Data Mining based Methodologies for Cardiac Risk Patterns

Identiﬁcation

V. G. Almeida

, J. Borba

, T. Pereira

, H. C. Pereira

1,2

, J. Cardoso

and C. Correia

Instrumentation Center (GEI-CI), University of Coimbra, Coimbra, Portugal

ISA-Intelligent Sensing Anywhere, Coimbra, Portugal

Keywords:

Data Mining, Artiﬁcial Neural Network, Clustering, Arterial Distension Waveform, Cardiovascular Diseases.

Abstract:

Cardiovascular diseases (CVDs) are the leading cause of death in the world. The pulse wave analysis provides

a new insight in the analysis of these pathologies, while data mining techniques can contribute for an efﬁcient

diagnostic method. Amongst the various available techniques, artiﬁcial neural networks (ANNs) are well es-

tablished in biomedical applications and have numerous successful classiﬁcation applications. Also, clustering

procedures have proven to be very useful in assessing different risk groups in terms of cardiovascular function

in healthy populations. In this paper, a robust data mining approach was performed for cardiac risk patterns

identiﬁcation. Eight classiﬁers were tested: C4.5, Random Forest, RIPPER, Naïve Bayes, Bayesian Network,

Multy-layer perceptron (MLP) (1 and 2-hidden layers) and radial basis function (RBF). As for clustering pro-

cedures, k-means clustering (using Euclidean distance) and expectation-maximization (EM) were the chosen

algorithms. Two datasets were used as case studies to perform classiﬁcation and clustering analysis. The accu-

racy values are good with intervals between 88.05% and 97.15%. The clustering techniques were essential in

the analysis of a dataset where little information was available, allowing the identiﬁcation of different clusters

that represent different risk group in terms cardiovascular function. The three cluster analysis has allowed the

characterization of distinctive features for each of the clusters. Reﬂected wave time (T_RP) and systolic wave

time (T_SP) were the selected features for clusters visualization. Data mining methodologies have proven

their usefulness in screening studies due to its descriptive and predictive power.

1 INTRODUCTION

According to World Health Organization (WHO),

Cardiovascular Diseases (CVDs) represent 31% of all

global deaths, making them the leading cause of death

in the world (Mendis, 2011). The global impact of

CVDs has motivated the development of low-cost di-

agnostic tools to be used in early stages of CVDs de-

velopment (Mendis, 2011).

The assumption that arterial stiffness is a marker

of CVDs has gained much relevance and support

during the last years. Several studies have reported

its predictive value in cardiovascular (CV) mortality

(Laurent, 2006).

The arterial distension waveform (ADW) contains

a signiﬁcant amount of physiological information hid-

den in its morphology. The development and subse-

quent validation of efﬁcient non-invasive ADW acqui-

sition devices has gained signiﬁcant importance in the

last years (Avolio et al., 2010). The use of piezoelec-

tric (PZ) sensors in ADW monitoring have been repo-

rted by some authors with good performance in in

vivo clinical trials (Almeida et al., 2011a; Clemente

et al., 2010). Augmentation Index (AI) and Pulse

Wave Velocity (PWV) are amongst the most impor-

tant features that can be extracted from the ADW.

Both of them have been widely studied by the sci-

entiﬁc community and proved their usefulness in as-

sessing local and regional arterial stiffness (Laurent,

2006).

Data mining techniques have attracted a great deal

of attention due to their ability in extract implicit

and potentially useful information from huge amounts

of data. Their feasible implementation in computer-

aided diagnosis methodologies has given new insights

for the development of innovative and effective deci-

sion support systems for CV premature risk assess-

ment.

Classiﬁcation and clustering routines in CV risk

assessment have already been reported in the litera-

ture, with good results. Shah, et. al. (2011) proved

the usefulness and feasibility of assessing different

127

G. Almeida V., Borba J., Pereira T., C. Pereira H., Cardoso J. and Correia C..

Data Mining based Methodologies for Cardiac Risk Patterns Identiﬁcation.

