Content uploaded by Luigi Ceccaroni
Author content
All content in this area was uploaded by Luigi Ceccaroni
Content may be subject to copyright.
Appl. Comput. Math., V.xx, N.xx, 20xx, pp.xx-xx
INTEROPERABLE KNOWLEDGE REPRESENTATION IN CLINICAL
DECISION SUPPORT SYSTEMS FOR REHABILITATION
L. CECCARONI AND L. SUBIRATS1
Abstract. International standards and standard proposals of nomenclatures, ontologies and
information models exist in medicine, but the scope of most of them is overly general to cope
with the specificities that characterize rehabilitation. Here we carry out an ontology-based
exploration of the concepts and relationships in the rehabilitation domain, integrating clinical
practice, the clinical investigator record ontology and international standards. The aim of the
analysis is to identify potential logical problems with the use of existing models and international
standards in representing and reasoning with real clinical data, and to understand whether and
how these data might be defined more formally than in the current practice. Our analysis of the
relationships among rehabilitation concepts revealed issues related to confusion among classes
and their properties, incorrect classifications, overlaps and loss of information. It also suggested
properties that should be included in a formal model suitable to be used by decision support
systems.
Keywords: knowledge representation, rehabilitation, functional diversity, clinical decision sup-
port systems, personalized medicine, evidence-based medicine.
AMS Subject Classification: 68T30
1. Introduction
Intelligent systems, ontologies and interoperable knowledge representations are used, in medicine
in general and in physical medicine and rehabilitation (from now on, rehabilitation) in particu-
lar, for the purpose of increasing the interoperability among representations of the knowledge
about patients’ status. Interoperable representations are especially valuable for certain artificial
intelligence systems whose reasoning is based on experience or analogy, because reasoning result
improves if it is based on knowledge accumulated through a wide and diverse range of cases
spanning as many healthcare organizations as possible. An example are clinical decision support
systems (CDSSs) using case-based reasoning (CBR). Nomenclatures, ontologies and informa-
tion models can be introduced to facilitate automatic knowledge integration when representing
patient’s clinical data in these systems.
Information models exist to organize the information about patients and to capture the com-
munication processes of rehabilitation professionals, such as: the virtual medical record (vMR)
standard proposal and the ISO/EN 13606 standard. Nomenclatures and ontologies exist to iden-
tify and classify the characteristics of a patient, and other related concepts, through standard
descriptors, such as: the systematized nomenclature of medicine - clinical terms (SNOMED
CT), the international classification of functioning, disability and health (ICF), and the inter-
national classification of diseases (ICD). Developing and mapping these information models,
nomenclatures (including upper nomenclatures) and ontologies to the medical health records
(MHRs) used in healthcare organizations, is the first step towards knowledge integration. Diffi-
culties arise in the development of these models and systems, especially in the standardization
of the clinical-knowledge acquisition process (mapping scales and questionnaires into standard
1Media-TIC building, Roc Boronat 117, 5th floor, 08018, Barcelona, Catalonia, Spain
e-mail: lceccaroni@bdigital.org and lsubirats@bdigital.org
Manuscript received xx.
1
2 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
terminologies) using current methodologies (and ICF and SNOMED CT as standards). In the
paper, as summarized in Fig. 1, the advantages of mapping electronic health records (EHRs),
or patient health records (PHRs), to terminologies and classifications are considered to share
and reuse knowledge in the domains of population health, clinical environment, administration
and report presentation. In the following sections, the results obtained consisting of advances in
the knowledge representation related to computer-based decision support in rehabilitation are
presented, using a real-world rehabilitation scenario as case study.
Figure 1. Role of classifications and terminologies in healthcare services
The rest of the paper is organized as follows. Section 2 describes a scenario of physical
rehabilitation. Sections 3 and 4 explain respectively for what purposes CDSSs and ontologies
are developed and used in medicine and rehabilitation. Section 5 describes how problems found
in the extraction of standard-based indicators are solved. Section 6 compares potential solutions
for EHRs interoperability, and Section 7 describes the architecture of a system which implements
the ontology and solves EHRs interoperability. Finally, Section 8 mentions remaining issues
about reasoning in the rehabilitation domain, and conclusions are drawn in Section 9.
