=Paper=
{{Paper
|id=Vol-1546/paper_18
|storemode=property
|title=ODM on FHIR: Towards Achieving Semantic Interoperability of Clinical Study Data
|pdfUrl=https://ceur-ws.org/Vol-1546/paper_18.pdf
|volume=Vol-1546
|authors=Hugo Leroux,Alejandro Metke,Michael John Lawley
|dblpUrl=https://dblp.org/rec/conf/swat4ls/LerouxML15
}}
==ODM on FHIR: Towards Achieving Semantic Interoperability of Clinical Study Data ==
https://ceur-ws.org/Vol-1546/paper_18.pdf
ODM on FHIR: Towards Achieving Semantic
Interoperability of Clinical Study Data
Hugo Leroux, Alejandro Metke-Jimenez and Michael J. Lawley
The Australian E-Health Research Centre, Health and Biosecurity, CSIRO
{firstname.lastname}@csiro.au
Abstract. Observational clinical studies play a pivotal role in advanc-
ing medical knowledge and patient healthcare. However, to lessen the
prohibitive costs of conducting these studies and support evidence-based
medicine, results emanating from these studies need to be shared and
compared with one another. This paper explores how semantic interop-
erability of clinical data can be achieved by integrating two prominent
standards for clinical data: ODM and FHIR. ODM lacks a rich-enough
information model to adequately capture the contextual information of
clinical study data. This is overcome by using FHIR’s information model
to achieve semantic interoperability of clinical data. This work outlines
our ongoing effort to integrate the ODM standard to the FHIR standard.
In particular, it demonstrates how the hierarchical ODM model lends it-
self to be mapped to the ubiquitous FHIR resources. We describe the
approach and provide insights into the assumptions made to fit the clin-
ical data extracted from the ODM standard into the FHIR resources.
Our focus is not only on mapping the data from ODM to the FHIR
models but on capturing the contextual information, present in other
sources, such as the study protocol, and which should have been made
available with the extracted data. Finally, we discuss the exceptions un-
der which the extracted ODM data does not adequately fit the targeted
FHIR resources and offer some insight into a suitable solution.
1 Introduction
It is increasingly important to share results from clinical studies. The challenge
when comparing results from clinical studies is to ensure that we compare corre-
sponding data sets. The CDISC ODM1 defines an XML-based standard that has
been mandated by the Food and Drug Administration (FDA) for the electronic
capture and reporting of clinical study data. While ODM provides a vehicle
to communicate the results back to the regulatory body, it lacks a rich-enough
information model to capture the innate contextual information of the clinical
study data [11]. As a result, ODM is ill-suited for advancing the semantic in-
teroperability solution that is required to achieve cross-study exploration of the
clinical studies.
1
Clinical Data Interchange Standards Consortium Operational Data Model
2
A recent report [12] has advocated the integration of clinical data from dis-
parate studies in a machine readable format to promote efficient evidence-based
medicines. By the same token, Hsu et al. [8] argue that to fulfil precision medicine
requires the mining and aggregation of clinical data from multiple sources and
novel approaches to obtaining contextual observations. The Fast Healthcare In-
teroperable Resources (FHIR) framework, which is a HL7 standard that has
been swiftly adopted by the health-care community, looks the likely candidate
for overcoming this challenge. It is geared towards communication of clinical
data using HL7 messaging protocols but is also supported by a rich information
model to achieve semantic interoperability of clinical data. This makes FHIR
the natural match to complement the ODM standard.
In this work, we present an approach to integrate the CDISC ODM standard
with the FHIR resources to enrich longitudinal clinical study data extracted from
ODM. The aim is to exploit the rich information model from FHIR to reintroduce
the contextual information that is often contained within the study protocol
documents but not contained within ODM and rarely made available alongside
the clinical data. We explore the hierarchical concepts within the ODM and
describe how these fit the pervasive FHIR resources. In so doing, we elaborate
on the assumptions that have been made to adapt the extracted clinical data
to the FHIR resources. The expected benefit of this approach is in facilitating a
richer exploration and querying of clinical data coupled with relevant contextual
information.
2 Background
2.1 CDISC ODM
The CDISC ODM data model [4] is specifically designed for a data capture
context. It consists of two main hierarchies: a Clinical Data and a Metadata
hierarchy, as depicted in Figure 1, that are referenced using the same object
identifier (OID). These two parallel hierarchies ensure that the clinical study
follows a predetermined structure of subject, event, form, item group and
item. An item corresponds to a single measurement or analysis result captured
during the study. An item group typically comprises a set of contextually-
related measurements or results. A form (or case report form) is a collection
of items, some grouped, for capturing and displaying clinical data. A study
event corresponds to a patient encounter in the course of the study whereby
data corresponding to one or more forms is collected.
