=Paper=
{{Paper
|id=None
|storemode=property
|title=Comparison of ECG Beat Classification methods on a Mobile System
|pdfUrl=https://ceur-ws.org/Vol-729/paper6.pdf
|volume=Vol-729
}}
==Comparison of ECG Beat Classification methods on a Mobile System==
https://ceur-ws.org/Vol-729/paper6.pdf
Comparison of ECG Beat Classification methods on a
Mobile System
Fernando Arena Varella1,2 Guilherme L. de Lima1 Cirano Iochpe1 Valter Roesler1
favarella@inf.ufrgs.br gllima@inf.ufrgs.br ciochpe@inf.ufrgs.br roesler@inf.ufgrs.br
1 Universidade Federal do Rio Grande do Sul
2 I9Access Tecnologia Ltda
Av. Bento Gonçalves, nº 9500, Setor 4 – Prédio 72 – Sala 211 - Agronomia –
Porto Alegre/RS - Brasil
ABSTRACT including Brazil where more than 30% of the deaths are
The constant evolution of mobile technologies has lead to several related to them [1]. An important characteristic of some
systems with focus on the mobile electrocardiogram (ECG) heart disease treatments is the need for constant monitoring
telemonitoring of patients. They usually present a client-server of the patient´s heart beat waves. This monitoring is
architecture, where a web server centralizes the storage, typically performed by either an electrocardiographer or a
management and processing of the patient’s signals, and a mobile holter. More important yet, in emergency situations, it is
phone is responsible for acquiring and transmitting the data over crucial the fast access to the patient’s vital signals by the
the internet (WAP or 3G) to the server. However, there are
countries, like Brazil, whose mobile network services do not
physician, allowing him or her to take preventive or
present full coverage, even in large urban areas. This leads to a reactive actions in order to save the patient’s life.
lack of security under certain situations like emergency calls or The evolution of telecommunications helped the
critic patient’s monitoring when the patient is in an area without
development of several systems with focus on the ECG
mobile network coverage. In these situations, the patient’s vital
signals must be analyzed as soon as possible, in order to allow fast
telemonitoring of patients, allowing them to stay at home,
preventive or reactive actions to be taken. An alternative way to instead of in a hospital in various situations. It helps to
deal with this problem is to support the analysis of vital signals in reduce the high cost with patients internment and transport
a mobile phone which can be either offline or communicating and to shorten the long waits for hospital beds. Another
through a cell phone network. Though, since mobile phones, usage for the ECG telemonitoring is in critical situations
compared to PCs, have slower CPUs and less internal memory, like strokes or heart stops, when the patient’s signals must
it’s important to select an ECG wave analysis method that is not be analyzed as soon as possible. Most of the systems used
costly with respect to computation time and memory usage. In this in such situations rely on a mobile phone to transmit the
paper, we present comparison study made in order to select a
data acquired by the electrocardiographer to a central
reliable method for ECG beat classification running in a mobile
phone. Three ECG beat classification methods were selected to be
server, where it can be analyzed by the physicians almost in
analyzed and implemented in a mobile phone. Tests were made in real time. However, in countries like Brazil, where mobile
regions with limited cell phone network coverage in south Brazil. network coverage is not ideal, even in large urban areas
Test results were compared in order to choose a method whose with dense people concentration, this strategy can be risky
implementation is efficient enough to run in a mobile phone and to the patients since network communication is not reliable
can achieve high accuracy on the classification of the ECGs, even in a number of regions. To deal with this problem, we
if running in a phone with limited CPU as well as memory suggest to perform the ECG analysis through an application
resources. running in the cell phone itself when it is impossible to
Keywords transmit the ECG waves to the remote server. This could
make the overall system fail-safe with respect to the lack of
Remote Home Care, ECG Beat Classification, Mobile good network coverage.
Systems, Emergency ECG Monitoring
Many computational methods have been proposed to
1. INTRODUCTION automate the process of ECG waves analysis, helping
Heart and cardiocirculatory diseases are the most common physicians to diagnose anomalies on ECG exams. Although
causes of death in the majority of western countries, these methods have achieved very good performances,
many of them rely on heavy computational algorithms and
can be very costly in terms of CPU and memory usage.
