Proceedings of International Joint Conference on Neural Networks, Montreal, Canada, July 31 - August 4, 2005

ECG Signal Classification using Block-based Neural NetworksWei Jiang, Seong G. Kong, and Gregory D. Peterson

Department of Electrical and Computer EngineeringThe University of Tennessee

Knoxville, TN 37996-2 100, U.S.A.E-mail: skong~utk.edu

Abstract - This paper investigates the application ofevolvable block-based neural networks (BbNNs) to ECGsignal classification. A BbNN consists of a two-dimensional (2-D) array of modular basic blocks that canbe easily implemented using reconfigurable digitalhardware. BbNNs are evolved for each patient in orderto provide personalized health monitoring. A geneticalgorithm evolves the internal structure and associatedweights of a BbNN using training patterns that consist ofmorphological and temporal features extracted from theECG signal of a patient. The remaining part of the ECGrecord serves as the test signal. The BbNN was tested forten records collected from different patients provided bythe MIT-BIH Arrhythmia database. The evolved BbNNsproduced higher than 90% classification accuracies.

I. INTRODUCTION

The electrocardiogram (ECG) makes a useful diagnostictechnique for monitoring heart activities. Analysis ofheartbeat patterns extracted from the ECG signal may reflectthe symptoms indicating that the heart needs immediateattention. Several methods have been proposed to classifyECG heartbeat patterns. Various features are obtained fromthe ECG signal such as morphological features [1], heartbeattemporal intervals [2], frequency domain features [3], andwavelet transform coefficients [4]. Classification techniquesof ECG patterns include linear discriminant analysis [1],support vector machines [5], and artificial neural network[6]. Unsupervised clustering of ECG complexes using self-organizing maps has been studied [7].

ECG signal patterns vary greatly for different individuals.Even for the same individual, heartbeat patterns significantlychange with the time of day and under different situations.Abnormal patterns result from a wide variety of heartproblems. As a consequence, a certain classification methodthat works well for a given dataset may produce inconsistentresults on different datasets. A fixed-structure classifiertrained with a limited number of data may not be able totrack the time-varying nature ofECG signals.

Evolvable classifiers change the structure andconfigurations as well as internal parameters to cope withdynamic environments. Block-based neural network (BbNN)

[8] models incorporate modular neural networks andevolutionary algorithms. BbNNs offer simultaneousoptimization of structure and weight with a geneticalgorithm (GA). The modular structure of BbNNs enableseasy expansion in size by adding more blocks. BbNNs canbe implemented by use of reconfigurable digital electronichardware such as FPGAs that allow on-line partialreorganization of internal structures. BbNNs have beenapplied to the problems such as mobile robot navigation [8],pattern recognition [9], and time series prediction [10].

This paper presents evolvable BbNNs for theclassification of abnormal heartbeat patterns from the ECGsignal. The internal structure and associated weights of aBbNN are evolved to classify heartbeat patterns of eachindividual for personalized health monitoring. The BbNNsuse both morphological and temporal features extracted fromthe ECG signal. The evolved BbNN produced an average of98% classification accuracy of the ECG signals from theMIT-BIH Arrhythmia database [ 15].

II. BLOCK-BASED NEURAL NETWORKS

A. Network Structures

A BbNN can be represented by a two-dimensional (2-D)array of blocks. Each block is a basic processing elementthat corresponds to a feedforward neural network with fourvariable input/output nodes. A block is connected to its fourneighboring blocks with signal flow represented by an arrowbetween the blocks. Leftmost and rightmost blocks arelaterally interconnected. Signal flow uniquely specifies theinternal configurations of a block as well as the overallnetwork structure. Figure 1 illustrates the network structureof an mxn BbNN ofm rows and n columns of blocks labeledas Bij. The first row of blocks BI1, B12, ..., Bin is an inputstage and the blocks Bml, Bm2, ..., Bmn form an output stage.BbNNs with n columns can have up to n inputs and noutputs.

2-7803-9048-2/05/$20.00 02005 IEEE 326

xi x2 ... tX

Yl Y2 Yn

Figure 1: Structure of block-based neural networks

1v3

of two inputs and two outputs (2/2), but with differentplacement of the nodes. The cases of all input nodes (4/0) orall output nodes (0/4) are considered as invalidconfigurations and therefore are not allowed.