DOI: 10.5220/0004222701270133

In Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms (BIOINFORMATICS-2013), pages 127-133

ISBN: 978-989-8565-35-8

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

CV risk groups in a healthy population. Jovic et. al.

(2011) addressed the electrocardiogram (ECG) clas-

siﬁcation problem using the combination of several

features in the analysis of heart rate (HR) variability.

Tsipouras et. al. (2008) developed a fuzzy rule-based

decision support system for the diagnosis of coronary

artery disease. The dataset included several features,

including demographic, clinical, and vascular data.

Furthermore, the application of Artiﬁcial Neural Net-

works (ANNs) in the cardiac ﬁeld reveals its potential

and affordability (Yang et al., 2012).

Previous established developments have con-

tributed for a non-invasive acquisition device that

can be a convenient and suitable solution to assess

the hemodynamic condition (Almeida et al., 2011a;

Almeida et al., 2011b; Almeida et al., 2011c). In

the present study, an approach of CVDs risk assess-

ment is presented by the analysis of interesting pat-

terns and relationships in the pulse wave morphology

using classiﬁcation and clustering procedures.

This paper is organized as follows: section 2 de-

scribes all the classiﬁcation and clustering algorithms.

Section 3 centers on the used methodology, including

ADW data acquisition and data pre - processing tasks.

Results and Discussion are presented in section 4, and

ﬁnally, Conclusions are reported in section 5.

2 DATA MINING ALGORITHMS

Among the many different deﬁnitions that are found

in the literature, data mining can be deﬁned as the

automatic or semi-automatic process of discovering

patterns in huge amounts of data. When data mining

techniques are efﬁciently applied, new and potentially

useful information can be extracted from a dataset.

Classiﬁcation is used to classify data into prede-

ﬁned categorical class labels, where the class is the

attributed most important for the model characteriza-

tion. Clustering consists of grouping similar data into

clusters. A cluster is deﬁned as a set of objects that

are closely related to each other, and less related to

objects outside the cluster.

A set of classiﬁcation and clustering algorithms

are described below.

2.1 Classiﬁcation Algorithms

2.1.1 Artiﬁcial Neural Networks

Artiﬁcial neural networks (ANNs) are sophisticated

analytical techniques, inspired by the biological ner-

vous system. In an ANN, a set of connected input and

output units, in which each connection has a speciﬁc

weight associated to it. The network learns by adjust-

ing the weights so as to be able to predict the correct

class label of the inputs.

Even though ANNs need high computational re-

sources, need rigorous parameter tuning (for example,

the number of hidden neurons in MLP) and have poor

interpretability, they present advantages that make

their effectiveness undeniable, as their outstanding ac-

curacy in general, ability to work with continuous at-

tributes and above average tolerance to redundant at-

tributes.

The most popular ANN architecture is the multi-

layer perceptron (MLP), which consists of input units,

an output layer and one or more hidden layers deter-

mined by the user. It uses back-propagation as the

learning algorithm. After the network inputs pass

through the input layer, they are weighted and fed to

a second layer, the hidden layer. The outputs of the

hidden layer can be inputs to another hidden layer or

inputs to the output layer, which will emit the net-

work’s prediction for the given tuples. Another pop-

ular approach is the radial basis function (RBF) net-

work, in which every hidden unit implements a radial

activation function and each output unit implements a

weighted sum of hidden unit outputs (Haykin, 1998).

2.1.2 Decision Trees Induction

Decision trees are very popular classiﬁcation meth-

ods, due to their simplicity and interpretability. They

also have no need of parameter tuning or high compu-

tational power while holding good accuracy in overall

(Kotsiantis, 2007).

Decision trees assume a ﬂowchart structure,

where each internal node represents a test in an at-

tribute that need to be classiﬁed, and each branch rep-

resents a possible outcome of this test. Popular deci-

sion tree induction techniques are the Random Forest

(Breiman, 2001) and C4.5 (Quinlan, 1993).

