EVOLUTIONARY COMPUTATION APPROACH TO

ECG SIGNAL CLASSIFICATION

Farid Melgani

Dept. of Information and Communicatrion technologies, University of Trento, Via Sommarive, 14-I-38050, Trento, Italy

Yakoub Bazi

College of Engineering, Al Jouf University, 2014 Sakaka, Saudi Arabia

Keywords: ECG classification, feature reduction, particle swarm optimization, support vector machine.

Abstract: In this paper, we propose a novel classification system for ECG signals based on particle swarm

optimization (PSO). The main objective of this system is to optimize the performance of the support vector

machine (SVM) classifier in terms of accuracy by automatically: i) searching for the best subset of features

where to carry out the classification task; and ii) solving the SVM model selection issue. Experiments

conducted on the basis of ECG data from the MIT-BIH arrhythmia database to classify five kinds of

abnormal waveforms and normal beats confirm the effectiveness of the proposed PSO-SVM classification

system.

1 INTRODUCTION

The recent literature reports different and interesting

methodologies for the automatic classification of

electrocardiogram (ECG) signals (e.g., de Chazal et

Reilly, 2006, and Inan et Giovangrandi, 2006).

However, in the design of an ECG classification

system, there are still some open issues, which if

suitably addressed may lead to the development of

more robust and efficient classifiers. One of these

issues is related to the choice of the classification

approach to be adopted. In particular, we think that,

despite its great potential, the SVM approach has not

received the attention it deserves in the ECG

classification literature compared to other research

fields. Indeed, the SVM classifier exhibits a

promising generalization capability thanks to the

maximal margin principle (MMP) it is based on

(Vapnik, 1998). Another important property is its

lower sensitivity to the curse of dimensionality

compared to traditional classification approaches.

This is explained by the fact that the MMP makes

unnecessary to estimate explicitly the statistical

distributions of classes in the hyperdimensional

feature space in order to carry out the classification

task. Thanks to these interesting properties, the SVM

classifier proved successful in numerous and

different application fields, such as 3D object

recognition (Pontil et Verri, 1998), biomedical

imaging (El-Naqa et al., 2002), remote sensing

(Melgani et Bruzzone, 2004 and Bazi et Melgani,

2006). Turning back to ECG classification, other

issues which can be identified are: 1) feature

selection is not performed in a completely automatic

way; and 2) the selection of the best free parameters

of the adopted classifier is generally performed

empirically (model selection issue).

In this paper, we present in a first step a thorough

experimental exploration of the SVM capabilities for

ECG classification. In a second step, in order to

address the aforementioned issues, we propose to

optimize further the performances of the SVM

approach in terms of classification accuracy by 1)

automatically detecting the best discriminating

features from the whole considered feature space

and 2) solving the model selection issue. The

detection process is implemented through a particle

swarm optimization (PSO) framework that exploits a

criterion intrinsically related to the SVM classifier

properties, namely the number of support vectors

(#SV).

Melgani F. and Bazi Y. (2008).

EVOLUTIONARY COMPUTATION APPROACH TO ECG SIGNAL CLASSIFICATION.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 19-24

DOI: 10.5220/0001060200190024

 SciTePress

2 PROPOSED APPROACH

2.1 Support Vector Machine (SVM)

Let us first for simplicity consider a supervised

binary classification problem. Let us assume that the

training set consists of N vectors x

∈ ℜ

(i = 1, 2,

…, N) from the d-dimensional feature space X. To

each vector x

, we associate a target y

∈ {-1, +1}.

The linear SVM classification approach consists of

looking for a separation between the two classes in

X by means of an optimal hyperplane that

maximizes the separating margin (Vapnik, 1998). In

the nonlinear case, which is the most commonly

used as data are often linearly nonseparable, they are

first mapped with a kernel method in a higher

dimensional feature space, i.e., Φ(X) ∈ ℜ

d’

(d’> d).

