3

Understanding Ontologies

Ontology is not a word used frequently in everyday discourse, perhaps even among scientists, and it can have slightly different meanings depending on the context. The committee sought to understand varying usages of the word and to identify a definition for our work. This chapter discusses the term’s etymology in the context of computer and information science, since work in these fields has been the basis for one of the most widely used definitions, and it explains the prevailing definition of the term. In considering how this definition applies in behavioral sciences, the committee recognized that existing means of classifying and structuring knowledge in these fields lie on a continuum, and that varied approaches have utility for their intended purposes. We applied this idea in examining several examples relevant to the behavioral sciences. The chapter closes with the committee’s conclusions about the nature of behavioral ontologies.

DEFINING ONTOLOGY

From the perspective of computer science, an ontology has been defined as a shared conceptualization (of the “objects, concepts, and other entities that are assumed to exist” in a particular domain) that is formally specified (Gruber, 1995, p. 908; Gruber, 1993; see also Studer et al., 1998). This definition emerged from a study by the U.S. Defense Advanced Research Projects Agency (DARPA) of how knowledge could be shared across computer systems (Neches et al., 1991). In the early 1990s, participants in this DARPA initiative argued that artificial intelligence (AI) would require the use of standard ontologies to ground content-specific agreements for the

sharing and reuse of knowledge among software systems. The application of ontologies in this context was led by Gruber, whose work on an ontology-representation language known as Ontolingua demonstrated how ontologies could be encoded in a logic-based, knowledge-representation system for use by intelligent computer programs. Although earlier computer scientists working on intelligent computer systems had identified the importance of ontology engineering as a component of their work (e.g., Regoczei and Plantinga, 1987), Gruber and his colleagues clarified the role of ontologies in knowledge engineering and offered a solution to one of the identified technical limitations of shared, reusable knowledge-based systems. As a result of this work, Gruber and his colleagues helped make the notion of ontology engineering an integral part of computer science, and his definition of ontology is the most widely referenced one.

In Gruber’s definition of ontology, “explicit” refers to the manner in which a developer carefully enumerates the types of concepts used and the constraints on their uses; “conceptualization” refers to an abstract view of the world consisting of the relevant concepts and the relationships among them that exist within a specific domain.¹ “Formal” refers to specifications that are machine readable and have well-defined semantics, and “shared” refers to the conceptualization being agreed on and accepted by those working in a discipline. Based on this definition, an ontology’s primary purpose is to represent the entities in a domain by providing sets of machine-understandable statements and linking the descriptions and classifications of the terms and relationships among them.

Thus, in the context of computer and information science, an ontology refers to a specification of entities within a domain, which loosely parallels the philosophical definition of ontology as the science or study of existence (e.g., Stanford Encyclopedia of Philosophy²). According to Gruber, what “exists” is that which can be represented (Gruber, 1995). Therefore, ontologies formally enumerate the entities in some discipline, their relationships, and their definitions. More precisely, they do so by means of a declarative formalism: a knowledge representation that defines classes (or types), attributes (or properties), individuals (or specific members of a class), and relationships among class members (Gruber, 2016). This enumeration of a set of concepts within a domain is intended to help solve real-world problems by providing a vocabulary to support statements about the knowledge in that domain. Those representational terms are typically defined with both human-readable text describing what the term means (for use by people) and formal axioms that constrain the interpretation and use of the terms (for use by computers) (Gruber, 1995).

___________________

¹ A concept is an abstract idea that represents a class of objects or events. Concepts and relationships among them are significant components of ontologies: see Chapter 2.

² See https://plato.stanford.edu/entries/logic-ontology/#Ont

The committee’s work was guided by Gruber’s definition of ontology as a “formal, explicit specification of a shared conceptualization.” This definition provides flexibility for considering the potential for various classification systems to accelerate progress in the behavioral sciences. The definition does not constrain ontologies to any particular kind of formalization, nor does it specify how widely shared a conceptualization must be to form the basis for an ontology, as long as the conceptualization is defensibly formal, explicit, and shared. In particular, as discussed below, an ontology need not be formalized in a logic such as the Web Ontology Language (OWL); see Box 3-1.

Not all ontology developers accept Gruber’s definition, although it is by far the most cited, with more than 21,000 references, according to Google Scholar. In particular, some ontologists, often called ontological realists, reject the idea that ontological terms reflect any kind of conceptualization. Instead, they construe the terms of an ontology as representing entities in an objective reality (Smith and Ceusters, 2010), and they argue that the terms in an ontology correspond to universal distinctions about the world that require no cognitive interpretation. The notion of ontological realism has faced significant challenges, although it has received considerable attention within the community of scientists building biological ontologies (Merrill, 2010).

The Gruber definition is important for thinking rigorously about how to strengthen ontologies in the behavioral sciences. However, determining precisely which possible ontologies meet its criteria is not straightforward. Much of the discussion in this report applies not only to systems that do meet those criteria, but also to the larger set of purposefully developed resources that enumerate the essential entities in a discipline. Therefore, unless otherwise specified, in this report the word ontology refers to that larger set.

A CONTINUUM OF SEMANTIC SPECIFICATION

Ontologies are used in many different kinds of applications, including those for information integration, knowledge management, Semantic Web services, and enterprise application integration. Ontologies can be used in different ways depending on the nature of the problem at hand. For example, ontologies can be applied to improve information retrieval systems by providing a common understanding of concepts that humans and computers can both use. Ontologies can also be applied to undergird automated reasoning systems by providing formal definitions for concepts and the relationships among them (Staab and Studer, 2016).

Many authors argue that ontologies may be specified in various ways, such as lists of controlled terms, thesauri, taxonomies, and formal representations in logic, as all of these can represent formal, explicit specifications of shared conceptualizations—although with different degrees of formality. Some authors suggest that folksonomies—collections of terms offered as descriptors by communities of users, such as hashtag terms used to classify postings to social media—can be viewed as something close to an ontology. The committee did not, however, include folksonomies as ontologies since they are not managed systems that formally specify definitions or relationships among the terms.

