See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221754562

Design of a Fuzzy-based Decision Support System for Coronary Heart Disease

Diagnosis

Article  in  Journal of Medical Systems · January 2012

DOI: 10.1007/s10916-012-9821-7 · Source: PubMed

CITATIONS READS

79 4,039

4 authors, including:

Adel Lahsasna Raja Noor Ainon

University of Malaya University of Malaya
15 PUBLICATIONS   289 CITATIONS    32 PUBLICATIONS   362 CITATIONS   

SEE PROFILE SEE PROFILE

Awang Bulgiba
University of Malaya
205 PUBLICATIONS   3,455 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The association between obesity and dengue severity among pediatric patients: A systematic review and meta-analysis. View project

Primary Care Case mix View project

All content following this page was uploaded by Awang Bulgiba on 31 December 2013.

The user has requested enhancement of the downloaded file.

J Med Syst
DOI 10.1007/s10916-012-9821-7

ORIGINAL PAPER

Design of a Fuzzy-based Decision Support System

for Coronary Heart Disease Diagnosis
Adel Lahsasna · Raja Noor Ainon ·
Roziati Zainuddin · Awang Bulgiba

Received: 26 October 2011 / Accepted: 3 January 2012

© Springer Science+Business Media, LLC 2012

Abstract In the present paper, a fuzzy rule-based sys- FRBS is able to identify the uncertainty cases so that
tem (FRBS) is designed to serve as a decision support the physician can give a special consideration to deal
system for Coronary heart disease (CHD) diagnosis with them and this will result in a better management
that not only considers the decision accuracy of the of efforts and tasks. Furthermore, employing ECS has
rules but also their transparency at the same time. To specifically improved the ability of FRBS to detect
achieve the two above mentioned objectives, we apply patients with CHD which is desirable feature for any
a multi-objective genetic algorithm to optimize both the CHD diagnosis system.
accuracy and transparency of the FRBS. In addition
and to help assess the certainty and the importance Keywords Coronary heart disease ·
of each rule by the physician, an extended format of Fuzzy rule-based system · Transparency ·
fuzzy rules that incorporates the degree of decision Data mining · Medical diagnosis
certainty and importance or support of each rule at
the consequent part of the rules is introduced. Further-
more, a new way for employing Ensemble Classifiers Introduction
Strategy (ECS) method is proposed to enhance the
classification ability of the FRBS. The results show that Background
the generated rules are humanly understandable while
their accuracy compared favorably with other bench- Coronary heart disease (CHD) refers to the narrowing
mark classification methods. In addition, the produced of arteries that supply blood and oxygen to the heart,
causing serious cardiovascular complications including
myocardial infarctions (heart attack) and angina [1].
It is the most common type of heart disease and the
A. Lahsasna (B) · R. N. Ainon · R. Zainuddin leading cause of death worldwide [2, 3]. Early detection
Faculty of Computer Science and Information of CHD in patients, however, can result in significant
of Technology, University of Malaya, 50603
life saving [4, 5] when followed by adequate medical
Kuala Lumpur, Malaysia
e-mail: lahsasna@siswa.um.edu.my treatments.
In fact, CHD diagnosis is a quite challenging task,
R. N. Ainon
e-mail: ainon@um.edu.my since the physician needs to carefully examine a com-
bination of symptoms and signs that may overlap with
R. Zainuddin
e-mail: roziati@um.edu.my other causes [5]. In addition, the information available
about the patient is inherently uncertain [6]. In order
A. Bulgiba to help deal with this problem in a more efficient
Julius Centre University of Malaya and CRYSTAL,
Faculty of Science, University of Malaya, 50603
way, many approaches have been proposed to assist
Kuala Lumpur, Malaysia the physician’s decision making about whether CHD is
e-mail: awang@um.edu.my likely to exist or not.
 J Med Syst

Invasive coronary angiography is the gold-standard Transparency or interpretability which refers to the
clinical method used for diagnosing CHD. Its use, ability of a system to express its behavior in an under-
however, is restricted to a limited population due to standable way, has recently gained more attention and
the high cost and high-level technical expertise and it is considered as an important criterion for medical
technological required to perform this procedure [7]. diagnosis systems [32–34]. One of the advantages that a
In addition, a substantial number of patients referred transparent diagnosis system may offer is its capability
to this modality appear free from CHD [8]. Conse- to give a response on how a particular decision of
quently, attention has shifted to the use and develop- diagnosis was made. This feature is of great importance
ment of relative cost-effective and reliable non-invasive as it may allow the physician to: (1) either consider or
methods so that all the patients with suspected CHD reject the decision made by the diagnosis system (2)
would be able to afford the medical test and those identify CHD risk factors which helps in its diagnosis
who have an active CHD can be accurately identified and treatment in early and reversible stages [35].
at early stage [9]. The noninvasive methods include Unlike ANNs and SVMs, Fuzzy rule-based systems
the widely used exercise electrocardiogram (ECG) test offer a convenient format for representing the knowl-
[10] and a number of noninvasive imaging modalities. edge underlying a system in the form of transparent and
These modalities include electron-beam computerized linguistic conditional statements. These statements are
tomography (EBCT) [11], single-photon emission com- in the form of “if condition(s) then action(s)” [36]. Such
puted tomography (SPECT) [12], positron emission kind of format is humanly understandable as it uses a
tomography (PET) [13], (contrast) stress echocardio- language close to the natural language which makes it
graphy [14], magnetic resonance angiography(CMRA) a suitable tool for interpretation and analysis. In addi-
[15] and most recently Multislice computed tomograph tion, FRBSs use a fuzzy logic-based reasoning scheme
(MSCT) [16]. to draw conclusions that simulates in some respects the
The data obtained from the above mentioned modal- human thinking mechanism [37]. The aforementioned
ities and other sources such as clinical and historical advantages make FRBSs a suitable method for devel-
data are employed by a computer based decision sup- oping diagnosis system for CHD.
port system (DSS) to classify the patients as either hav- Typically, fuzzy rules are generated using two meth-
ing or not having evidence of CHD. Several techniques ods: expert-driven approach and data-driven approach.
were proposed to build these DSSs that can help the In the first approach, the values of fuzzy rule para-
physician in the decision making process by providing a meters such as interval boundaries and membership
rapid and accurate diagnosis of CHD. functions are defined and set manually by the expert,
while these values are defined automatically from a set
of representative examples using a learning method in
Related Works the other approach [37].
Recently, data-driven rule generation methods have
Soft computing techniques which indicate a number dominated the development of fuzzy-based CHD di-
of methodologies used to find approximate solutions agnosis systems partly due to the cost and difficulty
are successfully applied in many real-world problems of manually setting the rules, and partly due to the
including medical diagnosis problem [17]. Their use availability of historical patient data and the recent
in developing disease diagnosis systems are motivated development of efficient machine learning algorithms.
by their ability to handle uncertain information upon [38–42]. Their only objective is however, getting higher
which medical diagnosis is usually based [6]. diagnostic accuracy, disregarding the transparency is-
Artificial neural networks (ANNs) [18–28] and re- sue. Hence, fuzzy approach adopted by these systems
cently SVMs [28–31] are the most commonly used soft is a kind of black-box system that is used only for
computing methods in developing CHD diagnosis sys- calculating rather than inferencing the diagnosis label.
tem. Overall, the performances obtained by these sys- The design of fuzzy rule-based diagnosis system has to
tems appear promising as they achieved relatively good consider both the accuracy and the transparency.
classification accuracy, and thus can serve as reliable While the transparency feature is maintained in
decision support system for CHD detection. expert-driven approach, it is usually lost during
Despite the advantage of being highly accurate, the learning process, and to preserve it, many ap-
ANNs and SVMs have been criticized due to their proaches have been proposed [43–47]. Genetic-based
lack of transparency as they are black box systems, i.e approaches are among the most successful ones. Basi-
the user is prevented from knowing about the decision cally, they try to find a fuzzy system that has the re-
process of their inner systems [32]. quired balance between the accuracy and transparency
J Med Syst