The four nodes inside a block are connected with eachother with the weights. The signal uj denotes the input and vkindicates the output of the block. A weight Wjk denotes aconnection between the jth input node uj and the kth outputnode Vk. A block can have up to six connection weightsincluding the bias. For the case of two inputs and twooutputs (2/2), there are four weights and two biases. The 1/3case has three weights and three biases, and the 3/1configuration has three weights and one bias. An outputnode produces an output Vk with an activation function g( ):

Vk = WOk + wkuj kk=1,2, ,K (1)j=1

where J and K denote the number of input and output nodesof the block. For the type 1/3 basic block, J = I and K = 3.For blocks of the type 3/1, J = 3 and K = 1. Type 2/2 blockshave J = 2 and K = 2. The term WOk is called a bias, or theweight corresponding to a constant input uo = 1. A bipolarlogistic sigmoid function in the range [-1,1] is chosen as anactivation function:

VI VI

g(u) = I

V2 U2VI V2

vI U2

(c) (d)

Figure 2: Four possible internal configuration types of ablock. (a) One input and three outputs (1/3), (b) Three inputsand one output (3/1), (c) Two inputs and two outputs (2/2),

(d) A different arrangement of 2/2

Internal configuration is characterized by the input-outputconnections of the nodes. A node can be either an input or an

output according to the internal configuration determined bythe signal flow. An incoming arrow to a block specifies thenode as an input, and output nodes are associated withoutgoing arrows. Generalization capability emerges throughvarious internal configurations of a block. A block can berepresented by one of the four different types of internalconfigurations. Figure 2(a) shows a block with one input andthree outputs (1/3). A block in (b) has three inputs and an

output (3/1). Both Figure 2(c) and (d) correspond to the type

Other popular choices of the activation function includesaturated linear or logistic sigmoid functions.

B. Evolutionary Optimization ofBbNN

Network structure and connection weights of an

individual BbNN are encoded to form a chromosome foroptimization using the GA. The overall structure of a BbNNcan be effectively encoded with binary directions of signalflow. Signal flow provides an integrated representation ofBbNN structure and internal configurations. The signal flowdetermines the structure and the internal configuration of a

BbNN using a sequence of binary numbers. Any connectionbetween the blocks is represented with either 0 or 1. Bit 0denotes down (1) and left (v-), and bit I indicates up (T)and right (-*) signal flows. The signal flow bits associatedwith the blocks in the input and output stages are all zeros

and therefore are not included in structure encoding.Each generation of the GA process involves the three

stages: fitness evaluation, selection, and genetic operation.The current population of BbNNs are evaluated and rankedin terms of fitness value. The population goes through theselection and the genetic operation until the maximumfitness reaches the desired value.

The selection schemes used in this paper are elitistmethod, binary tournament with disruptive pressure. Popular

327

(a) (b) (2)

selection methods such as roulette-wheel often showpremature convergence in early phases of evolution orgenetic drift in later phases. Elitist methods ensure that thebest individuals pass their traits to the next generation inorder to reduce a genetic drift. The individual with thehighest fitness value in the previous generation is preservedin the next generation. The elitist method increases selectionpressure by preventing the loss of important genes. Elitistmethods can also speed up the convergence. In the elitistmethod, individuals with the lowest and the highest fitnessvalues are exchanged only when the individual with thehighest fitness value in the previous generation does notexist in the current generation. Tournament selection[11 ][ 12] has the same effects as both fitness proportionalsampling and selection probability adjustment [13][14]. Thebinary tournament scheme finds one of the two individualsselected by a fitness comparison. The tournament size canadjust the selection pressure to reduce disruptive effect ofuniform crossover. A larger tournament size usuallyincreases the selection pressure.

Crossover and mutation serve as basic genetic operatorsused to evolve the structure and weights of the BbNN.Connection weights and biases are represented with real-valued numbers. For a crossover operation of a pair ofBbNNs, a group of signal flow bits are selected with acrossover probability. The selected signal flow bits areexchanged. After the exchange, the internal structure of ablock is reconfigured according to the new signal flow bits.Corresponding weights will be updated by a weightedcombination of the two weights.

wJ awj +(I )w (3)

(I(-aX) w, + aw,where cc denotes a uniform random number in [0, 1]. Newconnection weights generated accordingly are initializedwith random numbers, and unconnected weights areremoved.