2. Rehabilitation scenario
A scenario of physical rehabilitation is considered here as case study, in which the main goal
is to improve the efficacy of rehabilitation of the upper extremity with respect to movement. To
achieve this, four sub-goals have been defined:
(1) Monitoring the movement of the upper extremity. In the monitoring, several technologies
are used, such as T-shirt–mounted sensors and orthotic devices.
(2) Performing specific activities of daily living (ADLs) with assistance as needed. In the
assistance of movement, the investigator plans the sequence of activities, while investiga-
tor agents monitor the movement of the upper extremity and assist patients only when
needed. These agents can be therapists or orthotic devices.
(3) Performing specific ADLs with continuous feedback. The feedback system gives real-time
information about the execution of the activity through virtual reality and interactive
video.
(4) Programming personalized rehabilitation activities. A knowledge-based system allows
personalized planning of sequences of activities on the basis of result indicators, whose
values are obtained from activities and other patient’s care information.
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 3
Figure 2. Representation of a physical-rehabilitation scenario
The rehabilitation-scenario concepts are depicted in Fig. 2 according to a consensus of experts.
The scenario is composed of five main steps (see Fig. 3): (1) diagnosis, (2) planning, (3)
preparation of ADLs, (4) execution of ADLs, and (5) activity reporting.
In this paper, patient is defined adopting Beale and Heard’s top-level ontology [1]: a patient
system (or patient ) is the object of care (typically one person, but could be more), seen es-
sentially as a biological or social system (depending on the perspective of the investigator); an
investigator system (or investigator ) is the investigating, healthcare-providing entity, the total-
ity of rehabilitation professionals and other agents who perform actions related to the care of
the patient (including the patient in the role of self-carer or self-medicator, as well as any family
members).
3. Decision support system
To automatically reason about real-world scenarios like this, and to assist professionals in
decision making within a standard framework and through so-called clinical decision-support
systems (CDSSs), a formal, standard knowledge representation is needed. To achieve it is
completely novel and challenging because very few tests, processes and methodologies were
designed to be consisted with current standard models [2]. Furthermore, CDSSs are tools which,
if used skillfully and respectfully by experienced rehabilitation and health professionals, can help
to cover the broad and complex areas of assessment and proposal, allowing professionals to focus
not exclusively on test data, but also on individual patients and their environment.
3.1. Categories of CDSSs. In fact, two main categories of CDSSs can be identified: those
oriented to assessment and those oriented to proposal. The objectives of the ones oriented to
assessment are: evaluation of a patient’s past, current and future status (this includes prognosis:
the likely outcome of an illness); quantification of risk; classification of patients according to
4 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
Figure 3. Description of the steps involved in a physical-rehabilitation scenario
their functional diversity. The objectives of the ones oriented to proposal are: risk prevention;
definition of therapeutic goals.
Here we deal with the design and development of a CDSS oriented to assessment and, more
specifically, to prognosis, a clinical-system environment which requires reasoning under uncer-
tainty. We use for the decision-making process a knowledge-based system (KBS) with case-based
reasoning (CBR), with standard indicators used to represent a case (e.g., Emotional functions
- b152). Fig. 4 shows a screenshot of the prognosis application’s graphical interface which
includes, as sections, the cause of the functional limitation, the rehabilitation treatment, and
the four branches of ICF (body functions,activities and participation,environmental factors and
body structures). The outcome of the prognosis CDSS is evaluated by investigators.
4. Medical knowledge modeling
Healthcare organizations use several tools to capture information. These tools make use of
specific terms, which are sometimes ambiguous: descriptor, grade, index, indicator, parameter,
questionnaire, scale, score and test. The terminology used in this paper is defined as follows
and is part of an ontology, which we defined (and encoded in OWL 2 [3]) based on standard
nomenclatures and ontologies:
•Index: a combination of indicators, questionnaires and possibly other indexes. The
function representing this combination gives as summarizing result a score.
•Indicator : a (sub jective or objective) parameter or descriptor used to measure or com-
pare activities and participation,body functions,body structures,environment factors,
processes, and results.
•Questionnaire (or test): a set of questions answered using a scale.
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 5
Figure 4. Interface of the CDSS oriented to assessment and more specifically, to
prognosis. ICF values of difficulty, deficiency or barrier are represented with red/4
in severe levels, orange/3 in grave levels, yellow/2 in moderate levels, green/1 in
mild levels and blue/0 in no difficulty, deficiency or environmental factors.
•Scale: a mapping between some ordered (qualitative or quantitative) values (or grades)
and their description. These values are used to answer questionnaires.