using the powerful search of genetic algorithm. This all the possible fuzzy rules in the second step, multi-
approach is, however, computationally costly when it objective genetic algorithms are utilized to obtain a
applies for our problem as the number of inputs is subset of a small number of rules with the highest
relatively high. Thus, it is desirable to apply a feature classification ability in the third step. The result of
selection method to select only the most relevant fea- this step is a set of fuzzy rules that incorporate at the
tures. This preprocessing step is used to reduce the consequent part in addition to their class labels, their
complexity and the time needed for the optimization degree of certainty and confidence.
process. To enhance further the accuracy capabilities of the
In addition, the fact that some approaches gener- produced fuzzy rules, we introduce in the fourth step a
ate highly transparent fuzzy rules with clear linguistic new method where the decision or classification made
meaning such in [46, 47], their consequent parts do by a rule which its degree of confidence is less than
not include enough information about the rules. While threshold value is supported by an ensemble of selected
most of these approaches include only the certainty classifiers in order to get a more reliable decision. This
(or confidence) degree which is important for assessing ensemble classifiers strategy (ECS) is inspired by com-
the classification ability of the rules, it is important to mon medical practice where more than one expert’s
incorporate also the degree of support of rules which opinion should be considered in complicated cases
indicates how often these rules occur in a data set. By that have low diagnostic certainty. Finally, the results
giving these two information measures about each rule, achieved by our FRBS are evaluated and compared
the decision maker can effectively assess the degree of with other benchmark methods.
importance and certainty of each rule. The rest of the paper is organized as follows.
Section “Data Set” describes the data set used in this
Objectives study while Section “Methodology” details the steps
and features of the proposed methodology. The re-
In this study, a transparent and a relatively accurate sults obtained and their discussions are given in
fuzzy rule-based system is designed for CHD diagnosis. Section “Results and Discussions” while conclusion is
The concept of transparency is extended to include not drawn in Section “Results and Discussions”.
only the traditional measures such as the number of
rules and fuzzy sets, but also to include the degree
of certainty and the support of each rule. The devel- Data Set
opment of our fuzzy rule-based system involves four
main steps. In the first step which is a preprocessing The data set used in this study was obtained from UCI
step aiming at reducing the complexity of the data Repository of Machine Learning Databases. It was col-
set, a feature selection method is applied to select the lected from Cleveland Clinic Foundation, Cleveland,
most discriminative subset of features. After generating Ohio and supplied by Robert Detrano, M.D., Ph.D. of

Table 1 Description of Cleveland heart disease data set

Attributes Description Type Value
Age Age Integer [29 77]
Sex Sex Integer 1 = male; 0 = female
Pc Chest pain type Integer 1 = typical angina; 2 = atypical angina;
3 = non-anginal pain; 4 = asymptomatic
Trestbps Resting blood pressure (in mm Hg on Integer [94,200]
admission to the hospital)
Chol Serum cholestoral in (mg/dl) Integer [126,564]
Fbs Fasting blood sugar (>120 mg/dl) Integer 1 = true; 0 = false
Restecg Resting electrocardiographic results (values 0, 1, 2) Integer [0, 2]
Thalach Maximum heart rate achieved Integer [71,202]
Exang Exercise induced angina Integer 1 = 4 yes; 0 = no
Oldpeak ST depression induced by exercise relative to rest Real [0.00, 62.00]
Slope The slope of the peak exercise ST segment Integer 1 = upsloping; 2 = flat; 3 = downsloping
 of MajorVessels Number of major vessels (0–3) colored by flourosopy Integer [0, 3]
Thal Thal Integer 3 = normal; 6 = fixed defect;
7 = reversable defect
CHD Coronary heart disease diagnosis Integer 0 = absent; 1 = present
 J Med Syst

the V.A. Medical Center, Long Beach, CA [48]. Orig- (3) evaluate the feature sets of size 3 that includes the
inally, the data had 76 attributes, out of which only 13 features from 2 and select the best one;
attributes were selected for use. The selected attributes (4) and so on until the stopping criteria is reached.
represent the clinical and noninvasive test results of 303
patients undergoing angiography. Removing the cases The number of selected features can be predefined
containing missing values, 270 cases were considered in by the user or can be automatically defined by the
study, out of which 120 cases are identified as patients algorithm. SFFS method improves SFS by dynamically
with CHD while 150 cases are diagnosed as patients changing the number of features included or eliminated
without CHD. at each step according to a certain criterion. In addi-
Table 1 gives a brief description of Cleveland data tion, this method allows the reconsideration of features
attributes and their respective values. included or removed at the previous steps [49].

Fuzzy Rule-based System Generation

Methodology Extending the Format of Fuzzy Rules

The proposed methodology composed of four main In medical diagnosis problem, the physician, based on
steps: the symptoms exhibited by the patients, classifies the
patient’s disease to one of the two classes, namely,
Feature Selection patient with CHD and patient without CHD (binary
classification). Thus, medical diagnosis problem falls
Feature selection is a preprocessing step aims at reduc- under the classification problems.
ing the number of features by selecting the subset of Formally, diagnosis problem can be solved by finding
features that can produce the best model according to a suitable classifier or mathematical function f that
a certain criterion. The literature shows that the most maps a set of symptoms X to a diagnosis class label c j.
effective feature selection methods are the Sequential This function can be written as follow:
Floating Forward Selection (SFFS) [49] and the genetic
f : x → cj
algorithms (GAs) [50]. While many studies [50–53] sug-
gest the superiority of SFFS over the sequential search where x = {x1 , x2 , ..., xn } is a set of symptoms, and c j
algorithms , there is no clear cut case study of which of where j ∈ {1, 2} is the class label. The classifier can
the two methods—SFFS and GAs—is better than the be ANNs or decision tree or any other classification
other [50]. In our study, SFFS is selected as it is much method.
faster than the GAs method. In fuzzy set theory, the classifier f is a set of fuzzy
Since we are dealing with classification problem, the rules where the kth rule has the following format:
criterion used for selection is the classification accu-
racy. For each candidate feature subset, SFFS performs Rk : if x1 is Ak1 and x2 is Ak2 and ...
10-fold cross-validation method and the subset that
and xn is Akn then Y is c j (1)
achieves the least misclassification error (or the best
classification accuracy) is selected. In the following is where Akj are fuzzy sets of the input variables
a brief description of the SFFS algorithm applied at (x1 , x2 , xn ) represented by linguistic values such as low,
this step. moderate and high while c j is the class label of the
class variable Y (which in this case is either Y = c1 for
Sequential Floating Forward Selection (SFFS) The present class or Y = c2 for absent class).
standard Sequential Forward Search (SFS) starts from Some studies (see for example [47]) used an ex-
an empty feature set and creates candidate feature tended format of (1) that includes the degree of cer-
subsets using the following procedure: tainty or confidence rk of the kth rule where rk ∈ [0 1].
The value rk represents also the certainty’s degree of
(1) evaluate the feature sets of size 1 and select the the decision made by this rule. This feature allows the
single best feature; physician to know about the degree of confidence of the
(2) evaluate the feature sets of size 2 that includes the decision made by the fuzzy classifier or more precisely
feature from 1 and select the best one; by the winner rule and whether to consider it or not.
J Med Syst

