Medical-treatment Recommendation and the Integration of Process

Models into Knowledge-based Systems

Laia Subirats

1,2

, Luigi Ceccaroni

, José María Maroto

, Carmen de Pablo

and Felip Miralles

Barcelona Digital Technology Centre, Roc Boronat, 117, 5th floor, 08018 Barcelona, Spain

Universitat Autònoma de Barcelona, Campus UAB, 08193 Bellaterra, Spain

Cardiac Rehabilitation Unit, Hospital Universitario Ramón y Cajal, 28034 Madrid, Spain

Keywords: Ontologies, Rule-based Reasoning, Cardiac Rehabilitation Processes, Clinical Decision Support Systems.

Abstract: Decision making based on evidence other than human reasoning is becoming increasingly important in

healthcare. Valuable evidence is in the form of treatment processes used by healthcare institutions and this

paper presents a new framework for representing and modeling knowledge from these processes.

Specifically, it presents the integration of data from literature, business processes and decision trees through

workflows that cover the full cycle of health care, from diagnosis to prognosis and treatment. With respect

to patient status, as single instants cannot convey sufficient information, time series are analyzed and

classified to improve decision-making ability. The elicitation of new knowledge takes into account

international standards, ontologies, information models, nomenclatures and multiple types of indicators. The

integration of formal process-modeling in knowledge-based systems is exemplified by a real-world

recommendation scenario. After evaluation with a medical-rehabilitation data set, results show a strong

correspondence between treatment recommended by the proposed system and clinical practice.

1 INTRODUCTION

In medical rehabilitation, to evolve from the current

situation to a more individualized one and to solve

interoperability problems, the involvement of all

stakeholders in the development of models for

rehabilitation processes is needed. Clinicians use

guidelines, workflows and experience in the form of

past cases.

Here, we focus the attention on the integration of

past cases and workflows as a valuable source of

implicit knowledge not previously recognized by

clinicians. Because workflows are often informally

described (e.g., in diagrammatic languages, such as

the business process modeling notation or the

unified modeling language activity diagram), they

cannot be automatically included into a

computational reasoning system, but need to go

through some sort of nontrivial formalization

process. For this, a variety of languages can be used,

viz. procedural languages (e.g., the business process

execution language) or highly-formal languages

(e.g., Petri nets). Existing knowledge, including case

libraries when a case-based reasoning (CBR)

system is used, constitutes the input source for

workflows, and the output data of these workflows

are then delivered back to knowledge containers as

well as case libraries.

Effective management of processes is critical to

the development as well as sustainment of

rehabilitation capabilities. In the context of this

paper, a process defines a description and ordering

of work activities across time and space that is

designed to yield specific results or services while

ensuring the rehabilitation's overall objectives. It

provides a conceptual basis for the integration and

coordination of distributed resources, tasks and

individuals (Cichocki et al., 1998).

Process models produced by healthcare experts

are represented as formal rules, providing a

framework that allows their semantic enrichment.

Whilst some progress in this direction has been

made in recent research (Jafarpour et al., 2011; Peek

et al., 2011 and Smith et al., 2012], the integration of

data from literature, business processes and decision

trees through workflows that cover the full cycle of

health care, from diagnosis to prognosis and

treatment, has not been presented before.

The rest of the paper is organized as follows.

Section 2 introduces a real-world scenario with 200

491

Subirats L., Ceccaroni L., María Maroto J., de Pablo C. and Miralles F..

Medical-treatment Recommendation and the Integration of Process Models into Knowledge-based Systems.

DOI: 10.5220/0004822004910498

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 491-498

ISBN: 978-989-758-015-4

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

cases, which illustrates the complexity of the domain

at hand. Section 3 describes the methods and the

architecture of a reasoning system that provides

recommendation about personalized rehabilitation

therapy and prognosis. Section 4 presents the results

of the application of the recommendation system to

the real-world scenario. Finally, conclusions and

future work are drawn in Section 5.

2 SCENARIO

This study is based on the anonymized data of 200

patients who performed cardiac rehabilitation at the

Hospital Universitario Ramón y Cajal that is

summarized in Table 1.

The map of processes

performed in cardiac rehabilitation programs at the

Hospital Universitario Ramón y Cajal is described in

Calvo et al., (2013). We will consider, as

representative examples, the Walking and the

Relaxation sessions activities. Each activity is part

of a hierarchy; the Walking activity, for example, is

a subprocess of Motor rehabilitation, which is a

subprocess of Cardiac rehabilitation.