Ontologies shape many aspects of human life, including media consumption, e-commerce, and the use of social media. For example, the ability to turn on Netflix and scroll through recommended movies depends in part on an ontology that the company has used to classify its content (Madrigal, 2018). Formal ontologies can be quite complex, with representation in logic that supports computer-based reasoning about the entities in the ontology. For example, the British Broadcasting Company uses a collection of publicly available ontologies to describe its content, including ontologies of journalism, politics, sports, and radio and television programs, all encoded in a logic that supports reasoning by computers.³ These representations allow computers to perform queries such as “Find all news programs that

___________________

³ See https://www.bbc.co.uk/blogs/internet/entries/78d4a720-8796-30bd-830d-648de6fc9508

discuss racquet sports,” using the ontologies to resolve which programs are news programs and which sports are racquet sports.

As these examples suggest, ontological systems may lie on a continuum of increasing semantic complexity. That is, classification systems designed for ontological purposes (the specification of definitions and relationships) may include weak semantics (such as a simple taxonomy that specifies only class—subclass relationships) or strong semantics (such as formal representation in a logic that allows developers to specify the properties of entities and constraints on those properties).⁴ Thus, we use the term ontological systems when referring to those that may or may not meet the definition of ontology or when that issue is relevant to the discussion.

Figure 3-1 illustrates the spectrum of semantic specification used in the context of the behavioral sciences, showing where controlled lists, thesauri, loose hierarchies, and taxonomies fall. Controlled lists, such as a list of social and behavioral determinants of health, are enumerations of specifically defined terms that help to provide consistency for users of the list. Thesauri organize terms so that the grouping reflects relationships among the terms (generally unstated): closely related terms are near one another, although exact relationships are unspecified. Taxonomies expand on thesauri by also showing hierarchical, class–subclass relationships, such as parent-child relationships, but the concepts are only enumerated: the relationships between concepts are not expressed in formal axioms. Table 3-1 provides more information about the examples shown Figure 3-1.

___________________

⁴ The committee benefited from a workshop presentation by Deborah McGuiness in which she discussed the idea that ontologies lie on a spectrum from a simple finite list of terms to expressive ontologies that specify logical constraints and detailed relationships (see, e.g., Lassila and McGuiness, 2001).

TABLE 3-1 Examples of Ontological Systems on a Continuum

NOTES: See Figure 3-1; see Appendix A for more information on the ontologies referred to in this report.

^a The five traits are extraversion (or extroversion), agreeableness, openness, conscientiousness, and neuroticism; see Goldberg (1981).

The committee considered several other ontological systems that are familiar to behavioral scientists but did not place them on the continuum because it was not clear how to characterize them in terms of the continuum. For example, the Hierarchical Taxonomy of Psychopathology (HiTOP) is a system for enumerating the behavioral phenotypic manifestation of psychiatric problems, organized through the use of factor analytic methods (covariance between symptoms). Hierarchical relationships in HiTOP have the potential to allow for improved classification of psychopathology dimensions to facilitate research and clinical practice. Another example is the Research Domain Criteria (RDoC), a framework of brain-based systems that may be associated with psychopathology. It integrates many levels of information, from genomics to behavioral processes. We discuss these two systems further below.

The point of the continuum is to demonstrate that a variety of representation systems are used in the behavioral sciences and that these representation systems lie on a continuum of semantic specification. Considering this

variety and the importance of designing classification structures to suit the needs of the researchers in a particular domain, the committee found that it was not useful to try to discern a strict cutoff below which a structure would not be considered an ontology or to classify known structures as ontologies or “non-ontologies.” Instead, we highlight that the structures that exist in the behavioral sciences serve ontological purposes that are scientifically valuable.

As the Gruber definition suggests, what ontologies have in common is that they provide a structure for the enumeration of the entities in some domain: they articulate formal decisions about what is known and, to varying degrees, about the relationships among the elements of what is known, and they provide a means for sharing these enumerations across diverse approaches and methodologies (Bilder et al., 2009; Poldrack and Yarkoni, 2016; Blanch et al., 2017). It is in this sense that a range of systems may serve ontological purposes. The terms enumerated by an ontology are symbols that take on their meaning through the shared conceptualization that relates the symbols to the entities in the world to which they refer. Much of the value of ontologies comes from following a common formalization that can be adopted for many tools and computational systems. But an ontological system with strong semantics is not necessarily better than one with weaker semantics. The use of strong or weak semantics will fit a specific purpose or set of purposes, and each inevitably will reflect the relative immaturity or maturity of the domain it is designed to systematize. However, there are often times when an ontology requires strong semantics and the attendant complexity of a formal logic to address a nuanced problem. For example, the use of strong semantics may be necessary to standardize and align related measures that may otherwise be unclear or imprecise.

The choice of an ontology-representation language may also hinge on the skills of the developer. For example, ontology engineers may choose to specify their conceptualization in a logic, such as OWL, that offers a rich semantics for description of how entities in the world relate to one another in subtle ways. However, it can be hard for many novices to use OWL without making simple mistakes. Instead, an ontology engineer might use a taxonomic representation in which it is easier for novices to encode class—subclass relationships, but this option obviously constrains what the developer can say about the entities being specified. Other languages that can be used include Excel, UML,⁵ object-oriented programming languages (e.g., Java), SKOS,⁶ RDFS,⁷ and Common Logic.⁸

___________________

⁵ See https://www.visual-paradigm.com/guide/uml-unified-modeling-language/what-is-uml/

⁶ See https://www.w3.org/TR/skos-primer/

⁷ See https://arxiv.org/ftp/arxiv/papers/1401/1401.3858.pdf

⁸ See https://www.w3.org/2004/12/rules-ws/slides/pathayes.pdf

When ontologies are represented in a logic such as OWL, the rich representation may make it straightforward to incorporate additional, non-ontological relations that enhance the scope of what the developer has represented. For example, one can add causal relations that allow the computer to identify how the presence of one entity may result in the presence of another, or to identify diagnostic relations that support problem solving to infer the presence of an entity based on the presence of other entities or on the properties of those entities. These complexities augment the representation from a specification of merely “what exists” to one that encodes knowledge about what inferences can be made if some entity is known to be present or known to have a property that takes on a particular value. Such an encoding, which goes beyond a representation of the essential elements in the domain and their basic relationships, is referred to as a knowledge base. Ontologies that are encoded in a semantically rich representation can provide the starting point for encoding full-fledged knowledge bases that support automated reasoning about the ontological entities, thereby providing a user with decision support, case analysis, or some other form of problem solving. The committee, however, focused on the representation of “what exists”—with the recognition that the creation of ontologies in the behavioral sciences would be the first step in creating a host of advanced computer systems that could reason about the knowledge bases derived from such ontologies.

