Pathology reports play an essential role in cancer treatment and research. They contain vital findings about a patient's cancer, such as cell histology and molecular markers, that are used to diagnose the type of cancer, determine treatment options, and enhance our understanding of the nature of the disease. At Roswell Park Comprehensive Cancer Center, 1 pathology reports are stored mainly as unstructured text in a relational database. 2 To find information within pathology reports, we use either string matching methods (e.g., regular expressions), or the TIES3 natural language processing (NLP) program. The drawback of string matching is that string variations need to be accounted for in order to find information. For instance, a search for patients whose tumors lack estrogen receptor (ER) proteins will also have to search for strings matching 'ER negative', 'estrogen receptor negative', 'hormone status negative', and the like.
TIES addresses some of this variability by mapping
multiple strings to the same ontology class, but some searches
consistently do not perform well.4 Moreover, a class identified
by TIES is not linked to the formal axioms that define the
class, which prevents researchers from fully leveraging the
formal relations that hold between classes within an ontology.
For example, the formal definition of Medullary Breast
Carcinoma (C17965 5 ) in the NCI Thesaurus (NCIt) [
To address these shortcomings, we are developing an ontology that we currently call the ‘Document Content Ontology’ 6 (DCO) to represent the terms, words, word 1 https://www.roswellpark.org 2 The database does contain some structured fields, but we find that most researchers are interested in the information contained in the unstructured text. 43 ThtItEpS:/c/otniessis. dtebnmtlyi.pfaititl.setdouidentify findings that the cells in the tumor lack estrogen-receptor proteins. 5 IRI: http://ncicb.nci.nih.gov/xml/owl/EVS/Thesaurus.owl#C17965 6 The name of the ontology may change! contexts, and their positions (i.e., indexes) within the documents. That is, we are using the DCO to represent the content of the document and where the content is located. It is important to make clear that while we are using an NLP program for named entity recognition (described in what follows), we are not developing an NLP program. Rather, we are augmenting the output of the NLP program so that we can more fully leverage the axioms contained within an ontology.
A high-level summary of the DCO is illustrated in Figure 1. Documents contain (i.e., has part) terms, and terms, which are composed of one more words, have meanings that are specified using the semantic type, sematic label, and semantic source annotation properties to reference an ontology class. The literal value data property associates the actual data (e.g., strings) with the term, and the polarity annotation represents whether the term has a positive or negative connotation (e.g., the patient does not have breast carcinoma). In some cases, we also represent the word context: the group of words surrounding the word or words that constitute a term. Word contexts are useful in aiding NLP programs to disambiguate the sense in which a word is being used. For brevity, not all properties and classes are discussed. Full details are available at https://github.com/RoswellParkResearch/document-contentontology.
We are aware that a number of other ontologies (such as the Information Artifact Ontology and Semanticscience Integrated * corresponding author Ontology) have terms similar to ours. However, these ontologies carry with them metaphysical commitments, such as a document being a type of generically dependent continuant. Since we are just beginning to develop the DCO, we wish (for now) to remain agnostic concerning such commitments.
At present, we are using the DCO to structure output from
the Noble Coder Named Recognition Engine [
We translate the output of Noble into OWL and load it along with the full ontology of association classes (which we call a named entity’s semantic type) into a GraphDB8 semantic triple store. This allows us to simultaneously query for pathology reports having a specified named entity and the ontology for other entities related to the named entity.
In addition to leveraging the ontologies axioms, we can also examine the word context surrounding a term. For instance, this is useful for addressing the aforementioned issue of searching pathology reports in which the cells are found to be ER negative.
We gratefully acknowledge support from the Roswell Park Comprehensive Cancer Center Biomedical Data Science Shared Resource that is funded in part by Cancer Center Support Grant NCI P30CA16056.