2.2 FHIR
FHIR2 , published as a Draft Standard for Trial Use by the Health Level
Seven (HL7) organisation is a specification designed to facilitate the exchange of
2
http://hl7.org/fhir/2015Sep/overview.html
3
ODM METADATA
Study
MetaDataVersion
Study Event Def
Form Def
DATA ItemGroup Def
Clinical Data Item Def
Subject Data
Study Event Data
Form Data
ItemGroup Data
Item Data
Fig. 1. Illustrates how the ODM data is organised into data and metadata
healthcare-related information. FHIR revolves around resources, which are snip-
pets of highly-focussed data. The specification defines a set of minimalistic and
generic elements and extensions are used to bridge the gap for the remaining
content. Figure 2 depicts the hierarchy for the proposed FHIR resources to model
the ODM data. The entities in red (CarePlan and Questionnaire) denote meta-
data concepts. The remaining entities model the clinical data at various levels.
Solid lines are used to denote the links between the entities. The original model
is depicted in Figure 4 in the Appendix. It was discarded in favour of Figure 2
so that there could be a clear demarcation between the metadata and the data
and to allow the questionnaire to be linked to the care plan. The original model
did not allow for a resource to contain the entire clinical data pertaining to a
patient. This role is fulfilled by the ClinicalImpression resource in the new
model and has necessitated the introduction of the EpisodeOfCare resource.
Clinical
CarePlan
plan Impression Episode of
investigations Care
episode of care
patient patient
Patient Encounter
patient
activity encounter
subject patient
Observation
Questionnaire encounter
Questionnaire
questionnaire Response value
Fig. 2. Depicts the metadata (red) and data (blue) FHIR resources and their links
4
3 Integrating ODM with FHIR
This section describes the approach to integrate ODM with FHIR. We describe
the components from the ODM hierarchy that we seek to map to FHIR. We
assume here that the person doing the mapping has access to all the contextual
information relating to the study. In this vein, it is not our aim to perform a
one-to-one mapping of the CDISC ODM with FHIR. Rather we seek a holis-
tic approach to mapping the hierarchical concepts between the two models as
depicted in Figure 3.
ODM
DATA CarePlan METADATA
Study
MetaDataVersion
Clinical Data
Pa@ent Study Event Def
Subject Data
Form Def
Clinical ItemGroup Def Ques@onnaire
Impression Study Event Data
Item Def
Episode of Care Form Data
Encounter ItemGroup Data
Ques@onnaire Item Data Observa@on
Response
Fig. 3. Illustrates how the ODM entities are mapped to the FHIR resources
3.1 Study
A study defines static information about the structure of an individual study. We
wish to not only capture this static information but also much of the contextual
information pertaining to the study that is contained in the study protocol. This
has resulted in the choice of the CarePlan resource to map the Study component
from ODM. It provides a link to the study coordinator through the participant
attribute and study protocol through the support attribute.
3.2 Subject
While the Subject represents a critical element of the study, its role is quite
subdued in ODM. In particular, the specification provides no functionality to
record the subject’s attributes such as gender, date of birth, recommending that
these be modelled as clinical data within the forms. The logical mapping for
the Subject in FHIR is the Patient resource. This allows us to include relevant
contextual information, such as the patient’s gender, date of birth and care
5
provider and allows the study subject to be linked to other FHIR resources
containing pertinent study-related information. The clinical data for each subject
is encapsulated within a ClinicalImpression resource that is linked to the
Patient resource.
3.3 Study Event
A study event comprises a StudyEventDef and a StudyEventData component
that are referenced using a common OID. The StudyEventDef manages the set
of forms to be completed at this phase of the study and represents an activity
within the CarePlan resource. StudyEventDef entities define scheduled and un-
scheduled events and these may be defined within the detail.scheduled at-
tribute of the activity. The StudyEventData entity contains clinical data col-
lected during a subject’s visit. We believe that the EpisodeOfCare resource is
appropriate for this entity because it provides details about the group of activ-
ities and their purpose pertaining directly to a patient. The care plan is linked
to this resource using the plan attribute. A study event may result in many
visits from a patient. Each individual visit is modelled as an Encounter and is
linked to the episode of care through the episodeOfCare attribute. The patient
attribute links the resource to the study subject while the assessor attribute
provides a link to the clinician conducting the clinical assessment.