This format of fuzzy rule-based system can be written antecedents of fuzzy rules, the consequent class of the
as follows: rule Rk , its degree of certainty [54], and support can
calculated as follows:
Rk : if x1 is Ak1 and x2 is Ak2 and ...
and xn is Akn then Y is c j with certainty rk (2) (a) Calculate the consequent class

where rk is the degree of certainty of the kth rule. (1) For each training pattern xt = (xt1 , xt2 , ...,
Although (2) gives more information about the kth xtn ), its compatibility grade μk (xt ) with the
rule than (1), it does not allow the decision maker rule Rk is calculated as
to know the degree of importance of the rule and μ Rk (xt ) = μk1 (xt1 ) ∗ μk2 (xt2 ) ∗ ... ∗ μkn (xtn ),
precisely how many times Rk rule holds true in the
training data. This concept is known in data mining field t = 1, ..., m (5)
as support and it is important criterion to assess the where μki (.) is the membership function of
rules. Using the degree of certainty alone might cause the fuzzy set Akj .
a mislead in the decision making, for example assume
that the degree of certainty of the Rk is rk = 1 (100% (2) for each class, calculate the sum of compati-
classification accuracy) but it covers only very few cases bility grades of the training patterns with the
in the training data. As a result, Rk has a low support fuzzy rule Rk as:
and cannot be used alone to generalize relations or 
draw conclusions. In addition, this value of support βClass h (Rk ) = μ Rk (xt ), h = 1, 2. (6)
helps also in handling partial contradictions between xt ∈Class h

the rules and considering special cases. In our study, (3) find the class c j that has the maximum value
in order to provide more information about rules, the of βClass h (Rk ):
format of fuzzy rules is extended to include the support
value of the rule. This extended format can be written βClass cj = max{βClass 1 (Rk ), βClass 2 (Rk )} (7)
as follows:
if the consequent class of the fuzzy rule can-
Rk : if x1 is Ak1 and x2 is Ak2 and ... and xn is Akn not be uniquely defined, we do not generate
then Y is c j with certainty rk and support sk (3) the fuzzy rule Rk .

where sk is the support value of the kth rule. (b) Calculate the certainty grade
Certainty grade can be calculated as follows:
Fuzzy Rule Generation Procedure
  2
CF j = βClass Cj (Rk ) − β̄ / βClass h (Rk ) (8)
Let Ti be the number of linguistic values associated h=1
with the input variable xi . One of these linguistic values
Where
is used as the fuzzy antecedent Aik for the input vari-
able xi in each rule Rk . In addition “dont care” is an 
2

additional linguistic value that represents the irrelevant β̄ = βClass h (Rk ) (9)
h=1,h=Ck
fuzzy antecedents that can be deleted without affecting
the fuzzy systems performance. Thus, the number N (c) Calculation of the support
of possible combinations of the antecedent part in this
case is: Calculate the support sk with the following for-
N = (T1 + 1) ∗ (T2 + 1) ∗ (T3 + 1) ∗ ... ∗ (Tn + 1) (4) mula:
NCC Rk
Using this method, all the possible antecedents of sk = , Rk ∈ FRBSc j (10)
mj
fuzzy rules can be generated, but when the number
of inputs is high (high-dimensional data set), the num- where NCC R j is the number of training patterns
ber of rules generated will be exponentially increased. correctly classified by the rule Rk , m j is the num-
Since we are applying feature selection procedure at ber of training patterns belonging to the class c j
the pre-processing stage, the number of inputs will be and FRBSc j is the set of fuzzy rules associated with
adequate to apply this approach. After defining the the class c j.
 J Med Syst

Fuzzy Reasoning Method the first approach, various objective functions are com-
bined into a single function using weight factors. The
The fuzzy reasoning adopted in this study is based difficulty of this approach lies in the determination
on a single rule winner [54]. A new pattern xt = (xt1 , of the proper weight values that characterize the user
xt2 , ..., xtn ) is assigned to the consequent class of the preference. The second approach searches for non-
winner Rw . That is the rule which has the highest com- dominated Pareto optimal compromises between the
bination between the matching degree of the pattern conflicting objectives. It is more practical, as it offers
with the antecedent-part and the certainty degree of the decision maker multiple alternatives that have the
classes. The winner rule can be determined as follows: same cost. There are a number of MOGAs proposed
in the literature [55–57], of which NSGA-II algorithm
μ Rw (xt ).CFw = max{μ R j (xt ).CFk |Rk ∈ FRBS} (11) [58] is the most commonly used multi-objective genetic
The classification of the new pattern is rejected in algorithms to handle this kind of problem.
the case where two or more rules with different classes Non-dominated genetic algorithm II NSGA-II is an
have the same maximum value in Eq. 11. efficient MOGAs introduced by Deb et al. [58] to
overcome some of the NSGAs [59] drawbacks such
Fuzzy Rule-based System Optimization as computation complexity, the need for specifying
a sharing parameter, and non-elitism approach [58].
After generating all possible rules, the optimization The advantages of NSGA-II with respect to other
procedure is applied to search for a subset of a small MOGAs is the preservation of diversity and the fast
number of rules (within all possible rules) with the high- non-dominated sorting of individuals. The concept of
est classification ability. The first objective represents non-dominated relation can be defined as follows. So-
the accuracy of FRBS while transparency (a subset with lution S A dominates S B if the following two conditions
small number of rules) defines the second objective. hold:
These two modeling objectives of our FRBS can be
written as follows: 1. S A is strictly better than S B in at least one objective,
and
Maximize facc (FRBS), Minimize frule (FRBS) (12) 2. S A is no worse than S B in all objectives.

Where facc (FRBS) is the accuracy of FRBS mea- Controlled elitist genetic algorithm It is an enhanced
sured by the rate of classification accuracy, frule (FRBS) version of NSGA-II proposed by Deb and Goel [60]
is the number of fuzzy rules of FRBS. To optimize for controlling the extent of elitism to certain portion
simultaneously the above mentioned objectives, con- defined by the user. This approach allows for better
trolled NSGA-II is utilized, and the results of the op- convergence comparing with original NSGA-II. The
timization are Pareto-front solutions that represent a following subsection describes the chromosome repre-
number of different FRBSs solutions with their corre- sentation used in our implementation.
sponding 2-tuple of values tFRBS where tFRBS = (accu-
racy of FRBS, rules’ number of FRBS). The selection Chromosome Representation
between these solutions usually depends on user pref-
erence and the type of problem under investigation. Let N be the total number of fuzzy rules generated
In our case, since the accuracy in medical diagnosis is and S is the subset of rules selected, where S ≤ N. The
critical, the FRBS with the highest accuracy is selected. chromosome is coded as binary string of length N and
In the following subsection, a brief description of the each binary bit represents one rule. When the value
multi-objective genetic algorithm used is introduced. of qth bit is set to 1, the qth rule is selected while in
the other case (qth bit is set to 0), the qth rule is not
Multi-objective Genetic Algorithms

