eXiTCDSS: A Framework for a Workﬂow-based CBR

for Interventional CDSS

Application to Transcatheter Aortic Valve Implantation (TAVI)

Andres El-Fakdi, Francisco Gamero, Joaquim Mel

endez and Roberto Petite

Control Engineering and Intelligent Systems Group eXIT, University of Girona, Campus Montilivi Building PIV, 17071

Girona, Spain

Keywords:

Clinical Decision Support Systems, Case-based Reasoning, TAVI, Clinical Workﬂow.

Abstract:

The work presented in this paper presents a workﬂow-based Clinical Decision Support System (CDSS) de-

signed to give case-speciﬁc assessment to clinicians during complex surgery or Minimally Invasive Surgerys

(MISs). Following a perioperative workﬂow, the designed software will use a Case-Based Reasoning (CBR)

methodology to retrieve similar past cases from a case base to give case-speciﬁc advice at any particular point

of the process. The graphical user interface allows easy navigation through the whole support progress, from

the initial conﬁguration steps to the ﬁnal results organized as sets of experiments easily visualized in a user-

friendly way. The tool is presented giving advice to an example application, a complex minimally invasive

surgery which is receiving growing attention lately, the Transcatheter Aortic Valve Implantation (TAVI).

1 INTRODUCTION

Nowadays, medicine and health ﬁelds are getting

more and more involved with computer science.

Among all branches, the main focus of the current re-

search points towards Artiﬁcial Intelligence (AI) to

improve the performance of Clinical Decision Sup-

port Systems (CDSSs). In a general term, CDSSs

comprise a large spectrum of systems which provide

clinicians, staff, patients, and other individuals with

knowledge and person-speciﬁc information, intelli-

gently ﬁltered and presented at appropriate times, to

enhance health and health care (Berner, 2009).

CDSSs are classiﬁed into two main groups,

depending on whether they are knowledge-based

systems, or nonknowledge-based systems (Berner,

2009). The knowledge-based CDSSs are the most

common type of CDSSs used in clinics and hospi-

tals. They are structured around rules mostly in the

form of IF-THEN statements. Most of knowledge-

based systems consist of three parts, the knowledge

base, inference engine, and mechanism to communi-

cate (Wyatt and Spiegelhalter, 1991). The rules are as-

sociated with compiled data extracted from a knowl-

edge base. The inference engine combines the rules

from the knowledge base with the patient’s data. The

communication mechanism will allow the system to

show the results to the user as well as have input

into the system. Nonknowledge-based CDSSs use AI

through machine learning techniques, which allow the

computer to learn from past experiences and to recog-

nize patterns in the clinical data (Marakas, 1999).

Artiﬁcial Neural Network (ANN) (Baxt, 1995) and

genetic algorithms (Laurikkala et al., 1999) are two

common types of nonknowledge-based systems. The

fusion of a knowledge base with nonknowledge-based

machine learning techniques results into an hybrid

system. Hybrid systems extract the best from both

methodologies, ﬁnally resulting into an overall im-

provement of the system performance and thus pro-

viding an optimal solution for clinical decision sup-

port systems (Demmer-Fushman and Lin, 2007). This

paper will focus on this kind of hybrid systems, pre-

senting a software framework for CDSS which uses

a well-known lazy learning technique called Case-

Based Reasoning (CBR) (Aamodt and Plaza, 1994)

and a past patients case base to asses clinicians.

This paper work presents a workﬂow-based CDSS

designed to give case-speciﬁc assessment to clini-

cians during complex surgery. Following a periop-

erative workﬂow, the designed software will use a

CBR methodology to retrieve similar past cases from

a case base to give case-speciﬁc advice at any partic-

ular point of the process. The graphical user inter-

face allows easy navigation through the whole sup-

port progress, from the initial conﬁguration steps to

201

El-Fakdi A., Gamero F., Meléndez J. and Petite R..

eXiTCDSS: A Framework for a Workﬂow-based CBR for Interventional CDSS - Application to Transcatheter Aortic Valve Implantation (TAVI).

DOI: 10.5220/0004233802010206

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2013), pages 201-206

ISBN: 978-989-8565-37-2

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

the ﬁnal results organized as sets of experiments eas-

ily visualized in a user-friendly way. The eXiTCDSS

tool is presented giving advice to an example appli-

cation, assisting a recent complex minimally invasive

surgery which is receiving growing attention lately,

the Transcatheter Aortic Valve Implantation (TAVI).

