Ontology customization for
                      biomedical informatics
             Author Pablo López-García
       Supervisors Arantza Illarramendi
     Studies/Stage PhD in Web Information Systems, expected Fall 2012
          Affiliation University of the Basque Country UPV/EHU (Spain)
             E-Mail pablo.lopez@ehu.es



                          Aims and Objectives of the Research
The overall aim of the research is to explore to which extent existing semantic technologies,
particularly biomedical ontologies and ontology mappings, can be customized to better serve
biomedical applications.

                           Justification for the Research Topic

Semantic technologies, such as comprehensive biomedical ontologies (e.g., SNOMED CT
[1]), and ontology mappings linking several ontologies (e.g., SNOMED CT, NDF-RT [2], and
RxNorm [3]), are offered as valuable tools to the biomedical community. However, due to the
high degree of specialization in health care and life sciences, two undesirable scenarios
might occur: (i) the whole breadth of an extended biomedical ontology might be excessive for
some applications and users or, on the other hand, (ii) information of interest might be
missed because it is scattered among several ontologies.

                                   Research Questions

   1. How useful are ontology modularization techniques to extract subsets from a large
      biomedical ontology to annotate medical records? Can the size of those subsets be
      reduced by using term frequency information from an authoritative corpus? How does
      size reduction affect coverage?

   2. Once a module from a reference ontology has been adopted for a domain, how useful
      are ontology mappings and ontology repositories as an approach to add new, related
      information from another domain to the module?

                                  Research Methodology

   1. Case study: Annotating discharge summaries in cardiology using SNOMED CT. Four
      graph-traversal heuristics [4], and one logic-based technique [5] were explored to
      extract subsets from SNOMED CT. Extracted subsets were subsequently filtered with
      term frequency information from MEDLINE. As signatures and test sets, twenty
      manually coded discharge summaries from cardiology patients were used. The
      coverage, size, and precision of extracted subsets were measured.

   2. Case study: Using disorders in SNOMED CORE to extract drugs from RxNorm. The
      CORE problem list subset of SNOMED CT [6] was used as signature in the domain of
      diseases to find related drugs in RxNorm, using NDF-RT as a cross-ontology, and the
      UMLS Metathesaurus to provide the mappings.
                                    Research Results to Date

1. Published results [7]: Graph-traversal heuristics provided high coverage (71-96% of
   terms in the test sets of discharge summaries) at the expense of subset size (17-51%
   of the size of SNOMED CT). Logic-based techniques extracted small subsets (1%),
   but coverage was limited (24-55%). Filtering reduced the size of large subsets to 10%
   while still providing 80% coverage. The study shows that the requirements of
   preserving ontology entailments (key to ontology modularization), and achieving high
   precision (key to annotation), necessarily conflict. Attempting to satisfy the two aims
   simultaneously can lead to unsatisfactory results.

2. Preliminary results: Using the CORE subset’s 5,814 SNOMED CT concepts as input,
   we were able to extract 7,499 related drugs in RxNorm. The CORE subset is roughly
   6%, and 9% the size of SNOMED CT’s Disease (ID 64572001), and Clinical finding
   (ID 404684003) hierarchies, respectively. The extracted RxNorm subset was 3.87%
   the size of the full RxNorm. The current research is focused on measuring the
   coverage of the new subset (CORE plus the extracted RxNorm) when annotating
   terms from a large, real dataset.

                                             References
1.   IHTSDO (2012), International Health Terminology Standards Development Organization,
     http://www.ihtsdo.org/.
2.   Brown SH, Elkin PL, Rosenbloom ST, Husser C, Bauer BA, Lincoln MJ, Carter J, Erlbaum M, Tuttle MS
     (2004), VA National Drug File Reference Terminology: a cross-institutional content coverage study, in:
     Studies in Health Technology and Informatics 107(Pt 1), 477-481.
3.   Liu S, Ma W, Moore R, Ganesan V, Nelson SJ (2005), RxNorm: prescription for electronic drug
     information exchange, in: IT Professional 7(5), 17-23.
4.   Seidenberg J, Rector A (2006), Web ontology segmentation: analysis, classification and use, in:
     Proceedings of the 15th International Conference on World Wide Web, Edinburgh, Scotland, UK:
     Association for Computing Machinery, 13-22.
5.   Cuenca Grau B, Horrocks I, Kazakov Y, Sattler U (2008), Modular reuse of ontologies: theory and
     practice, in: Journal of Artificial Intelligence Research 31, 273-318.
6.   Fung KW, McDonald C, Srinivasan S (2010), The UMLS-CORE project: a study of the problem list
     terminologies used in large healthcare institutions, in: Journal of the American Medical Informatics
     Association 17(6), 675-680.
7.   López-García P, Boeker M, Illarramendi A, Schulz S (2012), Usability-driven pruning of large ontologies:
     the case of SNOMED CT, in: Journal of the American Medical Informatics Association 19(e1), e102-
     e109.