Let us consider a 50-year-old man from Spain

(the patient), who suffered an acute myocardial

infarction. (This refers to the condition when blood

supply to the heart is stopped. When the heart does

not get enough oxygen, heart muscles die or get

permanently damaged.) Now, two weeks after his

hospital discharge, he starts a cardiac rehabilitation

program. His initial evaluation shows he has some

risks factors, such as smoking and hypertension, and

some problems with his functional capacity,

measured in metabolic equivalent of task (MET).

His left ventricular ejection force is normal, and he

does not have problems with anxiety, depression,

dyslipidemia, sedentary lifestyle, diabetes, alcohol,

abdominal perimeter or body mass index. His

therapist wants to plan a personalized rehabilitation

process for him. The aim of the clinical decision

support system (CDSS), apart from providing him a

summary of his main indicators, is to provide her:

 Personalization of the Treatment. It

provides a recommendation about the

personalization of a treatment, such as the

intensity of the Walking activity, if the

Relaxation activity should be performed.

 Prognosis. The state of risk factors such as

hypertension and other indicators after

rehabilitation treatment can be predicted.

Therefore, his therapist, based on her knowledge

and experience, decides to prescribe the patient a

Cardiac rehabilitation process including the

Walking activity and the Relaxation activity. His

therapist documents her decision in the rehabilitation

platform which the CDSS is integrated to. The next

week, our patient starts the first rehabilitation

session of the Walking activity. The aim of the

CDSS in this stage of the rehabilitation process is to

personalize the session providing the therapist

support about when to stop it. Figures that

summarize the map of processes performed in the

cardiac rehabilitation scenario, the methodology and

the main structure of the ontology can be found at

https://code.google.com/p/cardiacrehabilitation/wiki/

Summary.

3 METHODOLOGY

The proposed rule-based framework is composed of

five components.

(1) An ontology, written in the ontology web

language (OWL) with rules implemented in the

semantic web rule language (SWRL). SWRL was

chosen because it is a de facto standard to extend

OWL and can be easily integrated in Protégé (the

editor used to define the ontology). Rules are

extracted from several sources with different levels

of interpretation: literature (requiring more

interpretation), business processes and decision trees

(requiring less interpretation):

Business Processes are modeled using business

process modeling notation 2.0 (BPMN). BPMN was

chosen for being a standard for business process

modeling. These rules are used to personalize a

session.

Literature rules can be extracted from Velasco

et al. (2009). These rules are used to provide

prognosis and to personalize a rehabilitation

treatment.

Decision Trees are obtained using Weka, a data

mining and machine learning software. Weka was

chosen for being a Java-based open source tool. The

J48 classifier is applied to the Hospital Universitario

Ramón y Cajal data set to obtain personalization of

psychotherapy and psychotropic therapy and predict

risk of hypertension, emotional functions and

exercise tolerance functions. Due to the limited

number of patients, only solutions which a certain

level of recall and precision set by the professional

are taken into account. These rules are used to

provide prognosis and to personalize a rehabilitation

treatment.

(2) A reasoner, Pellet, which infers properties

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

492

and relationships from rules. Pellet was chosen

because it is an open-source Java OWL DL reasoner

with SWRL-support.

(3) A querying system, SPARQL Protocol and

RDF Query Language (SPARQL), to perform the

queries.

(4) A Java framework, Jena, to link with data

applications.

(5) A Java-based content management system

(CMS), Liferay, to provide an interface to interact

with users. The CMS includes the medical health

record (MHR) were data from patients is stored.

SPARQL, Jena and Liferay were chosen because

they are open-source Java-based systems.

This section includes a description of the

ontology’s structure and examples of rules and

queries.

3.1 Ontology

Data obtained from the MHR of the hospital are

expressed as virtual medical record (vMR) classes.

The vMR is an information model developed within

HL7, which is designed to solve interoperability

problems in the electronic exchange of clinical

information and to reduce development costs and

time responses in CDSSs that use it. The vMR is

used by the CDSS and stored in an ontology.