The main metrics of the ontology (which can be accessed at http://bioportal.bioontology.
org/ontologies/47011) are as follows: 137 classes; maximum depth of 9; maximum number
of siblings of 16; 20 classes with a single subclass, 16 object properties. There are not multiple
primitive representations of the same concepts as it is specified in Cimino [4]’s guidelines.
4.1. Classes. Most top-level concepts come from SNOMED CT [5], e.g.: Assessment scales
(273249006, subclass of Staging and scales), Physical object (260787004), Process (415178003,
subclass of Observable entity). Some of them come from ICF, e.g.: Environmental factors (e),
Activities and participation (d). ICF codes have a letter b,d,eor spossibly followed by a
number while SNOMED CT codes are composed by a 9-digit number. In selecting and reusing
concepts, we always maintained the hierarchical organization and consistency of the original
standard ontologies. In any case, top-level categories are overly general to characterize the
ontology, which can be better framed, conceptually, through other categories, more related to
rehabilitation scenarios, such as: Diagnosis (439401001), Functional independence measure
(FIM) (273469003), Orthotic devices (224898003), Rehabilitation therapy (52052004).
6 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
Figure 5. Summary of ontology’s classes and relationships
Figure 6. Ontology’s modifies relationship
Figure 7. Ontology’s is composed of relationship
In Fig. 5, 6 and 7 we show how the ontology is related to and integrate the state of the
art. Concepts are encoded as SNOMED CT classes if not otherwise specified. Concepts in
bold-frame boxes are encoded as ICF classes (e.g., Body structures). Concepts in grey boxes
are reused from Beale and Heard [1]’s clinical investigator record ontology (e.g., Instruction).
Underlined relations are reused from the ICD (specifically its 11th revision, ICD-11) (e.g., has
localization). Concepts in white boxes and non-underlined relations represent an extension by
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 7
the authors of existing approaches based on clinical practice (e.g., has indicator). Relations in
the figure, if not otherwise specified, are is a . At the top-level of the ontology, there are the
following concepts, as shown in Fig. 5.
•Historical (following Sowa [6]’s top-level categories), which subsumes:
–Observation: information created by an act of observation, measurement, question-
ing, or testing of the patient or related substance;
–Action: a record of intervention actions that have occurred, due to instructions or
otherwise;
–Environment: information on the context of the patient;
•Interpretation of findings: inferences of the investigator using the personal and published
knowledge bases about what the observations mean, and what to do about them; includes
all diagnoses, assessments, plans, goals;
•Instruction: instructions, based on interpretations of findings, sufficiently detailed so
as to be directly executable by investigator agents (people or machines), in order to
accomplish a desired intervention.
The Interpretation of findings category corresponds to Sowa [6]’s Description category, to the
notion of hypothesis in general science and to Rector [7]’s Meta-observations. The Instruction
category corresponds to Sowa [6]’s Script category. Assessment (see Fig. 5) relates to past,
current or projected states of affairs. Proposal relates to desired ones. A Diagnosis is the
attachment of a label to a group of observed signs and symptoms, which designates it (in the
understanding of the investigator) as being a particular phenomenon. A Differential diagnosis
allows for multiple possibilities, due to the lack of sufficient information or understanding to
attach one label. A Goal, such as monitoring the upper limb while performing the Activity of
daily living (ADL) (129025006) Dressing (129003000) (d540), is a statement about what the
desired state of the patient system should be, while a Plan is a statement about how to get
there.
4.2. Object properties. Object properties represent relationships between two classes or in-
stances. Apart from properties is a,is composed of and modifies, object properties of the
proposed ontology are described in Table 1. Some of them are based on ICD-11 [7].
Table 1. Object properties of the ontology
Object property Domain Range
has contraindication ADL Substance
has disease (ICD-11) Person Disease
has goal Person, ADL Participation, Body func-
tions, Body structures
has indicator Diagnosis, ADL Observation parameter
has interpretation Diagnosis Indicator
has location (ICD-11) Observation Body structures
has occupation Person Occupation
has technology ADL Physical object
is assisted by, is assisted as needed by ADL Person, Physical object
is executed by ADL Person
is manifestation of (ICD-11) Disease Observation parameter
is programmed by ADL Person
uses ADL Observation parameter
8 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
Fig. 5, 6 and 7 show the most used relationships among concepts, and the context in which
they are used in a typical scenario. Continuous lines represent the relation has subclass. Dis-
continuous lines represent: the object properties defined in Table 1, the relation composed of or
the relation modifies.