EXAMPLES OF ONTOLOGICAL SYSTEMS

To gain a more detailed understanding of how ontological systems function in the behavioral sciences the committee explored several examples. We look first at an example of a formal and explicit ontology that lies on the far right of the continuum of semantic formality shown in Figure 3-1 (above). We then examine three different systems for classifying mental health problems.

A Formally Specified Ontology: The Behavioral Change Intervention Ontology

There are very few ontological systems in the behavioral sciences that have been developed using standard representation languages such as OWL. While the Behavioral Change Intervention Ontology (BCIO) is still a work in progress (as is the case with many ontologies), it provides an example of what a behavioral science ontology looks like when constructed by practitioners in the field using a standard representation approach.

The BCIO is being developed as part of the Human Behaviour-Change Project at University College London (Michie et al., 2017). The developers

were responding to a problem in the study of behavior change interventions: such interventions are heterogenous in context, content, and methods of evaluation, and this heterogeneity makes it difficult to synthesize evidence, develop real-world policy, and design practical applications (Elliott et al., 2014; Michie et al., 2021). Especially concerning was the lack of common terms for such interventions, which is evident in the proliferation of theories and concepts in the relevant research (Larsen et al., 2017). For example, a multidisciplinary literature review of theories of behavior change, with strict inclusion criteria in relation to theory and behavior, identified 83 different theories with a total of 1,725 component constructs (Davis et al., 2015). The review showed that these theories tended to be overlapping and underspecified, often sharing constructs with other theories, using different names for the same constructs, measuring the same constructs using differing items, and inadequately defining constructs and relationships.

Thus, the overall aim of the Human Behaviour-Change Project has been to automate evidence searching, synthesis, and interpretation to make it easy to rapidly search for the answers to questions from clinicians, researchers, and policy makers who want to know: What works? Compared with what? How well? With what exposure? With what behaviors (for how long)? For whom? In what settings? Why? (Michie et al., 2021).

To achieve rapid retrieval of this sort, it was necessary to organize evidence on behavior change interventions ontologically, in a formal and explicit way. That is, a shared formal description of entities and relationships capturing domain knowledge was needed to support aggregation and semantic querying (i.e., finding information using not only the presence of words, but also their meaning) (Michie et al., 2021). Modeling their work on the Gene Ontology,⁹ the developers of the BCIO created an ontology with a formal, rich, and explicit specification of concepts, which is why it falls on the far-right side of the continuum depicted in Figure 3-1 (above). Their ontology defines and organizes entities and the relationships among them in terms of a hierarchy using a common language that can cross disciplinary boundaries and topic domains.

The ontology for behavior change interventions was developed in accordance with the principles of the Open Biological and Biomedical Ontology (OBO) Foundry (Smith et al., 2007). The OBO Foundry represents the work of a group of investigators who promote collaboration and interoperability of ontologies in the biomedical sciences by providing a common framework for ontology development, with a commitment to the use of standards and good ontology engineering practices (Michie et al., 2021). The members of the OBO Foundry recommend the use of an upper

___________________

⁹ See http://geneontology.org/docs/introduction-to-go-resource/; Ashburner et al. (2000); also see Appendix A.

ontology, one that models extremely general distinctions about entities in the world (e.g., whether something is a temporally limited process or whether it persists in time), captured in the Basic Formal Ontology (BFO)¹⁰ (Arp et al., 2015; Grenon et al., 2004; Smith and Grenon, 2005).

The BFO provides a formal ontology beneath which other ontologies, such as the BCIO, are developed (Michie et al., 2021). The BFO is intended to promote clarity and interoperability among ontologies, which naturally promotes further synthesis and integration, not only within disciplines or subdisciplines, but also across them. Empirical evidence suggests that upper-level ontologies, such as the BFO, can be hard to use, however, even by developers who are well versed in their constructs (Stevens et al., 2018). The committee appreciates the stringent principles for ontology engineering advocated by the OBO Foundry, but we recognize that some of these guidelines are not uniformly accepted, and scientific ontologies that do not adhere to OBO Foundry criteria may still be valid.

The BCIO developers used a multistep, rigorous, and iterative process to build their ontology (Mitchie et al., 2021). Although the approach taken by the BCIO team was tailored to the particular challenges of creating this ontology, it is representative of ontology engineering practices commonly used in the sciences, and it provides insight into the tremendous human effort needed to develop a practical ontology of even modest scope.

The BCIO team identified 12 entities as central and common to all behavioral interventions, which are presumably familiar to intervention scientists: intervention, content, delivery, mechanism of action, exposure, reach, engagement, context, population, setting, behavior and outcome. Using these entities, further entities were defined, including such concepts as source, mode, schedule, dose, fidelity, and adherence.

The next step in the development was an expert feedback phase, during which experts were asked to rate the extent to which entity names were clear, definitions were nonoverlapping and without redundancy, relationships were suitable, and the overall structure was clear.¹¹ This feedback was discussed and incorporated into further refinements of definitions until consensus was reached, and the results were shared with a wider team that included systems architects and computer scientists.¹² The resulting behavioral interventions ontology has two types of behavioral change interventions, each with its own set of associated entities. The entities are related by 19 ontological relationships, such as has part, subclass of, has attribute,

___________________

¹⁰ See https://basic-formal-ontology.org/

¹¹ The feedback report is available as extended data at https://osf.io/yj235/; also see West et al. (2020).

¹² The full report of this phase is available at https://github.com/HumanBehaviourChangeProject/ontologies; also see Norris et al. (2020).

has disposition, has process part, evaluates, has output, is about, difference between (Mitchie et al., 2021). Each of the entities in the final version of the ontology has a parent class that relates to concepts from external ontologies developed by other groups, including the Information Artifact Ontology (Ceusters, 2012), which provides entities of relevance for describing data and information, and the Ontology for Biomedical Investigations (OBI, Bandrowski et al., 2016).