Mutation operator randomly adds a perturbation in anindividual according to the mutation probability. A BbNNcan have different probabilities for the mutation of structureand weight bit strings. When signal flow is reversed aftermutation, all the irrelevant weights are removed and createdwith a random value on a proper direction. A weight selectedfor mutation will be updated with:

III. ECG SIGNAL CLASSIFICATION

A. DatasetThe MIT-BIH Arrhythmia Database [15] provides the

ECG signals used in the experiment. The database contains48 records obtained from 47 different individuals (Tworecords came from the same patient). Each record contains2-channel ECG signals measured for 30 minutes. Twenty-three records serve as representative samples of routineclinical recordings. The remaining 25 records includeheartbeat waveforms such as complex ventricular, junctional,and supra-ventricular arrhythmias. Continuous ECG signalswere filtered using a bandpass filter with a passband from0.1 to 100 Hz. Filtered signals were then digitized at 360 Hz.The beat locations are automatically labeled at first andverified later by independent experts to reach consensus. Thewhole database contains more than 109,000 annotations ofnormal and 14 different types of abnormal heartbeats. Fourrecords containing paced heartbeats were excluded in thisstudy according to Association for the Advancement ofMedical Instrumentation recommended practices [16].Several noisy records were also excluded. Channel 1 signalsof the 10 records (100, 106, 119, 202, 205, 209, 210, 212,213, and 215) were chosen from the dataset for experiment.

B. Feature ExtractionA heartbeat template defines a partial waveform of an

ECG signal that includes an R-peak [6]. Figure 3 showsheartbeat templates Hk-I, Hk, and Hk+,. Let Tk and Tk+I denotethe time intervals between the neighboring R-peaks of thecurrent heartbeat template Hk. The template Hk includes allthe samples in the interval [t-Tkl2, t+ Tk+1/2]. The intervalsrange from 132 to 819 ms in the selected MIT-BIH database.Heartbeat templates are then normalized to have an equalnumber of samples. In each heartbeat template, the R peak isshifted to the center and then linearly interpolated to have auniform number of samples among all the heartbeattemplates of a record.

Hkil IHIM,

t- Tlkv t-T. tv t+ T,/2 t+TrV

Figure 3: Definition of an ECG heartbeat templatewhere r denotes a zero mean, unit variance Gaussian randomvariable. Mutation probabilities for weight and structurewere 0.005 and 0.001 in the experiments.

Heartbeat templates are processed by a dimensionalityreduction technique to reduce the amount of data to beprocessed each time. Principal components analysis (PCA)finds a projection that best represents the data in a leastmean-squared error sense for feature dimensionalityreduction [17]. The PCA reduces the original dimensions of

328

Wk = Wk +r (4)

Tk 1. r. I.:k+l

heartbeat templates (greater than 150) to three principalcomponents as the morphological features of the heartbeattemplate. The time interval Tk between neighboring R peaksgives useful information for the classification of abnormalheartbeat patterns. A feature vector is composed of threeprincipal components (xI, x-2, x3) as morphological featuresand the time interval Tk (x4) between the R peaks as atemporal feature. The 4-dimensional features serve as theinput to the BbNN for the classification of ECG heartbeatpatterns after proper normalization.

C. Evolution ofBbNNFor the BbNN training patterns we use the heartbeat

templates corresponding to the first five minutes of a record.The remaining templates of the record are used for testpatterns.A 2x4 BbNN was chosen for classification of ECG

heartbeat patterns. The GA optimizes network structure andconnection weights by maximizing the fitness function givenby:

Fitness PI P2I N n(i 2

N E (dJk Yjk )j=l k=l

premature, ventricular escape, right bundle branch block,and the fusion of ventricular and normal beats. Figure 4shows example waveforms of normal and two abnormalheartbeats (atrial premature and premature ventricularcontraction). There exists a morphological similaritybetween the normal and atrial premature pattern. Thetemporal feature gives higher discrimination of atrialpremature pattern since the time interval (652.8 ms) issmaller than that of normal pattern (813.9 ms). For apremature ventricular contraction pattern, the time intervalsof both patterns are similar (644.4 ms), while the shapes arequite different. A combined use of morphological andtemporal features provides successful classification ofabnormal heartbeat patterns.

