DESCRIPTION (provided by applicant): We will develop a family-based Quality Assurance (QA) framework for biomedical ontologies. Ontology QA is critical for increasing the use of ontologies in interdisciplinary research and in electronic health records (EHRs). We will develop computational techniques for identifying concepts with high probability of errors to improve efficiency and effectiveness of ontology QA. Biomedical ontologies are large, complex knowledge representation systems that enable the integration of knowledge from different fields. The largest, best-known ontology repository is the Bioportal of the National Center for Biomedical Ontologies, containing more than 300 ontologies and tools for editing, browsing, and visualizing these ontologies. However, many errors have been discovered in BioPortal's ontologies. QA in BioPortal has been mostly focused on use-cases and ad hoc techniques. Our computational techniques will automatically identify sets of concepts with a high likelihood of errors to empower ontology QA. In past research, we have designed many QA techniques for single ontologies and have shown that sets of complex and uncommonly classified concepts have significantly higher percentages of errors. The theoretical bases for our QA are Abstraction Networks (AbNs), which summarize ontologies in a compact way. Using AbNs, we identified many error-prone concepts. In this project, we will perform QA for whole families of ontologies. We have already identified seven preliminary families, based on structural properties. If a classification of concepts yields higher than usual error rates in several ontologies of a family F then we hypothesize that this will be true for such classifications for most ontologies of F. We will build a prototype software tool (BLUOWL) for determining AbNs for each family, to support QA of its ontologies. Our primary test beds will be seven cancer-related ontologies, e.g., the National Cancer Institute thesaurus (NCIt), with different properties and purposes. Some non-cancer ontologies will also be included. We have published preliminary QA results for four such ontologies. In evaluation studies, we will formulate and test hypotheses, statistically expressingthe error expectations for various kinds of concepts. Ontologies' curators were recruited to review the suspicious concepts we will identify as part of their regular QA efforts (outside of ourbudget). In summary, we will: Identify families of BioPortal ontologies based on ontology structure and design a unified methodology for deriving their abstraction networks; Build a software tool (BLUOWL) for QA of each family; Investigate concept classifications more likely to be erroneous in each family; Perform evaluation of our QA methodologies and usability studies for BLUOWL.
|Effective start/end date||3/4/15 → 2/28/19|
- National Institutes of Health