to the logical definitions provided by SNOMED CT to represent biomedical knowledge, but rather to the definitions we generate from the lexical features of the terms of SNOMED CT concepts.
Approaches to quality assurance in biomedical ontologies
can be classified into lexical, structural and semantic
approaches [
Of note, while DL techniques are generally used in the context of semantic approaches, in this work, we leverage a DL reasoner for the implementation of a lexical approach to QA, since our logical definitions are created on the basis of lexical features.
The compositionality of terms in biomedical ontologies is
well documented and has been exploited for quality assurance
purposes (e.g., [
The specific contribution of this work is not in leveraging the compositionality of biomedical terms for suggesting relations, but rather in proposing a description logics approach to doing so. While ad hoc programming is usually necessary for comparing bags of words, our work demonstrates it can also be supported effectively by a DL reasoner. To our knowledge, this is the first attempt to generate logical definitions based on the lexical features of concept names in SNOMED CT for quality assurance purposes.
Our method for identifying missing hierarchical relations from SNOMED CT can be summarized as follows. We first create logical definitions from the lexical features of concept names, which we represent in the web ontology language, OWL. We infer hierarchical (subClassOf) relations among these concepts using a reasoner. Finally, we compare the hierarchy obtained from lexical features to the original SNOMED CT hierarchy. We review the differences manually for evaluation purposes. In this preliminary investigation, we applied this approach to a significant subset of the Clinical Finding hierarchy rooted with the concept Disorder of head (disorder) (118934005) and a smaller subset of the Procedure hierarchy rooted with the concept Operative procedure on head (procedure) (89901005).
For each concept under investigation, we extract the fully
specified name, which consists of the preferred term (e.g.,
“Disorder of head”) followed by a semantic tag in parentheses
(e.g. “disorder”). For each concept C with fully specified name
“w1 w2 … wn (T)”, where {w1, w2, … wn} is the set of words in
the preferred term and where T is the semantic tag, we create a
logical definition of the following form (expressed in the
simplified OWL syntax known as Manchester syntax [
T and (has_word some w1) and (has_word some w2) … and (has_word some wn)
For example, the class definition for the concept Complete ablepharon (disorder) (708541009) is shown in Fig. 1.
In practice, we use a simple script to create an OWL file that contains the class definitions for all the concepts under investigation. The words “the” and “of”, present in a large proportion of terms, are omitted when generating the class definitions.
Of note, the OWL constructs used in these definitions
(namely class equivalence and existential quantification to a
class expression) are compatible with the OWL 2 EL profile
[
Fig. 2. Asserted hierarchy – Ablepharon (disorder) prior to running the reasoner (no inferred subclasses)
We load this OWL file in the Protégé ontology editor (5.0
beta), in which we have installed the plugin for the ELK
reasoner [
(disorder) are subclasses of Ablepharon (disorder) in the inferred hierarchy (Fig. 3).
Since the subClassOf relations are inferred from lexical features, we need to filter out complex terms with prepositional phrases to avoid generating wrong subClassOf relations. For example, for Dementia due to Parkinson's disease (disorder) (101421000119107), a subClassOf relation is inferred to both
(disorder) (49049000). Similarly, for Goniopuncture without goniotomy (procedure) (202727004), a subClassOf relation is inferred to both Goniopuncture (procedure) (265293008) and Goniotomy (procedure) (265292003). While this behavior is expected from the reasoner, it is not desirable, because
prepositional expression “due to”. Similarly, Goniopuncture without goniotomy (procedure) specifically excludes Goniotomy (procedure). In practice, to avoid generating such wrong subClassOf relations, we filter out the relations generated when the name of the most specific (“child”) concept contains any of the following words: “and”, “or”, “and/or”, “with”, “without”, “from”, “due to”, “secondary to”, “except”, “by”, “after”, “revision” and “ligation for”.
To analyze which relations from the inferred hierarchy are not already in the original SNOMED CT hierarchy (i.e., the hierarchy found in the SNOMED CT distribution), we need to generate these two sets of hierarchical relations and compute the difference between them. Using Protégé, we export the inferred subClassOf axioms to a file in RDF format for comparison to the original hierarchical relations in SNOMED CT. Using a simple script, we write the original hierarchical relations in SNOMED CT to RDF for the subhierarchies under investigation. In practice, because the inferred relations can be between any two classes, we enrich the original hierarchy with the transitive closure of subClassOf relations. We load the files for the two sets of relations, inferred and original, into the triple store Virtuoso and use a SPARQL query to compute the set of hierarchical relations from the inferred set that is not part of the hierarchical relations originally in SNOMED CT (transitively closed). The SPARQL 1.1 operator MINUS makes such comparison between two graphs extremely easy.
We manually review for validity a random subset of 100 inferred relations that are not present in the original SNOMED CT hierarchy (transitively closed).
