MiDas: Automatic Extraction of a Common Domain of Discourse in Sleep Medicine for Multi-center Data Integration

Abstract

Clinical studies often use data dictionaries with controlled sets of terms to facilitate data collection, limited interoperability and sharing at a local site. Multi-center retrospective clinical studies require that these data dictionaries, originating from individual participating centers, be harmonized in preparation for the integration of the corresponding clinical research data. Domain ontologies are often used to facilitate multi-center data integration by modeling terms from data dictionaries in a logic-based language, but interoperability among domain ontologies (using automated techniques) is an unresolved issue. Although many upper-level reference ontologies have been proposed to address this challenge, our experience in integrating multi-center sleep medicine data highlights the need for an upper level ontology that models a common set of terms at multiple-levels of abstraction, which is not covered by the existing upper-level ontologies. We introduce a methodology underpinned by a Minimal Domain of Discourse (MiDas) algorithm to automatically extract a minimal common domain of discourse (upper-domain ontology) from an existing domain ontology. Using the Multi-Modality, Multi-Resource Environment for Physiological and Clinical Research (Physio-MIMI) multi-center project in sleep medicine as a use case, we demonstrate the use of MiDas in extracting a minimal domain of discourse for sleep medicine, from Physio-MIMI’s Sleep Domain Ontology (SDO). We then extend the resulting domain of discourse with terms from the data dictionary of the Sleep Heart and Health Study (SHHS) to validate MiDas. To illustrate the wider applicability of MiDas, we automatically extract the respective domains of discourse from 6 sample domain ontologies from the National Center for Biomedical Ontologies (NCBO) and the OBO Foundry.

Introduction

A key informatics challenge for multi-center comparative effectiveness trials in clinical research is the integration of heterogeneous datasets from different study centers^[1]. Clinical studies often use a set of controlled vocabularies, called a “data dictionary,” to provide a naming convention for data capture, database design and construction. Each data dictionary consists of a set of terms with its specific semantics in the context of a particular study. Hence, multi-center studies require harmonization of the multiple existing data dictionaries to facilitate subsequent data integration. Domain ontologies are often created to model the data dictionaries in a formal language and serve as a common schema for integrating data^[2,3] from multiple centers. Using a single domain ontology to model terms from all possible data dictionaries, however, is impractical due to the rich variety of clinical data types. Hence, a suite of domain ontologies is needed. A common domain of discourse can facilitate the construction of interoperable domain ontologies which may share a significant portion of common terms but also require terms unique to a particular study.

For the purpose of this paper, by “domain of discourse” we mean a collection of terms at the juncture of transition from terms in a domain ontology to terms in an upper-level ontology. “Upper-level ontologies,” such as the Basic Formal Ontology (BFO) or Descriptive Ontology for Linguistic and Cognitive Engineering (DOLCE), model concepts at a high-level of abstraction^[4,5], while “Upper-level domain ontologies,” such as BioTop^[6] and Ontology for General Medical Science (OGMS)^[7], model concepts that are specific to a domain. Upper-level domain ontologies have been traditionally used to provide: (1) a common modeling basis, (2) conceptual clarity of the domain-specific terms, and (3) ontological design patterns for consistent modeling^[8]. The existing upper-level domain ontologies in biomedicine, created manually using a top-down approach, require an ontology engineer to identify relevant terms at an appropriate level of abstraction for use across different projects. In addition, the existing upper-level ontologies model concepts at a single level of abstraction, while modeling terms from multiple data dictionaries require concepts at multiple levels of abstraction.