The membership decision rule is based on the

function sign[f(x)], where f(x) represents the

discriminant function associated with the hyperplane

in the transformed space and is defined as:

f(x) = w

⋅Φ(x) + b

(1)

The optimal hyperplane defined by the weight

vector w

∈

ℜ

d’

and the bias b

∈

ℜ

is the one that

minimizes a cost function that expresses a

combination of two criteria: margin maximization

and error minimization. It is expressed as:

∑

ξC

)(

wξw,Ψ

(2)

This cost function minimization is subject to the

following constraints:

ξ1b

−

≥+⋅ ))(( xw , i = 1,…, N

(3)

and

0≥

ξ , i = 1, 2, …, N

(4)

where the

’s are slack variables introduced to

account for non-separable data. The constant C

represents a regularization parameter that allows to

control the shape of the discriminant function. The

above optimization problem can be reformulated

through a Lagrange functional, for which the

Lagrange multipliers can be found by means of a

dual optimization leading to a Quadratic

Programming (QP) solution, i.e.,

∑∑

−

ji,

jijijii

,Kyyααα

)(

max xx

(5)

under the constraints:

0≥

for i = 1, 2, …, N

(6)

and

∑

0yα

(7)

where α=[α

, α

,…, α

] is the vector of Lagrange

multipliers and

)(

⋅

,K is a kernel function. The final

result is a discriminant function conveniently

expressed as a function of the data in the original

(lower) dimensional feature space X:

∑

∈

b,Kyα)f( )( xxx

(8)

The set S is a subset of the indices {1, 2, …, N}

corresponding to the non-zero Lagrange multipliers

’s, which define the so-called support vectors.

As described above, SVMs are intrinsically binary

classifiers. But the classification of ECG signals

often involves the simultaneous discrimination of

numerous information classes. In order to face this

issue, different multiclass classification strategies

can be adopted (Melgani et Bruzzone, 2004). In this

paper, we shall consider the commonly used one-

against-all strategy.

2.2 PSO Principles

Particle swarm optimization (PSO) is a stochastic

optimization technique introduced recently by

Kennedy and Eberhart, which is inspired by social

behavior of bird flocking and fish schooling

(Kennedy et Eberhart, 2001). Similarly to other

evolutionary computation algorithms such as genetic

algorithms (GAs) (Bazi et Melgani, 2006), PSO is a

population-based search method, which exploits the

concept of social sharing of information. This means

that each individual (called particle) of a given

population (called swarm) can profit from the

previous experiences of all other individuals from

the same population. During the iterative search

process in the d-dimensional solution space, each

particle (i.e., candidate solution) will adjust its flying

velocity and position according to its own flying

experience as well as the experiences of the other

companion particles of the swarm.

Let us consider a swarm of size S. Each

particle

) ,...,2 ,1( SiP

from the swarm is

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

characterized by: 1) its current position

t ℜ∈)(p ,

which refers to a candidate solution of the

optimization problem at iteration t; 2) its velocity

t ℜ∈)(v ; and 3) the best position

t ℜ∈)(p

identified during its past trajectory. Let

t ℜ∈)(p

be the best global position found over all trajectories

traveled by the particles of the swarm. The position

optimality is measured by means of one or more

fitness functions defined in relation to the considered

optimization problem. During the search process, the

particles move according to the following equations:

()

)()()(

)()()()()1(

tttrc

tttrctwt

ibiii

ppvv

−⋅+

−⋅+=+

)()()1( ttt

iii

vpp +=+

(9)

(10)

where r

(⋅) and r

(⋅) are random variables drawn

from a uniform distribution in the range [0, 1] so that

to provide a stochastic weighting of the different

components participating in the particle velocity

definition. c

and c

are two acceleration constants

regulating the relative velocities with respect to the

best global and local positions, respectively. The

inertia weight w is used as a tradeoff between global

and local exploration capabilities of the swarm.

2.3 PSO Setup

The position

ℜ∈

p of each particle P

from the

swarm is viewed as a vector encoding: 1) a

candidate subset F of features among the d available

input features, and 2) the value of the two SVM

classifier parameters, which are the regularization

and the kernel parameters C and γ, respectively.