Genetic algorithms are heuristic techniques inspired

by natural evolution for searching for optimum solu-
tion. Multi-objective genetic algorithms (MOGAs) are
classes of genetic algorithms which are mainly applied
for optimization problems that have multiple and even
conflicting objectives [55]. MOGAs use two approaches Fig. 1 Chromosome coding of rule selection procedure with N =
in handling multi-objectives optimization problems. In 10, S = 5
J Med Syst

considered. Figure 1 shows an example of chromosome some properties of transparency related to the fuzzy
coding of rule selection procedure with total number systems such as the number of rules, and the num-
of rules N = 10, and number of selected rules S = 5. ber of antecedent conditions in each rule [46, 47]. In
The rules number 1,5,6,7 and 10 are selected to form a addition, there is a special emphasis on the use of
candidate fuzzy rule-based system while rules number linguistic variables that have clear meaning [44, 45]. In
2,3,4,8 and 9 are ignored. this study, we consider all the above stated measures
and we highlight also the importance of providing more
information about degree of certainty and support of
Enhancing the Classification Accuracy of the Fuzzy
each rule, so that transparent and informative rules will
Rule-based System by an Ensemble Classifiers
be produced.
Strategy (ECS)

After producing the fuzzy rule-based system, we pro- Performance Measures

pose an approach to enhance the performance of the
FRBS by supporting its decision (classification) by an To assess the discriminating power of the fuzzy rule-
ensemble of classifiers in the case where the clas- based system, two well known measures are employed.
sification decision is made by a rule (winner rule) which The first and conventional method is the percentage
its degree of certainty is below a threshold α where α is of correctly classified testing patterns (PCC) while the
defined by the user, in this study α = 0.5. second is the area under the Receiver operating char-
Basically, Ensemble method uses a set of relatively acteristic (ROC) curve or AUC for short. The latter
accurate and diverse classifiers to classify all new cases measure is commonly used in the medical commu-
by taking a weighted vote of their classifications [61]. nity to evaluate the diagnostic power of tests for dis-
For our case, we are using FRBS for classification and eases [69]. In addition, 10-fold cross-validation method
we are not calling Ensemble method in all the cases (as is used to estimate the classification accuracy of our
the traditional way) but only on the uncertainty cases, fuzzy rule-based system [70]. The results achieved by
i.e. where the certainty of the FRBS classification is our system are compared with existing benchmark
low. And the reason behind proposing this method can methods.
be resumed as the following: in the cases where the
certainty of the FRBS classification is high (FRBS clas-
sification is most likely to be correct), we prefer to keep Results and Discussions
using FRBS to get the interpretation of the cases, while
we use Ensemble method (Bagging method with simple Feature Selection Procedure
majority vote) in the uncertainty cases (the certainty
degree of FRBS is low) to enhance the likelihood of In this step, SFFS method was applied to select the sub-
their correct classification. set of features that can produce the best classification
To compose the ensemble of classifiers, we se- model. The results of this step are shown in Table 2.
lect, in addition to our FRBS, six other popular clas- As can be noted from the table, SFFS selects the subset
sifiers, out of which 3 are non-fuzzy based methods that includes the following four features: sex, chest pain
and the others are fuzzy-based methods. The non- type,  of MajorVessels and Thal. In addition, while the
fuzzy based methods are: Artificial Neural Networks number of features decreases from 13 to only 4, the
(ANNs) [62], the popular decision tree algorithm C4.5 classification accuracy of the model is still maintained at
[63] and statistical method Linear Discriminant Analy- almost 83% which reflects the ability of SFFS method
sis (LDA) [64] while the fuzzy based methods are: in handling feature selection problem.
FH-GBML [46], SLAVE [43], and GP-FCS [65]. These
algorithms are included in many non-commercial tools
such as Weka [66], Keel [67], and Orange [68].
Table 2 Results of feature selection procedure
All features Selected features
Fuzzy Rule-based System Evaluation  of features 13 4
Description See Table 1 2 (sex), 3 (chest pain type),
Transparency Measures 12 ( of major vessels),
13 (Thal)
While the transparency concept is difficult to convert Misclassification
into measurable metrics, many researchers agree on error 17.07% 17.04%
 J Med Syst

Table 3 the number of fuzzy rules and fuzzy sets selected in the optimization step
Subsets S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 Average
 of rules 7 8 8 10 8 10 6 7 7 9 8
 of initial fuzzy sets 28 32 32 40 32 40 24 28 28 36 32
 of selected fuzzy sets 17 22 20 27 20 25 15 15 17 21 19.9
 of fuzzy sets/rule 2.4 2.8 2.5 2.7 2.5 2.5 2.5 2.1 2.4 2.5 2.5

Fuzzy Optimization symptoms and the diagnosis outcome while degree of

certainty degree helps to know how well this relation
This step aims at: (1) maximize the classification ac- is accurate. This kind of reasoning matches with the
curacy and (2) minimize the number of fuzzy rules in human thinking which allows partial truth to exist.
the fuzzy rule-based system. The results of this step are Rule’s support at the other hand allows knowing
shown in Tables 3, 4, 5, 6, 7 and 8. to what extent this relation can be generalized. For
example, in Table 4, which contains information about
the rules of fold2, the 8th rule has a certainty degree
Transparency Criterion
of 1 which means perfect classification ability while it
covers only 2 cases in the training data. This kind of
The total number of generated fuzzy rules are N = 300
relations can be seen as special cases and cannot be
rules, from which S rules are selected to represent the
generalized. On the other hand, rules 1, 2, and 3 have
fuzzy rule-based system. The results of this selection are
correctly classified 82, 56 and 30 cases respectively out
Pareto non-dominated solutions that represent fuzzy
of 243 cases (the total cases of training data). These
rule-based systems with different levels of accuracy-
rules should have special importance in the analysis
number of rule values. As stated before, the priority
and in drawing conclusions when comparing with the
is given to the accuracy of the systems, that is, the
others as they have dominant position in the diagnostic
FRBS with the higher accuracy is selected. Table 3
process.
shows the number of rules, fuzzy sets and number of
sets per rule for different folds. As can be seen from
Table 3, fuzzy rules selected from 300 rules have rel- Classif ication Accuracy Criterion
atively small number with an average of 8 rules per
fold. The difference in the number of rules between Tables 5 and 6 show respectively the PCC and AUC
folds is due to the changing of training data from a fold values of our FRBS before applying ensemble clas-
to another. Another improvement is in the number of sifiers strategy (ECS) and it is named in the table as
fuzzy sets selected per rule which decreases from 4 to B-FRBS and after using this strategy (A-FRBS). In
2.5 per rule. In addition, as can be seen in Fig. 2— addition, the two Tables 5 and 6 show also the PCC
which depicts the fuzzy rules of fold2-, the fuzzy rules and AUC values of a number of benchmark methods
use understandable description of the antecedent con- namely, (1) the decision tree C4.5, (2) Multilayer Per-
ditions in the form of if-then rules and they include in ceptron (MLP) which is the most popular supervised
addition to their class labels, their degree of certainty neural networks architecture, (3) HF-GBML [46], a
and support. The linguistic description helps the physi- fuzzy hybrid genetic-based machine learning for gen-
cian to understand the relations between the factors or erating interpretable fuzzy rules, (4) SLAVE, a GA

Table 4 list of classification rules of subset 2 and their corresponding values of support, confidence and classification accuracy
Training Testing
 of rule Class label Confidence Support Total Correct Incorrect Total Correct Incorrect
1 0 0.28 0.337 101 82 19 9 9 0
2 1 0.80 0.231 62 56 6 9 8 1
3 0 0.72 0.124 35 30 5 3 3 0
4 1 0.27 0.074 20 18 2 3 3 0
5 1 0.74 0.037 12 9 3 1 1 0
6 0 1.0 0.033 8 8 0 2 2 0
7 0 1.0 0.012 3 3 0 / / /
8 1 0.20 0.004 2 1 1 / / /
243 207 36 27 26 /
J Med Syst