3.4 Form
A form defines a collection of data items collected during the study and termed a
case report form. A form comprises a FormDef and a FormData component that
are referenced using a common OID. The form is linked to CarePlan through the
activity.actionResulting attribute. The FormDef defines the form structure
and its questions. The logical mapping of forms in FHIR is the Questionnaire
(Q) resource. This resource contains the typical attributes for questionnaires,
such as an identifier, version, publisher and status, but can also be customised
using the extension mechanism in FHIR. The FormData entity contains the clin-
ical data associated with the form. The logical mapping for the FormData in
FHIR is the QuestionnaireResponse (QR) resource. The benefits of using the
QR resource are that the order of the responses is maintained and these can
be linked and validated against the questions asked. Conversely, however, while
CDISC defines the CDASH3 model to standardise the generation of CRFs for
clinical studies, its use is not enforced. Consequently, in our experience, few
study coordinators choose to use them [11]. As a result, CRFs often contain
contextually unrelated questions grouped together because they match the way
in which the data entry person collects the data. Choosing to model this in a
FHIR resource will only serve to perpetuate a bad practice [11].
3
Clinical Data Acquisition Standards Harmonization
6
3.5 Item Group
The ItemGroupDef and ItemGroupData entities constitute an item group ref-
erenced using a common OID. The ItemGroupDef entity defines the optional
grouping of questions on a form. Groups are defined using the Q.group at-
tribute. The FHIR specification stipulates that a group attribute define either
a question or a group but not both. The ItemGroupData contains the clinical
data detailing the responses for the item group. FHIR organises these grouped
responses within the QR.group attribute. The FHIR specifications requires the
order of the responses within the group to be maintained. This is a very impor-
tant constraint. Consider the situation where the heart rate measurement of a
patient indicates that the patient might have had a slight malaise during data
collection and that subsequently a blood pressure measurement was taken. In
this case, it would be prudent to analyse the blood pressure observation in this
context and make inferences accordingly. On the flip side, questions are often
grouped together on a form to match the collection habit of the data entry per-
son and not necessarily because of their semantic similarity. Grouping responses
in this manner does not advance the semantic interoperability principles.
3.6 Item
At the item level, the ItemDef and ItemData entities define each question and
its subsequent response. The ItemDef entity defines the question asked dur-
ing the study along with defining attributes such as the datatype, data size,
measurement unit, permissible range and code list. The Q.group.question at-
tribute is the most appropriate to define the ItemDef entity. The logical mapping
for the ItemData entity is the QR.group.question attribute. The response to
the question is then contained within the question.answer sub-attribute. This
model works best in a lifestyle study scenario using questionnaires in the tradi-
tional question-answer mode. In the case of longitudinal clinical studies where
the responses are analogous to a patient’s observations during an episode of
care, we believe the ItemData entity to be more appropriately represented using
the Observation resource. Furthermore, as outlined in the FHIR specifications,
data captured in questionnaires can be difficult to query after the fact. Individ-
ual items within a QR or an Observation are subsequently linked back to the
Encounter in which they occur.
4 Discussion and Related Work
The FHIR resources provide a good fit to semantically enrich the extracted
data from the CDISC ODM. In spite of its shortcomings in providing context
to the clinical data, the CDISC ODM provides a sound hierarchical framework
for capturing the clinical data that needs to be replicated in the new model.
Several assumptions have been made when mapping the ODM data to the FHIR
resources because their objectives differ and this work represents one view of
7
how the mapping can be achieved. We chose to model the study as a CarePlan
because we want to model the activities planned for the patient during the study
in the context of the study protocol. The CarePlan resource offers a number
of attributes, such as context, category and description that can provide
additional context to the care plan. The clinical data pertinent to a patient
is modelled using the ClinicalImpression resource. The ClinicalImpression
permits very pertinent information to be associated to the patient’s data through
the use of the trigger, investigations and summary attributes. Furthermore,
it makes it possible to explicitly link the protocol followed and to associate the
findings to the clinical data.
The StudyEventDef is modelled as an activity within the care plan. At
a macro level, the study event data is categorised as an EpisodeOfCare. The
EpisodeOfCare resource provides broad context to the study event. At a micro
level, each visit within the study event is represented as an Encounter. This
provides a richer summary of the activity performed, allowing each visit to be
described atomically and linked back to the event using the episodeOfCare
attribute. However, this requires the ODM data to be rich and accurate enough.
This is particularly important for study event activities that occur at differing
times in a particular day.