(5)

where N denotes the number of training data, no is thenumber of actual output nodes. dJk and Yjk are desired andactual outputs of the kth output block referred tojth pattern.The term p, prevents an invalid internal block configurationwith all input nodes (4/0) or all output nodes (0/4), while P2excludes invalid structures that all the outputs of a middlestage are composed of only upward signal flow. Both termsare initially set to 1. An invalid configuration results in asmaller value of 0.9. Table I lists detailed parameter settingsused in the GA evolution.

TABLE 1: PARAMETERS FOR GA EVOLUTION

Parameter ValuePopulation 100Crossover Probability 0.30Mutation for Weight 0.005Mutation for Structure 0.001Max. Fitness Value 0.999Max. Generations 50

Tim:e (ins)

(a)

0;~200 a.O I2=2.00 o.:0 I40Z . I; ;t.0Time (ns)

(b)

Time (ms)

(C)

Figure 4: Sample ECG heartbeats, (a) Normal, (b) Atrialpremature, and (c) Premature ventricular contraction

IV. EXPERIMENT RESULTS

Six different types of abnormalities are considered: atrialpremature, premature ventricular contraction, aberrated atrial

Figure 5 shows a scatter plot of normal and abnormalheartbeats in terms of the first principal component (xi) andthe time interval (x4). The temporal feature increases theseparability of normal and abnormal patterns.

329

U fAX~~~~~~

iz

x

-4

xAbnorral beatso Norrral beats|

-12 -10 -8 -6 -4 -2 0 2 4 6

Xi

0

00

Ez

Figure 5: Scatter plot of normal and abnormal heartbeats interms of the first principal component (xi) and the time

interval (x4)

Figure 6 displays a typical evolution trend of BbNNs forECG heartbeat classification. The solid line indicates themaximum fitness trend and dotted line shows the averagefitness values. Evolution stops if the maximum fitness valuedoes not reach the limit of 0.999 in 50 generations.

5 10 15 20 25 30 35 40Generaton

Figure 7: A structure evolution trend

Figure 8 shows the structure and weights of the evolvedBbNN for classification of ECG heartbeat classification. Thelast output block was chosen as the output y. All the otherredundant output nodes marked as * are ignored. Inputsxl,..., x4 to the BbNN correspond to the three morphologicalfeatures and the temporal feature. The output indicates eithera normal or an abnormal heartbeat class.

6 10 15 20 26 30Generation

35 40 45 50

Figure 6: Evolution trend ofBbNNs

A near-optimal BbNN structure is found during theevolution process. In Figure 7, a BbNN with near-optimalstructure survives the competition. The number of BbNNswith a near-optimal structure increases, while the number ofall the other structures gradually decreases during theevolution. The solid line indicates the near-optimal structureand the dotted lines represent five non-optimal structures.Finally all the BbNN individuals are evolved to the near-

optimal structure.

Figure 8: Evolved BbNN for ECG signal classification after50 generations

Classification of ECG heartbeat patterns using the BbNNis summarized in Table 2. True positive (TP) and truenegative (TN) indicate the correct classification of normaland abnormal patterns. False negative (FN) refers tomisclassification of normal patterns as abnormality and falsepositive (FP) defines misclassification of abnormal patternsinto normality. Classification accuracy is then defined as theratio of the number of correctly classified patterns (TP andTN) and the total number of patterns. Each experiment was

averaged over five repeated trials. An average classificationrate obtained by the evolved BbNNs was approximately98%. A two-layer multiplayer neural network of the size 51-25-2 reported an average accuracy of 90% [6]. Although adirect comparison of the results is not possible because of

330

00

0°

IC3 .

#e. 4z>

10(D

U.0.7

0.6- 177 Max F* ess

the difference in the datasets selected, the evolved BbNNdemonstrates its potential for ECG signal classification.

TABLE 2: SUMMARY OF CLASSIFICATION ACCURACIES

Record TP TN FP - Accuracy

100 1799 30 0 72 96.21106 1234 460 0 1 99.94119 1296 364 0 0 100.00202 1789 25 46 10 97.01205 2120 54 24 2 98.82209 2100 348 25 45 97.22210 2001 174 20 8 98.73212 787 1482 9 6 99.34213 2135 405 86 73 94.11215 2661 131 2 0 99.93Total 17922 3473 ;212 217 98.03

V. CONCLUSION

This paper presents ECG signal classification withevolvable block-based neural network. Heartbeat signalsvary significantly depending on many factors such asindividual differences and time. A variety of heart problemsare responsible for abnormal heartbeat signals. Block-basedneural networks offer an evolvable ECG signal classifier thatcan change structure and internal parameters forpersonalized health monitoring system. The internalstructure and associated weights are optimizedsimultaneously with the GA. A 2x4 BbNN evolved with asubset of ECG signals produced an average of 98%classification accuracy for the ECG signals from the MIT-BIH database.