Ontologies are not static, and they are best viewed as reflecting continuing processes of dynamic adjustment to changes in the scientific consensus, rather than as capturing immutable ideas. Thus, the final step in the construction of the BCIO ontology was the establishment of a change-management and version-tracking strategy, which is one of the OBO Foundry principles of good practice. The most general classes of BCIO have been made available in OWL and are stored in the Human Behaviour-Change Project’s GitHub repository for open access. The BCIO lower-level ontology terms containing more granular concepts are reported to be under development at present. Thus, several behavioral change concepts are not yet specified in the BCIO (e.g., different subclasses of behavioral change techniques).

The developers of the BCIO hoped that a systematic approach to describing behavioral interventions and the contexts in which they have been used and studied will support researchers and clinicians in a number of ways. The BCIO, they suggest, helps structure thinking and communication about behavior change interventions. This structure can help researchers to identify knowledge gaps, to develop fruitful lines of inquiry, and to evaluate protocols. It can facilitate the synthesis of evidence and theories within and across intervention disciplines. An ontology such as the BCIO can also facilitate research that harnesses the power of artificial intelligence-based approaches such as machine learning and natural language processing for the purpose of searching databases and synthesizing, interpreting, and generating evidence and insights. Indeed, the developers suggest that the BCIO can be used as part of computer systems to evaluate the published literature and even help to generate research papers.

The BCIO illustrates the value of a formal and explicit ontology in the behavioral sciences because it provides a way to move beyond a prevalent challenge in the discipline: teams of researchers who work in silos, using data that are incompatible with data used by others in similar domains. It allows for integration of evidence when BCIO terms are shared by diverse groups of investigators to describe their data, and it makes the evidence searchable when the same ontological terms appear in both the underlying data descriptions and in the queries with which users perform their searches. Computers potentially can reason about behavioral intervention data to discover new relationships, to develop novel hypotheses, and to

expose gaps in the evidence (Larsen et al., 2017). We consider these opportunities more generally in Chapters 4 and 5.

Certain challenges also emerged in the building of the BCIO. For instance, the developers reported that, given the dearth of ontologies in social and behavioral sciences, there were few existing ontologies of behavior change to draw on (Mitchie et al., 2021). It was also challenging at times for the developers to clarify subtle distinctions in the definition of entities.

There are tradeoffs in choosing the degree of “richness” in semantic expressiveness to use when developing any ontology. It might be relatively easy to define “simple” classification systems with only controlled vocabularies and loose hierarchical structures. This type of structure, however, lacks the capacity to support automated reasoning about the entities that are represented—beyond perhaps reasoning about how those entities are classified—given the lack of formal definitions and distinct relationships among the entities. While “richer” specifications allow for increased use of semantic reasoning for important tasks, such as abstraction, consistency checking, and automatic classification, the use of such approaches also requires more effort—and more skill—in defining the semantics using appropriate constraints and axioms in the ontology.

Classification Systems for Mental Health Problems

In this section we explore three influential classification systems that are frequently used in the behavioral sciences: (1) the Diagnostic and Statistical Manual of Mental Disorders (DSM), (2) the Research Domain Criteria (RDoC), and (3) the hierarchical taxonomy of psychopathology (HiTOP). While each of these is an ontological system, none has the level of formal specification that would make it compatible with representation in a logic that supports computer-based reasoning. A review of these systems highlights several of the challenges inherent in moving toward the use of more standard and more computational ontologies in the behavioral sciences.

A Categorical Classification System: The Diagnostic and Statistical Manual of Mental Disorders

The DSM has been the primary tool for classifying mental disorders since the early 1950s.¹³ Its purpose is to support diagnosis and treatment of, and research about, mental disorders. The development and updating of the DSM are based on expert review carried out by workgroups, and

___________________

¹³ See https://www.psychiatry.org/psychiatrists/practice/dsm/history-of-the-dsm for an account of the history of the classification of mental disorders in the United States (American Psychiatric Association, n.d.).

a process for establishing consensus led by the American Psychiatric Association’s board of trustees. Over the years, this approach to establishing diagnostic criteria for mental disorders has generated controversy about many topics, including evolving understanding of human sexuality and the definitions of many disorders.

The DSM is a classification system for the diagnosis of psychiatric disorders: it can be used to identify whether or not individuals meet specified criteria for a disorder. That is, it treats disorders as categorically (qualitatively) distinct from one another. The DSM provides a disease code for each diagnosis in order to group sets of related disorders or manifestations of a disorder. For example, “anxiety disorders” are defined as sharing “features of excessive fear and anxiety and related behavioral disturbances” (American Psychiatric Association, 2013, p. 189). The DSM distinguishes among subtypes of anxiety in terms of the types of situations and objects that cause fear, anxiety, and avoidance behavior: for example, social phobias are distinguished from general anxiety disorder based on the idea that in the former, social situations cause anxiety, while in the latter, nonspecified stimuli cause anxiety.

All the diagnoses in the anxiety disorder class differ from diagnoses in the other classes on the basis of distinguishing characteristics of each class. For instance, the anxiety disorders class is distinguished from the class of depressive disorders based on the fact that the central defining feature of depressive disorders is the “presence of sad, empty or irritable mood, accompanied by somatic and cognitive changes that significantly affect the individual’s capacity to function” (American Psychiatric Association, 2013, p. 155). These distinguishing features of classes or disorders, or disorders themselves, are viewed as pathognomonic markers of the disorder or class, such that their presence alone would render a diagnosis definitive.

The most recent version of the DSM, the DSM-5 (American Psychiatric Association, 2013), contains 22 classes or categories of disorders and other conditions that may be a focus of clinical attention: neurodevelopmental; schizophrenia spectrum; bipolar; anxiety; obsessive-compulsive-trauma- and stressor-related; dissociative; somatic; feeding/eating; elimination; sleep/wake; sexual; gender dysphoria; disruptive/impulse/conduct; substance/additive; neurocognitive; personality; paraphilic; other; and medication induced disorders. Under each of these major classes of disorders is a set of hierarchically organized disorders, each with its own code. For instance, depressive disorders include: disruptive mood dysregulation disorder, major depressive disorder, persistent depressive disorder, premenstrual dysphoric disorder, substance/medication-induced depressive disorder, depressive disorder due to another medical condition, other specified depressive disorder, and unspecified disorder.