Evaluated Person (subclass of Person) contains

patient’s static information such as birth time and

sex. Problem is used to represent diseases, following

the ICD-10, such as myocardial infarction, chronic

rheumatic heart diseases, heart failure,

atherosclerosis, other peripheral vascular diseases

and congenital malformations of the circulatory

system. Observation Base (subclass of Observation)

is coded following the ICF and is used to represent

indicators, such as risk factors (e.g., tobacco,

hypertension, alcoholism, body mass index,

sedentarism, abdominal perimeter, arrhythmia,

creatinine, glucose), range of emotion, walking

activity, functional capacity, left ventricular ejection

fraction, state-trait anxiety inventory, Beck

depression inventory, and work situation. Data are

available in three instants of time: before disease,

before rehabilitation, and after rehabilitation. Goal is

used to represent the indicators whose value we

want to predict. Finally, General Knowledge is used

to represent the recommended therapeutic plan and

includes: (1) medical and surgical procedures, such

as surgical revascularization, valve prosthesis,

corrected congenital pathology, defibrillator

implantation, pacemaker implantation or

arrhythmias correction, (2) processes, such as

intense aerobic exercise, motor rehabilitation,

psychological rehabilitation, and (3) specific

activities, such as walking or psychotherapy.

These standard annotations of processes and result

indicators from biomedical ontologies and

terminologies (ICF, SNOMED CT, ICD-10 and

ICD-11) are used in a semantic, rule-based

framework.

3.2 Rules

Rules are extracted from literature, a fragment of

BPMN 2.0 and decision trees from classification

algorithms; needing different degrees of human

interpretation during the formalization process. In

Table 1: Summary of some attributes of the cardiac data set (SD = standard deviation).

Attribute: options (separated by /)

Values

# Missing

values

Gender: man / woman 163 / 37 0

Age [11, 78], Mean: 53, SD: 14.3 0

Left ventricular ejection fraction (LVEF) [0, 89], Mean: 35, SD: 29.4 35

METs before rehabilitation [2, 12], Mean: 7, SD: 2.3 57

Beck index before rehabilitation [0, 34], Mean: 8, SD: 6.2 35

Psychotherapy: yes / no 15 / 157 15

Psychotropic therapy: yes / no 6 / 166 28

Exercise tolerance after rehabilitation: no / mild / moderate

/ severe / complete deficiency

19 / 58 / 51 / 20 / 4 48

Emotional functions after rehabilitation: no / mild /

moderate / severe / complete deficiency

93 / 48 / 18 / 4 / 1 36

Hypertension risk after rehabilitation: low / medium / high 94 / 39 / 43 24

Medical-treatmentRecommendationandtheIntegrationofProcessModelsintoKnowledge-basedSystems

493

this section, rules to execute therapies, define

activities, execute activities and evaluate patients are

specified using SWRL. All rules are available at

[http://code.google.com/p/cardiacrehabilitation]; in

the following paragraphs a representative set of

examples is described in details.

Rules from Rehabilitation Processes Written

in BPMN 2.0. In Fig. 1 it is modeled the Walking

activity as an example. The tasks carried out (by

different actors) in the Walking activity are: routine

personalization; time personalization; pulsometer

placement and heart rate measurement; warm up;

walk; cardiac frequency observation; cool down;

basal heart rate measurement; patient’s activity

evaluation. The core tasks of the Walking activity

consist of the three typical phases of a walking

physical exercise in an open environment outside of

a healthcare institution: Warm up, Walk and Cool

down, each of which has a predefined duration,

which can be respected or not by the patient. In the

Walking activity, three indicators are monitored and

their values are stored in a MHR. (1) Treatment-

result indicators, which are periodically quantified.

METs are extracted from an exercise testing is a

treatment-result indicator. (2) Process indicators are

execution, security or end-session. In this activity

heart rate is a process indicator. (3) Session

indicators to indicate the end of the session. In the

walking activity maximum heart rate, Borg score,

walked distance and walked time are session

indicators. As described in Section 1, decision

support is provided to the Walking activity in

prognosis and session’s personalization. The

walking activity ends when one of the following

conditions is reached in the patient’s evaluation if a

certain level of maximal oxygen consumption (VO

max); the maximum heart rate below a predefined

limit and the walked distance above a predefined

limit.

The Relaxation activity is included in

Psychological rehabilitation, which is a subprocess

of Cardiac rehabilitation. It is modeled in a similar

way as the Walking activity in Fig. 1. The tasks

performed in the Relaxation activity are: place the

patient, relaxation tasks and patient assessment. The

core tasks of the Relaxation activity consist of

inhaling, exhaling, muscle contraction and muscle

extensions. In the Relaxation activity, several

indicators are monitored and their values are stored

in a MHR. Treatment result-indicators are

occupational status and Beck depression inventory

(BDI); process indicators are depth of inhalation,

depth of exhalation and muscle activity; and activity

result-indicators are therapist visualization,

measuring tape, Borg scale and anxiety assessment.

