Capturing the Relationship between Evolving Biomedical Concepts via Background Knowledge

Abstract

Semantic Web applications and knowledge-based systems heavily rely on the use of up-to-date ontologies. To support their adequate evolution, methods must detect the changes of meanings in concepts over time. This article proposes exploiting domain-speci c external source of knowledge to characterize the evolution of concepts in dynamic ontologies. Our original technique analyses the evolution of values in concept attributes. The approach uses ontological properties and mappings between ontologies from online repositories to deduce the nature of the relationship between a concept and its successive version. The proposed algorithm is experimentally evaluated comparing di erent congurations by using various successive versions of biomedical ontologies. The obtained results reveal the bene ts of considering external sources of knowledge to identify the correct semantic relation in ontology evolution.

Keywords

• Biomedical ontologies
• Ontology evolution
• Background knowledge

-

Software applications using Semantic Web technologies have gained interest in Life Sciences over the past years [ 10 ]. For example, clinical decision support systems use ontologies and their associated mappings for enhancing their reasoning capabilities [ 3 ]. However, the dynamic nature of medical knowledge forces ontology engineers to constantly revise the content of ontologies to keep it upto-date. Also, these modi cations might be propagated to depending artefacts (like mappings and annotations) to keep the underlying systems reliable.

Maintaining these artefacts up-to-date according to ontology evolution is a complex task, specially in highly dynamic domains. For example, the release of new versions of ontologies, like SNOMED CT, can impact a huge amount of mappings [ 9,11 ]. The development of automatic tools to assist domain experts in these maintenance tasks requires adequate characterization of ontology changes. Nevertheless, the detection of changes between di erent ontology versions remains an open issue. The limitations refer to understanding the evolution of meanings for a given concept. For instance, if it has become more or less speci c from the semantic point of view, or if the change does not a ect its semantics.

We have shown the importance of attribute values (e.g., concept label, synonyms, etc.) de ning concepts in the establishment of mappings [ 8 ] and, in turn, for their maintenance [ 7 ]. Consider the evolution of \poisoning central nervous system stimulants", which is the title of the concept `970' of ICD-9-CM version 2009, to \central nervous system stimulants" in ICD-9-CM version 2011. This shows that the concept became more general since the term \poisoning" was deleted. So, if one assumes that this concept was interrelated to another ontology with an \equivalent" mapping, the evolution previously described shall transform the \equivalent" mapping to an \is-a".

In this paper, we propose an approach to detect the semantic evolution of concept attribute values. Our method determines the relationship between the two concepts (one of each version) having the considered attributes as label or synonyms. To this end, the approach explores background knowledge (e.g., external source of knowledge like Bioportal [ 15 ]) to characterize ontology evolution by identifying speci c changes of attribute values de ning the concepts in di erent versions. It aims at determining if the evolved information has become more generic, more speci c, remains equivalent or if it is somehow related. We assessed the approach over a corpus of reference. It was built by domain experts, made up of two successive versions of concepts' attributes retrieved from SNOMED CT, ICD-9-CM and MeSH. To demonstrate the bene ts of considering domainspeci c background knowledge, we compare the obtained results with those from the constructed corpus.

The remainder of this article is organized as follows: Section 2 formalizes the addressed problem and describes the related work. Afterwards, we present the proposed method and the implemented algorithm (Section 3). Section 4 describes the evaluation including the obtained results. In the sequence, we present the discussion (Section 5), which is followed by the conclusion in Section 6. 2

Problem Statement and Related Work

This section introduces de nitions and the addressed problem. We describe the related work to clarify the originality of our approach. 2.1

De nitions

An ontology O is a set of concepts interrelated by semantic relationships [ 12 ]. j We de ne a set of concepts of O at time j, such that j ∈ N, as C(Oj ) = {ci Si ∈ N}. Each concept is characterized by a set of attributes. The set of attributes de ning a concept c ∈ C(Oj ) refers to the function A(cj ) = {Attj1; Attj2; :::; Attjn} (e.g., concept label, de nition, synonym, etc.). The attributes can di er from one ontology to another, but in general each attribute describing a concept has a name and an associated string value. For example, the attribute value \avian u" is a synonym of the concept \avian in uenza" from SNOMED CT version 2014AB. We de ne Attij :name (e.g., synonym) and Attij :value (e.g., \avian u"), but from now on we use Attij to denote Attij :value to simplify. We de ne Reli ∈ Oj ; Reli = {(a; b; ri)Sa; b ∈ C(Oj ); ri ∈ RelSymb} where RelSymb ∈ {–; ≡; ≤; ≥; ≈}.