ACKNOWLEDGEMENT

This work was supported in part by the National ScienceFoundation under grant No. ECS-03 19002.

REFERENCES

Classification," Proc. Int'l. Conf on Engineering in Mledicineand Biology Society (EMBS'2001). pp.2006-2008, October2001.

[41 P. de Chazal and R. B. Reilly, "A Comparison of the ECGClassification Performance of Different Feature Sets,"Computers in Cardiology, Vol. 27, pp.327-330, 2000.

[51 S. Osowski, L. T. Hoai, and T. Markiewicz, "Support VectorMachine-based Expert System for Reliable HeartbeatRecognition," IEEE Trans. on Biomedical Engineering, Vol.51, No. 4, pp.582-589, April 2004.

[6] Y. H. Hu, W. J. Tompkins, J. L. Urrusti, and V. X. Afonso,"Applications of Artificial Neural Networks for ECG SignalDetection and Classification," Journal of Electrocardiology,Vol. 26. (Suppl.), pp.66-73, 1994.

[7] M. Lagerholm, C. Peterson, G. Braccini, L. Edenbrandt, andL. S6rnmo, "Clustering ECG Complexes Using HermiteFunctions and Self-organizing Maps," IEEE Trans. onBiomedical Engineering, Vol. 47, No. 7, pp.838-848, July2000.

[8] S. W. Moon and S. G. Kong, "Block-based Neural Networks,"IEEE Trans. on Neural Networks, Vol. 12, No. 2, pp.307-317,March 200 1.

[91 S. W. Moon and S. G. Kong, "'Pattern Recognition withBlock-based Neural Networks," Proc. Int'I J. Conf on NeuralNetworks (IJCNN-2002), pp.992-996, May 2002.

[101 S. G. Kong, "Time Series Prediction with Evolvable Block-based Neural Networks," Proc. Int'l J. Conf on NeuralNetworks (IJCNN-2004), July 2004.

[11] T. Back, "Generalized Convergence Models for Tournament-and (j, X)-Selection," Proc. Int'l Conf on Genetic Algorithms(ICGA-95), pp.2-8. 1995.

[121 T. Blickle and L. Thiele, "A Mathematical Analysis ofTournament Selection," Proc. Int'l Conf on GeneticAlgorithms (ICGA-95), pp.9-16 1995.

[13] B. A, Julstrom, "It's All the Same to Me: Revisiting Rank-based Probabilities and Tournaments," Proc. Congress onEvolutionarv Computation (CEC-99), Vol. 2, pp.1501-1505,1999.

[14] J. Sarma and K. De Jong, "An Analysis of Local SelectionAlgorithms in a Spatially Structured Evolutionary Algorithm,"Proc. Int'l Conf on Genetic Algorithms (ICGA-97), pp.181-186, 1997.

[15] R. Mark and G. Moody, MIT-BIH Arrhythmia DatabaseDirectory (http://ecg.mit.eduldbinfo.html).

[16] Association for the Advancement of Medical Instrumentation,Recommended Practice for Testing and ReportingPerformance Results of Ventricular Arrhythmia DetectionAlgorithms, 1987.

[171 R. 0. Duda, P. E. Hart, and D. G. Stork, Pattern Recognition,2nd Edition, New York: Wiley-lnterscience, 2000.

[1] P. de Chazal, M. O'Dwyer, and R. B. Reilly, "AutomaticClassification of Heartbeats Using ECG Morphology andHeartbeat Interval Features,"' IEEE Trans. on BiomedicalEngineering, Vol. 5 1, No. 7, pp. 1 196-1206, July 2004.

[2] C. Alexakis, H. 0. Nyongesa, R. Saatchi, N. D. Harris, C.Davis, C. Emery, R. H. Ireland, and S. R. Heller, "FeatureExtraction and Classification of Electrocardiogram (ECG)Signals Related to Hypoglycaemia," Computers inCardiology, Vol. 30, pp.537-540, 2003.

[3] 1. Romero and L. Serrano, "ECG Frequency Domain FeaturesExtraction: A New Characteristic for Arrhythmias

331