The DSM is an ontological system because it defines the entities it enumerates, although it is not as rigorously formalized as the BCIO. For a

classification system to fall on the right side of the continuum, it would have clear, named relationships among the entities in the ontology and the definitions for the entities would be expressed in a computable logic.¹⁴ The DSM is placed to the left of the BCIO in Figure 3-1 (above) because it represents classification of mental disorders using a common language and standard criteria.

For decades, the DSM has been the primary system for the classification of mental health problems. DSM descriptions are taught in undergraduate courses in abnormal psychology, and they remain a core component of graduate training in clinical psychology, social work, and psychiatry. In addition, the DSM drives decisions about which medications or treatments patients receive and about how mental health providers are paid for their services. The DSM is the de facto structure for managing mental health conditions and arranging reimbursement for mental health care in the United States, and it is worth noting that those purposes—rather than the needs of scientists—drove the development of its categories.

Despite the DSM’s usefulness for clinical and other purposes, many of its categories do not have a firm scientific base. For example, the diagnostic categories in DSM do not align with findings from clinical neuroscience or genetics (Insel et al., 2010). Furthermore, DSM classification is based on behavioral phenotypic manifestations of underlying disorder (i.e., symptoms), not on pathophysiology (the DSM-5, released in 2013, includes 265 categories based on signs and symptoms). People who meet the criteria for many DSM disorders may in fact share very few common symptoms. For example, the DSM criteria for diagnosis of major depressive disorder specifies that a patient must have five of nine specified symptoms. Thus, two people both diagnosed with the disorder may have only one symptom in common.

Another issue is that many patients meet criteria for more than one mental disorder, leading to concerns that disorders are inappropriately studied in isolation from each other. Put differently, researchers have developed their thinking about disorders, no longer regarding them as distinct categories but rather as sharing underlying dimensions that explain their covariation. The categorical emphasis still present in the DSM has led to problems, such as researchers or clinicians choosing not to enroll patients with more complex diagnoses in research trials because they wish to focus on a single diagnosis rather than the relationships among diagnoses.

___________________

¹⁴ In contrast with the DSM, the Ontology of General Medical Science organizes medical diseases using such terms as bodily component, bodily feature, bodily process, bodily quantity, and the like (Ceusters and Smith, 2010). For classifications of mental disorders to follow this approach, it would be necessary to link the observed signs and symptoms to bodily (brain) processes.

For these reasons, the DSM’s dominance in the field has likely hindered scientific progress. However, the DSM system is still considered the authoritative system by which mental disorders are classified in the United States and many other countries.

New methodologies in genetics, neuroimaging, and other behavioral sciences have challenged the validity of the DSM categories. The work groups in charge of the DSM-5 revisions were aware of developments in biology and neuroscience but found that the new data were not sufficiently sensitive and specific to be useful in the DSM (Cuthbert, 2014). Observers of the DSM, however, have become increasingly concerned that, although useful in a practical sense, the DSM classifications do not reflect a supportable model of mental disorders and their causes. An additional concern is with the process by which the DSM was developed, a consensus process based on the judgments of a group of primarily White, male, upper- and middle-class clinicians.

Some of the limitations of the DSM are exemplified in research on phenotypes for mental disorders. The Bipolar and Schizophrenia Network on Intermediate Phenotypes consortium collected data on cohorts of patients diagnosed with schizophrenia, schizoaffective disorder, and bipolar disorder with psychosis (Clementz et al., 2016). The research examined whether and how patients could be grouped based on their cognitive control and sensorimotor activity, regardless of diagnostic category. The data showed that the patients could be divided into three biotypes that crossed diagnostic boundaries. A similar approach was taken with patients diagnosed with a major depressive disorder (Drysdale et al., 2017). Researchers used a machine learning approach to look at biomarkers for these patients in order to determine whether there was one coherent group or whether there were subtypes within the group. The data revealed four individual biotypes: patients in each of the biotype categories responded differently to treatment. It is this kind of research that has called into question the categorical approach to diagnosing mental disorder and suggested that symptoms may be best understood (and treated) if the underlying dimensions that cause them to covary can be identified. This approach is called the dimensional approach to understanding and defining problems in mental health.

A Dimensional Classification System: Research Domain Criteria

Tying observed or self-reported signs and symptoms at the behavioral phenotypic level more closely to biological processes was the objective when the National Institute of Mental Health (NIMH) sought to develop a new system for categorizing mental disorders. Its 2008 strategic plan called for the agency to “develop, for research purposes, new ways of

classifying mental disorders based on dimensions of observable behavior and neurobiological measures” (NIMH, 2008, p. 9). The resulting system, the Research Domain Criteria (RDoC), is not intended as an alternative or replacement for the DSM, but rather is a research framework that integrates many levels of information about human functioning (NIMH, n.d.-a).

The impetus for the development of RDoC was a growing awareness of the limitations of the DSM, discussed above (Cuthbert, 2014; NIMH, n.d.-a; Insel et al., 2010). The goal of RDoC was not to create a new diagnostic system that would compete with the DSM, but rather to devise alternative criteria to be used for purposes such as peer review for grant applications. This new research-focused approach would support investigators and reviewers in thinking outside the boundaries of the DSM. For example, an investigator could research underlying dimensions (e.g., the symptom anhedonia, which is a feature of more than one category) that cut across multiple DSM categories, rather than being constrained by the need to design research that aligned within DSM categories. From the perspective of NIMH, RDoC is intended to facilitate research that investigates fundamental dimensions, grounded in biology, that span multiple disorders (e.g., response to threat, attention, social processes); examines the full range of variation, from normal to abnormal; integrates genetic, neurobiological, behavioral, environmental, and self-report measures; and develops reliable and valid measures of components for basic and clinical studies (NIMH, n.d.-b). We note, however, that the assumption that fundamental dimensions are those grounded in biology implies that there is a single causal mechanism for the phenomena being accounted for. This assumption is too limiting because behavioral phenomena may be caused by a number of mechanisms (Edelman and Gally, 2001).