2.1.3 Bayesian Classiﬁcation

Bayesian classiﬁers are based on Bayes’ theorem, and

predict class membership probabilities, such as the

probability of a given tuple belonging to a particular

class. Usual Bayesian classiﬁers are the Naïve Bayes

and Bayesian networks.

The major advantage of the Bayesian classiﬁers

is their short computational time for training. They

also present good tolerance to missing values. How-

ever, their accuracy is usually lower than other classi-

ﬁers, which makes them good ’baseline’ classiﬁcation

methods (Kotsiantis, 2007).

BIOINFORMATICS2013-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

128

Figure 1: Schematic representation of main tasks performed along this work: ADW acquisition (where, DAQ is the data

acquistion sytem), feature points selection and database storage.

2.1.4 Associative Classiﬁcation

In associative classiﬁcation, association rules are gen-

erated and analysed, showing strong associations be-

tween attributes - value pairs that occur frequently in

a given dataset. A typical associative classiﬁer is the

RIPPER algorithm, which was proposed by Cohen

(1995) and is based on association rules with reduced

error pruning.

2.2 Clustering Algorithms

2.2.1 k-Means Algorithm

The k-Means algorithm is a type of partitional cluster-

ing. It acknowledges the number of desired clusters

input k and partitions a set of n objects into k clusters,

so that the resulting intra-cluster similarity is high but

the inter-cluster similarity is low. Cluster similarity is

measured in regard to the mean value of the objects

contained in a cluster, which is, in fact, the cluster

centroid. The k-means algorithm is an iterative pro-

cess, and continuously iterates until a speciﬁc crite-

rion function (usually the square error) converges.

K-means clustering is relatively scalable and efﬁ-

cient in processing large datasets. However, it cannot

handle categorical attributes and is unsuitable for dis-

covering clusters with non-convex shapes. Also, it is

quite sensitive to noise and outliers, so an efﬁcient

pre-processing of the data could and should be per-

formed before applying the method (Han and Kam-

ber, 2006).

2.2.2 Expectation - Maximization

The expectation-maximization (EM) clustering algo-

rithm is a complex probabilistic extension of the k-

means method. Instead of assigning each object to a

cluster with which it is most similar, EM assigns each

object to a cluster according to a weight represent-

ing the probability of membership, which means that

there are no strict boundaries between clusters, and

new means are determined based on weighted mea-

sures (Han and Kamber, 2006).

3 METHODS

This section describes the used methodology, includ-

ing data acquisition procedures, feature selection, pre-

processing routines and respective data mining analy-

sis.

3.1 Data Acquisition

The non-invasive PZ probe was placed over the

carotid artery during data acquisitions performed at

sampling rate of 1kHz. The architecture of the acqui-

sition system is shown in Figure 1.

Age, gender, weight, height, smoking habits and

diabetes history were registered and Body Mass In-

dex (BMI) was later calculated for all of the subjects.

Systolic and diastolic blood pressure values (SBP and

DBP, respectively) and heart rate (HR) were measured

before the ADW recording with an automatic digital

DataMiningbasedMethodologiesforCardiacRiskPatternsIdentification

129

Table 1: Non-demographic attributes list.

Attribute Description Units

T_SP Upstroke time [ms]

T_RP Time at reﬂection point [ms]

T_DW Starting time of dicrotic wave [ms]

H_SP Systolic amplitude [-]

H_RP Reﬂected wave amplitude [-]

H_DW Dicrotic wave amplitude [-]

R1 Downstroke time between systolic and dicrotic wave [ms]

R2 Amplitude quotient between H_DW and H_SP [-]

R3 Amplitude difference between H_SP and H_RP [-]

R4 Amplitude quotient between H_SP and H_RP [-]

AI Augmentation Index [%]

FWHM Full width at half maximum [ms]

RMSE Root mean square error (RMSE) between each pulse and the av-

erage pulse

[%]

RMSSD Root mean square of successive differences (RMSSD) of SP,

DW and RP (T-time and A-amplitude)

[ms or amp (a.u.)]

oscillometric sphygmomanometer (Omron M6 Com-

fort, Kyoto, Japan).

