Biomed Eng Lett (2013) 3:74-79

DOI 10.1007/s13534-013-0096-9

Fuzzy Unordered Rule Induction for Evaluating Cardiac Arrhythmia

V. S. R. Kumari and P. Rajesh Kumar

Received: 2 April 2013 / Revised: 29 June 2013 / Accepted: 30 June 2013

© The Korean Society of Medical & Biological Engineering and Springer 2013

Abstract

Heart rate variations reveal current or impending heart/

cardiac diseases. A non-stationary signal – heart rate - is

measured using Electrocardiogram (ECG) to assess cardiac

Arrhythmia. But studying ECG reports is both tedious and

time consuming when you have to locate abnormalities from

the collected data. This is overcome by computer based

analytical tools for in-depth data study/classification for

diagnosis. Automatic arrhythmia assessment is easy due to

the existence of image processing techniques. Many

algorithms exist for ECG signals detection/classification.

This paper investigates RR interval based ECG classification

procedures for arrhythmic beat classification, a process

based on RR interval beat extraction using Symlet on ECG

data. Extracted RR data is used as a classification feature

with beats being classified through a boosting algorithm and

Fuzzy Unordered Rule Induction Algorithm (FURIA).

Classification efficiency evaluation was through the MIT-

BIH arrhythmia database.

Keywords Arrhythmia classification, Electrocardiogram

(ECG), RR interval, Symlet, Boosting, Fuzzy Unordered

Rule Induction Algorithm (FURIA), MIT-BIH ECG data

INTRODUCTION

ECG records the heart’s electrical activity for heart disease

diagnosis. ECG is a non-invasive technique where voltage

plots are measured by leads against time. The signal is

measured on the human body’s surface. Heart rate/ rhythm

disorder or morphological pattern changes reveal cardiac

arrhythmia, detected by a recorded ECG waveform [1, 2]

analysis. High mortality due to heart diseases emphasises the

need for proper ECG arrhythmias detection/classification for

correct treatment. Though ECG’s visual inspection is tiring,

recent image processing advances have led to automatic

arrhythmia [3] assessment research resulting in algorithms

recently being developed for automatic ECG signals

detection/classification.

ECG measures electrical signals where every heart beat is

displayed as electrical waves with peaks and valleys. ECGs

provide two types of information. First, an electrical wave’s

duration in crossing the heart is determined by measuring

time intervals on the ECG with signal frequency ranging

from 0.05-100 Hz and dynamic ranges from 1-10 mV. Five

peaks and valleys labelled by successive letters of the

alphabet as P, Q, R, S and T characterize ECG signals. A

good ECG performance analysing system depends on

accurate detection of QRS complex as also T and P waves.

The heart’s upper chamber activation is represented by the P

wave, the atria; while QRS wave (or complex) and T waves

represent ventricular (lower chambers) excitation.QRS detection

is of great importance in automatic ECG signal analysis [4].

Once identified, a detailed ECG signal examination including

heart rate and ST segment are undertaken. A P-QRS-T wave’s

amplitude and duration includes heart afflicting disease

information, but ECGs do not provide data on cardiac

contraction/ pumping function. Cardiac diseases diagnosis is

through position and magnitudes of PR interval and segment,

ST interval and segment, QRS interval and QT interval. [5].

Fig. 1 reveals normal ECG pattern and components.

ECG determines Bundle branch block (BBB), a form of

heart block involving a conduction delay/failure in a branch

in the bundle of the heart. It may be complete/incomplete,

transient, permanent, or intermittent, and hence is named

based on the involvement of left or right bundle branch. It

cannot be discovered whether a bundle branch block is

complete. When linked to acute anterior wall myocardial

infarction, bundle branch block identifies high-risk patients.

V. S. R. Kumari ( ) Department of ECE, SMCE, Tummalapalem, Guntur, A. P NDIATel : +918632344402 / Fax : +91863244403E-mail : Vsrk46@gmail.com

P. Rajesh Kumar College of Engineering, Andhra University, Visakhapatnam, INDIA

REVIEW ARTICLE

Biomed Eng Lett (2013) 3:74-79 75

Cardiac impulses inability to be conducted down bundle

branches, cause abnormally shaped QRS complexes. BBB is

visible in high-risk, acute, anterior wall myocardial infarction,

due to ischemia or necrosis of the bundle branches, trauma

(as in surgical manipulation), or mechanical branch compression

due to a tumor. Further condition deterioration leads to

installation of a pacemaker.