The Questionnaire resource is a suitable match for mapping ODM forms.
However, as outlined in section 3.4, forms are often ill-conceived in ODM and
as discussed in sections 3.5 and 3.6, the tendency is not to organise questions in
a contextual manner but in one that befits the data capture process. The impli-
cation is that tremendous effort, which grows exponentially with the size of the
study, must be expended to semantically enrich the clinical data by regrouping
it contextually and integrating it with the relevant domain ontologies [10, 9].
As outlined in section 3.6, we advocate the use of the Observation resource
to model the responses from item data. FHIR considers the Questionnaire re-
source to be a specialisation of the Observation resource. The appeal in adopting
the Observation resource to store the ItemData responses is the ability to store
important contextual information alongside the clinical data. Specifically, the
dataAbsentReason attribute, which enables some justification to be provided
as to the absence of a measurement, is very important in a clinical study set-
ting, as is the ability to interpret the observation in the context of a controlled
vocabulary or ontology. Furthermore, an Observation allows pertinent informa-
tion such the method, specimen and performer but also the device, bodySite
and related attributes to be coupled with the data, which is very useful when
observing clinical data such as vital signs.
Ultimately, the aim of this work is to stimulate a debate on the most effec-
tive way to model clinical study data. While mapping ODM to FHIR is, in our
view, a suitable solution, we do not regard it as a permanent one. Fundamen-
tally, FHIR has the potential to manage clinical data in its own right. This has
numerous advantages because as discussed in [7], a mapping process invariably
leads to the loss of pertinent information. Furthermore, the process involves the
reintroduction of critical domain information into the model. A more efficient
8
process would be to include this information in casu during the data collection
phase.
4.1 Related Work
Several researchers have initiated approaches to address the semantic enrichment
of clinical data with a view to achieving interoperability. One such approach, the
Linked Clinical Data Cube (LCDC) [10, 9, 11] is a set of modularised data cubes
that helps manage the multi-dimensional and multi-disciplinary nature of clin-
ical data. A comprehensive comparison between the LCDC and this work is
outside the scope of this paper. They have similar aims in trying to semantically
enrich clinical study data and provide additional dimensions to overcome the
monolithic nature of the ODM data and facilitate the exploration and querying
of the data. The LCDC achieves this by introducing specialised cubes, slices and
observations. In this work, all the chosen resources, except for the care plan and
the questionnaires, have a link back to the patient. Furthermore, there are links
between the Observation, Encounter, Questionnaire and QuestionnaireResponse
resources to augment the dimensions offered. The patient and the observation
are the main focus in both cases and they both provide the mechanisms to in-
terpret the observations in the context of externally controlled vocabularies or
ontologies. Furthermore, they both offer the ability to specify that the observa-
tion data is missing, although the Observation resource in FHIR also includes
the justification as to why the result is missing. The main difference is in the ap-
proach. The LCDC requires mapping to the RDF Data Cube [5] and DDI-RDF
Discovery [3] vocabularies to organise the data and links to domain ontologies
to enrich it. In this work, the organisation, management and enrichment are
performed by FHIR using Codeable concepts to link it externally.
Dugas [7] describes two tools to convert forms between the CDISC ODM and
HL7 CDA4 formats to facilitate the sharing of electronic health records (EHRs)
and clinical data to address the problem of redundant documentation in both sys-
tems. He concluded that the conversion process is lossy because the CDISC and
HL7 models serve different purposes and hence have different properties. Simi-
larly, the SALUS project [6] aims to address the interoperability between clinical
care and the clinical research domain. More specifically, it looks at combining
the strengths of CRFs with those of EHRs to address adverse drug reactions.
Abler et al. [1] discuss the need for a language for forms that can effectively
record the logical relationships between questions or sets of questions asked in
the forms. While the natural inclination would be to look to the Questionnaire
resource to fulfil this need, a more encompassing solution would be to integrate
the capabilities of the Observation resource with the questionnaire.
The Pharmaceutical Users Software Exchange5 community, in concert with
the FDA, has started work on RDF representations of various CDISC models6 ,
4
Clinical Document Architecture
5
http://www.phusewiki.org/wiki/index.php?title=Semantic Technology
6
https://github.com/phuse-org/rdf.cdisc.org
9
including the terminologies published by the National Cancer Institute (NCI)
Enterprise Vocabulary Services7 . This community has started to evaluate the
RDF Data Cube [2, 13] for the publication of clinical study data.
