Motivation: The previously developed Search Ontology (SO) allows domain experts to formally specify domain concepts, search terms associated to a domain, and rules describing domain concepts. So far, Lucene search queries can be generated from information contained in the SO and can be used for querying literature data bases or PubMed. However, this is still insufficient, since these queries are not well suited for querying XML documents because they are not following their structure. However, in the medical domain, many information items are coded in XML. Thus, querying structured XML documents is crucial for retrieving similar cases or for identifying potential study participants. For example, information items of patients with a similar tumor classification documented in a certain section of the respective pathology report need to be retrieved. This requires a precise definition of queries. In this paper, we introduce a concept for the generation of such queries using a Search Ontology XML extension to enable semantic searches on structured data. Results: For a gain of precision, the paragraph of a document need to be specified, in which a specific information item expressed in a query is expected to appear. The Search Ontology XML Extension (SOX) connects search terms to certain sections in XML documents. The extension consists of a class which represents the XML structure and a relation between search terms and this XML structure. This enables an automatic generation of XPath expressions, which makes an efficient and precise search of structured pathology reports in XML databases possible. The combination of standardized Electronic Health Records with an ontology based query method promises a gain of precision, a high degree of interoperability and long term durability of both, XML documents and queries on XML documents.
Since untagged information in health information systems (HISs) is common, information access supported by automatic methods is difficult. It is still an open question how to accelerate the access to information captured in these systems or in Electronic Health Records (EHRs). On the one hand, content must be structured by automatic recognition processes. On the other hand, the structured data has to be queried in a structured way.
This paper will focus on the query side, by introducing a new
solution of semantic meaningful queries on structured XML
documents, defined by the Search Ontology (SO) XML Extension. The
SO
In this work, we introduce an extension of the SO that enables the definition of queries on structured XML documents. Assuming that we have structured and standardized XML documents, then we can query certain parts of the XML document by XPath expressions. The development of such XPaths is time consuming for domain experts, but also for computer scientists. We suggest to use ontologies to support domain experts in modelling XML queries.
Out of the ontology based query models, XPaths can be generated automatically, which in turn can be applied to document corpora on XML database systems for searching similar cases or for the identification of potential study participants. Even though the approach is inherent independent from the underlying XML structure, we will demonstrate the approach on an example of querying standardized Electronic Health Records (EHRs) in the pathology domain.
To address the problem of creating structured queries for
retrieving documents, previous work considered the unification of different
XML structures on the conceptual level, on the one hand by the
introduction of new query languages, e.g. CXPath
Figure 1 gives an overview on the basic approach presented in this paper.
Fig. 1. (1) The domain expert models the queries by the usage of SOX in
Protégé. (2) Using an extended version of the OntoQueryBuilder Plugin,
Protége generates XPath expressions out of the ontology. (3) The expert applies
the generated XPath expressions to an XML database, that (4) returns the
relevant documents.
A domain expert is in the middle of the query formulation and
retrieval process. He uses Protégé, the ontology editor of the Stanford
University
The formulation of structured queries can be very time consuming, especially in safety-relevant domains like post market surveillance. A concept can be described in different ways, on the one hand by synonyms, on the other hand by complex phrases, which in turn consist of multiple terms. Because of that we distinguish Simple_Terms from Composite_Terms. Composite_Terms. are made up of Simple_Terms, related by the Object Property has_part and are constrained by the additional Data Property max_distance, which defines the word distance between two Simple_Terms, where max_distance=0 represents that one word immediately follows after another word. Writing variations, synonyms of abbreviations of the Simple_Terms can be handled by the assignment of multiple labels to the concrete individual of a Simple_Term.
For instance, the complication of a medical device (e.g. occluder) is a reusable Search_Concept, which can be described_by several Search_Terms.
To such descriptions belongs among other things adjective phrases like incomplete closure. Instead of the adjective, other terms with the same semantic meaning could be used; the noun could be replaced by any term which represents the meaning of closure. Out of this definition, a query can be generated, which is in this example the disjunction of all combinations of adjectives and nouns (cf. second disjunction in Listing 1). Listing 1. Lucene Query for an occlusion device complication; the expression was generated by the plugin OntoQueryBuilder.
("occlusion device" OR occluder) AND (("insufficient sealing"~2 OR "insufficient closure"~2 OR "incomplete sealing"~2 OR "incomplete closure"~2 OR "inadequate sealing"~2 OR "inadequate closure"~2)) The latter example indicates that the formulation of a query can become a complex task; the cross-product of only 10 adjectives with 10 nouns results in 100 adjective substantive combinations. Hence, we have to manage Concepts and Terms by an appropriate ontology, especially if we want to reuse concepts or if we want to generate cross-products of certain term combinations.