Table 5 Classification accuracy of FRBS before (B-FRBS) and after (A-FRBS) applying Ensemble Classifier Strategy (ECS) and a
number of benchmark classifier algorithms using 10-fold cross-validation method
Methods S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 Average
Training data
C4.5 86.80 85.10 86.40 86.40 86.00 86.80 86.80 85.10 87.20 86.40 86.30
ANNs 87.60 85.50 86.80 87.20 86.00 87.20 87.60 85.50 87.60 87.20 86.80
HF-GBML 86.83 85.19 85.19 86.42 86.01 86.01 87.24 85.60 86.83 86.83 86.21
SLAVE 82.70 85.50 86.80 87.20 82.70 86.80 86.80 84.30 85.10 82.30 84.80
GP-FCS 78.60 78.10 78.10 79.80 78.60 77.70 79.40 81.40 83.90 79.00 79.50
LDA 81.4 83.50 83.90 84.70 83.50 83.50 86.00 83.50 85.10 85.10 83.90
B-FRBS 83.54 85.19 85.54 87.24 86.01 81.89 86.42 85.19 84.77 87.24 85.30
A-FRBS 86.83 85.19 85.54 86.42 86.01 86.83 86.83 85.19 85.19 86.42 86.04
Testing data
C4.5 77.70 88.80 85.10 85.10 88.80 81.40 74.00 96.20 77.70 77.70 83.30
ANNs 77.70 96.20 85.10 74.50 77.70 77.70 77.70 96.20 70.30 81.40 81.40
HF-GBML 77.78 96.30 85.19 70.37 92.59 77.78 77.78 93.10 70.37 81.48 82.27
SLAVE 74.00 96.20 81.40 77.70 81.40 70.30 74.00 92.50 77.70 77.70 80.30
GP-FCS 70.30 85.10 88.80 77.70 77.70 81.40 74.00 88.80 59.20 77.70 78.10
LDA 77.70 92.50 88.80 81.40 92.50 74.00 70.30 92.50 77.70 77.70 82.50
B-FRBS 70.37 96.30 85.19 85.19 88.89 66.67 77.78 96.30 74.07 77.78 81.85
A-FRBS 77.78 96.30 85.19 85.19 88.89 81.48 77.78 96.30 77.78 77.78 84.44

learning method for generating fuzzy rules, (5) GP- result also indicates that more reliable classification
FCS, a genetic programming algorithm for optimizing decision can be obtained when combined with other
fuzzy classifiers, and (6) the statistical method Linear classifiers known by their precision ability. This conclu-
Discriminant Analysis (LDA). sion is shared by West et al. [71] who suggested that
As can be seen from Tables 5 and 6, the initial results the accuracy of a set of classifiers is generally better
show that the ensemble classifiers strategy applied to than the best single classifier. In fact, it is common in
support the classification of rules with certainty value medical practice to have more than one opinion in cases
less than 0.5 has improved the accuracy from 81.85% where the degrees certainty of the diagnosis is weak.
to 84.44% and from 0.812 to 0.839 for PCC and AUC Thus, these classifiers are playing the role of experts
respectively. This accuracy is comparable or even better who provide their opinions about complicated cases
than the other classifiers listed in Tables 5 and 6. This that have low degree of certainty.

Table 6 AUC Values of FRBS before (B-FRBS) and after (A-FRBS) applying Ensemble Classifier Strategy (ECS) and a number of
benchmark classifier algorithms using 10-fold cross-validation method
Methods S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 Average
Training data
C4.5 0.863 0.844 0.859 0.858 0.853 0.864 0.864 0.846 0.869 0.859 0.858
ANNs 0.843 0.837 0.854 0.835 0.819 0.856 0.852 0.824 0.802 0.847 0.837
HF-GBML 0.863 0.844 0.849 0.858 0.853 0.860 0.868 0.849 0.864 0.863 0.857
SLAVE 0.819 0.848 0.862 0.866 0.815 0.836 0.865 0.834 0.844 0.811 0.840
GP-FCS 0.795 0.844 0.805 0.797 0.831 0.800 0.854 0.764 0.815 0.860 0.817
LDA 0.809 0.834 0.840 0.846 0.834 0.817 0.859 0.835 0.854 0.849 0.838
B-FRBS 0.826 0.844 0.847 0.866 0.853 0.814 0.860 0.846 0.841 0.867 0.846
A-FRBS 0.863 0.844 0.847 0.858 0.853 0.864 0.865 0.846 0.844 0.859 0.854
Testing data
C4.5 0.767 0.892 0.842 0.850 0.900 0.800 0.742 0.958 0.750 0.783 0.828
ANNs 0.883 0.933 0.842 0.825 0.833 0.800 0.675 0.925 0.692 0.783 0.819
HF-GBML 0.767 0.967 0.842 0.692 0.933 0.767 0.775 0.917 0.675 0.817 0.815
SLAVE 0.717 0.967 0.800 0.767 0.808 0.675 0.742 0.917 0.750 0.783 0.793
GP-FCS 0.750 0.892 0.842 0.742 0.833 0.708 0.642 0.842 0.633 0.817 0.770
LDA 0.767 0.933 0.883 0.825 0.933 0.717 0.708 0.925 0.758 0.783 0.823
B-FRBS 0.683 0.967 0.842 0.850 0.900 0.642 0.775 0.958 0.717 0.783 0.812
A-FRBS 0.767 0.967 0.842 0.850 0.900 0.800 0.775 0.958 0.750 0.783 0.839
 J Med Syst

Table 7 Sensitivity values of FRBS before (B-FRBS) and after (A-FRBS) applying Ensemble Classifier Strategy (ECS) and a number
of benchmark classifier algorithms using 10-fold cross-validation method
Methods S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 Average
Training data
C4.5 0.815 0.778 0.815 0.806 0.787 0.824 0.824 0.796 0.843 0.815 0.810
ANNs 0.833 0.815 0.833 0.833 0.815 0.852 0.852 0.833 0.833 0.806 0.831
HF-GBML 0.815 0.778 0.776 0.806 0.789 0.861 0.824 0.787 0.824 0.815 0.807
SLAVE 0.741 0.778 0.806 0.806 0.704 0.769 0.833 0.750 0.778 0.704 0.767
GP-FCS 0.694 0.778 0.676 0.750 0.833 0.704 0.870 0.676 0.741 0.787 0.751
LDA 0.759 0.824 0.843 0.833 0.824 0.796 0.852 0.833 0.870 0.824 0.826
B-FRBS 0.741 0.778 0.785 0.806 0.787 0.769 0.824 0.796 0.778 0.815 0.788
A-FRBS 0.815 0.778 0.785 0.806 0.787 0.824 0.833 0.796 0.778 0.815 0.802
Testing data
C4.5 0.667 0.917 0.750 0.833 1.000 0.667 0.750 0.917 0.500 0.833 0.783
ANNs 0.833 1.000 0.750 0.917 1.000 0.667 0.750 0.917 0.583 0.833 0.825
HF-GBML 0.667 0.875 0.750 0.583 1.000 0.667 0.750 0.833 0.417 0.833 0.738
SLAVE 0.500 1.000 0.667 0.667 0.750 0.417 0.750 0.833 0.500 0.833 0.692
GP-FCS 0.500 0.917 0.750 0.750 1.000 0.417 0.750 0.750 0.333 0.833 0.700
LDA 0.667 1.000 0.833 0.917 1.000 0.500 0.750 0.917 0.583 0.833 0.800
B-FRBS 0.500 1.000 0.750 0.833 1.000 0.417 0.750 0.917 0.500 0.833 0.750
A-FRBS 0.667 1.000 0.750 0.833 1.000 0.667 0.750 0.917 0.500 0.833 0.792