The primary motivation for the work presented in this paper is the Multi-Modality, Multi-Resource Environment for Physiological and Clinical Research (Physio-MIMI^[1]) project for the integration of multi-center retrospectively collected sleep medicine data. We utilize sleep medicine data as an exemplar for complex clinical data that may be collected from a variety of sources within and between clinical centers (e.g., sleep clinic, hospital record, sleep laboratory) and in a variety of formats (including variables generated from proprietary data collection programs and which produce files of summary and raw physiological data). A key feature of Physio-MIMI platform, built on Ruby on Rails framework, is the innovative uses of ontology simultaneously for two purposes:

• for directly driving the federated query interface VISAGE (VISual Aggregator and Explorer^[9]) and

• for integrating autonomous data resources through the database to ontology mapper called PhysioMap^[10].

The Physio-MIMI platform uses a domain ontology as a “plug-and-play” component, which is capable of harmonizing disparate data sources without requiring the data sources to conform to a single data model. In addition, the Physio-MIMI platform can be reused in other application domains by simply switching the domain ontology. The Sleep Domain Ontology (SDO) is the domain ontology created for the Physio-MIMI project to drive data integration, access and query across the participating sleep research centers^[11]. A semi-automated auditing of the current SDO revealed that 82.5% (1,188) terms out of the total 1,467 terms are from external ontologies, many of which are not specific to the Physio-MIMI project. This unbalanced structure of SDO is due to the reuse of complete class structures from external ontologies. For example, SDO imported many terms in the Foundational Model of Anatomy (FMA), though only a few FMA terms are relevant in sleep research^[1]. In addition, only a few external (upper) ontology terms are used as parent terms for sleep domain-specific terms. Given the limited use of a majority of the external (upper domain ontology) terms in sleep research, we needed to identify the minimal set of external terms with closest proximity to the domain of interest. These terms constitute a minimal common domain of discourse for sleep medicine, from which multiple data dictionaries for sleep research projects can be constructed by extending in the direction of concrete data elements needed for a study.

The main contribution of this paper is the Minimal Domain of discourse (MiDas) algorithm to automatically extract a minimal set of upper bounds from existing ontologies using a bottom-up approach. The resulting domain of discourse can be used to create a suite of interoperable (through the sharing of a set of core terms) domain ontologies, which in turn can be used to systematically migrate multiple data dictionaries used in clinical research projects. MiDas provides knowledge engineers a theoretically sound and automated methodology to create a domain upper level ontology, modeling concepts at multiple levels of abstraction and suitable in use for specific clinical application domains. Using the Physio-MIMI project as a real world use case, we demonstrate the use of MiDas in the creation of a sleep medicine minimal domain of discourse, which has been leveraged to facilitate the integration of multiple existing data dictionaries for sleep research.

1. Background

MiDas draws on background knowledge from three areas. One is the description logic profile of the Web Ontology Language (OWL-DL) that provides the formal logic foundation for implementing MiDas and obtaining the results presented in Section 4. The second is the Sleep Domain Ontology (SDO) that forms the input resource for MiDas. The third is the SHHS data dictionary^[12] in the Physio-MIMI project that is used to evaluate the application and effectiveness of MiDas. We provide an overview of these areas for subsequent use in the latter sections.

OWL

OWL^[13] is a logic-based language that is used to create expressive knowledge models or ontologies (for a domain of interest) and which also supports automated reasoning to discover implicit knowledge^[13]. An OWL ontology consists of a set of assertions or statements that are composed of real world objects (e.g. Mary), category of objects (e.g. person), and relations linking objects and/or categories (e.g. parentOf)^[13]. In addition, OWL allows the above three entities to be combined together in form of an expression to represent complex information (e.g. brother of parent is uncle).

Classes in an OWL can be organized into a class hierarchy with more generic classes modeled as parents of more specific classes using the subClassOf property (e.g. sleepApnea subClassOf sleepDisorder). The subClassOf relation is transitive^[13], that is if in addition to the above example statement we consider the statement is sleepDisorder subClassOf disorder then it can be inferred that sleepApnea subClassOf disorder. Further, classes can be defined using a set of OWL properties, namely equivalentClass, complementOf, someValuesFrom (existential quantification), allValuesFrom (universal quantification), disjointWith, intersectionOf, and unionOf that refer to other classes (Figure 2).