Bundle branch block fall under a group of heart problems

called intraventricular conduction defects (IVCD). Two

bundle branches, right and left exist. The right bundle carries

nerve impulses that contract the right ventricle (the heart’s

lower chamber) and the left bundle’s nerve impulses contract

left ventricle. The two bundles join together initially at a

place named the bundle of His. Nerve impulses pass through

the heart’s sinus node to the bundle of His and then into right

and left bundle branches. Bundle branch block refers to a

slowing/interruption of nerve impulses. Patients with right

bundle branch block (RBBB) have limited or nil conducted

nerve impulses. The right ventricle receives impulse through

muscle-to-muscle spread, outside a regular nerve pathway.

This impulse transmission mechanism is slow leading to

delayed right ventricle contraction. There are many types of

left bundle branch block (LBBB) each having its own

characteristic failure mechanism.

ECG features are represented as statistical measures or

extracted from time and frequency domains. Though this

procedure provides impressive results in some classification

tasks, it still fails to discriminate all ECG beats. Wavelet

transformation representing signals in various translations/

scales [6] is another feature extraction method. ECG signals

based cardiac arrhythmia classification methods developed

were weak and inaccurate. So ECG features are extracted for

cardiac arrhythmia classification. Many machine learning and

data mining methods were sought to improve ECG arrhythmia

detection accuracy.

Automated heartbeat classification based on various

features representing ECG and other classification methods

were studied by other researchers [7-12] earlier. They include

frequency-based features [7], ECG morphology [8, 10], and

heartbeat interval features [10, 11], neural networks [12],

Support Vector Machine [11], and self-organizing networks

[8] are some common classification methods.

KhoureichKa [13] proposed a waveform similarity and

RR interval ECG beat classification procedure. This method

classified 6 heart beat types (normal beat, atrial premature

beat, paced beat, premature ventricular beat, left bundle

branch block beat and right bundle branch block beat).

Wavelet transform techniques denoised ECG signal and

extracted RR intervals were used as feature. The classifier

has an annotated beats training database compiled for it as it

used waveform comparison of unknown beats. The MIT/

BIH arrhythmia database’s 46 records were used for evaluation

achieving 97.52% classification rate.

M.G. Tsipouras et al., [14] suggested a knowledge-based

arrhythmic beat classification method and arrhythmic episode

detection and classification using only RR-interval signals

from ECG recordings. Arrhythmic beat classification algorithm

uses a three RR-interval sliding window. Classification is

undertaken through 4 beat categories: normal, premature

ventricular contractions, ventricular flutter/fibrillation and 28

heart block. Input used is beat classification of a knowledge-

based deterministic automation to ensure classification and

arrhythmic episode detection. The six rhythm types classified

include ventricular couplet, ventricular bigeminy, ventricular

tachycardia, ventricular trigeminy, ventricular flutter/fibrillation

and 28 heart block. The MIT-BIH arrhythmia database

evaluates the proposed methods revealing 94% accuracy for

arrhythmic episode detection and classification and 98%

accuracy for arrhythmic beat classification. This method is

better than other complicated techniques like RR-interval

signal used for arrhythmia beat and episode classification.

Bashir, et al., [15] proposed a nested ensemble technique

for real time cardiac arrhythmia classification. The method

includes a training data set and feature selection manipulation,

with the former being manipulated for classifier learning

through updating training data, and feature selection for

improved performance and accuracy during classification.

Experiments proved that all ECG features should be considered

for the proposed model’s evaluation and accuracy.

In this paper, classification methods for arrhythmic beat

classification based on RR interval in ECG are investigated.

The RR beat interval is extracted using Symlet conversion

on ECG data is used [16] as in our previous investigations.

Extracted RR data is used as classification feature for

boosting algorithm and Fuzzy Unordered Rule Induction

Algorithm (FURIA). Classification efficiency was evaluated

Fig. 1. Normal ECG signal and its P, Q, R, S and T peaks.

76 Biomed Eng Lett (2013) 3:74-79

using MIT-BIH arrhythmia database. Classification efficiency

is evaluated with the MIT-BIH arrhythmia database. The

remainder of the paper is organized as follows: section 2

reviews related work in the literature, section 3 details the

methodology. Section 4 provides results and discussion, and

section 5 concludes the paper.

METHODOLOGY

Symlet wavelet

Waveforms bound in time and frequency are wavelets and

the latter’s analysis splits signals into the mother wavelet’s

shifted and scaled versions. The Continuous Wavelet

Transform (CWT) is known by the wavelet function ψ

adding signal times multiplied by scaled and shifted

versions. The continuous wavelet is defined mathematically

by

CWT is the reason for many wavelet coefficients C, a

function of scale and position. The original signal constituent

wavelets are got through multiplying all coefficients by

applicable scaled and shifted wavelets. Daubechies proposed

Symlet, almost symmetrical wavelets to get modifications of

the db family [17]. Both wavelet families are similar, with

difference of db wavelets having maximal phase while

Symlet have minimal phase. Symlet are compactly supported

wavelets with slight asymmetry and its wavelet coefficient is

any positive even number and highest number of vanishing

moments for a specific support width.