IV. RESULTS
We created logical definitions based on the lexical features of concept names for the 12,088 concepts (4871 distinct words) of the subhierarchy rooted with the concept Disorder of head (disorder) (118934005) and for the 3795 concepts (1899 distinct words) of the subhierarchy rooted with the concept
Running the ELK reasoner took a few seconds and resulted in the creation of 7079 inferred subClassOf relations among the concepts of the subhierarchy rooted with the concept Disorder of head (disorder). Similarly, 1357 relations were inferred in the subhierarchy rooted with the concept Operative procedure on head (procedure).
After subtracting from the inferred subClassOf relations created by the reasoner those subClassOf relations already present in the original version of SNOMED CT (transitively closed), we obtained 1210 inferred subClassOf relations for the
subClassOf relations for the Operative procedure on head (procedure) hierarchy. Of these, 469 subClassOf relations for disorders and 90 for procedures met our criteria for review (i.e., the name of the child concept does not contain any of the prepositional and other expressions listed earlier).
The random subset of 100 inferred subClassOf relations we reviewed comprises 83 disorders and 17 procedures. Overall, 78 relations were deemed valid, 19 invalid and 3 questionable (i.e., these relations seem to have face validity, but may not be compliant with SNOMED CT editorial policies). Examples of such relations are listed in Table I.
As expected, a vast majority of the hierarchical relations suggested lexically were already present in the original SNOMED CT hierarchy (transitively closed). Specifically, only 1210 of the 7079 hierarchical relations for disorders (17%) and 242 of the 1357 hierarchical relations for procedures (18%) were not already represented in SNOMED CT.
However, it was somewhat surprising to us to see that a large number of potentially missing hierarchical relations had been generated from this simple technique based on lexical features. Assuming 80% of the 559 hierarchical relations generated are correct, we discovered 447 missing hierarchical relations among the 15,883 concepts under investigation. Interestingly, the proportion is roughly the same for disorders and procedures.
In addition to the evaluation, we performed a cursory review of the 559 potentially missing hierarchical relations, among which we identified a few patterns. In 31 cases, the missing relation was between “carcinoma in situ of <some anatomical structure>” and “carcinoma of <some anatomical structure>” (or “<some anatomical structure> carcinoma”), for example, between Carcinoma in situ of palate (disorder) (92670007) and Palate carcinoma (disorder) (274084007). Another such patterns was found in 23 cases between “congenital <some disorder>” and the unqualified disorder, for example, between Congenital anterior staphyloma (disorder) (253230008) and Anterior staphyloma (disorder) (231888000).
The novel aspect of this work is to use a DL approach to lexical similarity. In practice, it means that no ad hoc programming is required for identifying partial ordering relations among sets of words for terms in an ontology reflecting hierarchical relations among the corresponding concepts. Instead, logical definitions created from lexical features can simply be represented in DL formalism and run through a reasoner to infer the relevant subClassOf relations. As shown here, this approach is easy to implement, efficient and scalable. The only programming required is for serializing the logical definitions in the appropriate DL format.
Moreover, given that SNOMED CT already uses DL techniques for representing its logical definitions based on biomedical knowledge and an EL reasoner for inferring its hierarchy, it can be expected that the IHTSDO could easily integrate the lexical approach to quality assurance proposed here.
Finally, having two kinds of logical definitions (from biomedical knowledge and from lexical features) represented in the same formalism would make it possible to integrate them into the same framework, for example to test the consistency between the two kinds of definitions.
This preliminary investigation is limited to two
subhierarchies of SNOMED CT for diseases and procedures.
However, we also generated definitions and inferred hierarchy
for the whole SNOMED CT and did not notice any scalability
issues. We did not leverage SNOMED CT synonyms for
creating logical definitions, but this should be a natural
extension of this investigation. In future work, we also would
like to normalize terms before creating the definitions, since
normalization is common approach to managing term variation
[
This bag-of-word approach to comparing terms tends to
generate more false positives than a linguistically motivated
approach, where the head of the noun phrase would be required
to be the same in two hierarchically related concepts, as we did
in other work [
In this preliminary investigation, we performed a limited evaluation. Given the encouraging results, we plan to extend the investigation to the entirety of SNOMED CT, evaluate the results more thoroughly, and share them with the SNOMED CT developers at the IHTSDO.
Finally, the lexical approach to quality assurance proposed
here could also complement structural approaches, such as the
lattice-based approach we proposed earlier [
This approach to identifying missing hierarchical relations would be applicable not only to the entirety of SNOMED CT, but to other biomedical ontologies as well. More specifically, it could be applied to any biomedical ontology for which concept names and hierarchical relations are available (i.e., most ontologies). The same approach could also be applied to the creation of partial mappings.
This work was supported by the Intramural Research Program of the NIH, National Library of Medicine. This work was conducted using the Protégé resource, which is supported by grant GM10331601 from the National Institute of General Medical Sciences of the United States National Institutes of Health. We would like to thank Dr. GQ Zhang for providing motivation and encouragement for this investigation. [9] IHTSDO, “SNOMED CT,” 2016.
TABLE I.
EXAMPLES OF SUBCLASSOF RELATIONS INFERRED FROM LEXICAL FEATURES yes yes yes yes yes yes