Figure 2:

(1) Overview of Resulting IPs in Taxonomy and Component Graph (2) Traversing Component Graph by OWL Assertions

OWL uses the Internationalized Resource Identifiers (IRIs) as names for classes and properties in an ontology. Entities in an ontology are usually prefixed by a namespace that refers to the full IRI of the ontology, for example owl:subClassOf is composed of the owl namespace representing the OWL IRI http://www.w3.org/2002/07/owl# that is followed by the subClassOf relation.

Sleep Domain Ontology (SDO)

SDO has been developed and used in Physio-MIMI to model terms representing sleep medication, disorders, clinical findings, and physiological phenotypes^[11]. The coverage of SDO includes, clinical findings, procedures and devices used in sleep medicine, anatomical descriptions, and information entities such as polysomnograms^[11]. The primary role of SDO is reconciling data heterogeneity in sleep medicine hence it defines additional constraints on the types of permissible data values. For example, SDO enumerates the set of permissible values for diagnosis and medication classes, while history cough can be assigned only Boolean values^[11].

The current version of SDO has 1,467 classes and 105 properties that includes classes defined in external ontologies, such as the Foundational Model of Anatomy (FMA)^[14], the Basic Formal Ontology (BFO)^[5], OGMS^[7], and the Computer-Based Patient Record (CPR) ontology^[15].

Physio-MIMI

The primary aim of Physio-MIMI project is an informatics infrastructure for the integration of, and access to, disparate physiological and clinical data. The Physio-MIMI project has two salient features: first, it uses a federated approach for data integration, allowing it to successfully integrate large datasets (polysomnograms can range from 1 to 20GB in size). Second, the Physio-MIMI project is tightly focused on a user-centered interface, coupled with the agile software development framework, to rapidly adapt to user and project requirements. The key challenges are to integrate, query, and visualize the datasets from multiple sleep studies conducted at different clinical centers, each of which may use different data dictionaries.

2. Methods

MiDas exhaustively computes the member terms of a minimal domain of discourse for any domain ontology. In the case of sleep medicine, we use the SDO as the input domain ontology and the output is a sleep domain of discourse, which facilitates integration of multiple data dictionaries in the multi-center Physio-MIMI project.

Our method for creating a domain of discourse involves three steps: 1) acquiring the SDO file; 2) creating the taxonomy graph; 3) computing the domain of discourse.

Acquiring SDO data

The SDO OWL file was retrieved from the Physio-MIMI public website^[1] and parsed while ignoring all owl:imports statements. Since, the MiDas algorithm does not require the complete external ontology, discarding the owl:imports statements allows us to reduce the total size of the OWL file to be parsed and loaded into memory for further computations.

Creating the taxonomy graph

In addition to the explicit assertion of class hierarchy (using rdfs:subClassOf), an OWL ontology may also include implicit assertions of class hierarchy using additional OWL properties. For example, Figure 1 illustrates how an explicit class hierarchy can be inferred from a set of OWL assertions.

Figure 1:

Example Taxonomy Graph Creation

The MiDas algorithm generalizes the above example to define a transformation step for converting an OWL ontology (with implicit assertions) into a taxonomy graph (with only explicit assertions using rdfs:subClassOf). Formally, a taxonomy graph is defined as:

Definition 1. Given an OWL RDF graph 𝒪. Another OWL RDF graph 𝒪^tax is a Taxonomy Graph of 𝒪 if it only has one relationship rdfs:subClassOf between classes and all of them are produced using Algorithm 1.