Sensitivity and Specif icity CHD. We notice from Tables 7 and 8 as well as Fig. 3
that the specificity value for FRBS is clearly higher than
ROC curves of FRBS before and after applying en- the sensitivity value. This observation is also true for all
semble classifiers strategy (ECS) are displayed in Fig. 3 classifiers (expect for MLP algorithm) and especially
with blue and green colors, respectively. These curves for fuzzy systems such as SLAVE, HF-GBML and
are graphical plots of the sensitivity vs. (1-specificity) B-FRBS where the difference between the specificity
values of the two systems. They also represent the and the sensitivity is very significant, for example the
values of true positives (TP), the false positives (FP), specificity in B-FRBS is 0.873 while the sensitivity is
the true negative (TN) and the false negatives (FN) 0.750. Another observation is that the non-fuzzy clas-
for each system. In our case the positive class is the sifiers generally have better sensitivity values than the
existence of CHD while the negative is the absence of fuzzy systems and as a result exhibits better ability in

Table 8 Specificity values of FRBS before (B-FRBS) and after (A-FRBS) applying Ensemble Classifier Strategy (ECS) and a number
of benchmark classifier algorithms using 10-fold cross-validation method
Methods S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 Average
Training data
C4.5 0.91 0.911 0.904 0.911 0.919 0.904 0.904 0.896 0.896 0.904 0.906
ANNs 0.852 0.859 0.874 0.837 0.822 0.859 0.852 0.815 0.770 0.889 0.843
HF-GBML 0.911 0.911 0.933 0.911 0.907 0.859 0.911 0.911 0.904 0.911 0.907
SLAVE 0.896 0.919 0.919 0.926 0.926 0.904 0.896 0.919 0.911 0.919 0.913
GP-FCS 0.896 0.911 0.933 0.844 0.830 0.896 0.837 0.852 0.889 0.933 0.882
LDA 0.859 0.844 0.837 0.859 0.844 0.837 0.867 0.837 0.837 0.874 0.850
B-FRBS 0.911 0.911 0.911 0.926 0.919 0.859 0.896 0.896 0.904 0.919 0.905
A-FRBS 0.911 0.911 0.911 0.911 0.919 0.904 0.896 0.896 0.911 0.904 0.907
Testing data
C4.5 0.867 0.867 0.933 0.867 0.800 0.933 0.733 1.000 1.000 0.733 0.873
ANNs 0.933 0.867 0.933 0.733 0.667 0.933 0.600 0.933 0.800 0.733 0.813
HF-GBML 0.867 0.917 0.933 0.800 0.867 0.867 0.800 1.000 0.933 0.800 0.878
SLAVE 0.933 0.933 0.933 0.867 0.867 0.933 0.733 1.000 1.000 0.733 0.893
GP-FCS 1.000 0.867 0.933 0.733 0.667 1.000 0.533 0.933 0.933 0.800 0.840
LDA 0.867 0.867 0.933 0.733 0.867 0.933 0.667 0.933 0.933 0.733 0.847
B-FRBS 0.867 0.933 0.933 0.867 0.800 0.867 0.800 1.000 0.933 0.733 0.873
A-FRBS 0.867 0.933 0.933 0.867 0.800 0.933 0.800 1.000 1.000 0.733 0.887
J Med Syst

Fig. 2 Linguistic fuzzy

rules of subset 2

detecting patients with CHD while fuzzy-based systems Thus, ECS has generally improved the ability of FRBS
achieve relatively better results in specificity which for CHD diagnosis and specifically for detecting the
means better accuracy in detecting patients without patient with CHD which is a desirable feature for any
CHD. These two observations show that the role of the CHD diagnosis system.
two sets of classifiers can be complementary and can be
used in an ensemble classifier strategy to improve the
quality of the FRBS. The results achieved by A-FRBS
(after applying ECS) shows that the improvement in Conclusions
sensitivity (from 0.750 to 0.792) is more significant than
the improvement in specificity (from 0.873 to 0.887). In this study, a fuzzy rule-based system for CHD diag-
nosis that considers both the accuracy and the trans-
parency is proposed. The produced fuzzy rules can be
easily understood by the physician as they use natural
linguistic terms to describe the relations between the
factors and the outcome of the diagnosis. In order to
enhance the transparency of FRBS, the format of the
fuzzy rules is extended to incorporate the certainty
degree and the support at the consequent part of each
rule so that the physician can check the importance and
validity of each rule and whether to consider it in the
decision process. In addition, the accuracy of diagnosis
decision is improved by employing an ensemble of
classifiers strategy (ECS) to support the decision of the
FRBS in the case where the degree of certainty of the
decision of the winner rule is low. These features make
the developed FRBS a suitable tool for CHD diagnosis
as it allows the physician not only to detect accurately
the existence of CHD but also understand the relations
between the factors and the CHD diagnosis. These re-
lations may reveal unexpected findings and knowledge
Fig. 3 ROC curve of FRBS before and after applying Ensemble that can be used to detect the CHD existence at early
Classifier Strategy (ECS) stage which may result in significant life saving.
 J Med Syst