As an example, let us consider rule (1), below,

which ends an activity if a process indicator is

greater than or equal to 2 (moderate deficiency).

ActivityEnd (?ae),

ProcessIndicator (?pi),

Patient (?p),

greaterThanOrEqual (?pi, 2),

hasIndicator (?p, ?pi)



hasNextTask (?p, ?ae) (1)

This codification means that, if a patient p has a

process indicator pi which is greater than or equal to

2, the activity is stopped and the next task is ae.

Similar rules are applied to several process

indicators of rehabilitation activities. Process

indicators are used to stop activities because of

severe alterations in body functions (such as a too

high heart rate) or in environmental factors (such as

a too high temperature or humidity).

Rules from Cardiac Rehabilitation Literature.

The risk of having another infarction after cardiac

rehabilitation is translated to rules using semantic

annotations. Problems appear when there are partial

matches in rules. To solve them, priorities among

attributes are established according to the proximity

to the root of the generated decision tree after

applying a classifier. These priorities are: disease,

tolerance to exercise, contraction force of ventricular

muscles and heart rhythm. A sample rule is (2),

which predicts that a person will have medium risk

of suffering a myocardial infarction, if a patient has

thallium stress test abnormal, a severe deficiency in

the treatment indicator of contraction force of

ventricular muscles and has suffered an angina

pectoris.

Patient (?p),

Thallium stress test abnormal (?th),

hasIndicator (?p, ?th),

Contraction force of ventricular muscles (?cfvm),

Deficiency (?severe),

hasDeficiency (?cfvm,?severe),

TreatmentIndicator (?ti),

hasType (?cfvm,?ti),

hasIndicator (?p, ?cfvm),

Angina Pectoris (?ap),

hasDisease (?p, ?ap),

Risk (?mediumRisk)



hasRisk (?p, ?mediumRisk) (2)

Rules from Decision Trees. The J48 classifier

from Weka is applied on data sets about

psychotherapy, psychotropic therapy, risk of

hypertension and prognosis of emotional functions

and exercise tolerance functions; and rules are

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

494

obtained from the resulting decision trees. An

example is the rule (4), which predicts that if a

patient does not suffer from arrhythmia, has as

treatment indicators a severe deficiency in emotional

functions and dyslipidemia less than or equal to 3

(severe deficiency) and follows a psychotropic

therapy; the prognosis of the treatment indicator of

emotional functions after cardiac rehabilitation is no

deficiency.

Patient (?p), Arrythmia (?a), hasnotDisease (?p,

?a), Dislipidemia (?di), TreatmentIndicator (?ti),

hasType (?di,?ti), lessThanOrEqual (3, ?di),

hasIndicator (?p, ?di), hasTherapy (?p,

?psychotropic), EmotionalFunctions (?ef),

Deficiency (?severe), hasDeficiency (?ef, ?severe),

hasType (?ef, ?ti), EmotionalFunctions (?efa),

Deficiency (?no), hasDeficiency (?efa, ?no),

CardiacRehabilitation (?cr), after (?efa, ?cr),

hasType (?efa,?ti)



hasIndicator (?p, ?ef) (4)

This rule is part of the implementation of the

decision tree in Fig. 2.

3.3 Querying

The proposed framework allows querying using

SPARQL. The following example shows how the

definition of the Walking activity is used by the

therapeutic plan. First, the system evaluates which

session indicators the patient should improve

through the following query:

Count(?i) where Patient (?p),

SessionIndicator (?i), Deficiency (?mild),

hasDeficiency (?i, ?mild),

has Indicator (?p, ?i) (5)

Then, activities which cover this objective are

shown. If Apollo wants to improve his exercise

tolerance functions, the query would be:

Select (?a) where Activity (?a),

Exercise Tolerance Functions (?i),

has Objective (?a, ?i) (6)

Finally, it is checked if the activity is not

contraindicated to the patient. In the Walking

activity, the corresponding query would be:

Count (?contraindication) where Patient (?p),

Moderate, severe or complete neuromusculoskeletal

and movement-related functions (?mscn),

has Contraindication (?p, ?mscn) (7)

4 RESULTS

Results are based on the application of the cardiac

rehabilitation scenario described in Section 1. Fig. 3

shows the interface of the CDSS provided to the

therapist. The interface shows an initial diagnosis

which includes causes of the functional limitation,

past procedures and functioning indicators.