Boosting

In boosting, k classifiers set is iteratively learned with every

training tuple being assigned weights [18]. The process

begins with a classifier Mi learning, and updating weights,

thereby helping the next classifier, Mi+1, to assign more

weights to misclassified Mi training tuples. After k classifiers

are learned iteratively, the final boosted classifier, M*, sums

up every classifier’s votes.

A higher weight is allotted to the classifier through

boosting with a lower error rate vote. So, accurate classifiers

have higher weightage. Classifier Mi’s vote weightage is

computed as follows:

For each class, c, weights of every classifier that assigned

class c to X is added up. The class of a tuple is returned

based on highest vote sum.

Boosting with decision stump

A decision stump includes a split decision tree. As a decision

stump is a weak learner it is combined with weak learners by

AdaBoost algorithm into an accurate classifier [19]. Decision

stumps form a committee. Base classifiers are built on weighted

examples with the committee deciding on a majority vote.

When this process starts, all examples get equal weights

which are increased in the second round for misclassified

examples but decreased for correctly classified examples by

the first decision stump. This procedure is repeated till a

particular number of stumps are created. Every committee

member has its own specialty because of special training.

Hence, the performance of a diversified committee is better

than that of a single member.

Boosting with REP tree

Reduced Error Pruning (REP) tree

The created tree is the best fit on training data in a decision

tree learning algorithm [20]. Usually built trees overfit training

samples and so are pruned to reduce training data dependency

and also to ensure that they fit other examples. Reduced

Error Pruning (REP) trees are procedures to prune decision

trees as it both easy and effective, but it does have a tendency

to over prune a tree. REP trees main disadvantage is that

they need a separate pruning dataset. But it is very powerful

with a large training dataset or through boosting. REP pruning

replaces a subtree with a leaf representing an examples

majority. Modifications are required only when it reduces

error through equal or lower number of misclassifications.

Boosting with J48

J48 algorithm is an implementation of C4.5 decision tree

learner. J48 uses greedy algorithm to prompt decision tree

classification. Based on a decision-tree built with a training

dataset unseen data is classified. J48 node generated decision

trees evaluate the significance of every feature, e.g., heart

rate.

ECG features are classified to the appropriate arrhythmia

class by following a root to tree leaves path resulting in class

decision. Decision trees are built top-down, with a suitable

attribute chosen at each level. An information-theoretic

measure evaluates features to provide “classification power”

for every feature. The training dataset is split into subsets on

feature selection, conforming to selected feature values. This

process is repeated for all subsets, till subset instances come

under one class.

Fuzzy Unordered Rule Induction Algorithm (FURIA)

Fuzzy Unordered Rule Induction Algorithm (FURIA) is a

rule-based classification method founded on RIPPER [21].

C scale, position( ) f t( )ψ scale, position, t( )dt

∞–

∞

∫=

log1 error M

i( )–

error Mi

( )--------------------------------

Biomed Eng Lett (2013) 3:74-79 77

FURIA learns fuzzy rules, - not conventional rules – and

unordered rule sets and not rule lists. Its advantage is

preserving simple and comprehensible rule sets. It includes

many modifications/extensions. It uses a rule stretching

method to deal with uncovered examples. Experiments prove

that FURIA outperforms original RIPPER and classifiers

like C4.5, significantly regarding classification accuracy.

FURIA specifically learns fuzzy rules instead of conventional

rules and unordered rule sets instead of rule lists. Uncovered

samples are handled with an efficient rule stretching method.

Fuzzy rules are general compared to conventional rules with

many advantages. For example, conventional (non-fuzzy)

rules produce “sharp” decision boundary models with

corresponding abrupt transitions among classes which is

questionable and not very intuitive. Also one expects support

for a class provided rule to reduce from “full” (inside rule

core) to “zero” (near boundary) slowly and not at once.

Fuzzy rules “soft” boundaries, is their characteristic, but they

convert to crisp boundaries when a classification decision is

planned. Boundaries are flexible in fuzzy rules. For example,

using suitable aggregation operators to combine fuzzy rules

which are not axis-parallel.

Conventional rule learners result in a decision list to

produce rules learned by turn for every class starting from

the smallest (in terms of relative occurrence frequency) and

ending with the second largest. Also, a default rule is added

to a majority class and a new query instance is classified by

the first rule in the list which covers it.