The HL7 Working group on Semantic Interoperability8 has initiated some
work on translating the XML or JSON version of FHIR into FHIR RDF. This
work is still in draft mode and we will look at translating our solution into FHIR
RDF once it has reached a more mature level.
5 Conclusion
As secondary use of clinical data gathers momentum, it will become increas-
ingly important to share and compare clinical studies. We have presented an
approach that integrates clinical data extracted from CDISC ODM to the FHIR
resources. This work takes advantage of the rich information model compris-
ing the FHIR resources to semantically enrich the clinical data and reintroduce
the domain information that was omitted during the data capture phase. The
objective is to achieve semantic interoperability of clinical study data by stan-
dardising and normalising the data along the same metrics. The main contri-
bution is a framework to organise clinical data in a manner that preserves its
organisation but captures its context. A sample of this mapping is available at:
http://healthinet.it.csiro.au/net/odmFhirMapping/.
References
1. Abler, D., Crichton, C., Welch, J., Davies, J., Harris, S.: Models for forms. In:
Proceedings of the compilation of the co-located workshops on DSM’11, TMC’11,
AGERE!’11, AOOPES’11, NEAT’11, & VMIL’11. pp. 13–18. ACM (2011)
2. Andersen, M.: Linked data to support clinical and non-clinical reporting. In: Pro-
ceedings of 2nd International Workshop on Semantic Statistics (SemStats 2014).
Riva del Garda, Italy (2014)
3. Bosch, T., Cyganiak, R., Gregory, A., Wackerow, J.: DDI-RDF Discovery Vocabu-
lary: A metadata vocabulary for documenting research and survey data. In: LDOW
(2013)
4. CDISC: Specification for the Operational Data Model (ODM) (2006),
http://www.cdisc.org/models/odm/v1.3/ODM1-3-0-Final.html
5. Cyganiak, R., Reynolds, D., Tennison, J.: The RDF Data Cube Vocabulary (2013),
http://www.w3.org/TR/2013/PR-vocab-data-cube-20131217/
6. Declerck, G., Hussain, S., Daniel, C., Yuksel, M., Laleci, G.B., Twagirumukiza, M.,
Jaulent, M.C.: Bridging data models and terminologies to support adverse drug
event reporting using EHR data. Methods of Information in Medicine 54(1), 24–31
(2015)
7. Dugas, M.: ODM2CDA and CDA2ODM: Tools to convert documentation forms
between EDC and EHR systems. BMC medical informatics and decision making
15(1), 40 (2015)
7
http://www.cancer.gov/cancertopics/cancerlibrary/terminologyresources/cdisc
8
http://wiki.hl7.org/index.php?title=RDF for Semantic Interoperability
10
8. Hsu, W., Gonzalez, N.R., Chien, A., Villablanca, J.P., Pajukanta, P., Viñuela, F.,
Bui, A.A.: An integrated, ontology-driven approach to constructing observational
databases for research. Journal of biomedical informatics 55, 132–142 (2015)
9. Lefort, L., Leroux, H.: Design and generation of linked clinical data cubes. In:
Proceedings of 1st International Workshop on Semantic Statistics (SemStats 2013).
Sydney, Australia (2013)
10. Leroux, H., Lefort, L.: Using CDISC ODM and the RDF Data Cube for the se-
mantic enrichment of longitudinal clinical trial data. In: Paschke, A., Burger, A.,
Romano, P., Marshall, M.S., Splendiani, A. (eds.) Semantic Web Applications and
Tools for the Life Sciences (SWAT4LS). CEUR Proceedings (2012)
11. Leroux, H., Lefort, L.: Semantic enrichment of longitudinal clinical study data
using the CDISC standards and the semantic statistics vocabularies. Journal of
Biomedical Semantics 6(1), 16 (2015)
12. van Valkenhoef, G., Tervonen, T., de Brock, B., Hillege, H.: Deficiencies in the
transfer and availability of clinical trials evidence: a review of existing systems and
standards. BMC medical informatics and decision making 12(1), 95 (2012)
13. Williams, T.: A primer on converting analysis results data to RDF Data Cubes
using free and open source tools. In: Proceedings of 10th Annual PhUSE confer-
ence(PhUSE 2014). London, UK (2014)
A Appendix
CarePlan
Encounter
subject activity
activity
patient
Patient
Clinical patient
Impression subject
investigation
Questionnaire encounter
Response
encounter
Questionnaire Observation
questionnaire value
Fig. 4. Original model depicting the metadata (red) and data (blue) FHIR resources
with actual (solid) and potential (dashed) links