RDoC includes six domains (dimensions) and seven levels of analysis: see Figure 3-2. RDoC provides a framework for understanding the nature of mental health and illness in terms of varying degrees of dysfunction in general psychological/biological systems, but it does not represent the same degree of formal specification as other examples with strong semantics.

Although RDoC is not intended for use in patient care, there is some evidence that its matrix may have value in capturing clinically valuable information (McCoy et al., 2015). Moreover, RDoC’s focus on systematizing normal human mental functioning and on guiding attempts to associate underlying psychological constructs of cognition and emotion with specific neural circuitry is considered an advance. However, because of its apparent grounding in psychophysiology, it is tempting to reify the RDoC dimensions as if they identify phenomena that exist in nature independent of human observation or interpretation. But the structure and elements of the RDoC matrix, like those of the DSM, were identified through a process in which

experts identified areas on which they could reach consensus; such a process is inherently arbitrary to some extent (Ross and Margolis, 2019).

While both the DSM and RDoC systems’ consensus frameworks are based on current and ongoing reviews of the empirical literature in mental health and psychophysiology, they continue to reflect concepts that originated decades ago. Those who work on the DSM rarely define new constructs, and the RDoC did not introduce any. Thus, the organizational principles of RDoC have remained largely untested, and the reproducibility of its circuit-function links is unknown—indeed there appears to be no one-to-one mapping between a dimension and a single circuit (Beam et al., 2021).

Researchers have used computational approaches to ontologies to explore how well the DSM and RDoC explain structure–function relationships and to contrast these frameworks with a data-driven one (Beam et al., 2021). The researchers applied natural language processing and machine learning techniques to the results of human neuroimaging collected over 25 years in order to redefine mental constructs in relation to brain activation data. They found that across multiple levels of domain specificity, the structure–function links (the idea that structure determines function) were better replicated in the mental constructs derived through data-driven analysis than those mapped from both the DSM and RDoC.

While several limitations in this approach were acknowledged by the authors, the use of computational approaches to neuroscience ontology demonstrated in their approach suggests the value of performing bottom-up ontological analyses. In contrast, reliance on expert-determined systems like the DSM and RDoC has the potential to introduce biases that may constrain research and limit scientific progress (Beam et al., 2021).

A Quantitative Approach: The Hierarchical Taxonomy of Psychopathology

A third approach to the classification of mental disorders uses factor analysis, which we discussed in Chapter 2 (also see Chapter 5).¹⁵ Rather than relying on the consensus of expert committees (as with the DSM), researchers using this quantitative approach seek consensus from studies of the “natural organization of mental health” (Kotov et al., 2021; p. 86). Rooted in a nearly century-long tradition of using factor analytic methods to identify empirical constellations of signs and symptoms (e.g., Achenbach, 1966; Eysenck, 1944), this approach is most recently exemplified in the work of the hierarchical taxonomy of psychopathology (HiTOP) consortium (Kotov et al., 2021; Lahey et al., 2017).

Three assumptions guide the quantitative nosology approach exemplified by HiTOP.¹⁶ First, mental disorders are defined as dimensions rather than categorical entities. Second, the natural organization of psychopathology can be discerned in the co-occurrence of its features: that is, using factor analytic methods, the underlying dimensions that organize psychopathology can be discovered by evaluating how signs and symptoms of psychopathology covary. Third, psychopathology can be organized hierarchically from narrow to broad dimensions so that specific psychopathology dimensions aggregate into more general factors.

The HiTOP model is premised on the idea that psychiatric signs and symptoms are captured by underlying dimensions of internalizing disorder, externalizing disorder, and thought disorder, rather than categorical diagnoses as described in the DSM (and other similar models). HiTOP aims to facilitate the search for causes and mechanisms of psychopathology by identifying coherent constructs that can be measured reliably (because the strong factorial measurement model has reliably established them) and are thus more usable and informative than traditional diagnoses. Thus, in the same way that RDoC supports use of a psychophysiological dimension (e.g., negative valence system to understand depression rather than a heterogeneous depression diagnosis), HiTOP also supports the use of a dimension, such as the dimension of negative symptoms rather than highly heterogeneous schizophrenia diagnosis (Kotov et al. 2021).

Hierarchical relationships in HiTOP allow for improved classification of psychopathology dimensions to facilitate research and clinical practice. However, HiTOP is mainly an empirical organization of psychopathology to support classification of psychiatric disorders. HiTOP, like the DSM and the RDoC, adheres to a hierarchical structure, which is an indicator that a

___________________

¹⁵ See https://dictionary.apa.org/factor-analysis

¹⁶ Nosology is the systematic classification of diseases.

classification system has specified relationships among the entities it defines. However, it should be noted that factor analysis, per se, does not result in a formal specification. While it describes the relationships between variables (through its covariance matrix), the empirical approach bypasses any kind of shared conceptualization. That is, use of factor analysis does not require shared conceptualization as an input, and the product of the factor analysis need not be anything that has a straightforward interpretation in terms of familar concepts: such interpretation involves the additional step of labeling the factors.

CONCLUSIONS

Both the DSM and RDoc are based in shared conceptualizations of phenomena, although they both lack formal specification of their content. In contrast, HiTOP is not based on any conceptualization. Rather, it offers a clear specification, not one with any formal semantics, but one derived from standard statistical inference. These three examples are each influential and demonstrate some points that arise in the development of ontological structures for the behavioral sciences.

First, in choosing a representation system—and the degree of formal specification needed—developers are choosing a language with the ex-pressivity needed for the purpose the ontology is designed to serve. For example, OWL is a computational logic that is designed to represent rich and complex domain knowledge and that allows automated reasoning. However, it would require expertise and efforts from the developers to formally define the ontologies. A simpler taxonomy does not impose the same demands. Neither is necessarily better than the other; each is more suitable in certain circumstances.

Ontologies that incorporate relatively strong semantics support use of the ontology to create a knowledge base that goes beyond the simple enumeration of classes and taxonomic relationships.¹⁷ The distinction between the two frameworks is important for understanding the essence of an ontology: a way of characterizing entities in the discipline, and the relationships among them. Thus, a structure that fails to formally specify terms and taxonomic relationships cannot be described as an ontological system, and one that goes distinctly beyond that function is something more complicated than an ontology.