Functional diversity levels of indicators are

represented as red/4 (complete), orange/3 (severe),

yellow/2 (moderate), green/1 (mild) and blue/0 (no

functional diversity) (Subirats et al., 2013). The

Figure 1: Cardiac rehabilitation processes.

Medical-treatmentRecommendationandtheIntegrationofProcessModelsintoKnowledge-basedSystems

495

Figure 2: Decision tree of emotional functions prognosis (br = before rehabilitation; ar = after rehabilitation); in brackets,

total/misclassified instances reaching the leaf.

Figure 3: CDSS's interface of patient's prognosis and personalization.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

496

Table 2: Evaluation of the CDSS.

Decision

support

Treatment or treatment indicator:

possible solutions (separated by /)

Accuracy Precision Recall

Personalization

of the treatment

Psychotherapy: yes / no 0.8 0.2 / 0.9 0.6 / 0.8

Psychotropic therapy: yes / no 0.9 0.2 / 0.9 0.3 / 0.9

Prognosis

Exercise tolerance

functions after

rehabilitation: no / mild / moderate /

severe / complete deficiency

0.5

0.2 / 0.6 / 0.5 / 0 /

0.1 / 0.8 / 0.6 / 0 / 0

Emotional functions: no / mild /

moderate / severe / complete deficiency

0.6

0.7 / 0.5 / 0.3 / 0 /

0.8 / 0.4 / 0.2 / 0.1 /

Hypertension risk after

rehabilitation: low / medium / high

0.7 0.8 / 0.8 / 0.6 1.0 / 0.4 / 0.8

Infarction risk: low / medium / high 0.9 0.9 / 0.9 /1.0 0.9 / 0.9 / 1.0

prognosis of the treatment indicators range of

emotion and exercise tolerance functions are

provided, and in the prognosis section it appears the

recommended therapy (that in this case is intense or

moderate aerobic exercise and hypertension risk

factor after rehabilitation).

Rules are evaluated through the correct

execution of the cardiac rehabilitation process in

order to verify completeness of the rule set.

However, the knowledge of the system can

increase/change over time due to (1) more patients

(rules from decision trees can change), (2) an update

of processes; and (3) an update of literature. As all

rules in this model are static, in the future, a Java

library which generates decision trees using Weka,

converts them automatically to SWRL rules, and

integrates them in the ontology will be created.

Table 2 shows the evaluation of the

personalization of therapies, prognosis and risk. In

order to evaluate the performance of prediction of

treatment indicators’ behavior, accuracy, recall (or

sensitivity) and precision (or positive predictive

value) are considered and defined as follows:

TNTP

accuracy







precision





recall





where TP are true positives, FN are false

negatives, FP are false positives and TN are true

negatives. Considering that low values of recall can

cause potential harm in the clinical domain,

hypothesis which higher values of recall are

generally considered better, while maintaining

precision above a specific threshold set by an expert

(Huang et al., 2006). In the personalization of the

treatment, the CDSS has recall values of 0.8 and 0.9

when recommending not to prescribe psychotherapy

and not to prescribe psychotropic therapy,

respectively. In the prognosis of exercise tolerance

functions after rehabilitation, recall values for mild

and moderate deficiency are 0.8 and 0.6,

respectively. In the prognosis of emotional

functions, recall values for mild and no deficiency

are 0.4 and 0.8, respectively. In the prognosis of

hypertension risk after rehabilitation, the recall

value for high risk is 1.0. In the prognosis of

infarction risk, recall values for low, medium and

high risk are 0.9, 0.9 and 1.0, respectively; however,

in this case, the CDSS was tested only on 19

patients. Some very low recall and precision values

are due to classes with a very low number of cases

(see Table 1). This is the case, for example, of

severe and complete deficiency in emotional

functions.

Paths of the decision trees which correspond to

predictions with a recall less than 0.8 or a precision

less than 0.6 are not encoded as rules. In these cases,

consequently, no support is provided to therapists.

Paths of the decision trees which are encoded as

rules correspond to bold values in the Treatment or

treatment indicator: possible solutions column of

Table 2.

In Van et al. (2011), a study that predicts risk

factors with decision trees with data from 3931

patients, the accuracy of the prediction of emotional

functions (which is the ICF encoding of the Hospital

Anxiety and Depression Score) is 0.5, while the

accuracy of the prediction of hypertension is 0.5.