This paper is structured as follows. In Section 2

a description of the workﬂow management during an

intervention is detailed. Also, recent applications of

CDSS for surgical processes are reviewed. Section 3

introduces the eXiTCDSS framework. In Section 4,

the eXiTCDSS demonstrates its performance with its

application to a TAVI procedure. Finally, conclusions

are included in Section 5.

2 CDSS INTEGRATION WITH

CLINICAL WORKFLOW

Examples of successful applications of CDSSs into

clinical workﬂows comprise computer based patient

record systems (Patel et al., 2000), knowledge man-

agement systems for biomedical engineering (Rinkus

et al., 2004) and computer based training systems in

pathology (Crowley et al., 2003). From the successful

applications mentioned before it can be extracted that

integration with workﬂow is key to success. How to

integrate the CDSS with clinician workﬂow, however,

remains a challenge, in part because there are no cur-

rent standards for clinical workﬂow (Das and Eichner,

2010).

Although there is no universally agreed upon deﬁ-

nition of the term workﬂow, for the purpose of this ar-

ticle, we have taken the workﬂow deﬁnition stated in

(Carayon et al., 2010) which deﬁnes a clinical work-

ﬂow as a modular sequence of tasks, with a distinct

beginning and end, performed for the speciﬁc pur-

pose of delivering clinical care. In order to implement

a workﬂow-based CDSS, tasks, timing and involved

subprocesses must by identiﬁed ﬁrst. Therefore, the

proposed workﬂow has been speciﬁed at up to four

level of detail: 1) clinical workﬂow, 2) phase, 3) task,

and 4) attribute. Figure 1 shows a schematic workﬂow

of an exempliﬁed operative process where the previ-

ously mentioned levels have been illustrated. The ﬁrst

level of the workﬂow represents the particular work-

ﬂow itself. The second level describes the phases,

being a phase the primordial division of the speciﬁc

clinical workﬂow. For the particular example shown

in Figure 1, each phase corresponds to the pre, intra,

and post-operative periods. In the same way, every

phase has been split into tasks, a task being any partic-

ular step taken during each phase e.g. apply anesthe-

sia, initial puncture location or valve ﬁnal placement.

Each task has a different number of distinguishable

items or attributes associated. These attributes refer

to all the important values or considerations that the

medical staff will take into account during the resolu-

tion of a task. The attributes can be described as nu-

merical data, text data, categorical data, and boolean

data. As numerical data it can be considered blood

count, coagulation parameters, age, size, or speciﬁc

physiologic measurements. The text data comprises

those textual items regarding the patient’s patholog-

ical or surgical history as well as possible allergies.

The categorical, in fact ordered categorical data, com-

prises attributes which measure a certain degree of

intensity, e. g. amount of calciﬁcation or valve re-

gurgitation while the boolean data conﬁrms or denies

the presence of an attribute, for example the vascu-

lar tortuosity or the existence of coronary ﬂow dam-

age. During the intervention execution and according

to the current information being generated, the CDSS

has to be capable to identify the phase, the task, and

the attributes involved. Then, the software will use the

CBR engine to retrieve the most similar cases to the

current one. The framework eXiTCDSS presented in

this article provides the required tools to deﬁne a case

structure for any clinical procedure based on a work-

ﬂow.

3 THE eXiTCDSS FRAMEWORK

Case-Based Reasoning (CBR) is a technique of arti-

ﬁcial intelligence that attempts to solve a given prob-

lem within a speciﬁc domain by adapting established

solutions to similar problems (Aamodt and Plaza,

1994). CBR has been formalized for purposes of rea-

soning and learning based on the exploitation of ex-

isting similar historical records as humans do. It has

been argued that CBR is not only a powerful method

for computer reasoning, but also a pervasive behavior

in everyday human problem solving; or, more radi-

cally, that all reasoning is based on past cases person-

ally experienced. These features make CBR a good

contender for any decision support system.

Four main phases of action are deﬁned in the CBR

methodology: retrieve, reuse, revise and retain. For

example, in TAVI, a case base contains information

about patients that have been operated in the past. Us-

ing this case base, a CBR system is able to give advice

to future TAVI cases by following the four phases: re-

trieve, reuse, revise and retain. First, in the retrieve

phase, the current case is compared with all the past

experiences in the case base, and the most similar are

recovered. Given a target problem, during the retrieve

step, cases from memory that are relevant to solving

HEALTHINF2013-InternationalConferenceonHealthInformatics

202

Figure 1: Schematic workﬂow showing an exempliﬁed operative process.