Finally, ontologies are important tools for identifying gaps in the way researchers talk about a discipline, for developing formal definitions when they are lacking, and for testing those definitions to see whether they are useful for their intended purpose. Many of the entities about which behavioral scientists

___________________

¹⁷ An ontology can incorporate logical relationships, such as causality, when they are part of the specification of terms. For example, the definition of “infectious disease” incorporates the causal relationship because the disease is defined as one caused by infection.

speak do not currently have formal definitions, and this is the problem that leads to the sorts of challenges discussed in Chapter 2. The development of an ontology is an opportunity to impose intellectual rigor on a research domain.

Based on our review of what ontologies are and how they function in the behavioral sciences, we offer three conclusions.

REFERENCES

Beam, E., Potts, C., Poldrack, R.A., and Etkin, A. (2021). A data-driven framework for mapping domains of human neurobiology. Nature News. https://www.nature.com/articles/s41593-021-00948-9

Berners-Lee, T., Hendler, J., and Lassila, O. (2001). The Semantic Web. Scientific American, 284, 29–37. https://doi.org/10.1038/scientificamerican0501-34

Bilder, R.M., Sabb, F.W., Parker, D.S., Kalar, D., Chu, W.W., Fox, J., Freimer, N.B., and Poldrack, R.A. (2009). Cognitive ontologies for neuropsychiatric phenomics research. Cognitive Neuropsychiatry, 14(4-5), 419–450. https://doi.org/10.1080/13546800902787180

Blanch, A., García, R., Planes, J., Gil, R., Balada, F., Blanco, E., and Aluja, A. (2017). Ontologies about human behavior: A review of knowledge modeling systems. European Psychologist, 22(3), 180–197. https://doi.org/10.1027/1016-9040/a000295

Ceusters, W. (2012). An information artifact ontology perspective on data collections and associated representational artifacts. Studies in Health Technology and Informatics, 180, 68–72.

Ceusters, W., and Smith, B. (2010). Foundations for a realist ontology of mental disease. Journal of Biomedical Semantics, 1(1), 10. https://doi.org/10.1186/2041-1480-1-10

Clementz, B.A., Sweeney, J.A., Hamm, J.P., Ivleva, E.I., Ethridge, L.E., Pearlson, G.D., Keshavan, M.S., and Tamminga, C.A. (2016). Identification of distinct psychosis biotypes using brain-based biomarkers. The American Journal of Psychiatry, 173(4), 373–384. https://doi.org/10.1176/appi.ajp.2015.14091200

Cuthbert, B.N. (2014). The RDoC framework: Facilitating transition from ICD/DSM to dimensional approaches that integrate neuroscience and psychopathology. World Psychiatry, 13(1), 28–35. https://doi.org/10.1002/wps.20087

Davis, R., Campbell, R., Hildon, Z., Hobbs, L., and Michie, S. (2015). Theories of behaviour and behaviour change across the social and behavioural sciences: A scoping review. Health Psychology Review, 9(3), 323–344. https://doi.org/10.1080/17437199.2014.941722

Drysdale, A.T., Grosenick, L., Downar, J., Dunlop, K., Mansouri, F., Meng, Y., Fetcho, R.N., Zebley, B., Oathes, D.J., Etkin, A., Schatzberg, A.F., Sudheimer, K., Keller, J., Mayberg, H.S., Gunning, F.M., Alexopoulos, G.S., Fox, M.D., Pascual-Leone, A., Voss, H.U., Casey, B.J., Dubin, M.J. and Liston, C. (2017). Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nature Medicine, 23(1), 28–38. https://doi.org/10.1038/nm.4246

Edelman, G.M., and Gally, J.A. (2001). Degeneracy and complexity in biological systems. Proceedings of the National Academy of Sciences of the United States of America, 98(24), 13763–13768. https://doi.org/10.1073/pnas.231499798

Elliott, J.H., Turner, T., Clavisi, O., Thomas, J., Higgins, J. P., Mavergames, C., and Gruen, R.L. (2014). Living systematic reviews: An emerging opportunity to narrow the evidence-practice gap. PLoS Medicine, 11(2), e1001603. https://doi.org/10.1371/journal.pmed.1001603

Eysenck, H.J. (1944). Types of personality: A factorial study of seven hundred neurotics. Journal of Mental Science, 90, 851–861. https://doi.org/10.1192/bjp.90.381.851

Goldberg, L.R. (1981). Language and individual differences: The search for universals in personality lexicons. In L. Wheeler (Ed.), Review of Personality and Social Psychology, 2, 141–165. https://projects.ori.org/lrg/PDFs_papers/universals.lexicon.81.pdf.

Grenon, P., Smith, B., and Goldberg, L. (2004). Biodynamic ontology: Applying BFO in the biomedical domain. Studies in Health Technology and Informatics, 102, 20–38.

Gruber, T.R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–220. https://doi.org/10.1006/knac.1993.1008

Gruber, T.R. (1995). Toward principles for the design of ontologies used for knowledge sharing? International Journal of Human-Computer Studies, 43(5-6), 907–928. https://doi.org/10.1006/ijhc.1995.1081

Gruber, T. (2016) Ontology. In L. Liu L. and M. Özsu (Eds.), Encyclopedia of Database Systems. Springer. https://doi.org/10.1007/978-1-4899-7993-3_1318-2

Hastings, J. (2021). Formal Representations of Ontologies for Automation of Analyses. Commissioned paper prepared for the Committee on Accelerating Behavioral Science Through Ontology Development and Use. Available: https://nap.nationalacademies.org/resource/26464/Hastings-comissioned-paper.pdf

Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D.S., Quinn, K., Sanislow, C., and Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. The American Journal of Psychiatry, 167(7), 748–751. https://doi.org/10.1176/appi.ajp.2010.09091379

Kotov, R. (2021). The Hierarchical Taxonomy of Psychopathology (HiTOP): A quantitative osology based on consensus of evidence. Annual Review of Clinical Psychology, 17, 83–108. https://doi.org/10.1146/annurev-clinpsy-081219-093304

Lahey, B.B., Krueger, R.F., Rathouz, P.J., Waldman, I.D., and Zald, D.H. (2017). A hierarchical causal taxonomy of psychopathology across the life span. Psychological Bulletin, 143(2), 142–186. https://doi.org/10.1037/bul0000069