4.3. Indicators. Indicators are the main classes used for representing the status of a patient
and potentially number in the thousands. The two main classes used in rehabilitation are
process indicators and result indicators. A process indicator is used to assess whether a task
is being performed correctly. A result indicator is used to assess the performance in carrying
out an activity or whether the objectives of the activity have been achieved. To facilitate
human practice, only a selection of indicators, grouped into core sets, are used. Core sets can
be formed according to functionality,pathology or rehabilitation process. Core sets are useful
because human investigators can process only a fraction of the categories found in relevant
ontologies such as ICF and SNOMED CT [8]. Core sets already exist for several pathologies,
such as multiple sclerosis,spinal cord injury (SCI) or traumatic brain injury, though finding
the core categories for rehabilitation processes and moving from a pathology-based approach to
one based on functionality and rehabilitation is needed. A methodology to define core sets of
standard-based, rehabilitation-based indicators from existing questionnaires and non-standard
indicators (observation parameters) would include at least the following steps:
(1) search for questionnaires and non-standard indicators about rehabilitation;
(2) prioritization and selection of questionnaires and non-standard indicators based on literature-
relevance, and coverage of rehabilitation processes and indicator types;
(3) extraction of standard-based indicators from selected questionnaires and non-standard
indicators;
(4) aggregation of indicators according to a taxonomy of rehabilitation processes;
(5) selection of a core set of these indicators.
Steps 1 and 2 have been already carried out by the authors and other researchers in previous
research; in this paper we contribute to step 3, while the fourth and fifth steps belong to the
medicine domain and to future work.
5. Extraction of standard-based indicators
Methodologies exist to extract indicators from clinical questionnaires and encode them using
international standards. If it is necessary to combine indicators and their values, additional
methodologies might be needed to carry out this combination. These methodologies for combi-
nation do not currently exist for rehabilitation or their scope is limited. To encode indicators
into international standards, the World Health Organization’s ICF is considered first because it
provides a comprehensive specification of health-related human functioning in the domains of (i)
body functions and structures (e.g., sensory,neuromusculoskeletal and movement-related func-
tions), (ii) activities and participation, ranging from basic (e.g., dressing and eating) to complex
(e.g., working and living independently), and (iii) environmental factors that provide a context
for understanding functioning, functional diversity and health. If no suitable category is found
in the ICF to define a concept, SNOMED CT, which includes top-level categories, is considered.
The methodology to encode indicators into ICF can be found in Cieza et al. [9][10], while the
methodology to encode them into SNOMED CT is described by Subirats and Ceccaroni [11]:
(1) Using any search-capable SNOMED CT interface (e.g., [12]), search for the concept that
you want to encode.
(2) If there is no exact match, search for a synonym.
(3) If there are no synonyms, use a combination of hypernyms and hyponyms to find concepts
that are modeled in SNOMED CT.
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 9
(4) Check if the type of the concepts found in SNOMED CT properly models the concept
to be encoded. There are 19 types of concepts in SNOMED CT, such as clinical finding,
physical object, social context, physical force, substance or procedure. In the previous
examples, the type should be procedure in all cases.
See Table 2 for an example.
Limits exist in these methodologies if resulting indicators need to be used by formal rea-
soning systems, because in several cases branches of standard classes are heterogeneous in the
relationships used and some of the leaves of a sub-tree represent properties or parts of the parent
concept rather than subclasses. In particular, an analysis of the relationships within ICF revealed
problems related to confusion between classes and their properties, incorrect classifications and
overemphasis on subsumption [13].
Table 2. Encoding of indicator values into SNOMED CT and ICF
Questionnaire (FIM) item SNOMED CT ICF
Dressing upper body Ability to dress (165235000) Dressing (d540)
Putting on clothes (d5400)
Dressing lower body Ability to dress (165235000) Dressing (d540)
Ability to put on footwear
(284978003)
Putting on footwear (d5402)
Toileting Toileting (129004006) Toileting (d530)
Transfers: bed/chair/ wheelchair Chair/bed transfer ability
(165236004)
Transferring oneself (d420)
Transfers: toilet Ability to transfer be-
tween wheelchair and toilet
(302274006)
Transferring oneself (d420)
Locomotion: stairs Climbing stairs (129016000) Climbing (d4551)
5.1. Difficulties of mapping clinical questionnaires into standard terminologies and
ontologies. Several problems exist with the standardization of observation parameters (ques-
tionnaires’ indicator names and values) into SNOMED CT and ICF. One main difficulty is that
there is a mismatch between the way SNOMED CT categorizes its concepts and the question-
naire items. SNOMED CT makes a distinction between Clinical finding and Observable entity.
Findings are observations that are meaningful by themselves whereas observables need to have
values to complete their meaning. A questionnaire item should be mapped only to SNOMED
CT observables (see Table 2). Combinations of a questionnaire item and its values map to find-
ings. If we take, for example, the Transfers: toilet item, the complete mapping to SNOMED
CT and ICF is shown in Table 3. (Here we see that the extremes of the scale can be mapped to
findings, as an alternative to a mapping using only observables.) The mapping to ICF doesn’t
suffer from the problem of category mismatch, as ICF categories are neutral with respect to
functional diversity.
Sections 5.1.1 and 5.1.2 deal with specific issues found in the standardization of indicators
to ICF and SNOMED CT, respectively. Observation parameters of questionnaires FIM, spinal
cord independence measure (SCIM), patient competency rating scale (PCRS), extended Glasgow
outcome scale (GOSE), hospital anxiety and depression scale (HAD), from the Historical class
of 350 patients who suffer from neurological diseases, are considered.
5.1.1. Problems found in the standardization of observation parameters to ICF. Overlaps and
loss of information are found when implementing the methodology of ICF-encoding of Cieza et
al. [9][10].
10 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
Table 3. Encoding of indicator values into SNOMED CT and ICF
Questionnaire item’s
value
SNOMED CT ICF
”Transfers: toilet” 7 Able to transfer between wheelchair and toilet
(302275007)
”Transferring one-
self” 0
”Transfers: toilet”
6/5/4/3/2
Ability to transfer between wheelchair and toilet
with modified independence /supervision /min-
imal assistance /moderate assistance /maximal
assistance
”Transferring one-
self” 1/1/2/2/3
”Transfers: toilet” 1 Unable to transfer between wheelchair and toilet
(302276008)
”Transferring one-
self” 4
When the content of an item is not explicitly named in the corresponding ICF category, but
at the same time is included, then the item is linked to this ICF category and the additional in-
formation not explicitly named by the ICF is documented. The problem here is that when, e.g.,
functions of body structures are linked only to the activity or body function, body structures
cannot be mapped and distinguished. For example, the ICF standardizations of Dressing the up-
per body and Dressing the lower body of FIM are the same (Dressing, d540) or they are not based
on the original body structures (Putting on clothes, d5400, and Putting on footwear, d5402). A
solution would be to add relations representing the additional information not explicitly named
by the ICF, e.g., the relation has localization to link to Body structures.
The response options of an item are linked if they refer to additional constructs. Depending on
the answer-option chosen, the item is standardized to some indicators or others. For example,
the standardization of Dressing the upper body of SCIM is: Support and relationships (e3) if
the answer is Requires total assistance or Requires partial assistance;Dressing (d540) if it is
Independent;Assistive products and technology for personal use in daily living (e1151) if it is
Independent but requires adaptive devices. Items, independently of their nature, are standardized
to e3 if the answer is Requires total assistance. This causes an aggregation of diverse content
into one indicator, with potential loss of semantics. A solution would be to keep the information
found in the question and answer, and add a relation representing the degree of assistance, as
shown for SNOMED CT in Table 3.
Items are linked into high level categories. If the content of an item is standardized into high-
level ICF categories, it can cause a loss of information as higher categories are a generalization
of the concept and also do not usually appear in core sets (see section 4.3). For example, the
option Requires total assistance of nearly all items of SCIM is standardized to e3. In the ICF core
sets of SCI, there is a lower-level category Health professionals (e355) (which is not a completely
satisfactory standardization), but the category e3 is not there. A solution would be to form new
core sets according to rehabilitation processes (an initiative which is already ongoing) and taking
into account this generality issue both in core-set definitions and in mappings of questionnaires.
If the information provided by the meaningful concept is not sufficient for making a decision
about the most precise ICF category it should be linked to, the meaningful concept is assigned
not definable (nd), personal factor (pf), not covered by ICF (nc) or health condition (hc). For
example, in PCRS, the item Problem in accepting criticism is standardized to nd. A solution
would be to use SNOMED CT to complement the ICF, e.g., in this case, with the class Tends
to be sensitive to criticism (286846009).
5.1.2. Problems found in the standardization of observation parameters to SNOMED CT. In the
mapping of MHR observation parameters to SNOMED CT, the main difficulty is the standard-
ization of the values of the indicators. Specifically, problems appear when different types of items
are mapped to the same standard concept and their values need to be combined. Combining
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 11
values, and changing data types, has the disadvantage that information is (potentially) lost. For
example, the observable entity Climbing stairs (129016000) is the standardization of SCIM and
FIM concepts. The range of values of FIM’s Locomotion: stairs is 1 to 7, while the range of
SCIM’s Stair management is 0 to 3. Another example is the concept Anxiety (48694002), which
is measured by GOSE as a Boolean, while in HAD it is presented as a subscale with a range of
values between 0 and 21.
6. Interoperability and information models
With respect to interoperability, to exchange EHRs in a standard way, we have taken into
account the Smart Open Services for European Patients (epSOS), the virtual medical record
(vMR) and the ISO/EN 13606 initiatives. The patient summary is part of epSOS [14] and
includes evaluated persons’ relevant health information. Relevant information is understood as
the minimal set of personal health data (160 descriptors) that are of interest to health pro-
fessionals to assist citizens, and the ignorance of which could pose a risk to the health of the
evaluated person. The epSOS’ main objective is to assist practitioners in unscheduled care and
its general policy is to adopt international standards, such as HL7 v2. vMR, a simplified view
of HL7 v3, is still in the process of improvement and evolution by the vMR Project Team [15],
and there are several CDSSs that use it [16]. vMR is designed to reduce development costs and
time responses in CDSSs. As a consequence, although it is not widely adopted in hospitals and
there are not many tools available to facilitate its implementation (unlike the information model
ISO/EN 13606 [17]), we considered it the most appropriate information model for CDSSs today.
Table 4 shows some examples of the mapping between CDSS concepts and vMR objects, where
the data types used are:
•concept descriptor (CD), a reference to a concept defined in an external code system,
terminology, or ontology;
•entity name (EN), a name for a person, organization, place or thing;
•timestamp (TS), a quantity specifying a point on the axis of natural time; a point in
time is most often represented as a calendar expression; and
•interval timestamp (IVLTS), a set of consecutive values of an ordered base datatype.
These data types are a simplified version of ISO 21090 data types, which is an implementable
specification based on the abstract HL7 v3 data types specification.
Table 4. Relationships between CDSS and vMR objects
CDSS object (SNOMED CT code) Reference informa-
tion model (RIM)
Attribute Data type
Name (371484003) Person Name EN [0..*]
Date of birth (184099003) Evaluated person Birth time TS [0..1]
Problem (or disorder, 64572001) Problem base Problem code CD
Event (272379006) Adverse event base Adverse event code CD
Date of diagnosis (432213005) Problem base Diagnostic event
time
IVLTS
Body functions, activities and par-
ticipation, environmental factors
and body structures
Observation base Observation focus CD
12 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
7. Architecture of the knowledge-based system
The architecture of a healthcare institution’s knowledge-representation system which uses the
described ontology and interchanges knowledge with one or more health institutions using vMR
is illustrated in Fig. 8. This architecture includes the following modules:
•MHR: Relevant information derived from each clinical event systematically and sequen-
tially recorded. A medical health record is a set of both written and graphic documents,
referring to episodes of illness of a patient, and the activity that is generated because
of these episodes, stored in electronic form. These clinical data records, mainly from
questionnaires, are stored in MHRs.
•Knowledge storage system: MHR’s observation parameters standardized into ICF and
SNOMED CT indicators.
•Interface engine: Engines which encode/decode into/from HL7 v3 messages among
healthcare institutions. HL7 v3 is object-oriented (OO), uses the Unified Modeling Lan-
guage (UML) and is based on a data model called the Reference Information Model
(RIM).
•Aggregator: Health-institutions knowledge aggregated before storage.
•Investigator system: System which analyzes the knowledge and interprets findings about
patients.
Figure 8. Architecture of a system which implements the ontology and vMR
8. Issues in reasoning
For the decision-making process, we use a knowledge-based system (KBS) with case-based
reasoning (CBR). The CBR reasoning cycle includes the typical retrieve, reuse, revise and
retain phases. In the first phase (retrieve), which is implemented using jColibri, the most similar
patients to the evaluated patient are retrieved and ranked, using a k-nearest neighbor (k-NN)
similarity measure. Applying reasoning such as CBR is always possible within a healthcare
institution because of data homogeneity. Efforts are ongoing (of which this research is part)
to adopt common, standard-based representations, therefore allowing the exploitation of data
among all compliant institutions. However, even with the multiplication of potential data-sets
to be used as case libraries, several issues remain with reasoning in the rehabilitation domain
and we mention here some of them:
•New standardized indicators need to be made compatible with existing, historical data
series.
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 13
•The use and weight of demographic variables (such as disease,years from the injury,
cause or age) needs to be considered.
•Even if standard indicators are used, different patients may be evaluated using different
indicators, making their comparison difficult to be defined.
•Indicators may not be independent of each other and this dependence in classes and
values needs to be taken into account.
•The most relevant categories for representing patients with respect to their functional
diversity (instead of diseases) need to be determined, using core sets or otherwise.
•Similarity calculation among patients is used in the retrieve phase of CBR to provide
decision support. About weight assignment for similarity calculation, after several exper-
iments and consultation with medical investigators, we can suggest that the more specific
a category, the more weight it should have; and categories which do not appear in core
sets should have weight equal to (or close to) 0. The weight given to ICF categories
which appear in core sets depends on their level: a relative weight of 1 should be given
to 2nd level categories, a weight of 2 to 3rd level categories, and a weight of 3 to 4th level
categories. Finally, in approaches using indicators standardized into SNOMED CT, as
there are no SNOMED CT core sets, ICF core sets could be used to assign weights.
9. Conclusion
The framework of the research presented in this paper is the progress in clinical decision
support using formal reasoning paradigms. To this aim, some previous steps need to be taken,
such as: having clear, standard, interoperable knowledge organized in ontologies, nomenclatures
(such as ICF and SNOMED CT) and information models (such as vMR); and having a mapping
system between clinical questionnaires and indicators expressed in standard terminology (both
in terms of classes and in terms of values). Interoperable representations are valuable for a
number of artificial intelligence systems whose reasoning is based on experience or analogy,
because reasoning result improves if it is based on knowledge accumulated through a wide and
diverse range of cases spanning as many healthcare organizations as possible. We analyzed
the mapping from observed parameters in the clinical practice to ICF and to SNOMED CT
observables and findings; we detected problematic issues and we provided suggestions towards
solutions that can improve the current situation, showing the potential of introducing artificial
intelligence techniques in medical assessment and proposal.
10. Acknowledgment
The research described in this paper arises from a Spanish research project called Rehabilita
(Disruptive technologies for the rehabilitation of the future), which is funded by the Centre for
Industrial Technological Development (CDTI), under the CENIT program, in the framework of
the Spanish government’s INGENIO 2010 initiative. This work is also partly supported by the
Catalonia Comp etitiveness Agency (ACC1 ´
O). The opinions expressed in this paper are those of
the authors and are not necessarily those of Rehabilita pro ject’s partners, CDTI or ACC1 ´
O.
References
[1] Beale, T., Heard, S. An Ontology-based Model of Clinical Information, Stud. Health Technol. Inform., N.129,
2007, pp.760-764.
[2] Mpofu, E., Oakland, T. Rehabilitation and health assessment: applying ICF guidelines, Springer Publishing
Company, New York, 2010, 760p.
[3] Cuenca Grau, B., Horrocks, I., Motik, B., Parsia, B., Patel-Schneider, P., Sattler, U. OWL 2: The next step
for OWL, Web Semantics: Science, Services and Agents on the World Wide Web, V.6, N.4, 2008, pp.309322.
[4] Cimino, J.J. Desiderata for Controlled Medical Vocabularies in the Twenty-First Century, Methods Inf. Med.,
V.37, N.4-5, 1998, pp.394-403.
14 APPL. COMPUT. MATH., V.XX, N.XX, 20XX
[5] Donnelly, K. SNOMED CT: The Advanced Terminology and Coding System for eHealth, Stud. Health
Technol. Inform., V.121, 2006, pp.279-290.
[6] Sowa, J.F. Knowledge Representation: Logical, Philosophical, and Computational Foundations, Brooks Cole
Publishing, Pacific Grove, 2000, 594p.