The addressed problem consists in de ning the semantic relationship that exists between two successive versions of an attribute of a given concept. Our technique must decide if the considered concept has become more or less speci c or if it remains equivalent after evolution (i.e., a new version of the ontology). 2.2

Related Work

Ontology matching is a research eld where background knowledge has been implemented. A rst signi cant tentative was proposed by Aleksovski et al. to align two ontologies that present poor lexical overlap and limited structural properties using a semantically rich knowledge source [ 1 ]. The approach consists in nding anchoring matches, using lexical heuristics, of the source and target ontology in the external one. The proposal uses its semantics to deduce the relationship that holds between the concepts.

Sabou et al. [ 16 ] proposed an ontology matching paradigm that could be complementary to existing classic ones. It automatically explores multiple and heterogeneous on-line knowledge sources to derive mappings. The approach aims at aligning ontologies' concepts by selecting the most appropriate knowledge distributed over several external ontologies. They explored formal properties of the background knowledge to infer the possible relationship that could exist between the concepts to be aligned.

TaxoMap uses WordNet as background knowledge [ 13 ]. The intervention of a domain expert is required to determine the best anchor in WordNet that corresponds to the concepts to align. This anchor delimits the usable sub-graph in WordNet to optimize the alignment phase. Once the appropriate sub-graph is identi ed, classic matching techniques interrelate the concepts of source and target ontologies with those previously de ned in the sub-graphs.

Mougin et al. [ 14 ] proposed the use of WordNet to disambiguate information contained in biomedical systems with the UMLS3. The goal was to validate obtained mappings in cases of unambiguous matches between the information content and UMLS or, disambiguate the obtained alignment in case of several correspondences using the information provided by WordNet. They showed that general knowledge can improve the validation of direct mappings and help in the identi cation of indirect mappings of concepts to the UMLS.

Zhang and Bodenreider [ 19 ] aimed at taking advantage of domain-speci c knowledge using the UMLS to improve the alignment between anatomical ontologies. They revealed that domain knowledge is a key factor behind the identi cation of additional mappings compared with the generic schema matching approach. The use of UMLS as an external resource was interesting for various aspects: (1) generating more mappings; (2) providing di erent synonyms for a given concept and (3) de ning relations between concepts in a semantic network. 3 www.nlm.nih.gov/research/umls

More recently, Arnold and Rahm [ 2 ] explored generic external resources and proposed a two-step enrichment technique to improve existing imprecise ontology mappings. They used linguistic techniques and resources like WordNet to re ne the semantic relation between aligned concepts. Their work aims to transform equivalence between concepts into a \is-a" or \part-of" which may further re ect the real semantics of mapped concepts. It was not applied to ontology evolution.

The use of background knowledge has also been investigated for ontology evolution. The background knowledge was used to assess new statements identied as relevant and must be included in an ontology at evolution time [ 18 ]. This work, part of the EVOLVA framework [ 17 ], aims at enriching an ontology with additional relevant knowledge (i.e., statements) by using background ontologies.

All the surveyed approaches fail in considering the impact of ontology evolution on dependent artifacts as ultimate focus, so several gaps remain open. First, in most of the approaches, the used background knowledge is usually of general nature (the knowledge is described at a higher level of abstraction). This might be interesting for disambiguating the context, but not to characterize the evolution of concepts, especially if their de nition become more nely described. Second, a single source of knowledge (i.e., only one ontology) is usually implemented, which limits its coverage. This is true mainly for domain speci c ontologies, which require a very precise description of the external knowledge. The use of multiple connected ontologies might optimize the coverage of the domain through a more precise description of the domain. 3

Determination of Semantic Relationship between

Changing Concepts

Our method aims to reach an accurate characterization of the semantic evolution of concepts by analysing multiple and domain speci c interconnected ontologies contained in Bioportal. We assume that performing a match between di erent domain-speci c ontologies might provide necessary and su cient facts to determine the relationship between evolving concepts. Mappings that link them might allow reasoning over several ontologies. This aspect, combined with the richness of the content of ontologies, provides a good support for ontology evolution, and in particular, the characterization of the modi cations that a ect entities.