Larsen, K.R., Michie, S., Hekler, E.B., Gibson, B., Spruijt-Metz, D., Ahern, D., Cole-Lewis, H., Bartlett Ellis, R.J., Hesse, B., Moser, R.P., and Yi, J. (2017). Behavior change interventions: The potential of ontologies for advancing science and practice. Journal of Behavioral Medicine, 40(1), 6–22. https://doi.org/10.1007/s10865-016-9768-0

Lassila, O., and McGuinnes, D. (2001). The role of frame-based representation on the Semantic Web. Linköping Electronic Articles in Computer and Information Science, Vol. 6 (2001), No. 005, Linköping University. Available: https://www.ida.liu.se/ext/epa/ej/etai/2001/018/01018-etaibody.pdf

Madrigal, A.C. (2018, December 22). How Netflix reverse-engineered Hollywood. The Atlantic. https://www.theatlantic.com/technology/archive/2014/01/how-netflix-reverse-engineered-hollywood/282679/

McCoy, T.H., Castro, V.M., Rosenfield, H.R., Cagan, A., Kohane, I.S., and Perlis, R.H. (2015). A clinical perspective on the relevance of research domain criteria in electronic health records. The American Journal of Psychiatry, 172(4), 316–320. https://doi.org/10.1176/appi.ajp.2014.14091177

Merrill, G.H. (2010). Ontological realism: Methodology or misdirection? Applied Ontology, 5(2), 79–108. https://doi.org/10.3233/ao-2010-0076

Michie, S., Thomas, J., Johnston, M., Aonghusa, P.M., Shawe-Taylor, J., Kelly, M.P., Deleris, L.A., Finnerty, A.N., Marques, M.M., Norris, E., O’Mara-Eves, A., and West, R. (2017). The Human Behaviour-Change Project: Harnessing the power of artificial intelligence and machine learning for evidence synthesis and interpretation. Implementation Science, 12(1), 121. https://doi.org/10.1186/s13012-017-0641-5

Michie, S., West, R., Finnerty, A.N., Norris, E., Wright, A.J., Marques, M.M., Johnston, M., Kelly, M.P., Thomas, J., and Hastings, J. (2021). Representation of behaviour change interventions and their evaluation: Development of the upper level of the Behaviour Change Intervention Ontology. Wellcome Open Research, 5, 123. https://doi.org/10.12688/wellcomeopenres.15902.2

National Institute of Mental Health (NIMH). (n.d.-a) About RDoC. https://www.nimh.nih.gov/research/research-funded-by-nimh/rdoc/about-rdoc

______. (n.d.-b) Developing an RDoC Study. https://www.nimh.nih.gov/research/research-funded-by-nimh/rdoc/developing-an-rdoc-study

______. (2008). Strategic Plan. https://www.hsdl.org/?view&did=755067

Neches, R., Fikes, R.E., Finin, T., Gruber, T., Patil, R., Senator, T., and Swartout, W.R. (1991). Enabling technology for knowledge sharing. AI Magazine, 12(3), 36. https://doi.org/10.1609/aimag.v12i3.902

Norris, E., Marques, M.M., Finnerty, A.N., Wright, A.J., West, R., Hastings, J., Williams, P., Carey, R.N., Kelly, M.P., Johnston, M., and Michie, S. (2020). Development of an Intervention Setting Ontology for behaviour change: Specifying where interventions take place. Wellcome Open Research, 5, 124. https://doi.org/10.12688/wellcomeopenres.15904.1

Poldrack, R., and Yarkoni, T. (2016). From brain maps to cognitive ontologies: Informatics and the search for mental structure. Annual Review of Psychology, 67, 587–612. https://doi.org/10.1146/annurev-psych-122414-033729

Regoczei, S., and Plantinga, E.P. (1987). Creating the domain of discourse: Ontology and inventory. International Journal of Man-Machine Studies, 27(3), 235–250. https://doi.org/10.1016/S0020-7373(87)80054-8

Ross, C.A., and Margolis, R.L. (2019). Research domain criteria: Strengths, weaknesses, and potential alternatives for future psychiatric research. Molecular Neuropsychiatry, 5(4), 218–236. https://doi.org/10.1159/000501797, 218-235

Smith, B., Ashburner, M., Rosse, C., Bard, J., Bug, W., Ceusters, W., Goldberg, L.J., Eilbeck, K., Ireland, A., Mungall, C.J., OBI Consortium, Leontis, N., Rocca-Serra, P., Ruttenberg, A., Sansone, S.A., Scheuermann, R.H., Shah, N., Whetzel, P.L., and Lewis, S. (2007). The OBO Foundry: Coordinated evolution of ontologies to support biomedical data integration. Nature Biotechnology, 25(11), 1251–1255. https://doi.org/10.1038/nbt1346

Smith, B., and Ceusters, W. (2010). Ontological realism: A methodology for coordinated evolution of scientific ontologies. Applied Ontology, 5(3-4), 139–188. https://doi.org/10.3233/ao-2010-0079

Smith, B., and Grenon, P. (2005). The cornucopia of formal-ontological relations. Dialectica, 58, 279–296. https://doi.org/10.1111/J.1746-8361.2004.TB00305.X

Staab, S., and Studer, R. (Eds.). (2016). Handbook on Ontologies. New York: Springer.

Stevens, R., Lord, P., Malone, J., and Matentzoglu, N. (2018). Measuring expert performance at manually classifying domain entities under Upper Ontology Classes. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3248494

Studer, R., Benjamins, V.R., and Fensel, D. (1998). Knowledge engineering: Principles and methods. Data & Knowledge Engineering, 25(1-2), 161–197. https://doi.org/10.1016/S0169-023X(97)00056-6

West, R., Michie, S., Shawe-Taylor, J., Thomas, J., Finnerty, A.N., Johnston, M., Aonghusa, P.M., O’Mara-Eves, A. Stokes, G., Norris, E., Marques, M., Kelly, M.P., Ganguly, D., Moore, C., Hou, Y., Bonin, F., Wright, A., Veally, C., Zink, S., and Schenk, P. (2020). Human Behaviour-Change Project. https://doi.org/10.17605/OSF.IO/UXWDB

This page intentionally left blank.