In RDF Schema (RDFS) and OWL, the rdfs:subClassOf property includes concept subsumption ⊑ and the owl:equivalentClass property includes concept equality ≡ as part of their logical underpinning in description logic. We use three basic entailments in the algorithm:

• Entailment (1) states that if two classes are equal, then each one is a subclass of the other one: C ≡ D |= (C ⊑ D) ∧ (D ⊑ C)

• Entailment (2) states that if class C is the union of a group of c_i, then every c_i is a subclass of C: (C ≡ c₁ ⊔ c₂ ⊔ · · · ⊔ c_i) |= ∧ⁱ(c_i ⊑ C)

• Entailment (3) states that if class C is the intersection of a group of c_i, then C is a subclass of every c_i: (C ≡ c₁ ⊓ c₂ ⊓ · · · ⊓ c_i) |= ∧ⁱ(C ⊑ c_i)

It is important to note that the taxonomy graph created using the transformation step should be consistent with the original ontology. Hence, we give a formal proof that the transformation step used to create a taxonomy graph from the original OWL ontology file is sound, that is, all possible entailments supported by the taxonomy graph are also supported in the original OWL file.

Theorem 2. Given an OWL RDF graph 𝒪 and its Taxonomy Graph 𝒪^tax, let 𝒪_𝒧 and ${𝒪}_{𝒧}^{\mathit{tax}}$ be the underpinning ontologies separately over description logic 𝒧. For every axiom α over 𝒧, we have ${𝒪}_{𝒧}^{\mathit{tax}}|=\alpha$ entails 𝒪_𝒧 |= α.

Proof. From the definition of Taxonomy Graph, the axiom only has one form as B ⊑ A. Suppose ${𝒪}_{𝒧}^{\mathit{tax}}|=B⊑A$, then we have 〈B rdfs : subClassOf A〉 in 𝒪^tax, which comes from four cases:

• If Case 1 in Algorithm 1 holds, it exists in the original 𝒪, then 𝒪_𝒧 |= B ⊑ A holds directly;

• If Case 2 in Algorithm 1 holds, then we have 〈B owl : equivalentClass A〉 in 𝒪, thus 𝒪_𝒧 |= A ≡ B holds, as a result, we have 𝒪_𝒧 |= B ⊑ A because of entailment (1);

• If Case 3 in Algorithm 1 holds, then we have 〈A owl : unionof A′〉 and 〈A′ rdf : item B〉 in 𝒪, thus 𝒪_𝒧 |= A ≡ (B₁ · · · ⊔ B ⊔ · · · B_n) holds, as a result, we have 𝒪_𝒧 |= B ⊑ A because of entailment (2);

• If Case 4 in Algorithm 1 holds, then we have 〈B owl : intersectionof A′〉 and 〈A′ rdf : item A〉 in 𝒪, thus 𝒪_𝒧 |= B (A₁ · · · ⊓ A ⊓ · · · A_n) holds, as a result, we have 𝒪_𝒧 |= B ⊑ A because of entailment (3).

Algorithm 1:

Taxonomy Graph Creation

Computing the domain of discourse

In the next step of the algorithm, all SDO classes are classified as either a Domain Term or a Non-Domain Term, using the following definitions:

Definition 3. Given an ontology 𝒪, a prefix or base IRI common to all the identifiers of terms in the domain, a term t ∈ Sig(𝒪) is a Domain Term if it begins with that prefix and Sig(𝒪) denotes the signature of ontology 𝒪.

Definition 4. Given an ontology 𝒪, a prefix or base IRI common to all the identifiers of terms in the domain, a term t ∈ Sig(𝒪) is a Non-domain Term if its identifier does not begin with that prefix.

First, the algorithm identifies the immediate parent of a domain term that is also a non-domain term in the taxonomy graph. These terms are called Immediate Parent (IP) classes. In the second phase, the algorithm computes any non-domain terms that were not identified in the first phase but are used in the definition of a domain class. These terms are called Component classes.

The set of non-domain terms that are present in the result set of the above algorithm constitute the common domain of discourse for a given domain (sleep medicine in this paper).

