Author ORCID Identifier

Year of Publication


Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation




Computer Science

First Advisor

Dr. Jinze Liu

Second Advisor

Dr. Licong Cui


Biomedical terminologies serve as knowledge sources for a wide variety of biomedical applications including information extraction and retrieval, data integration and management, and decision support. Quality issues of biomedical terminologies, if not addressed, could affect all downstream applications that use them as knowledge sources. Therefore, Terminology Quality Assurance (TQA) has become an integral part of the terminology management lifecycle. However, identification of potential quality issues is challenging due to the ever-growing size and complexity of biomedical terminologies. It is time-consuming and labor-intensive to manually audit them and hence, automated TQA methods are highly desirable. In this dissertation, systematic and scalable methods to audit biomedical terminologies utilizing their structural as well as lexical information are proposed. Two inference-based methods, two non-lattice-based methods and a deep learning-based method are developed to identify potentially missing hierarchical (or is-a) relations, erroneous is-a relations, and missing concepts in biomedical terminologies including the Gene Ontology (GO), the National Cancer Institute thesaurus (NCIt), and SNOMED CT.

In the first inference-based method, the GO concept names are represented using set-of-words model and sequence-of-words model, respectively. Inconsistencies derived between hierarchical linked and unlinked concept pairs are leveraged to detect potentially missing or erroneous is-a relations. The set-of-words model detects a total of 5,359 potential inconsistencies in the 03/28/2017 release of GO and the sequence-of-words model detects 4,959. Domain experts’ evaluation shows that the set-of-words model achieves a precision of 53.78% (128 out of 238) and the sequence-of-words model achieves a precision of 57.55% (122 out of 212) in identifying inconsistencies.

In the second inference-based method, a Subsumption-based Sub-term Inference Framework (SSIF) is developed by introducing a novel term-algebra on top of a sequence-based representation of GO concepts. The sequence-based representation utilizes the part of speech of concept names, sub-concepts (concept names appearing inside another concept name), and antonyms appearing in concept names. Three conditional rules (monotonicity, intersection, and sub-concept rules) are developed for backward subsumption inference. Applying SSIF to the 10/03/2018 release of GO suggests 1,938 potentially missing is-a relations. Domain experts’ evaluation of randomly selected 210 potentially missing is-a relations shows that SSIF achieves a precision of 60.61%, 60.49%, and 46.03% for the monotonicity, intersection, and sub-concept rules, respectively.

In the first non-lattice-based method, lexical patterns of concepts in Non-Lattice Subgraphs (NLSs: graph fragments with a higher tendency to contain quality issues), are mined to detect potentially missing is-a relations and missing concepts in NCIt. Six lexical patterns: containment, union, intersection, union-intersection, inference-contradiction, and inference-union are leveraged. Each pattern indicates a potential specific type of error and suggests a potential type of remediation. This method identifies 809 NLSs exhibiting these patterns in the 16.12d version of NCIt, achieving a precision of 66% (33 out of 50).

In the second non-lattice-based method, enriched lexical attributes from concept ancestors are leveraged to identify potentially missing is-a relations in NLSs. The lexical attributes of a concept are inherited in two ways: from ancestors within the NLS, and from all the ancestors. For a pair of concepts without a hierarchical relation, if the lexical attributes of one concept is a subset of that of the other, a potentially missing is-a relation between the two concepts is suggested. This method identifies a total of 1,022 potentially missing is-a relations in the 19.01d release of NCIt with a precision of 84.44% (76 out of 90) for inheriting lexical attributes from ancestors within the NLS and 89.02% (73 out of 82) for inheriting from all the ancestors.

For the non-lattice-based methods, similar NLSs may contain similar quality issues, and thus exhaustive examination of NLSs would involve redundant work. A hybrid method is introduced to identify similar NLSs to avoid redundant analyses. Given an input NLS, a graph isomorphism algorithm is used to obtain its structurally identical NLSs. A similarity score between the input NLS and each of its structurally identical NLSs is computed based on semantic similarity between their corresponding concept names. To compute the similarity between concept names, the concept names are converted to vectors using the Doc2Vec document embedding model and then the cosine similarity of the two vectors is computed. All the structurally identical NLSs with a similarity score above 0.85 is considered to be similar to the input NLS. Applying this method to 10 different structures of NLSs in the 02/12/2018 release of GO reveals that 38.43% of these NLSs have at least one similar NLS.

Finally, a deep learning-based method is explored to facilitate the suggestion of missing is-a relations in NCIt and SNOMED CT. Concept pairs exhibiting a containment pattern is the focus here. The problem is framed as a binary classification task, where given a pair of concepts, the deep learning model learns to predict whether the two concepts have an is-a relation or not. Positive training samples are existing is-a relations in the terminology exhibiting containment pattern. Negative training samples are concept-pairs without is-a relations that are also exhibiting containment pattern. A graph neural network model is constructed for this task and trained with subgraphs generated enclosing the pairs of concepts in the samples. To evaluate each model trained by the two terminologies, two evaluation sets are created considering newer releases of each terminology as a partial reference standard. The model trained on NCIt achieves a precision of 0.5, a recall of 0.75, and an F1 score of 0.6. The model trained on SNOMED CT achieves a precision of 0.51, a recall of 0.64 and an F1 score of 0.56.

Digital Object Identifier (DOI)