We de ne the Algorithm 1 that exploits search modules, the structure and properties of ontologies and mappings stored in Bioportal. Concept attributes play a key role in the de nition of mappings and for their maintenance [ 4 ]. In consequence, we assume that the modi cations in attribute values of concepts directly impact their semantics, which in turn, might in uence dependent mappings. For this reason, given a changed concept, our algorithm takes as input the value of the attribute before evolution and its value after evolution. In the following, we explain the algorithm. 1. Search for concepts (statement 1 to 3). The goal is to nd the attribute values Att1 and Att2 in the description of concepts in ontologies di erent Algorithm 1: Background knowledge-based semantic relationship identi cation between evolving concepts (b) The path only contains subsumed concepts then the concept located at the highest level of the hierarchy is considered as less speci c. (c) If the two concepts are siblings then it returns ≈ \partially related to".

This relation is imprecise, but of interest for mapping maintenance [ 6 ]. (d) Otherwise, no relationship can be precisely determined and the \undened" value is returned.

Consider the example of Fig. 1. We observe that the concept code `M0006899' whose label is \Pituitary dwar sm" in MeSH evolved from version released in 2012 to \Pituitary dwar sm II " ('M0452907') in version 2013. Algorithm 1 identi es that these concepts are siblings in SNOMED CT, by searching Bioportal, therefore the relationship symbol \partially related to" is returned. “Pituitary  dwarfism”   (MeSH)   1 Search  in  ontologies   SNOMED  CT,   ICD9CM,  MEDDRA,   NCIT,  DOID,  RCD,  HP,   DERMLEX,  NATPRO,   CRISP,  SOPHARM,   BDO,  SNMI  

(Direct  method)   No  common  ontologies  

Use  mappings   15  mappings  available   (OMIM  ontology)   “Pituitary  dwarfism  II”  

(MeSH)   Search  in  ontologies   1

OMIM  

NDFRT   2 (Indirect  method)   “Pituitary  dwarfism  II”  (OMIM)  

Mapped_to   “Laron-­‐type  isolated  somatotropin  defect”  (SNOMED  CT)  

SNOMED  CT  is  the  common  ontology   “Laron-­‐type  isolated  somatotropin  defect”  and   “Pituitary  dwarfism”  have  the  same  super  concept   3 (“short  stature  disorder”)  they  are  siblings   We assessed the e ectiveness of our approach on realistic case studies of the biomedical domain. The goal is to show the ability of the algorithm to infer the semantic relations for characterizing ontology evolution in modi cations of concept attribute values. Experiments relied on successive versions of ICD-9-CM, SNOMED CT and MeSH. The evaluation consists in characterizing the evolution of concepts of these ontologies by analysing the evolution of their attributes value. However, as no Gold Standard for such an evaluation exists, we needed to construct our own corpus of reference to compare the obtained results. We conducted the following three steps method to obtain our corpus of reference: 1. We selected 1.000 couples of attributes from SNOMED CT, ICD-9-CM and MeSH from concepts a ected by change operations. One attribute of each couple comes from a concept at time j and the other attribute comes from the context of the concept at time j + 1. We chose these couples based on the similarity score between the attribute values (i.e., we excluded attributes with very low similarity and unchanged attributes at time j +1). The context of a concept denotes all its subconcepts, superconcepts and siblings and the similarity refers to syntactic and word-based distances [ 5 ]. 2. Three experts evaluated the selected couples of attributes to determine \equivalent", \more speci c", \less speci c", \partially matched to" or \no relation" as the relationship that link them. 3. They performed one round of evaluation and we merged the answers that are the same for all reviewers. The experts collaborated and re-evaluated a second time those attribute couples for which no agreement was found. We achieved an average agreement rate of 86% for the concerned attributes. Finally, we retained 675 pairs of attributes that have the consensus of opinion about their semantic relationship. 4.2

Results

Fig. 2 presents two sets of results. The graphic on the left side depicts all the cases tested with our algorithm including those for which the algorithm did not detect a relationship (e.g., cases where the pairs of attribute were not found in external ontologies). Aiming to better evaluate the precision of the algorithm, the right side of Fig. 2 presents the results excluding all cases of disagreement, where the algorithm returns no relation. The main reason for this disagreement happens when the analyzed attributes do not exist in external ontologies.

The obtained results are convincing since a general precision of 77% is obtained (CR=BK in Fig.2, i.e., experts and algorithm agree on the semantic relationship between the concept attributes). The most signi cant results are obtained for \no relation" (unmappable in Fig. 2, left side), \equivalent" and \partial match" relations, reaching an average of 88% of precision. However, this number decreases to 53% when considering the subsumption relationship (More specif. and Less specif.).