Since the first part of the position vector implements

a feature detection task, each component

(coordinate) of this part will assume either a “0”

(feature discarded) or a “1” (feature selected) value.

The conversion from real to binary values will be

done by a simple thresholding operation at the 0.5

value.

Let f(i) be the fitness function value associated

with the ith particle P

. The choice of the fitness

function is important since, on its basis, the PSO

evaluates the goodness of each candidate solution p

for designing our SVM classification system. A

possible choice is to adopt the class of criteria that

estimates the leave-one-out error bound, which

exhibits the interesting property of representing an

unbiased estimation of the generalization

performance of classifiers. In particular, for SVM

classifiers, different measures of this error bound

have been derived, such as the radius-margin bound

and the simple support vector (SV) count (Vapnik,

1998). In this paper, we will explore the simple SV

count as fitness criterion in the PSO optimization

framework because of its simplicity and

effectiveness as shown in the context of the

classification of hyperspectral remote sensing

images (Bazi et Melgani, 2006).

2.4 SVM Classification with PSO

• Initialization

Step 1: Generate randomly an initial swarm of size

Step 2: Set to zero the velocity vectors v

(i= 1, 2,...,

S) associated with the S particles.

Step 3: For each position

ℜ∈

p of the particle

) ,...,2 ,1( SiP

from the swarm, train an SVM

classifier and compute the corresponding fitness

function f(i) (i.e., the #SV measure).

Step 4: Set the best position of each particle with its

initial position, i.e.,

ibi

, (i=1, 2,.., S)

(11)

• Search process

Step 5: Detect the best global position

p in the

swarm exhibiting the minimal value of the

considered fitness function over all explored

trajectories.

Step 6: Update the speed of each particle using

Equation (9).

Step 7: Update the position of each particle using

Equation (10). If a particle goes beyond the

predefined boundaries of the search space, truncate

the updating by setting the position of the particle at

the space boundary and reverse its search direction

(i.e., multiply its speed vector by -1). This will

permit to forbid the particles from further attempting

to go outside the allowed search space.

Step 8: For each candidate particle p

(i= 1, 2,..., S),

train an SVM classifier and compute the

corresponding fitness function.

Step 9: Update the best position

p of each particle

if its new current position

p (i= 1, 2,..., S) has a

smaller fitness function.

• Convergence

Step 10: If the maximal number of iterations is not

yet reached, return to Step 5.

• Classification

Step 11: Select the best global position

p in the

swarm and train an SVM classifier fed with the

subset of detected features mapped by

and

modeled with the values of the two parameters C

and γ encoded in the same position.

Step 12: Classify the ECG signals with the trained

SVM classifier.

EVOLUTIONARY COMPUTATION APPROACH TO ECG SIGNAL CLASSIFICATION

3 EXPERIMENTAL RESULTS

Our experiments were conducted on the basis of

ECG data from the MIT-BIH arrhythmia database

(Mark et Moody, 1997). In particular, the considered

beats make reference to the following classes:

normal sinus rhythm (N), atrial premature beat (A),

ventricular premature beat (V), right bundle branch

block (RB), left bundle branch block (LB), and

paced beat (/). Similarly to (Inan et al., 2006), the

beats were selected from the recordings of 18

patients, which correspond to the following files:

100, 102, 104, 105, 106, 107, 118, 119, 200, 201,

203, 205, 208, 212, 213, 214, 215, and 217. For

feeding the classification process, we adopted in this

study the two following kinds of features: i) ECG

morphology features; and ii) three ECG temporal

features that are the QRS complex duration, the RR

interval (i.e., time span between two consecutive R

points representing the distance between the QRS

peaks of the present and previous beats), and the RR

interval averaged over the ten last beats (de Chazal

et Reilly, 2006). The total number of morphology

and temporal features is equal to 303 for each beat.