This has advantages and disadvantages. For example, it

might have an unwanted bias as classes are not treated

symmetrically. Sorting rules on a priority basis compromises

comprehensibility (condition part of every rule includes

negated conditions of earlier rules). To offset this FURIA

learns an unordered rule set. In other words, a rule set for

each class in a one-versus-rest scheme where the resultant

model is incomplete, i.e., a new query may be uncovered by

any rule (this way decision lists faces less problems). This

algorithm and results are extension of boosting techniques

for cardiac arrhythymia prediction [22].

EXPERIMENTAL SETUP

A MIT-BIH arrhythmia database is used to evaluate classifiers

performance. The evaluation dataset has 165 instance; 55

events each for Right bunch bundle block, Left bunch bundle

block and Normal RR interval. Continuous wavelet transforms

using symlet2 filters are applied. Figs. 2-4 show event

output when continuous wavelet transforms using symlet2 is

applied.

RESULTS AND DISCUSSION

Instances are classified as Right bunch bundle block (R ),

Left bunch bundle block (L) and Normal RR interval (N).

FURIA learns fuzzy rules for each class from the training

Fig. 2. Left bunch bundle block output on applying continuouswavelet transforms using symlet2.

Fig. 3. Right bunch bundle block output on applying continuouswavelet transforms using symlet2.

Fig. 4. Normal RR interval output on applying continuous wavelettransforms using symlet2.

78 Biomed Eng Lett (2013) 3:74-79

instances. Following are the five rules obtained by FURIA:

(a2 in [-78.404, -69.722, inf, inf]) =>class=L (CF = 0.94)

(a118 in [-inf, -inf, -17.512, -12.86]) and (a213 in [-inf,

-inf, 1.1767, 1.2272]) => class=N (CF = 0.97)

(a118 in [-inf, -inf, -17.512, -14.293]) and (a238 in [-inf,

-inf, -0.54844, -0.35637]) => class=N (CF = 0.95)

(a233 in [-0.84513, -0.15514, inf, inf]) and (a6 in [-inf,

-inf, -80.307, -76.38]) => class=N (CF = 0.92)

(a156 in [-inf, -inf, -0.75362, 2.7266]) and (a82 in [2.2714,

3.0664, inf, inf]) => class=R (CF = 0.98)

Fuzzy Unordered Rule Induction Algorithm (FURIA) and

Boosting classify instances with a decision stump, boosting

with both J48 and REP tree. Classification accuracy

experiments results are provided in the following Tables/

Figures. Table 1 tabulates the result summary for varied

techniques. Fig. 5 shows plotted classification accuracy and

Fig. 6 shows plotted Root mean squared error.

According to observations the obtained results of the Root

Mean Square Error was minimised with FURIA algoritham

compared to previous algoritham i.e. boossting with decision

stump.

Table 2 tabulates the precision, recall and f-Measure. The

precision and recall is computed as

Precision =

Recall =

For a good classification system the values of precision

and recall should be high which is seen in the boosting with

decision stump and FURIA technique.

CONCLUSION

This paper investigates RR interval based ECG classification

method for arrhythmic beat classification. The methodology

is based on RR interval beat extraction using Symlet conversion

on ECG data. Extracted RR data is a classification feature.

Number of relevant images retrieved

Total number of images retrieved------------------------------------------------------------------------------------------

Number of relevant images retrieved

Total number of relevant images in the Database------------------------------------------------------------------------------------------------------------------------

f measure 2precision recall×

precision recall+----------------------------------------×=

Table 1. Summary of the results.

Boosting with decision stump Boosting with J48 Boosting with REP tree FURIA

Classification accuracy 92.125% 91.52% 91.52% 92.121%

Root mean squared error 0.211(21.1%)

0.1947(19.47%)

0.1891(18.91%)

0.1897(18.97%)

Fig. 5. Classification accuracy. Fig. 6. Root mean squared error.

Table 2. Precision, recall and F Measure.

Precision Recall F-Measure

Boosting with decision stump 0.911 0.921 0.916

Boosting with J48 0.91 0.915 0.912

Boosting with REP tree 0.915 0.915 0.914

FURIA 0.911 0.921 0.916

Biomed Eng Lett (2013) 3:74-79 79

Boosting algorithms and FURIA classify beats. MIT-BIH

arrhythmia database evaluates classification efficiency. Instances

are classified as Right bunch bundle block, Left bunch

bundle block and Normal RR interval. FURIA achieved a

classification accuracy of 92.12% and Root mean square

error with boosting with decision stump was 0.211 (21.1%)

and with FURIA was 0.1893 (18.97%).