The graphic on right side of Fig. 2 reveals the general e ciency of our algorithm in a subset of the cases (by excluding disagreement on the \no relation"). Globally, the results are signi cant since 96% of precision is reached. We even obtain 100% precision for the identi cation of \equivalent" relation (i.e., all the cases returned by our algorithm are correct according to domain experts). The ability of the algorithm to identify the subsumption relationships is less convincing as shown by the results (13% of the cases were not correctly identi ed).

CorrectlyidenJfied NotidenJfied

7 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 506 119

CorrectlyidenJfied NotidenJfied 0 106 44 70 167 CR=BK Unmappable Equivalent

MoreSpecif. LessSpecif. ParJalMatch

CR=BK Unmappable Equivalent

MoreSpecif. LessSpecif. ParJalMatch The proposed method is able to correctly detect 95% of the semantic relationship in concept evolution when the concepts exist in ontologies stored in Bioportal. The remaining 6% di ers mainly for two reasons: { The level of granularity to describe one concept can di er from one ontology to another. This situation impacts on the outcome of our method when one ontology includes as synonyms a list of terms that have subsumption relations in another ontology. For instance, the term \Chondrosarcoma of bone" is de ned in SNOMED CT as a sub-concept of \Chondrosarcoma", but in CTV3 these two terms are synonyms. While building the corpus of reference, the experts adopted the de nition of SNOMED CT for these terms (i.e., more speci c than) so the algorithm found a di erent result. { Domain experts were more precise than existing ontologies. This was, for instance, observed when using the terms \autonomic peripheral nervous system diseases" and \autonomic central nervous system diseases". MeSH, National Drug File { Reference Terminology, and Neuroscience Information Framework Standard Ontology de ne these two terms as synonyms of \autonomic nervous system diseases". Our method detects that these terms are equivalent, but domain experts considered these two terms as siblings. Consequently, the outcome of the algorithm di ers from the corpus of reference.

The use of background knowledge shows the possibility of automatically capturing the impact of changes in concepts from a semantic viewpoint. However, this approach contains some limitations: { In our method, semantic changes can only be measured if the considered attribute is the label (or synonym) of a concept in another external ontology.

This situation was observed in 83% of the concepts in the corpus. { Mappings between ontologies in the repository must be correct and up-todate. Our method uses these mappings to select the ontologies that are analysed, since we assumed that unmapped ontologies do not describe the same domain (i.e., they do not have overlapping concepts). { Non-equivalent relation need to be inferred by our method. Bioportal only contains equivalent mappings, i.e., if there is a mapping between two concepts from di erent ontologies, the interrelation between these concepts is always an equivalence. This can lead to cases where subsumed concepts (in ontology A) are mapped to the same concept (in ontology B). For instance, \left bundle branch blocks" and \right bundle branch blocks" are sibling concepts in ICD9 and CTV3, but they are interrelated to the same concept of MeSH (where these two terms are described as synonyms). The outcome of our method could be even more accurate if other types of relationships exist in available mappings from the repository.

To complement the proposed technique, we are currently working on an algorithm to determine non-equivalent relations between concepts. We are studying the use of several mappings to navigate from one ontology to another relying on domain-speci c background knowledge. This can allow collecting more information about the attribute to select the appropriated relation according to the level of granularity used to describe the concept in the original ontology. 6

Conclusion

Dealing with the evolution of ontologies and of their dependant artefacts relies on appropriate ontology changes identi cation. It demands characterization of the semantic relation between evolving concepts. In this paper, we proposed an approach exploiting domain-speci c background knowledge to determine the semantic evolution of concepts. Our algorithm analyzed the modi cations in the values of concept attributes and the experiments showed the e ectiveness of the technique with large life sciences ontologies from Bioportal. Future work involves the re nement of the algorithm and further experiments with additional datasets.

Acknowledgements

This work is supported by the National Research Fund (FNR) of Luxembourg and S~ao Paulo Research Foundation (FAPESP) (Grant #2014/14890-0).