In order to train the classification process and to

assess its accuracy, we selected randomly from the

considered recordings 500 beats for the training set,

whereas 42185 beats were used as test set (thus, the

training set represents just 1.18% of the test set). The

detailed numbers of training and test beats are

reported for each class in Table 1. Classification

performance was evaluated in terms of three

accuracy measures, which are: 1) the overall

accuracy (OA); 2) the accuracy of each class; and 3)

the average accuracy (AA).

Due to the good performances generally

achieved by the nonlinear SVM classifier based on

the Gaussian kernel [6], we adopted this kernel in all

experiments. The parameters C and γ were varied in

the ranges [10

-3

, 200] and [10

-3

, 2], respectively. The

k value and the number of hidden nodes (h) of the

kNN and the RBF classifiers were tuned in the

intervals [1, 15] and [10, 60], respectively.

Concerning the PSO algorithm, we considered the

following standard parameters: swarm size S=40,

inertia weight w=0.4, acceleration constants c

and

equal to the unity, and maximum number of

iterations fixed to 40.

3.1 Experiment 1: Classification in the

Original Feature Space

In this experiment, we applied the SVM classifier

directly on the whole original hyperdimensional

feature space which is composed of 303 features.

During the training phase, the SVM parameters (i.e.,

C and γ) were selected according to a m-fold cross-

validation (CV) procedure. In all experiments

reported in this paper, we adopted a 5-fold CV. The

same procedure was adopted to find the best

parameters for the kNN and the RBF classifiers. The

best values obtained for the three investigated

classifiers are C=25, γ=0.5, k=3 and h=20. As

reported in Table 2, the OA and AA accuracies

achieved by the SVM classifier on the test set are

equal to 87.95% and 87.60%, respectively. These

results are much better than those achieved by the

RBF and the kNN classifiers. Indeed, the OA and

AA accuracies are equal to 82.78% and 82.34% for

the RBF classifier, and 78.21% and 79.34% for the

kNN classifier, respectively. This experiment

appears to confirm what observed in other

application fields, i.e., the superiority of SVM with

respect to traditional classifiers when dealing with

feature spaces of very high dimensionality.

3.2 Experiment 2: Classification based

on Feature Reduction

In this experiment, we trained the SVM classifier in

feature subspaces of various dimensionalities. The

desired number of features was varied from 10 to 50

with a step of 10, namely from small to high

dimensional feature subspaces. Feature reduction

was achieved by means of the traditional Principal

Component Analysis (PCA) algorithm. Figure 1-a

depicts the results obtained in terms of OA by the

three considered classifiers combined with the PCA

algorithm, namely the PCA-SVM, the PCA-RBF

and the PCA-kNN classifiers. In particular, it can be

seen that, for all feature subspace dimensionalities

except the lowest one (i.e., 10 features), the PCA-

SVM classifier maintains a clear superiority over the

two other classifiers. Its best accuracy was found by

using a feature subspace composed of the first 40

components. The corresponding OA and AA

accuracies are equal to 88.98% and 88%,

respectively. Comparing these results with those

obtained by the SVM classifier in the original

feature space (i.e., without feature reduction), a

slight increase of 1.03% and 0.4% in terms of OA

and AA, respectively, was achieved (see Table 2).

3.3 Experiment 3: Classification with

PSO-SVM

In this experiment, we applied the PSO-SVM

classifier on the available training beats. At

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

convergence of the optimization process, we

assessed the PSO-SVM classifier accuracy on the

test samples. The achieved overall and average

accuracies are equal to 91.44% and 91.19%

corresponding to substantial accuracy gains with

respect to what yielded either by the SVM classifier

applied on all available features (+3.49% and

+3.59%, respectively) or by the PCA-SVM classifier

(+2.46% and +3.19%, respectively) (see Table 2 and

Figure 1). Its worst class accuracy was obtained for

atrial premature beats (A) (88.16%) while that of the

SVM and the PCA-SVM classifiers corresponded to

paced beats (/) (73.43%) and ventricular premature

beats (V) (78.06%), respectively. This shows the

capability of the PSO-SVM classifier to reduce

significantly the gap between the worst and best

class accuracies (8.25% against 15.43% and 20.21%

for the PCA-SVM and the SVM classifiers,

respectively) while keeping the overall accuracy to a

high level.