Figure 2: Case structure and retrieval process along the op-

erative workﬂow.

it are recovered. A case is a whole although usu-

ally two parts can be identiﬁed according to the prob-

lem to be solved: problem and solution space. Prob-

lem space can incorporate speciﬁc case information

and also contextual information useful for solving the

problem whereas solution space incorporates infor-

mation directly related with the solution for solving

the problem. Next, in the reuse phase, a solution to

the current case is determined based on the solutions

found in the retrieved cases, which are mapped to tar-

get the actual problem. This may involve adapting the

solution as needed to ﬁt the new situation. Third, the

computed solution is evaluated in the revise phase. Fi-

nally, the retain phase analyzes whether to retain the

case in the case base.

When the intervention begins, parallel to the

workﬂow progression described in Figure 2, the CBR

mechanism retrieves speciﬁc and contextual problem

information of similar past cases. Such assessment

step determines which cases address problems most

similar to the current problem, to identify them as

starting points for solving the new problem. Once the

operation ends, clinicians proceed with the new case

revision, which evaluates and adjusts the adapted so-

lutions and, if suitable, the new case will be retained

with the system learning from the situation by stor-

ing the result as a new case for future use. The soft-

ware has been designed to work either on-line inside

the Operation Room (OR) or off-line, as a knowledge

database for patient study.

As the eXiTCDSS framework goes beyond pure

CBR prototyping and aims to support workﬂow-

oriented decision support, other elements are required

in addition to the basic CBR modules. In order to ma-

nipulate the data, a common representation of cases

is required. Also, three main components are distin-

guished: the workﬂow editor, the CBR engine and

the results navigator. These components are described

next.

3.1 Case Representation

eXiTCDSS requires a plain Coma Separated Value

(CSV) ﬁle to handle the data. Each row corresponds

to a case, and each column to attributes of the cases.

The ﬁrst four rows describe the attributes as follows.

The ﬁrst row corresponds to the attribute descriptions

(for example, ”Annular calciﬁcation”). The second

row corresponds to the attribute name (usually in a

compressed form, as for example, ”Annularcalciﬁca-

tion”). The third row corresponds to the attribute type

(-1 ignore, 0 boolean, 1 numerical, 2 textual, 3 cate-

gorical). The fourth row corresponds to the attribute

weight (relevance). This representation covers most

of the data used in medical applications and is easy to

manage and general enough to be used by any of the

eXiTCDSS:AFrameworkforaWorkflow-basedCBRforInterventionalCDSS-ApplicationtoTranscatheterAorticValve

Implantation(TAVI)

203

current CBR techniques (mainly distance functions).

3.2 The Workﬂow Editor

To support a workﬂow-oriented CDSS software, a

workﬂow editor tool has been developed inside the

main application so that users can deﬁne a workﬂow

ﬁle. All the needed features to describe any type of

workﬂow are displayed in a window independent in-

terface that can be accessed from the tools label in the

top area of the main window frame of the platform.

This functionality includes a shapes menu with all the

types of boxes, containers and arrows used to deﬁne

the workﬂow structure with all the needed phases and

tasks. The created ﬁgures are editable and colorable,

with a label for its identiﬁcation. After the structure

deﬁnition, on the right side of the window, the user

proceeds with the attributes-to-task association. By

loading the CSV ﬁle presented in Section 3.1, the user

can access to all the attributes which deﬁne a patient’s

case and associate them to its correspondent task of

the workﬂow. Once the workﬂow creation is com-

pleted, the workﬂow ﬁle is saved as an Extensible

Markup Language (XML) ﬁle which can be loaded

later on by the eXiTCDSS main application in order

to proceed with the project development.

3.3 The CBR Engine

All the information required to set up a CBR system

according to user requirements is stored in the conﬁg-

uration ﬁle. The CBR engine is responsible for read-

ing this ﬁle, extracting the selected XML and CSV

ﬁles, methods and parameters and, ﬁnally, calling and

executing the related CBR algorithms. Therefore,

once the conﬁguration ﬁle is set, the eXiTCDSS loads

the patient case base and the associated workﬂow ﬁle.

Then, the engine compares the stored cases with the

current patient data, and selects the most similar cases

from the case base. There are local and global simi-

larity measures. Local similarity measures compare

two attribute values. Global similarity measures com-

bine different local similarity outcomes to determine

the similarity between two cases. The application al-

lows the user to track the different tasks of the inter-

vention workﬂow with its associated attributes while

consulting the case base. In addition, the software

offers the possibility to load/save different attributes

pre-selection. These presets allow clinicians to look

for similarities between cases by just working with

a speciﬁc set of attributes or only considering a sin-

gle phase of the workﬂow, for example for carrying

out patient studies during the pre-operative phase. As

output of the retrieve phase, the CBR engine creates

a distance matrix that depicts the similitude between

the new case and the cases in the case base. This ma-

trix is shown in the results navigator window.