References

1. Aleksovski , Z. , Klein , M. , ten Kate , W., van Harmelen , F. : Matching unstructured vocabularies using a background ontology . In: Proceedings of the 15th International Conference on Managing Knowledge in a World of Networks . pp. 182 { 197 . EKAW' 06 , Springer-Verlag, Berlin, Heidelberg ( 2006 )

2. Arnold , P. , Rahm , E.: Semantic enrichment of ontology mappings: a linguisticbased approach . In: Advances in Databases and Information Systems . pp. 42 { 55 . Springer ( 2013 )

3. Bodenreider , O. : Biomedical ontologies in action: role in knowledge management, data integration and decision support. Yearbook of medical informatics p. 67 ( 2008 )

4. Dinh , D. , Dos Reis , J.C. , Pruski , C. , Da Silveira , M. , Reynaud-Dela^tre, C.: Identifying relevant concept attributes to support mapping maintenance under ontology evolution . Web Semantics: Science, Services and Agents on the World Wide Web 29 , 53 { 66 ( 2014 )

5. Dos Reis , J.C. , Dinh , D. , Da Silveira , M. , Pruski , C. , Reynaud-Dela^tre, C.: Recognizing lexical and semantic change patterns in evolving life science ontologies to inform mapping adaptation . Arti cial Intelligence in Medicine 63 ( 3 ), 153 { 170 ( 2015 )

6. Dos Reis , J.C. , Dinh , D. , Pruski , C. , Da Silveira , M. , Reynaud-Dela^tre, C.: Mapping adaptation rules for the automatic reconciliation of dynamic ontologies . In: Proceedings of the ACM International Conference on Information and Knowledge Management (CIKM) ( 2013 )

7. Dos Reis , J.C. , Pruski , C. , Da Silveira , M. , Reynaud-Dela^tre, C.: Dykosmap: A framework for mapping adaptation between biomedical knowledge organization systems . Journal of biomedical informatics 55 , 153 { 173 ( 2015 )

8. Dos Reis , J.C. , Pruski , C. , Da Silveira , M. , Reynaud-Dela^tre, C.: Characterizing semantic mappings adaptation via biomedical kos evolution: A case study investigating snomed ct and icd . In: Proceedings of the Annual AMIA Symposium ( 2013 )

9. Dos Reis , J.C. , Pruski , C. , Da Silveira , M. , Reynaud-Dela^tre, C.: Understanding semantic mapping evolution by observing changes in biomedical ontologies . Journal of Biomedical Informatics ( 2013 )

10. Feigenbaum , L. , Herman , I. , Hongsermeier , T. , Neumann , E. , Stephens , S.: The semantic web in action . Scienti c American 297 ( 6 ), 90 { 97 ( 2007 )

11. Gross , A. , Hartung , M. , Thor , A. , Rahm , E.: How do computed ontology mappings evolve?-a case study for life science ontologies . In: Joint Workshop on Knowledge Evolution and Ontology Dynamics ( 2012 )

12. Gruber , T.R.: A translation approach to portable ontology speci cations . Knowledge Acquisition 5 ( 2 ), 199 {220 (Jun 1993 )

13. Hamdi , F. , Safar , B. , Reynaud , C. , Niraula , N.: TaxoMap alignment and re nement modules: Results for OAEI 2010 . In: The Fifth International Workshop on Ontology Matching. pp. 212 { 219 ( 2010 )

14. Mougin , F. , Burgun , A. , Bodenreider , O. : Using wordnet to improve the mapping of data elements to umls for data sources integration . In: Proceedings of the AMIA annual Symposium . pp. 574 { 578 ( 2006 )

15. Noy , N.F. , Shah , N.H. , Whetzel , P.L. , Dai , B. , Dorf , M. , Gri th , N., Jonquet , C. , Rubin , D.L. , Storey , M.A. , Chute , C.G. , Musen , M. : Bioportal: ontologies and integrated data resources at the click of a mouse . Nucleic acids research (37) ( 2009 )

16. Sabou , M., d'Aquin , M. , Motta , E.: Exploring the semantic web as background knowledge for ontology matching . Journal on data semantics XI pp. 156 { 190 ( 2008 )

17. Zablith , F. : Evolva: A comprehensive approach to ontology evolution . In: The Semantic Web: Research and Applications, the 6th European Semantic Web Conference , ESWC, pp. 944 { 948 . Springer, Heraklion, Crete, Greece ( 2009 )

18. Zablith , F., d'Aquin , M. , Sabou , M. , Motta , E. : Using ontological contexts to assess the relevance of statements in ontology evolution . In: Proceedings of EKAW 2010 - Knowledge Engineering and Knowledge Management by the Masses ( 2010 )

19. Zhang , S. , Bodenreider , O. : Experience in aligning anatomical ontologies . International journal on Semantic Web and information systems 3(2) , 1 ( 2007 )