The patient measurement protocol was mainly

based on the subject condition standardization of Van

Bortel, et. al., (2002) . All measurements were made

in a similar time of the day, at the same temperature

controlled room (22-23

◦

Celsius). Subjects remained

seated on a comfortable chair during the measure-

ments. At least 3 measurements of roughly 30-40 sec-

onds were made per subject. A morphological analy-

sis was performed to identify pulse artefacts caused

by voluntary and involuntary movements. Subse-

quently, the pulses were segmented and feature point

extraction by pulse wave analysis was performed.

The dataset is composed by tuples where each tu-

ple represents one segmented pulse. Each pulse con-

tains a maximum of 29 attributes. Most of these at-

tributes are numerical, but some of them are cate-

gorical/nominal attributes. All non-demographic at-

tributes are described in Table 1.

• Dataset I: This group is constituted by 25 healthy

(class 1) subjects and 25 unhealthy subjects (class

2). A set of 2947 pulses was analysed. The num-

ber of pulses is approximately equal for each class

(1425/1522). Eight classiﬁers were trained us-

ing this dataset. Demographic data (age, gender,

smoker, diabetes history, height, weight, BMI,

SBP, DBP, HR) were removed.

• Dataset II: This dataset consists of 93 healthy and

young subjects between 18-30 years. A set of

4471 pulses was analysed. The full list of 29 pa-

rameters was used for clustering procedures.

3.2 Data Pre-processing

Before applying the desired data mining techniques,

all data must be rigorously pre-processed to avoid

low-quality mining results. Usual data cleaning rou-

tines were executed to discover and correct discrepan-

cies in data, consequently ensuring high-quality data.

These routines include missing value removal, outlier

identiﬁcation, irrelevancy and redundancy analysis.

Data normalization procedures to transform the at-

tributes into a 0 to 1 range were applied to the data

used in ANN classiﬁcation tasks, due to their need of

standardized inputs to perform efﬁciently.

3.3 Data Mining

The Weka (Waikato Environment for Knowledge

Analysis) free machine learning system was selected

as the data mining tool for use during classiﬁcation

and clustering analysis due to its comproved efﬁ-

ciency, versatility and affordability (WEKA, 2012).

3.3.1 Classiﬁcation

Classiﬁcation methods can be evaluated according

to their general accuracy, computation cost, robust-

ness, scalability and interpretability (Han and Kam-

ber, 2006). The selected classiﬁcation algorithms

were: C4.5, Random Forest, RIPPER, Naïve Bayes,

Bayesian Network, MLP (1HL), MLP (2HL) and

RBF. Accuracy (ACC), sensitivity (SST), speciﬁcity

(SPT) and precision (PRC) measures were deter-

mined with the number of true positives (TP), true

BIOINFORMATICS2013-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

130

negatives (TN), false positives (FP) and false nega-

tives (FN). Training speed was also measured, accord-

ing the following rank: * [>60s], ** [20-60s], ***

[10-20s], **** [2-10s], ***** [0-2s]. The classiﬁca-

tion task was performed in Weka using 10-fold cross-

validation.

3.3.2 Clustering

Clustering methods can be evaluated by their vi-

sual interpretability, scalability, dimensionality, ro-

bustness and ability to deal with different types of

attributes. A wide range of clustering methods ex-

ists on the literature. K-means clustering (using Eu-

clidean distance) and expectation-maximization (EM)

were the algorithms selected to test. Reﬂected wave

time (T_RP) and systolic wave time (T_SP) were the

selected features for clusters visualization.

4 RESULTS AND DISCUSSION

4.1 Dataset I

Results from the classiﬁcation performance are ex-

pressed in Table 2. All classiﬁcation methods ex-

cept Bayesian based - classiﬁcation exhibited accu-

racy values >95%. Random forest was the best

classiﬁer, in overall, by having the highest accuracy

and sensitivity. Comparing with the other classiﬁers,

ANN methods require high computational resources,

as they can take more than 30 seconds to be trained.