Algorithm 2 is implemented using a single Python command-line script. The RDFLib 2.4 library was used along with the InfixOWL library^[16] for parsing and serializing the SDO ontology into memory and navigating the ontology structure, respectively. Rather than produce a separate taxonomy graph, the implementation included an Application Programming Interface (API) that allowed direct navigation of an ontology’s taxonomy graph. The InfixOWL library includes a function for navigating the components of an OWL class definition and this was used for navigating the ontology structure.

Implementation

MiDas has been implemented as standalone generic tool that can use any domain ontology (expressed in OWL) as input to extract a minimal domain of discourse. In addition to its application in sleep medicine, in the Discussion section we describe the use of MiDas to create domain of discourse for 6 domain ontologies selected from the OBO Foundry^[17] and the National Center for Biomedical Ontologies (NCBO).

In addition to creating the taxonomy graph and implementing Algorithm 2, the computing of transitive closure over the taxonomy graph (based on the rdfs:subClassof property) represents the third computation task of MiDas. The computations were performed on a machine running the Mac OSX operating system with a quad-core Intel i7 processor and 4GB of memory.

Algorithm 2:

BottomUpDomainBoundary Algorithm

3. Results

Results from SDO

From the 1,467 original classes in SDO, the first phase of Algorithm 2 resulted in 17 IP classes, while the second phase of the Algorithm identified 12 Component classes. Hence, a total 29 non-domain terms were identified to form the minimal domain of discourse for sleep medicine. The 29 terms extracted from a total of 1188 non-domain terms in the original SDO represents a 97.6% reduction in the size, which is easier to manage (both for ontology engineers and users). Table 1 summarizes the distribution of the domain of discourse terms according to their source ontology.

Table 1:

Distribution of External Ontology Classes in the Domain of Discourse Identified by MiDas

Non-domain Classes	Ontology for General Medicine Science (OGMS)	Computer-Based Patient Record(CPR) Ontology	Foundational Model of Anatomy (FMA)	BioTop Ontology	Basic Formal Ontology (BFO)	Drug Ontology
Immediate Parent classes	6	7	0	3	1	0
Component classes	0	2	6	2	1	1

Evaluation

As we discussed earlier, the primary application of MiDas is in creation of a domain of discourse for facilitating integration of multiple data dictionaries and thereby enable data integration in multi-center clinical studies. We used the results of applying MiDas on SDO to integrate terms from the SHHS data dictionary^[12] as part of the Physio-MIMI project.

The resulting SDO modeled classes representing terms from the Physio-MIMI project (on physiological variations in individuals with different genotypes, disease risk factors and health outcomes). This validated the applicability of the MiDas algorithm to create a minimal domain of discourse for integrating multiple data dictionaries that were created for different clinical studies.

Validation of extended SDO by domain expert: We are collaborating with sleep domain expert to validate the extended SDO. Using structured feedback forms(Figure 4), the domain experts can review the new concepts and their positions in the SDO class hierarchy.

Figure 4:

Review form for eliciting domain expert feedback on extended SDO

To further validate the generic approach of the MiDas algorithm, we applied it on 4 ontologies hosted at the OBO Foundry^[17] and 2 NCBO ontologies.

Applying MiDas to OBO Foundry and NCBO ontologies

The 6 ontologies selected for this additional study are briefly described:

1. Expressed Sequence Annotation for Humans (eVOC) ontology: There are four core eVOC ontologies that represent terms used for describing sources of samples used in tissue expression profile, differential gene expression levels and distribution of expression across the genome.

2. Infectious Disease Ontology (IDO). The IDO ontologies are a set of ontologies representing terms related to various diseases, including dengue fever, influenza, malaria, and tuberculosis among others.

3. Ontology for Biomedical Investigations (OBI). The ontology is a comprehensive provenance ontology covering more than 18 communities, including proteomics, transcriptomics, imaging, and toxigenomics. OBI represents metadata terms for tracking the provenance of experiment data and published results in biomedicine.