3.4 The Results Navigator

The results navigator window contains a table with

the most similar cases to the current case. The re-

sults table contains some features to enhance decision

support. Thus, clinicians can rearrange the resultant

table to see the most and the least similar cases to the

current case. Also, if the medical staff needs addi-

tional information for decision making, every case of

the case base is linked to its complete clinical history

where clinicians can consult any image or ﬁle. Fi-

nally, the results window allows the user to choose the

number of similar cases to the current case to retrieve.

Possibilities include selecting the k-nearest neighbors

or selecting the cases with a similarity degree higher

than a pre-ﬁxed threshold.

One of the most signiﬁcant advances of using a

CBR engine is that the knowledge database is con-

tinuously updating. Thus, new, revised cases will be

added to the case base for future use. eXiTCDSS has

been developed using the Java language. It is compat-

ible with the Linux and Windows OS.

4 eXiTCDSS SUPPORTS

TRANSCATHETER VALVE

IMPLANTATION

In Transcatheter Aortic Valve Implantation (TAVI)

(Webb and Cribier, 2011), a synthetic valve is trans-

ported to the heart through a small hole made in groin.

This procedure can be compared to that performed

when placing a stent, or performing balloon angio-

plasty. This technique was ﬁrst developed in Europe,

where it was initially performed in 2002. Since then,

more than 10000 patients have beneﬁted from it and

the results have shown the procedure to be effective

in improving functioning in the patients with severe

aortic stenosis. In the recent years TAVI is assuming

a major role in the routine management of patients

with aortic stenosis and now TAVI is considered the

standard in patients who are not candidates for con-

ventional surgical Aortic Valve Replacement (AVR).

On the basis of almost 10 years of experience TAVI

also appears to be a reasonable option for some opera-

ble, but high-risk patients. Nevertheless considerable

work needs to be done before TAVI is expanded into

lower risk groups.

Next lines describe an example application of the

HEALTHINF2013-InternationalConferenceonHealthInformatics

204

Figure 3: TAVI workﬂow showing tasks and data source where attributes are extracted from.

Figure 4: The main working window of the eXiTCDSS sup-

port tool.

eXiTCDSS performance. Figure 3 shows the speciﬁc

workﬂow for the TAVI intervention. As can be ap-

preciated, the procedure (from pre-operative to post-

operative treatment) is subdivided in 8 different tasks,

starting with a study of patient suitability for TAVI in

Task 0, to Task 7 where the patient is ﬁnally trans-

ferred to the Intensive Care Unit (ICU). Parallel to

the workﬂow progress, all the tasks are linked with

the different information sources which will provide

the necessary attribute data required for deﬁning the

patient case model. The TAVI workﬂow together with

the CSV ﬁle containing the patient case base are both

loaded into the workﬂow editor in order to produce

the correspondent XML workﬂow ﬁle. Pre-selected

workmods mentioned in Section 3.3 are available at

this point so the user can load or save a desired selec-

tion of whatever attributes and/or operation tasks he

wishes to receive support from.

For demonstration purposes, the application

shows its capabilities while giving support to a one

particular aspect discussed during Tasks 1 and 2 of

a TAVI intervention, which aims to answer a ba-

sic question: Which access? Nowadays, the most

common approaches in TAVI are the transfemoral,

the transapical, and the subclavian way. These ap-

proaches are selected after studying the patient’s pro-

Figure 5: The table shows an ordered list of similar cases to

the query case 0.

ﬁle. The eXiTCDSS support for this demonstration

starts by, from all the attributes list available at Task

1 and 2, selecting those ones considered by clinicians

as the most relevant for deciding the vascular access

(see Figure 4). Next, the case number 0 is selected as

the new query case from a case base of 20 cases. It

is important to remark that, in a real clinical situation,

the vascular access of the query case is not known and

thus, this attribute has not been selected so it remains

unknown for the program. The results offered by the

eXiTCDSS application after the retrieve phase are il-

lustrated in Figure 5. The table depicts the 10 most

similar cases to the query case based on the similarity

distance described in Section 3.3. Also, the vascular

access selected in every case is shown. As can be ap-

preciated, the closest cases to the query case used the

transfemoral approach, the same approach that was

used with the query case. The eXiTCDSS engine re-

trieves similar transfemoral approaches when asked

for a new access way, advising clinicians about which

access way is the best based on similarities with past

patients. Next, experts will be able to study all the

proﬁles of the retrieved cases or launch another re-

trieve search based on some other attributes.

eXiTCDSS:AFrameworkforaWorkflow-basedCBRforInterventionalCDSS-ApplicationtoTranscatheterAorticValve

Implantation(TAVI)

205

5 CONCLUSIONS

This paper presents eXiTCDSS, a workﬂow-based

CDSS designed to give case-speciﬁc assessment to

clinicians during complex surgery or MISs. The

framework facilitates interaction with physicians,

which are guided along the application in a user-

friendly way. Its workﬂow structure offers high ver-

satility allowing the clinicians to decide in which

steps of the procedure they wish to receive sup-

port. The tasks and attributes selection can easily

be saved/loaded into independent ﬁles for future use.