4 CONCLUSIONS

The main novelty of this paper is to be found in the

proposed PSO-based approach that aims at

optimizing the performances of SVM classifiers in

terms of classification accuracy by detecting the best

subset of available features and by solving the tricky

model selection issue. Its completely automatic

nature renders it particularly useful and attractive.

The results confirm that the PSO-SVM classification

system boosts up significantly the generalization

capability achievable with the SVM classifier.

Finally, it is noteworthy that the general nature of

the proposed PSO-SVM system makes it applicable

not just to morphology and temporal features but to

other kinds of features such as those based on

wavelets and high-order statistics. Finally, other

optimization criteria could be considered as well

individually or jointly depending on the application

requirements.

REFERENCES

Bazi Y., Melgani F., (2006). Toward an Optimal SVM

Classification System for Hyperspectral Remote

Sensing Images. IEEE Trans. Geosci. Remote Sens.,

44, 3374-3385.

De Chazal F., Reilly R.B., (2006). A patient adapting heart

beat classifier using ECG morphology and heartbeat

interval features. IEEE Trans. Biomedical

Engineering, 43, 2535-2543.

El-Naqa I., Yongyi Y., Wernick M.N., Galatsanos N.P.,

Nishikawa R.M., (2002). A support vector machine

approach for detection of microcalcifications. IEEE

Trans. on Medical Imaging, 21, 1552-1563.

Inan O.T., Giovangrandi L., Kovacs J.T.A., (2006).

Robust neural network based classification of

premature ventricular contractions using wavelet

transform and timing interval features. IEEE Trans.

Biomedical Engineering, 53, 2507-2515.

Kennedy J., Eberhart R.C., (2001). Swarm Intelligence.

San Mateo, CA: Morgan Kaufmann.

Mark R., Moody G., (1997). MIT-BIH Arrhythmia

Database 1997 [Online]. Available:

http://ecg.mit.edu/dbinfo.html.

Melgani F., Bruzzone L., (2004). Classification of

hyperspectral remote sensing images with support

vector machine. IEEE Trans. on Geosci. Remote Sens.,

42, 1778-1790.

Pontil M., Verri A., (1998). Support vector machines for

3D object recognition”, IEEE Trans. on Pattern

Analysis and Machine Intelligence, 20, 637-646.

Vapnik V., (1998). Statistical Learning Theory. New

York: Wiley.

75,00

80,00

85,00

90,00

95,00

10 20 30 40 50

# Selected Features

OA [%]

PCA-SVM

PCA-RBF

PCA-KNN

PSO-SVM

Figure 1: Overall percentage accuracy (OA) versus

number of selected features (first principal components)

achieved on the test beats by the different classifiers.

Table 1: Numbers of training and test beats used in the experiments.

Class N A V RB / LB Total

Training beats 150 100 100 50 50 50 500

Test beats 24966 119 4239 3939 6971 1951 42185

EVOLUTIONARY COMPUTATION APPROACH TO ECG SIGNAL CLASSIFICATION

Table 2: Overall (OA), average (AA), and class percentage accuracies achieved on the test beats by the different

investigated classifiers.

Method OA AA N

A V RB / LB

SVM 87.95 87.60 90.05 83.19 92.12 93.15 73.43 93.64

RBF 82.78 82.34 85.14 78.99 90.39 86.74 66.53 86.26

kNN 78.21 79.34 76.50 66.38 71.99 93.27 75.92 92.00

PCA-SVM 88.98 88.00 89.36 83.19 78.06 93.50 90.60 93.28

PCA-RBF 83.04 82.11 85.86 80.67 87.85 83.87 68.85 85.54

PCA-kNN 83.91 82.02 85.62 69.74 79.05 93.04 73.89 90.77

PSO-SVM 91.44 91.19 91.12 88.16 93.70 93.70 92.01 96.41

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing