- Ontological realism aims at the development of high quality ontologies that faithfully represent what is general in reality and to use these ontologies to render heterogeneous data collections comparable. To achieve this second goal for clinical research datasets presupposes not merely (1) that the requisite ontologies already exist, but also (2) that the datasets in question are faithful to reality in the dual sense that (a) they denote only particulars and relationships between particulars that do in fact exist and (b) they do this in terms of the types and type-level relationships described in these ontologies. While much attention has been devoted to (1), work on (2), which is the topic of this paper, is comparatively rare. Using Referent Tracking as basis, we describe a technical data wrangling strategy which consists in creating for each dataset a template that, when applied to each particular record in the dataset, leads to the generation of a collection of Referent Tracking Tuples (RTT) built out of unique identifiers for the entities described by means of the data items in the record. The proposed strategy is based on (i) the distinction between data and what data are about, and (ii) the explicit descriptions of portions of reality which RTTs provide and which range not only over the particulars described by data items in a dataset, but also over these data items themselves. This last feature allows us to describe particulars that are only implicitly referred to by the dataset; to provide information about correspondences between data items in a dataset; and to assert which data items are unjustifiably or redundantly present in or absent from the dataset. The approach has been tested on a dataset collected from patients seeking treatment for orofacial pain at two German universities and made available for the NIDCR-funded OPMQoL project.
One goal of ontology-based research is the integration of information residing in heterogeneous data collections in the hope that by running queries over the resultant combined data collections we will be able to answer questions that would otherwise remain unanswered [1]. Such integration can be achieved through different paradigms, including: mediation [2], federation [3], data warehousing [4], and, most recently, the Ontology-Based Data Access (OBDA) paradigm [5], which is distinguished by the fact that it keeps the data sources and conceptual layer of an information system separate and independent.
To be effective, all such paradigms require ontologybased mappings ranging not only over the database schemas but also over the data types by means of which the data are stored [6]. Research in OBDA revealed that successful information integration requires much more detail than is standardly provided: it requires also suitable mechanisms for mapping individual data values – rather than merely data fields – to corresponding instances of ontology classes – for example to patients in a clinical study. This in turn requires the specification of how identifiers for such instances can be generated from such data values in order to enable creation of an ABox suitable for answering queries relating to such instances [7]. Such specification, we believe, may well be a critical issue in the context of clinical research datasets, where (as we shall discover below) data values do not always denote what is suggested by the variable or fieldname under which they appear.
Suppose, for example, that in the record of some patient the variable phenotypic gender is associated with a value of either ‘0’ or ‘1’ – meaning ‘male’ or ‘female,’ respectively. It is then safe to create an ABox statement to the effect that this patient’s phenotypic gender is an instance of the corresponding ontology class. If no data value is found, however, then it should not be assumed that the patient in question does not have a phenotypic gender. If, on the other hand a value of ‘2’ – documented as meaning ‘unknown’ – is found, then this should not lead to an ABox assertion to the effect that the given patient’s phenotypic gender is an instance of a special kind which is neither male nor female. The value ‘unknown’ provides information not about the patient, but rather about the data we have about the patient.
The problem we face in creating data value to ontology mappings from clinical research data repositories is that the information needed for such mappings is not explicitly represented in the datasets. Rather, it is scattered through various data dictionaries and instruction manuals (relating for example on how to extract and process data from responses to standardized questionnaires).
The explicit representation that is pursued by the
Referent Tracking (RT) methodology is based on
Ontological Realism as described in [8], and on the thesis
that explicit representation can best be achieved by
generating unique identifiers to all instances of ontology
classes which are described – whether explicitly and
implicitly – in our data. In [9] we described an algorithm to
achieve explicit representation of this sort from highly
structured electronic health record (EHR) data. The research
questions we address here are:
(
project Ontology for Pain-related Mental Health and Quality of Life (OPMQoL) which involves the integration of five datasets which – although collected independently – cover similar sorts of information about patients who experienced one or other form of orofacial pain [10]. All datasets are made available as spreadsheet tables (from here on referred to as ‘source tables’). Each row in the body of each such table is a collection of data items obtained from a single patient; each column is a collection of data items resulting from some specific type of observation. If a header row is present, its cells indicate what sorts of observations are reported on in the respective columns.
The de-identified dataset used for the work described here – from here on referred to as the ‘study set’ – was collected from 390 patients seeking treatment for orofacial pain [11]. Inclusion criteria were that patients had at least one diagnosis according to the Research Diagnostic Criteria for Temporomandibular Disorders (RDC/TMD) [12]. The study set comes with a variable (n=161) codebook and a technical report explaining certain dependencies and implicit assumptions [13].
be expressed as a collection of Referent Tracking Tuples (RTT) [14]. An RTT is an assertion about a particular, i.e. an entity in reality that exists in space and time [15]. Each
used for instance-level relationships in the definitions of the
Relation Ontology [16]. Ignoring here certain housekeeping
parameters we can assert that RTT assertions about
continuants (entities such as patients, hospitals, teeth, jaws
which endure through time, as contrasted with occurrents or
processes), are of the form ‘x p-rel y t-rel t’, where:
‘x’ is the (ideally) singular and globally unique instance
identifier (IUI) denoting the particular described,
‘y’ is either: (
RTT assertions that do not mention a continuant have the form ‘x p-rel y,’ where ‘x ’, ‘p-rel’ and ‘y’ are otherwise treated in the way described above.
RT aims to do away with the ambiguity in assertions such as ‘John has a benign duodenal polyp’. This assertion tells us that there exists some instance of a given type, but not which one in particular. This ambiguity is preserved in John’s EHR, where diagnostic codes drawn from some terminology or ontology are used to assert existence in John at some time t1 of polyps of a given type. The consequence is that, when a later assertion is added to John’s EHR to the effect that he has a malignant duodenal polyp, the data provides no basis for inferences concerning whether it is the very same polyp as the one referred to at t1 that has turned malignant or some other polyp appearing at some later time t 2 [14]. This ambiguity disappears when we represent the first-described situation using the following RTTs: #1 part-of #2 at t1 #1 instance-of benign duodenal polyp at t1 #1 instance-of malignant duodenal polyp at t1 where ‘#1’ denotes the polyp and ‘#2’ John. The alternative situation, would be represented by using distinct IUIs for each polyp as follows, where ‘#3’ denotes a second polyp:
IM patient_study_record id LV patient_identifier id IM patient sex CV gender sex CV male sex CV female sex UA sex q3 CV no_pain_in_ lower_face q3 CV pain_in_ lower_face q3 IM in_the_past_month q3 IM lower_face q3 IM time_of_q3_concretization q3 RP an_8_gcps_1 q3 UP an_8_gcps_1 q3 UA an_8_gcps_1 q3 JA an_8_gcps_1
BLANK BLANK 0 0 1 10 BLANK BLANK BLANK BLANK A further goal of RT is to make explicit all the implicit assumptions that need to be taken into account to interpret given data correctly. Some of these assumptions result from the use of broken information models or from practices such as registering ICD-9-CM code 659.7 – ‘Abnormality in fetal heart rate or rhythm’ – in the diagnosis field of the mother’s EHR. The RT method is most effective when its principles are applied at the time of data collection and registration, though as shown in [17] post-hoc translations are also possible.
The work reported here involved the following steps:
(
IV.
Research questions (
CV IM UA JA RP UP
Template Av. (SD) information type (IT), the detailed semantics of which is described in section V. Common to all information types is that part of the template that appears to the left of the dashed vertical line in Table 1. This specifies the conditions which must be satisfied if RTTs are to be generated on the basis of the information provided to the right of this line.
Table 2 answers research questions (
should be maximally explicit and maximally self-explanatory. By ‘maximally explicit’, we mean that each such repository should contain explicit reference to any and all the entities, including their interrelationships, that must exist for an assertion encoded in the repository to be a faithful representation of the corresponding part of reality. By ‘maximally self-explanatory’ we mean that the data in the repository should be presented in such a way that a researcher seeking to query the repository does not need to concern himself with any idiosyncrasies of and between datasets, or codes or formats, that were combined or used to build the repository. A strategy to achieve this is to submit to such a repository only individual datasets which are themselves maximally explicit and self-explanatory.
Our approach is based on the – to us – obvious distinction between data and what data are about. It then takes advantage of the fact that RTTs can be used to describe in explicit fashion not merely the portions of reality described by data items in a dataset, but also these data items themselves. This allows us to describe explicitly even those particulars that are only implicitly referred to in a dataset by generating suitable unique identifiers. It also allows us to provide information about correspondences (such as co-reference) between data items in a dataset, and also to assert which data items are redundant, or unjustifiably absent, and so forth.
The study set contains some explicit data items which are about particulars on the side of the patient such as gender, facial pains experienced, clicking noises heard when opening their mouths, and so forth. Referent Tracking requires each of these particulars to be assigned an IUI; Ontological Realism tells us that each one of them is instance of at least one universal. What universals these particulars are instances of is typically only very indirectly represented in the study set.