The SO is used in practice in the OntoVigilance project
In this paper, we will concentrate on the special domain of
pathology, where a lot of semi-structured information occurs in terms of
pathology reports. We consider this information semi-structured,
because the pathologists structure their information by headers and
keywords, but these structure is usually not technically
implemented. In fact, pathology reports are based on certain section
patterns and section-introducing keywords, like material, macroscopy
or microscopy. We verified manually that documents originated
from the Institute of Pathology of Leipzig, the sections introducing
keywords like Material, Makroskopie or Mikroskopie were
constantly used for section tagging. Therefore, the reports can be
structured in sections by section boundary detection, which is not the
main focus of this article. Consequently, legacy data can be
transformed into a structured format. Suitable standards for long term
persistence are EN 14822 a.k.a. HL7 RIM
In this work, we will use pathology reports, mapped to
standardized EHRs by the usage of the openEHR archetypes
openEHREHR-OBSERVATION.lab_test-histopathology.v1 for structuring
pathology data and openEHR-EHR-CLUSTER.tnm_staging_7th.v1
for structuring the TNM classification
<value>Makroskopisch</value> </name> <Overall_macroscopic_description> <name>
<value>Makroskopisch</value> </name> <value> The TNM structured classification string in the XML snippet is “pT2”. The sentence which introduces the overall macroscopy description contains a noun Hautexzidat (HE) (en: excised skin material), marked by a bold font. Due to linguistic variability, this noun can vary, i.e. synonyms or abbreviations such as H.E. are used in practice. In front of the noun there is an underlined adjective keilförmig (en: cuneiform) for specifying the shape. Again, semantic and linguistic variants of the term exist (e.g. rundlich (en: roundish)). Furthermore, the order of the adjectives in the phrase could change: the order “Ein ff. rundliches fadenmarkiertes H.E.” was found and is also valid.
When EHRs are stored in structured XML, another query language
is more suitable than classical free text retrieval methods such as
Lucene
The output of the SO are Lucene queries, but they do not follow the structure of XML documents, thus, they are not applicable to XML documents. However, the SO delivers already a reusable framework which only has to be extended for enabling structured queries in XML. By extending the SO with the SOX, queries are automatically producible out of the ontology, which can be executed on XML documents. For this purpose, two elements were added to the SO, on the top level of the ontology the class XML_Structure and the Object Property in. XML_Structure represents the XML document structure. Namespaces and tag names of the XML document are defined by XML_Structure class labels. Figure 5 illustrates the SO for the described use case, where the documents follow the structure of Listing 2 and XPaths of Listing 3 have to be generated as output.
The modelling of the SO (illustrated in fig. 5) has to be done manually by the domain expert. For querying HEs with different kinds of shapes the Search_Concept HE_Shapes was defined, described_by HE_Phrase, which consists of two Simple_Terms (HE_Form and HE_Term). In the example of this paper, individuals of HE_Form can be adjectives like rundlich or keilförmig; HE_Term has only one individual, the different writing variations (Hautexzidat, H.E.) can be handled by multiple label assignments. Figure 5 illustrates also the usage of the in relation for the specification of the position inside the XML document, for instance is the term T2_Term expected in the associated section Primary_Tumor. In a similar way, HE_Shapes are bound to the XML_Structure. The following figure 4 shows the class definition of HE_Shapes in Protégé. In summary, out of the DSO XML extension, it is possible to generate automatically the required XPath expressions. 3.2
For each Search_Concept one XPath can be generated
automatically. This can be done by a scheduled adaption of the Lucene
export plugin, the OntoQueryBuilder, which is already developed as
part of the OntoVigilance
The generated XPaths can be used for structured queries and for the integration in other XML techniques (XSLT or XQuery). 4
With an example on querying structured EHRs, we introduced an extension of the Search Ontology to support querying structured XML documents. The SOX approach can simplify the managing of a big pool of XPath expressions in one overarching DSO in practice. 4.1
Indeed, SPARQL queries on OWL based patient data would be more powerful than XPath expressions on XML, but a comprehensive and long term persistence storage of pathology data within semantic web technologies is only partially solved and still an open research question. Therefore, until there is no standardized domain ontology available, queries on standardized XML will be more stable and long term durable. To put it in brief, the first requirement is a layer of standardized EHRs and tools which work at this layer, like the introduced SOX. After that step a more powerful ontology layer is demandable.
We used the EN 13606 standardized XML in this work. However, another option would be EN 14822 or even any other proprietary XML format. When the community comes to an agreement, which EHR standard will be used in German Health Information Systems in future, not only the EHR would be interoperable, the usage of a standardized query language implies that queries could be interoperable too. Presupposed standardized EHRs would be used, it is imaginable that queries on such EHRs can be interoperable and therefore used in different hospitals. When openEHR is used, the depending SOX or the resulting queries could be stored in a repository and they could be linked to the belonging archetypes. Consequently, the query can be reused like the archetype itself, this would save development time and lead to quality intensification. 4.2
Research in Natural Language Processing (NLP) delivers methods
Listing 3 already respects declension forms of adjectives which have a stable word stem. But for the development of a minimal SO, NLP methods like stemming are necessary too. Because of that, we have to think about the integration of such methods into ontological contemplations about querying structured information in the near future. 4.3
XML elements are more than symbolic structures, they have to be
considered in detail, last but not least they should have be bound to
a top level ontology like the General Formal Ontology (GFO)
When EHRs are persisted in standardized XML, it is possible to query them in a structured way. The introduced Search Ontology XML extension connects search terms to certain parts in XML documents and enables an ontology based definition of semantic searches. Out of this, XPath expressions can be generated for querying XML database systems. Our solution supports the reuse oriented specification of complex and powerful XPath expressions without deep syntactic knowledge about XPath. The approach is open for additional extensions; parts of the ontology can be reused and adapted easily for other use cases.
Thanks to Claire Chalopin, Wolf Müller, Katrin Schierle, Lars Voitel, Christian Wittekind for their support, to the reviewers of ODLS for their constructive feedback, especially Dagmar Waltemath, and the organizers, among which we mention Frank Loebe and Daniel Schober, for arranging ODLS. This work was conducted using the Protégé resource, which is supported by grant GM10331601 from the National Institute of General Medical Sciences of the United States National Institutes of Health.