Acknowledgement This research was supported by a funda- myocardial infarcts: an imaging study. Lancet 361:374–379,
mental research grant scheme No. FP093/2007C from the Min- 2003.
istry of Higher Education, Malaysia. 13. Di Carli, M.F., Davidson, M., Little, R., Khanna, S., Mody,
F.V., Brunken, R.C., et al., Value of metabolic imaging with
positron emission tomography for evaluating prognosis in
patients with coronary artery disease and left ventricular dys-
Conflict of Interest function. Am. J. Cardiol. 73:527–533, 1994.
14. Rainbird, A.J., Mulvagh, S.L., Oh, J.K., McCully, R.B.,
The authors declare that they have no conflict of Klarich, K.W., Shub, C., et al., Contrast dobutamine stress
echocardiography: clinical practice assessment in 300 consec-
interest. utive patients. J. Am. Soc. Echocardiogr 14:378–385, 2001.
15. Kim, W.Y., Danias, P.G., Stuber, M., Flamm, S.D., Plein,
S., Nagel, E., et al., Coronary magnetic resonance angiogra-
phy for the detection of coronary stenoses. N. Engl. J. Med.
References 345:1863–1869, 2001.
16. Nieman, K., Cademartiri, F., Lemos, P.A., Raaijmakers, R.,
1. Lilly, L., Pathophysiology of Heart Disease: A Collaborative Pattynama, P.M., and de Feyter, P.J., Reliable noninvasive
Project of Medical Students and Faculty. 5 ed. (Lippincott coronary angiography with fast submillimeter multislice spi-
Williams and Wilkins, Philadelphia, 2010). ral computed tomography. Circulation 106:2051–2054, 2002.
2. Roger, V.L., Go, A.S., Lloyd-Jones, D.M., Adams, R.J., 17. Yardimci, A., Soft computing in medicine. Applied Soft
Berry, J.D., Brown, T.M., et al., Heart disease and stroke Computing 9:1029–1043, 2009.
statistics 2011 update: a report from the American Heart 18. Akay, M., Noninvasive diagnosis of coronary artery disease
Association. Circulation 123:e18–e209, 2011. using a neural network algorithm. Biol. Cybern. 67:361–367,
3. Mathers C.D., and Loncar, D., Projections of global mortality 1992.
and burden of disease from 2002 to 2030. PLoS Med 3:2011– 19. Lapuerta, P., Azen, S. P., and Labree, L., Use of neural
2030, 2006. networks in predicting the risk of coronary artery disease.
4. Reddy, K.S., Cardiovascular diseases in the developing coun- Comput. Biomed. Res. 28:38–52, 1995.
tries: Dimensions, determinants, dynamics and directions 20. Goodenday, L. S., Cios, K. J., and Shin, I., Identifying coro-
for public health action. Public Health Nutr. 5:231–237, nary stenosis using an image recognition neural network.
2002. IEEE Eng. Med. Biol. Mag. 16:139–144, 1997.
5. Yan, H.M., Jiang, Y.T., Zheng, J., Peng, C.L., and Li, Q.H., A 21. Reategui, E.B., Campbell, J.A., and Leao, B.F., Combining
multilayer perceptron-based medical decision support system a neural network with case-based reasoning in a diagnostic
for heart disease diagnosis. Expert Syst. Appl. 30:272–281, system. Artif. Intell. Med. 9:5–27, 1997.
2006. 22. Azuaje, F., Dubitzky, W., Lopes, P., Black, N., and
6. Adlassnig, K.P. Fuzzy set theory in medical diagnosis. IEEE Adamsom, K., Predicting coronary disease risk based on
Trans. Systems Man Cybernet. 16:260–265, 1986. short-term RR interval measurements: A neural network ap-
7. Escolar, E., Weigold, G., Fuisz, A., and Weissman, N.J., New proach. Artif. Intell. Med. 15:275–297, 1999.
imaging techniques for diagnosing coronary artery disease. 23. Mobley, B. A., Schechter, E., Moore, W. E., McKee, P. A.,
Can. Med. Assoc. J. 174:487–495, 2006. and Eichner, J. E., Predictions of coronary artery stenosis by
8. Henneman, M. M., Schuijf, J. D., van der Wall, E. E., and artificial neural network. Artif. Intell. Med. 18:187–203, 2000.
Bax, J. J., Non-invasive anatomical and functional imaging 24. Lewenstein, K., Radial basis function neural network ap-
for the detection of coronary artery disease, Br. Med. Bull. proach for the diagnosis of coronary artery disease based on
79-80:187–202, 2006. the standard electrocardiogram exercise test. Med. Biol. Eng.
9. Naghavi, M., Falk, E., Hecht, H.S., Jamieson, M.J., Kaul, S., Comput. 39:1–6, 2001.
Daniel, D., Berman et al., From vulnerable plaque to vul- 25. Scott, J. A., Aziz, K., Yasuda, T., and Gewirtz, H., Integration
nerable patient: Part III. Introducing a new paradigm for the of clinical and imaging data to predict the presence of coro-
prevention of heart attack; identification and treatment of the nary artery disease with the use of neural networks. Coron.
asymptomatic vulnerable patient. Screening for Heart Attack Artery Dis. 15:427–434, 2004.
Prevention And Education (SHAPE) task force report. Ex- 26. Mobley, B. A., Schechter, E., Moore, W. E., McKee, P. A.,
ecutive Summary, Am. J. Cardiol. 2006. doi:10.1016. and Eichner, J. E., Neural network predictions of significant
10. Gibbons, R. J., Balady, G. J., Bricker, J. T., Chaitman, B. R., coronary artery stenosis in men. Artif. Intell. Med. 34:151–
Fletcher, G. F., Froelicher, V. F., et al., ACC/AHA 2002 161, 2005.
guideline update for exercise testing: summary article: a re- 27. Babaoglu, I., Baykan, O.K., Aygul, N., Ozdemir, K., Bayrak,
port of the American college of cardiology/American heart M., Assessment of exercise stress testing with artificial neural
association task force on practice guidelines. Circulation network in determining coronary artery disease and predict-
106:1883–1892, 2002. ing lesion localization, Expert Syst. Appl. 36:2562–2566, 2009.
11. Raggi, P., Callister, T.Q., Cooil, B., He, Z.X., Lippolis, N.J., 28. Babaoǧlu, İ., Baykan, Ö. K., Aygul, N., Ozdemir, K., and
Russo, D.J., et al., Identification of patients at increased Bayrak, M., A Comparison of Artificial Intelligence Methods
risk of first unheralded acute myocardial infarction by elec- on Determining Coronary Artery Disease, In: Papasratorn,
tron beam computed tomography. Circulation 101:850–855, K., Lavangnananda, W., Chutimaskul, Vanijja, V. (Eds.), Ad-
2000. vances in Information Technology, Communications in Com-
12. Wagner, A., Mahrholdt, H., Holly, T.A., Elliott, M.D., puter and Information Science. Vol. 114, pp. 18–26. Berlin
Regenfus, M., Parker, M., et al., Contrast-enhanced MRI Heidelberg: Springer, 2010.
and routine single photon emission computed tomography 29. Zhidong, Z., Noninvasive Diagnosis of Coronary Artery
(SPECT) perfusion imaging for detection of subendocardial Disease Based on Instantaneous Frequency of Diastolic Mur-
J Med Syst