Although it has been designed to give support to a

wide range of interventions, the eXiTCDSS has been

initially applied to give support to TAVI interven-

tions. The tool has demonstrated its performance giv-

ing support to a speciﬁc step of a TAVI procedure with

good results. Current work is focused with improv-

ing the user interface, specially during the OR phase.

Software architecture modiﬁcations will provide the

application with voice activation and gesture recog-

nition tools. Finally, the feedback received from the

numerous physicians consulted has been very positive

and the interest showed in this project is encouraging.

Clinicians agree that CDSS are of high value specially

in recently growing MIS like TAVI where the number

of interventions per year is still low compared to com-

mon surgery and which also require expert hands due

to its complex procedure.

ACKNOWLEDGEMENTS

We would like to give our special thanks to the Hos-

pital Clinic of Barcelona, the Centre Hospitalier Uni-

versitaire de Rennes and the Laboratoire Traitement

du Signal et de l’Image from the University of Rennes

for helping us with the case model deﬁnition and giv-

ing us example cases of TAVI patients. This work

has been ﬁnanced by the Spanish Government Com-

mission Ministerio de Industria, Turismo y Comer-

cio (MITyC) under the project PLAN AVANZA 2 la-

beled by Information Technology for European Ad-

vancement 2 (ITEA2).

REFERENCES

Aamodt, A. and Plaza, E. (1994). Case-based reasoning:

Foundational issues, methodological variations, and

system approaches. Artiﬁcial Intelligence Communi-

cations, 7:39–52.

Baxt, W. (1995). Application of artiﬁcial neural networks

to clinical medicine. Lancet, 346(8983):1135–1138.

Berner, E. (2009). Clinical decision support systems:

State of the art. AHQR publication no. 09-0069-EF,

Agency for Healthcare Research and Quality.

Carayon, P., Karsh, B., Cartmill, C., Hoonakker, P., Hundt,

A., Krueger, D., and Thuemling, T. (2010). Incor-

porating health information technology into workﬂow

redesign. Technical Report 10-0098-EF, Center for

Quality and Productivity Improvement, University of

Wisconsin-Madison.

Crowley, R., Naus, G., and Friedman, C. (2003). Develop-

ment of visual diagnostic expertise in pathology: An

information processing study. Journal of American

Medical Informatics Association, 10(1):39–51.

Das, M. and Eichner, J. (2010). Challenges and barriers

to clinical decision support design and implementa-

tion experienced in the agency for healthcare research

and quality cds demonstrations. AHQR publication

no. 10-0064-EF, Agency for Healthcare Research and

Quality.

Demmer-Fushman, D. and Lin, J. (2007). Answering

clinical question with knowledge-based and statisti-

cal techniques. In Association for Computational Lin-

guistics, pages 63–103.

Laurikkala, J., Juhola, M., Lammi, S., and Viikki, K.

(1999). Comparion of genetic algorithms and other

classiﬁcation methods in the diagnosis of female

urinary incontinence. Methods of Information in

Medicine, 38:125–131.

Marakas, G. (1999). Decision support systems in the 21st

century. Prentice Hall, Upper Saddle River, New Jer-

sey.

Patel, V., Kushniruk, A., Yang, S., and Yale, J. (2000). Im-

pact of a computer-based patient record system on

data collection,knowledge organisation, and reason-

ing. Journal of American Medical Informatics Asso-

ciation, 7(6):569–85.

Rinkus, S., Walji, M., Johnson-Throop, K., Malin, J., Tur-

ley, J., and Smith, J. (2004). Human-centerd design of

a distributed knowledge management system. Journal

of Biomedical Informatics, 38:4–17.

Webb, J. and Cribier, A. (2011). Percutaneous transarterial

aortic valve implantation: what do we know? Euro-

pean Heart Journal, 32:140–147.

Wyatt, J. and Spiegelhalter, D. (1991). Field trials of medi-

cal decision-aids: potential problems and solutions. In

15th Symposium on Computer Applications in Medi-

cal Care, pages 3–7, Washington, USA.

HEALTHINF2013-InternationalConferenceonHealthInformatics

206