However, ANN algorithms have shown superior re-

sults in terms of speciﬁcity, which means that the FP

error rate is low. The three classiﬁers with highest

accuracy scores were: random forest, 1-hidden layer

MLP and 2-hidden layers MLP.

4.2 Dataset II

4.2.1 Two Clusters

Figure 2 displays the clusters performance for EM

and k-means algorithms by the RP time and SP time

analysis. Different colored points represent different

pulse labels (blue = class 1; red = class 2). Categor-

ical features (gender, smoker) were eliminated in the

k-means clustering method due to its limitations with

categorical attributes.

The EM plot presents unsatisfactory splitness be-

tween the two classes, as visible by the dashed ar-

row that indicates a point from class 2 inside the class

1 group. Visually, the results of k-means clustering

are more desirable and satisfying, as it partitioned

Figure 2: Clusters performance obtained for EM (top) and

k-means algorithm (bottom) using two clusters, Cluster

1=blue, Cluster 2=red.

the dataset in two perfectly homogeneous risk groups.

For both EM and k-means, the blue group (class 1)

presents itself as a more healthier partition than the

red group (class 2), as it represents the cases where

the reﬂection wave arrives after the systolic point, and

class 2 represents a group with higher risk for CV

complications due to the early RP time.

4.2.2 Three Clusters

The k-means clustering was the chosen algorithm

for three risk group clustering, due to its better per-

formance in two-cluster analysis in comparison with

EM.

During the clustering procedures, RMSSD_RPT

negatively interfered in primary cluster measure-

ments. For this reason, this parameter was removed

from the subsequent analysis alongside categorical

features. The clustering groups are presented in Fig-

ure 3. The dataset is split in three clusters, which are

not completely split due to clusters 2 and 3. This

could be due to irrelevant features taking part in the

clustering process.

It can be noted by Table 3 that cluster 3 (green ho-

mogeneous zone in Figure 3) is mostly represented by

ADW type C pulses, where T_RP > T_SP. Adding to

DataMiningbasedMethodologiesforCardiacRiskPatternsIdentification

131

Table 2: Classiﬁer selection results.

Classiﬁer ACC SST SPT PRC Training Speed

C4.5 95.72 95.58 95.86 95.58 ****

Random Forest 97.15 97.47 96.85 96.66 ****

RIPPER 95.72 95.72 95.34 95.07 ***

Naïve Bayes 88.05 88.05 86.60 89.42 *****

Bayesian Network 89.01 84.60 92.90 91.80 *****

MLP (1 HL) 96.98 96.84 97.11 96.91 **

MLP (2 HL) 96.74 96.56 96.91 96.70 *

RBF 96.23 96.23 95.23 96.93 ***

Figure 3: Clusters performance obtained for k-means algo-

rithm using three clusters, Cluster 1=blue, Cluster 2=red,

Cluster 3=green.

that, the mean AI of the cluster centroid is negative (-

11.2%), which represents cluster 3 as a low CVD risk

group. Cluster 2 (red points) pulses are mostly type

B pulses, where T_SP > T_RP. Cluster 2 represents

an intermediate group in terms of CV risk. Cluster 1

pulses (blue points) represent the less homogeneous

group, as the are also scattered in the cluster 2 area.

These pulses are mainly ADW type A pulses, with

some ponctual type B pulses. This group has higher

CV risk in comparison with Cluster 2 and 3.

5 CONCLUSIONS

In the current paper, data mining techniques were ap-

plied to the analysis of ADW features using classiﬁ-

cation and clustering techniques.

In spite of low speed during training, ANNs al-

gorithms have shown high performance in terms of

successfully assessing correct classes (MLP (1-HL)

and MLP (2-HL)). Random Forest also presented su-

perior results in terms of accuracy and sensitivity. The

better training speed was achieved by Bayesian clas-

siﬁers, but, on the other hand, they presented low per-

formance.

In future trials, the use of multiple classiﬁers in the

Table 3: Three clusters features mean.