murs and SVM. In: Proceedings of the 2005 IEEE Engi- 47. Ishibuchi, H., Yamamoto, T., Nakashima, T., Hybridiza-
neering in Medicine and Biology 27th Annual Conference, tion of Fuzzy GBML Approaches for Pattern Classification
Shanghai, China. 5651–5654, 2005. Problems, IEEE Trans. Syst. Man Cybern., Part B, Cybern.
30. Comak, E., Arslan, A., Turkoglu, I., A Decision Support 35:359–365, 2005.
System Based on Support Vector Machines for Diagnosis of 48. Detrano, R., Janosi, A., Steinbrunn, W., Pfisterer, M.,
The Heart Valve Diseases. Comput. Biol. Med. 37:21–27, 2007. Schmid, J., Sandhu, S., Guppy, K., Lee, S., and Froelicher,
31. Babaoglu, I., Findik, O., and Bayrak, M., Effects of prin- V., International application of a new probability algorithm
ciple component analysis on assessment of coronary artery for the diagnosis of coronary artery disease. Am. J. Cardiol.
diseases using support vector machine. Expert Syst. Appl. 64:304–310, 1989.
37:2182–2185, 2010. 49. Pudil, P., Novovicova, J., and Kittler, J., Floating Search
32. Ainon, R.N., Bulgiba, A.M., and Lahsasna, A., AMI Screen- Methods in Feature Selection. Pattern Recogn. Lett. 15:1119–
ing Using Linguistic Fuzzy Rules, J. Med. Syst., Springer 1125, 1994.
Netherlands. doi:10.1007/s10916-010-9491-2. Published on- 50. Oh, I.S., Lee, J.S., and Moon, B.R., Hybrid genetic algorithms
line 2 May, 2010. for feature selection, IEEE Trans. Pattern Anal. Mach. Intell.
33. Casillas, J., Cordn, O., Herrera, F., and Magdalena, L., Inter- 26:1424–1437, 2004.
pretability Issues in Fuzzy Modeling. Studies in Fuzziness and 51. Ferri, F.J., Pudil, P., Hatef, M., and Kittler, J., Compara-
Soft Computing vol. 128, Springer, Heidelberg 2003. tive Study of Techniques for Large-Scale Feature Selection,
34. Bojarczuk, C.C., Lopes, H.S., and Freitas, A.A., Genetic pro- Pattern Recognition in Practice IV, In: Gelsema, E.S., and
gramming for knowledge discovery in chest pain diagnosis, Kanal, L.N., (Eds.), 403–413, 1994.
IEEE Eng. Med. Biol. Mag. (Special issue on data mining and 52. Kudo, M., and Sklansky, J., Comparison of Algorithms that
knowledge discovery) 19:38–44, 2000. Select Features for Pattern Recognition. Pattern Recogn.
35. De Backer, G., Ambrosioni, E., Borch-Johnsen, K., Brotons, 33:25–41, 2000.
C., Cifkova, R., Dallongeville, J., et al., European guide- 53. Jain, A., and Zongker, D., Feature Selection: Evaluation,
lines on cardiovascular disease prevention in clinical practice: Application, and Small Sample Performance. IEEE Trans.
Third joint task force of European and other societies on Pattern Anal. Mach. Intell. 19:153–158, 1997.
cardiovascular disease prevention in clinical practice (con- 54. Ishibuchi, H., Nozaki, K., and Tanaka, H., Distributed Rep-
stituted by representatives of eight societies and by invited resentation of Fuzzy Rules and Its Application to Pattern
experts), Eur. Heart J. 24:1601–1610, 2003. Classification, Fuzzy Sets Syst. 52:21–32, 1992.
36. Dubois, D., and Prade, H., What are fuzzy rules and how to 55. Konak, A., Coit, D.W., and Smith, A.E., Multi-objective op-
use them. Fuzzy Sets Syst. 84:169–185, 1996. timization using genetic algorithms: a tutorial. Reliab. Eng.
37. Bates, J.H.T., and Young, M.P., Applying fuzzy logic to med- Syst. Saf. 91:992–1007, 2006.
ical decision making in the intensive care unit, Am. J. Respir. 56. Coello, C.A.C., A comprehensive survey of evolutionary
Crit. Care Med. 167:948–952, 2003. based multi-objective optimization techniques. Knowl. Inf.
38. Akay, Y.M., Akay, M., Welkowitz, W., Kostis, J., Noninva- Syst. 1:269–308, 1999.
sive detection of coronary artery disease, IEEE Eng. Med. 57. Van Veldhuizen, D.A., and Lamont, G.B., Multi-objective
Biol. Mag. 13:761–764, 1994. evolutionary algorithms: analyzing the state-of-the-art.
39. Cios, K.J., Goodenday, L.S., Shah, K.K., Serpen, G., A novel Evol.Comput. 8:125–147, 2000.
algorithm for classification of SPECT images of a human 58. Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T., A fast
heart.In: Proc. 9th IEEE Symp. on Computer-Based Medical and elitist multi-objective genetic algorithm: NSGA-II. IEEE
Systems, IEEE Comput. Soc. Press, Los Alamitos, CA, USA, Trans. Evol. Comput. 6:182–197, 2002.
1–5, 1996. 59. Srinivas, N., and Deb, K., Multi-objective optimization using
40. Jain, R., Mazumdar, J., Moran, W., Application of fuzzy clas- non-dominated sorting in genetic algorithms. Evol. Comput.
sification system to coronary artery disease and breast cancer, 2:221–248, 1994.
Australasian Phys. Eng. Sci. Med. 21:141–147, 1998. 60. Deb, K., and Goel, T., Controlled elitist non-dominated sort-
41. Lovelace, J.J., Cios, K.J., Sala, D.M., Goodenday, L.S., ing genetic algorithms for better convergence, In: Zitzler, E.,
Internet-based system for diagnosis of coronary artery dis- Deb, K., Thiele, L., Coello, C.A.C., and Corne, D., (Eds.), J.
ease. Comput. Cardiol. 2:45–48, 1998. Med. Syst. Proceedings of the First International Conference
42. Hudson, D.L., Cohen, M.E., Deedwania, P.C., A hybrid sys- on Evolutionary Multi-Criterion Optimization EMO 2001
tem for diagnosis and treatment of heart disease. In: Proc. (Berlin: Springer, 2001) 67–81.
16th Annual Internat. Conf. of the IEEE Engineering in 61. Dietterich, T.G., An experimental comparison of three
Medicine and Biology Society, Engineering Advances: New methods for constructing ensembles of decision trees: Bag-
Opportunities for Biomedical, IEEE, New York, NY, USA, ging, boosting, and randomization. Mach. Learn. 40:139–158,
2 1368–1369, 1994. 2000.
43. Polat, K., Gunes, S., and Tosun, S., Diagnosis of heart dis- 62. Haykin, S., Neural Networks: A Comprehensive Foundation
ease using artificial immune recognition system and fuzzy (Prentice-Hall, 1999).
weighted pre-processing, Pattern Recogn. 39:2186–2193, 2006. 63. Quinlan, J.R., C4.5: Programs for Machine Learning (Morgan
44. Verlinde, H., De Cock, M., and Boute, R., Fuzzy versus quan- Kaufmann Publishers, 1993).
titative association rules: a fair data driven comparison, IEEE 64. Fisher, R.A., The use of multiple measurements in taxonomic
Trans. Syst. Man Cybern. B Cybern. 36:679–684, 2006. problems. Annals of Eugenics 7:179–188, 1936.
45. Gonzalez, A., and Perez, R., SLAVE: a genetic learning sys- 65. Snchez, L., Couso, I., Corrales, J.A., Combining GP Opera-
tem based on an iterative approach. IEEE Trans. Fuzz. Syst. tors With SA Search To Evolve Fuzzy Rule Based Classifiers.
7:176–191, 1999. Inf. Sci. 136:175–192, 2001.
46. Nauck, D., Data Analysis with Neuro Fuzzy Methods (Habil- 66. Witten, H., and Frank, E., Data Mining: Practical Machine
itation thesis. Otto-von-Guericke University of Magdeburg, Learning Tools and Techniques with Java Implementations
Faculty of Computer Science, Magdeburg, Germany, 2000). (Morgan Kaufmann, 1999).
 J Med Syst

67. Alcal-Fdez, J., Snchez, L., Garcia, S., del Jesus, M.J., Ventura, 69. Altman, D.G., and Bland, J.M., Diagnostic tests 3: receiver
S., Garrell, J.M., Otero, J., Romero, C., Bacardit, J., Rivas, operating characteristic plots. Br. Med. J. 309:188, 1994.
V.M., Fernndez, J.C., Herrera, F., KEEL: A Software Tool 70. Kohavi, R., A study of cross-validation and bootstrap for
to Assess Evolutionary Algorithms to Data Mining Problems. accuracy estimation and model selection, In: Mellish, C.S.,
Soft Comput. 13:307–318, 2009. ed. Proceedings IJCAI-95 Montreal, Quebec, (Morgan Kauf-
68. Demsar, J., Zupan, B., Leban, G., and Curk, T., Orange: from mann, Los Altos, CA, 1995) 1137–1143.
experimental machine learning to interactive data mining, 71. West, D., Rampal, P.R., and West, V., Ensemble strategies for
European Conference of Machine Learning (Springer, Pisa, a medical diagnostic decision support system: a breast cancer
Italy) 537–539, 2004. diagnosis application, Eur. J. Oper. Res. 162:532–551, (2005).

View publication stats