The strategy for translating explicit data items into RTTs is covered by the Literal Value (LV) and Coded Value (CV) records in the template (Table 1). Template lines of either type have under ‘REF’ the label obtained or constructed from the relevant data dictionary or other supporting information associated with the code value. The template shows, for example, that if, for a patient in the study set, the value for the variable ‘sex’ is ‘0’ (L5), then the gender of this patient is described as ‘male.’ This can be translated in RT terms into a assertion that the given patient’s gender is an instance of the universal male gender (or, in case gender does not qualify as a universal [18], that it is a member of the defined class ‘male gender’ – we will ignore this distinction in the remainder of this paper).
The IUIs assigned through application of our method are in reality very large numbers generated by an RT system to ensure the needed high probability of uniqueness. For the sake of readability, however, we provide simple abbreviations to stand in for these IUIs. We also leave out full specification of time-related information (which would be needed, for example, to deal with cases where a patient’s gender changes from one time to the next), and certain housekeeping details required by syntactically and semantically correct RTTs [15]).
To see how IUI assignment works, now, we will suppose
that, while processing the study set on the basis of the
template illustrated in Table 1, the IUI #pat-1 is assigned to
the first patient described and that #patg-1 is assigned to his
gender. Then the following collection of assertions would be
generated as part of a faithful RT-like representation of the
corresponding portion of reality (POR) on the basis of lines L3
and L5 of the template:
#pat-1 instance-of PATIENT at t
#patg-1 instance-of MALE-GENDER at t
#patg-1 inheres-in #pat-1 at t
(
L7, for instance, brings it about that when, for patient #pat1 in the study set, no value for the variable ‘sex’ is provided – expressed by the appearance of ‘BLANK’ in the template under both ‘Min’ and ‘Max’ – an RTT is generated that declares the data item #patgL-1 to be an instance of an underspecified ICE. This assertion does not mean that the data item itself is absent; rather it means that certain information is missing.
An absence or presence of a value for some variable may be justified or unjustified depending on the value of some other variable. The last four lines in Table 1, for example, describe dependencies between the variables ‘q3’ (for which the possible values ‘1’ and ‘0’ mean, respectively, current presence or absence of pain) and ‘an_8_gcps_1’, the latter containing answers to the question ‘How would you rate your facial pain on a 0 to 10 scale at the present time, that is right now, where 0 is “no pain” and 10 is “pain as bad as could be”?’ L13 states that when the values for both ‘q3’ and ‘an_8_gcps_1’ are ‘0’, then the two ICEs of which the coding for the answers are concretizations enjoy a corresponds-to relation to the same portion of reality.
L16 asserts that, if a record in the dataset has a ‘0’ value for the variable q3, and if there is no value for the variable ‘an_8_gcps_1’, then the absence of a value for ‘an_8_gcps_1’ is justified. This is then documented by means of an RTT to the effect that the corresponding ICE is justifiably blank (as concretized by, for instance, an empty cell in that part of the spreadsheet). As a last example, L14 asserts that if the value given for ‘an_8_gcps_1’ is between 1 and 10 while the value for q3 is 0, then the value for the former is unjustifiably present (the corresponding ICE must thus be classified as disinformation – as dictated by the coding guidelines for the corresponding pair of questions).
correspondence, and a better theory of ICEs, for instance concerning the various types that exist and how they relate to concretizations and to each other; these issue are currently not addressed in the Information Artifact Ontology or any other realism-based ontology.
We have presented the beginnings of a methodology that allows a clinical research dataset to be translated into a set of of Referent Tracking Tuples that has the following features: not only the portion of reality described by the dataset and the dataset itself are represented in a way that mimics the structure of reality, but so also are the relations between components of this dataset on the one hand and the corresponding portions of reality on the other. Applying the methodology to a concrete dataset and performing some basic exploratory statistics revealed that all of the relations we distinguished between data items and what they are about (if, indeed, they are about anything at all) do indeed occur in our study data. A set of
complicated exchange information models that are used in message-based paradigms or in the Extract – Transform – Load (ETL) analyses and procedures used in data warehousing. Although the syntax and semantics of RTTs seems to us to be powerful enough to represent what is required, a current limitation is the insufficient development of the Information Artifact Ontology. A second limitation is that not all RTTs can easily be translated into OWL-based languages. Where the former is a job to be done by ontologists, the latter is a task for computer science.
from the National Institute of Dental and Craniofacial
responsibility of the authors and does not necessarily represent the official views of the NIDCR or the NIH.