Attribute

Cluster

1 2 3

Pulses 458 1550 2463

Age 21.6 21.0 21.7

Weight (Kg) 63.0 55.4 63.2

Height (m) 1.7 1.6 1.7

BMI (Kg/m

) 21.3 20.8 21.8

SBP (mmHg) 109.9 106.1 108.6

DBP (mmHg) 69.6 70.3 68.9

HR (bpm) 72.8 67.5 72.9

SPT (ms) 172.7 234.4 117.1

RPT (ms) 103.0 143.2 179.9

DWT (ms) 240.3 306.4 274.4

SPA (amp) 1.0 1.0 1.0

RPA (amp) 0.8 0.9 0.9

DWA (amp) 0.8 0.8 0.7

R1 (-) 67.6 72.0 157.3

R2 (-) 0.8 0.8 0.7

R3 (-) 0.2 0.1 -0.1

R4 (-) 0.8 0.9 -0.9

AI (%) 21.0 12.6 -11.2

RMSSD_SPT (ms) 34.9 25.5 21.7

RMSSD_SPA (amp) 0.0 0.0 0.0

RMSSD_RPA (amp) 0.5 0.0 0.2

RMSSD_DWT (ms) 54.2 18.5 42.8

RMSSD_DWA (amp) 0.1 0.1 0.1

FWHM (ms) 459.0 450.6 466.3

RMSE (%) 0.1 0.0 0.1

development of predictive models as, for example, a

weighted multiple classiﬁer predictive methodology

as suggested by Gorunescu et. al. (2011) could be an

interesting approach to develop.

In clustering methods, more than the usefulness

when there is no, or little, information available, the

information that is extracted from them can be cru-

cial in fully understanding the data. The three cluster

analysis has allowed the characterization of distinc-

tive features for each of the clusters. It is also im-

BIOINFORMATICS2013-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

132

portant to focus the number of pulses in each one of

the clusters. Cluster 1 presents only 458 pulses, a re-

duced value in comparison with pulses in cluster 2

and 3, 1550 and 2463, respectively. However, this

fact occurs due to the population characteristics, only

young subjects between 18 and 30 years. Clustering

methodologies can be further improved by perform-

ing a progressive attribute removal for a subsequent

visual correction.

The developed methodology has proven its usage

in screening studies due to its descriptive and predic-

tive power. The successful application of data mining

techniques can help to predict under-diagnosed pa-

tients, and identify and classify at-risk people in terms

of health with the goal of reducing healthcare cost.

ACKNOWLEDGEMENTS

Authors acknowledge Fundação para a Ciência e

Tecnologia for funding SFRH/BD/61356/2009 and

PTDC/SAU-BEB/100650/2008, project developed

under the initiative of QREN, funding by UE/FEDER,

through COMPETE-Programa Operacional Factores

de Competitividade. The authors also thanks to Coim-

bra University Hospital Centre (C.H.U.C.) and Dr.

Rui Providência, SCDSOS-Sudden Cardiac Death

Screening of Risk factors, for the support in clinical

trials.

REFERENCES

Almeida, V. G., Pereira, H. C., Pereira, T., Figueiras, E.,

Borges, E., Cardoso, J. M. R., and Correia, C. (2011a).

Piezoelectric probe for pressure waveform estimation

in ﬂexible tubes and its application to the cardiovas-

cular system. Sensors and Actuators A: Physical,

169(1):217–226.

Almeida, V. G., Pereira, T., Borges, E., Pereira, H. C., Car-

doso, J. M. R., and Correia, C. (2011b). A real time

cardiac monitoring system-arterial pressure waveform

capture and analysis. Proceedings of the PECCS 2011.

Algarve, Portugal.

Almeida, V. G., Santos, P., Figueiras, E., Borges, E.,

Pereira, T., Pereira, H. C., Cardoso, J. M. R., and

Correia, C. (2011c). Hemodynamic features extrac-

tion from a new arterial pressure waveform probe.

Proceedings of the BIOSTEC (BIOSIGNALS 2011).

Rome, Italy.