CONFLICT OF INTEREST STATEMENTS

Kumari VSR declares that s/he has no conflict of interest in

relation to the work in this article. Kumar PR declares that

s/he has no conflict of interest in relation to the work in this

article.

REFERENCES

[1] Couderc JP, Xia X, Peterson DR, McNitt S, Zhao H, PolonskyS, Zareba W. T-wave morphology abnormalities in benign,potent, and arrhythmogenic I< sub>kr</sub> inhibition. HeartRhythm. 2011; 8(7):1036-43.

[2] El-Sherif N, Caref EB, Yin H, Restivo M. Theelectrophysiological mechanism of ventricular arrhythmias inthe long QT syndrome: tridimensional mapping of activationand recovery patterns. Circ Res. 1996; 79(3):474-92.

[3] Saxena SC, Kumar V, Hamde ST. Feature extraction from ECGsignals using wavelet transforms for disease diagnostics. Int JSyst Sci. 2002; 33(13):1073-85.

[4] Raghavendra BS, Bera D, Bopardikar AS, Narayanan R.Cardiac arrhythmia detection using dynamic time warping ofECG beats in e-healthcare systems. IEEE Int Sym Wireless,Mobile, Multimedia Networks. 2011. pp. 1-6.

[5] McSharry PE, Clifford GD, Terassenko L, Smith LA. Adynamical model for generating synthetic electrocardiogramsignals. IEEE T Bio-med Eng. 2003; 50(3):289-94.

[6] Özbay Y, Ceylan R, Karlik B. Integration of type-2 fuzzyclustering and wavelet transform in a neural network basedECG classifier. Expert Syst Appl. 2011; 38(1):1004-10.

[7] Ince T, Kiranyaz S, Gabbouj M. A generic and robust system forautomated patient-specific classification of ECG signals. IEEET Bio-med Eng. 2009; 56(5):1415-26.

[8] Kaneko M, Iseri F, Gotoh T, Yoneyama T, Yamauchi T,Takeshita K, Sueda N. QRS morphological analysis using twolayered self-organizing map for heartbeat classification. ComputCardiol. 2010. pp. 975-8. IEEE.

[9] Basil T, Chandra BS, Lakshminarayan C. A comparison ofstatistical machine learning methods in heartbeat detection andclassification. Big Data Analytics. 2012. pp. 16-25.

[10] De Chazal P, O'Dwyer M, Reilly RB. Automatic classificationof heartbeats using ECG morphology and heartbeat intervalfeatures. IEEE T Bio-med Eng. 2004; 51(7):1196-206.

[11] Kampouraki A, Manis G, Nikou C. Heartbeat time seriesclassification with support vector machines. IEEE T Inf TechBiomed. 2009; 13(4):512-8.

[12] Jiang W, Seong KG. Block-based neural networks forpersonalized ECG signal classification. IEEE T Neural Networ.2007; 18(6):1750-61.

[13] Ahmad K. ECG beats classification using waveform similarityand RR interval. 2011. http://arxiv.org/abs/1101.1836.

[14] Tsipouras MG, Fotiadis DI, Sideris D. An arrhythmiaclassification system based on the RR interval signal. Artif IntellMed. 2005; 33:237-50.

[15] Bashir MEA, Akasha M, Lee DG, Yi GM, Ryu KH, Cha EJ, BaeJW, Cho MC, Yoo CW. Nested ensemble technique forexcellence real time cardiac health monitoring. BioComp. 2010.pp.519-25.

[16] Kumari R. Performance evaluation of boosting techniques forcardiac arrhythmia prediction. Int J Comput Appl. 2012; 57(17).

[17] Boultadakis G, Skrapas K, Frangos P. Time-frequency analysisof radar signals. RTO/SET, 80. 2004. pp. 7.1-20.

[18] Friedman J, Hastie T, Tibshirani R. Additive logistic regression:a statistical view of boosting. Ann Stat. 2000; 28(2):337-407.

[19] Rätsch G, Onoda T, Müller KR. Soft margins for adaBoost.Mach Learn. 2001; 42(3):287-320.

[20] Quinlan JR. Simplifying decision trees. In Gaines B, Boose J,editors. Knowledge acquisition for knowledge-based systems.1987. pp. 239-52.

[21] Hühn J, Hüllermeier E. FURIA: an algorithm for unorderedfuzzy rule induction. Data Min Knowl Disc. 2009; 19(3):293-319.

[22] Kumari VSR, Kumar PR. Performance of evaluation ofboosting technique for cardiac arrhythmia prediction. Int JComput Appl. 2012; 57(17):18-223.