Although the accuracy of the prediction of emotional

functions and hypertension is a little bit higher in the

proposed approach than in the study of Van et al.

(2011), this approach is not able to predict minor

classes. On the other hand, the study of Chang et al.

(2011), predicts the risk factors of hypertension and

hyperlipidemia with a combined accuracy of 0.9.

5 CONCLUSIONS

This paper introduced a framework, based on rules

from three sources (business processes, literature

Medical-treatmentRecommendationandtheIntegrationofProcessModelsintoKnowledge-basedSystems

497

and decision trees) and including biomedical

annotations, to provide formal semantics to cardiac

rehabilitation processes represented in BPMN 2.0.

This semantics, together with rules obtained from

literature and data mining, can be used to introduce

automatic reasoning in decision support. The rule-

based framework uses existing ontologies and

formal process notation to enhance interoperability

and reasoning in rehabilitation and, specifically, to

allow working with different types of indicators, and

to merge medical ontologies and terminologies such

as ICF, SNOMED CT, ICD-10 and ICD-11.

Different issues appeared when implementing

rules obtained from sources needing different

degrees of human interpretation. Difficulties due to

the high number of rules coming from BPMN 2.0

processes were solved generalizing them. And only

rules from decision trees which have recall and

precision above a threshold of 0.8 and 0.6,

respectively, were encoded.

The proposed CDSS, which uses data from 200

patients carrying out cardiac rehabilitation at a

hospital, has high values of recall in the

personalization of therapies, and in the prognosis of

activities and risk factors.

Future work includes making the framework

more dynamic and able to automatically reorganize

itself when new or updated data or processes are

available. In addition, more cases will be analyzed to

cover classes which are less represented in the

knowledge base.

ACKNOWLEDGEMENTS

The research described in this paper is partly

supported by a Spanish research project called

Rehabilita (Disruptive technologies for the

rehabilitation of the future) (CEN-2009-1043). This

work is also partly supported by the Catalonia

Competitiveness Agency (ACC1Ó).

REFERENCES

Calvo, M., Subirats, L., Ceccaroni, L., Maroto J. M., de

Pablo, C., Miralles, F. Automatic assessment of

socioeconomic impact in cardiac rehabilitation. Int. J.

Environ. Res. Public Health, 10(11), 5266-5283, 2013.

C. D. Chang, C. C. Wang and B. C. Jiang. Using DM

techniques for multi-diseases prediction modeling of

hypertension and hyperlipidemia by common risk

factors. Expert Syst. Appl. 38(1):5507-5513, 2011.

A. Cichocki, A. Helal, M. Rusinkiewicz and D. Woelk.

Workflow and Process Automation. Kluwer Academic

Publishers, 1998.

K. Huang, H. Yang, I. King and M. R. Lyu. Maximizing

sensitivity in medical diagnosis using biased minimax

probability machine. IEEE Transactions on

Biomedical Engineering, 53(5):821–831, 2006.

B. Jafarpour, S. Raza Abidi and S. S. Raza Abidi.

Exploiting OWL reasoning services to execute

ontologically-modeled clinical practice guidelines.

AIME 2011, LNCS, 6747: 307-311, 2011.

N. Peek, R. Goud, N. de Keizer, M. van Engen-Verheul,

H. Kemps and A. Hasman. CARDSS: Development

and evaluation of a guideline based decision support

system for cardiac rehabilitation. AIME 2011, LNCS,

6747:109-118, 2011.

F. Smith, M. Missiko and M. Proietti. Ontology-based

querying of composite services. Business System

Manag. and Engineering, 7350:159-180, 2012.

Subirats, L., Ceccaroni, L., Lopez-Blazquez, R., Miralles,

F., García-Rudolph, A. and Tormos, J.M. Circles of

Health: towards an advanced social network about

disabilities of neurological origin. J. Biomed. Inform.,

46(6), 1006-1029, 2013.

A. Van, V. Gay, P. J. Kennedy, E. Barin, P. Leijdekkers.

Understanding Risk Factors in Cardiac Rehabilitation

Patients with Random Forests and Decision Trees. 9

ADM, 121: 11-22, 2011.

Velasco, J. A, Cosín, J., Maroto, J. M., Muñiz, J.,

Casasnovas, J. A., Plaza, I. and Abadal, L. T. GPC en

Prevención Cardiovascular y Rehabilitación Cardiaca

de la Sociedad Esp. de Cardiología. Rev. Esp. Cardiol.

53: 8 1095-1120.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

498