4. Xenopus anatomy and development (XAO) ontology. XAO is a set of controlled vocabularies of anatomy and development of the African clawed frog (Xenopuslaevis).

5. Zebrafish anatomy and development (ZFA) ontology. ZFA a set of controlled vocabularies of anatomy and development of the Zebrafish (Danio rerio).

6. Phenotypic quality ontology (PATO). The ontology models terms used in phenotype annotation.

Table 2 summarizes the result of applying the MiDas on the 6 ontologies. Community driven ontologies such as OBI and IDO have a large number of non-Domain terms, primarily from the upper-level ontology such as BFO. The results for the two OBO Foundry ontologies, namely XAO and PATO, are surprising as similar to OBI they were also expected to use BFO classes as parent classes for their domain terms. A manual review of the ontologies using the NCBO ontology visualization service confirmed that XAO and PATO do not include BFO terms (unlike OBI and IDO).

Table 2:

Result of Minimal Common Domain of Discourse Extraction for 2 NCBO Ontologies and 4 OBO Foundry Ontologies

	Ontology	Domain	Domain Classes	Immediate Parent classes	Component classes
NCBO	eVOC	Expressed Sequence Annotation for Humans	1000	0	0
	IDO	Infectious Disease Ontology	403	13	0
OBO	OBI	Ontology for Biomedical Investigations	1504	38	66
	XAO	Xenopus Anatomy and Development	817	1	0
	ZFA	Zebrafish anatomy and development	2686	1	38
	PATO	Phenotypic quality	2203	1	0

4. Discussions

As the number of collaborative research projects increases, exemplified by the 60 Clinical and Translational Medicine Awards (CTSA) since 2006 by the National Institutes of Health (NIH), the already difficult issue of large-scale data integration will increasingly become a critical challenge. Hence, a bottom-up, automated approach to create a upper-level domain ontology in place of a top-down prospective approach (e.g. OGMS ontology) represents a scalable and flexible approach. In addition, the use of an existing domain ontologies in the MiDas algorithm leverages a structured knowledge model that is often created a community and represents a consensus among the community members.

A manual review of the sleep domain of discourse terms identified by MiDas reveal a set of classes from multiple upper-level ontologies with different levels of abstraction. This reflects that terms in a common domain of discourse spans multiple levels of abstraction, for example the BFO class ProcessualEntity is at a higher level of abstraction in comparison to the BioTop ontology class HumanPopulation. In addition, terms from multiple domain upper level ontologies such as the OGMS ontology class OGMS_0000015 (clinical history) means that a single upper-level ontology cannot address the requirements of integrating multiple data dictionary. Hence, the ability for creating a domain of discourse that includes terms from multiple upper-level or domain upper level ontologies, which is already modeled in the input domain ontology, is an important feature of the MiDas algorithm.

One of the prerequisites and potential limitation of the MiDas algorithm is the existence of a domain ontology that can be used as input. Creation of a domain ontology is an expensive and time consuming endeavor that may not be possible for many biomedical domain^[18]. But, the rapid increase in the number of ontologies submitted to the NCBO, with more than 260 listed ontologies, points to possible solution to the issue of availability of domain ontologies.

5. Conclusion

MiDas automatically creates an effective domain of discourse to facilitate integration of data dictionaries in multi-center clinical studies. In comparison to existing domain upper level ontologies, MiDas creates a minimal domain of discourse and includes terms from multiple domain upper level ontologies are different levels of abstraction. This flexible, bottom up approach in MiDas, in contrast to the top-down approach used in creation of existing domain upper level ontologies, is necessary to address the requirements of clinical studies such as Physio-MIMI project. The integration of the SHHS data dictionary into the domain of discourse, extracted by MiDas from SDO, validates the usability and effectiveness of the MiDas algorithm.

Figure 3:

Integration of SHHS Data dictionary terms into SDO as Children of sleep Domain of Discourse Terms