[7] Rector, A.L. Nowlan, W.A., Kay, S. Foundations of an Electronic Medical Record, Methods Inf. Med., V.30,
1991, pp.179-186. http://apps.who.int/classifications/icd11.
[8] Rodrigues, J.M., Kumar, A. Bousquet, C., Trombert, B. Using the CEN/ISO standard for categorial structure
to harmonise the development of WHO international terminologies, Stud. Health Technol. Inform., N.150,
2009, pp.255-259.
[9] Cieza, A., Brockow, T., Ewert, T., Amman, E., Kollerits, B., Chatterji, S., ¨
Ust¨un, B., Stucki, G. Linking
health-status measurements to the international classification of functioning, disability and health, Rehabil.
Med., V.34, N.5, 2002, pp.205-210.
[10] Cieza, A., Geyh, S., Chatterji, S., Kostanjsek, N., ¨
Ust¨un, B., Stucki, G. ICF linking rules: an update based
on lessons learned, Rehabil. Med., V.37, 2005, pp.212-218.
[11] Subirats, L., Ceccaroni, L. An ontology for computer-based decision support in rehabilitation, The Tenth
Mexican International Conference on Artificial Intelligence (MICAI 2011), Lecture Notes in Computer Sci-
ence, Springer, V.1, 2011, pp. 549-559.
[12] College of American Pathologists, National Health Service. Systematized Nomenclature of Medicine Clinical
Terms (SNOMED CT). http://bioportal.bioontology.org/ontologies/1416.
[13] Della Mea, V., Simoncello, A. An ontology-based exploration of the concepts and relationships in the Activ-
ities and Participation component of the International Classification of Functioning, Disability and Health,
Biomed. Semantics, V.3, N.1 in production, 2012.
[14] Smart Open Services for European Patients (epSOS) http://www.epsos.eu/.
[15] VMR Project Team. Virtual Medical Record (vMR) for Clinical Decision Support Domain Analysis Model
2011, Release 1, September. http://wiki.hl7.org/index.php?title=Virtual_Medical_Record_(vMR) (ac-
cessed on 15-04-2012).
[16] Kashfi, H. The Intersection of Clinical Decision Support and Electronic Health Record: A Literature Review,
Proc. of the Federated Conf. on Comp. Science and Inf. Syst., Szczecin, Poland, 18-21 Sept. 2011, pp. 347353.
[17] Mart´ınez-Costa, C., Men´arguez-Tortosa, M., Fern´andez-Breis, J.T. An approach for the semantic interoper-
ability of ISO/EN 13606 and openEHR archetypes. Biomed. Inform., V.43 N.5, 2010, pp. 736-746.
Luigi Ceccaroni - obtained an M.Sc. de-
gree in Information-Technology Languages and Sys-
tems by the Universitat Polit`ecnica de Catalunya
- BarcelonaTech (UPC)(Spain); and completed his
PhD in Artificial Intelligence (first class honors) at
UPC. His main research interests combine: ontolo-
gies; recommendation systems; semantic tools; Web
services; the semantic Web; personalization; and ap-
plication of artificial intelligence to healthcare and
environmental sciences. During his career he has
worked as a member of research staff at the Network
Agent Research (NAR) group at Fujitsu Laboratories
of America; as director of research at TMT Factory
(Barcelona, Spain) and as a senior member of research
staff of the Software Department (LSI) at UPC. He
has also been adjunct professor of Artificial Intelli-
gence at UPC. He is currently working as a senior
member of research staff at Barcelona Digital Tech-
nology Centre.
L. CECCARONI AND L. SUBIRATS: INTEROPERABLE KNOWL. REPRESENT. IN CDSSS FOR REHABIL. 15
Laia Subirats - holds a B.Sc. degree in Telecom-
munications Engineering (first class honors) from
Pompeu Fabra University (Spain) and a M.Sc. in
Telematics Engineering by Universitat Polit`ecnica de
Catalunya - BarcelonaTech (UPC) (Spain) and Car-
los III University of Madrid (Spain). During her de-
gree she has been working in some projects in disci-
plines such as ontologies, software engineering, social
networks, reputation and e-learning; both in national
and international centers (Telef´onica I+D and Euro-
pean Organization of Nuclear Research in Geneva,
Switzerland) and initiatives such as Google Summer
of Code. She is currently working in Barcelona Digital
Technology Centre in her Ph.D. in Artificial Intelli-
gence about intelligent systems for decision support
in rehabilitation.