Avolio, A. P., Butlin, M., and Walsh, A. (2010). Ar-

terial blood pressure measurement and pulse wave

analysis–their role in enhancing cardiovascular as-

sessment. Physiol Meas, 31(1):R1–47. Avolio, Al-

berto P Butlin, Mark Walsh, Andrew England Physiol

Meas. 2010 Jan;31(1):R1-47. Epub 2009 Nov 26.

Bortel, L. M. V., Duprez, D., Starmans-Kool, M. J., Safar,

M. E., Giannattasio, C., Cockcroft, J., Kaiser, D. R.,

and Thuillez, C. (2002). Clinical applications of arte-

rial stiffness, task force iii: Recommendations for user

procedures. AJH, 15:445–452.

Breiman, L. (2001). Random forests. Machine Learning,

45:5–32.

Clemente, F., Arpaia, P., and Cimmino, P. (2010).

A piezo-ﬁlm-based measurement system for

global haemodynamic assessment. Physiol Meas,

31(5):697–714. Clemente, Fabrizio Arpaia, Pasquale

Cimmino, Pasquale England Physiol Meas. 2010

May;31(5):697-714. Epub 2010 Apr 16.

Gorunescu, F., Gorunescu, M., Saftoiu, A., Vilmann, P., and

Belciug, S. (2011). Competitive/collaborative neural

computing system for medical diagnosis in pancreatic

cancer detection. Expert Systems, 28(1):33–48.

Han, J. and Kamber, M. (2006). Data Mining:Concept and

Techniques. Elsevier, San Francisco.

Haykin, S. (1998). Neural Networks - A Comprehensive

Foundation. Pearson education, India.

Jovic, A. and Bogunovic, N. (2011). Electrocardio-

gram analysis using a combination of statistical, ge-

ometric, and nonlinear heart rate variability features.

Artif Intell Med, 51(3):175–86. Jovic, Alan Bo-

gunovic, Nikola Netherlands Artif Intell Med. 2011

Mar;51(3):175-86. Epub 2010 Oct 25.

Kotsiantis, S. B. (2007). Supervised machine learning:

A review of classiﬁcation techniques. Informatica,

31:249–268.

Laurent, S., e. a. (2006). Expert consensus document on ar-

terial stiffness:methodological issues and clinical ap-

plications. European Heart Journal, 27:2588–2605.

Mendis, S. e. a. (2011). Global Atlas on Cardiovascular

Disease Prevention and Control. World Health Orga-

nization, Geneva.

Quinlan, J. R. (1993). C4.5: Programs for Machine Learn-

ing. Morgan Kaufmann Publishers, San Mateo, Cali-

fornia.

Shah, A. S., Dolan, L. M., Gao, Z., Kimball, T. R., and

Urbina, E. M. (2011). Clustering of risk factors: a

simple method of detecting cardiovascular disease in

youth. Pediatrics, 127(2):e312–8.

Tsipouras, M. G., ThemisP.Exarchos, Fotiadis, D. I., Kot-

sia, A. P., Vakalis, K. V., Naka, K. K., and Michalis,

L. K. (2008). Automated diagnosis of coronary

artery disease based on data mining and fuzzy mod-

eling. IEEE Transactions On Information Technology

In Biomedicine.

WEKA (2012). Weka 3: Data mining software in java.

Yang, J., Singh, H., Hines, E. L., Schlaghecken, F., Iliescu,

D. D., Leeson, M. S., and Stocks, N. G. (2012). Chan-

nel selection and classiﬁcation of electroencephalo-

gram signals: An artiﬁcial neural network and ge-

netic algorithm-based approach. Artif Intell Med,

55(2):117–26. Yang, Jianhua Singh, Harsimrat Hines,

Evor L Schlaghecken, Friederike Iliescu, Daciana D

Leeson, Mark S Stocks, Nigel G Netherlands Artif In-

tell Med. 2012 Jun;55(2):117-26. Epub 2012 Apr 12.

DataMiningbasedMethodologiesforCardiacRiskPatternsIdentification

133