Considering this, we present, in this paper, a comparison of
some of the most reliable ECG beat classification methods.
Our main objective is to point out one or more methods that
can successfully classify ECG beats with high accuracy and
fast response even when implemented and running in a
mobile phone with limited CPU as well as memory the classification task and has its feature extraction step
resources. based on the DWT. In this case, the data is first classified
by a set of MLPs, and then classified again by a final MLP.
The next sections are organized as follows. Section 2
Yu and Chen [5] also used the DWT for the feature
introduces the ECG beat classification problem and the
extraction, but the classification step is made by a
most commonly used techniques to solve it. Section 3 is
Probability Neural Network. On their other work [6], they
reports on the implementation of some classification
allied the DWT with the Higher Order Statistics (HOS) to
methods in mobile devices. In section 4 we comment on the
build the feature set, and used a MLP as the classifier. Yu
results obtained in the evaluation of the implemented
and Chou [7] introduced a switchable scheme for the
methods. Finally, in Section 5 we present our conclusions
classification task. They have 2 feature sets, both extracted
and point out to future work.
using the DWT allied to the Independent Component
2. ECG ANALYSIS BACKGROUND Analysis (ICA). Depending on some characteristics of the
The automatic ECG classification methods apply feature sets, they switch between the minimum Euclidean
techniques from several computing areas, like Distance classifier, minimum Mahalanobis Distance
computational statistics, pattern recognition, support classifier or a Bayesian classifier. Finally, Minhas and Arif
vectors machines, etc. However, there is a basic set of [8] developed a method similar to the Chen and Yu’s
common steps between these methods: Beat Detection, method, they rely on the HOS and the DWT to build the
Feature Extraction and Beat Classification. The next feature set and perform the classification with the K-
sections will introduce the ECG classification process and Nearest Neighbor (KNN) Neural Network.
review the most used techniques. However there is a basic set of common steps between
2.1 ECG Analysis Process these methods: Beat Detection, Feature Extraction and
Classification.
The Beat Detection step is usually made using common
annotated ECG database, easing the detection process and The Beat Detection is a signal processing consisting of
unifying the data sets used amongst the majority of the QRS complex detection. The QRS complex detection
authors. When applied to real ECGs, this stage must be methods already have excellent solutions and are not
made by a QRS detector algorithm, like the Pan Tompkins addressed in this work. What is true highlight is that the
method [2]. following step will process only segments of the QRS
complexes. The feature extraction stage consists in
The Feature Extraction follows the beat detection, and it is
transforming the signal into a suitable format to perform
responsible for transforming the raw ECG signal, mapping
the classification. Finally we perform classification of
the original values to meaningful features. This is a very
extracted features in a certain class amongst the classes that
important stage, and the most successful classification
the method is able to identify, such as normal beat,
methods rely on a good feature extraction stage to achieve
pacemaker beat, etc.
high performances.
Finally, the Beat Classification stage will be performed 2.2 Materials and Methods
over the features extracted on the second stage. The authors The three methods that this work analyzes uses discrete
employ several classification and pattern recognition wavelet transform together with mathematical procedures
techniques, allowing them to distinguish the type of each to extract the features that will be classified.
beat that is being classified. 2.2.1 Discrete Wavelet Transform
Many authors have already employed these steps on the The wavelet functions are widely used in signal processing.
past, building classifiers that can achieve high accuracy While the Fourier Transforms work only in frequency
rates on the classification task. We analyzed the results and domain, the wavelet analysis works also in the time
implementation of many methods, aiming the selection of domain, a fact which facilitates the analysis of local
one of them for the usage on a mobile system. Thus, the characteristics of signal, what is very important for
method must be efficient in time and in memory usage, at classification of ECG beats. Both the Fourier analysis and
the same time as it must achieve high accuracy to guarantee wavelet analysis intend to approximate a signal by a linear
that it can improve the capacity of the system to save combination of sines and cosines (Fourier analysis) or
human lives. wavelets. The approximation of the signal in wavelet
Mehmet Engin [3] developed a method that unifies a analysis is performed by a process called wavelet transform
Multilayer Perceptron (MLP) Artificial Neural Network which has the continuous and discrete variants.
(ANN) with the fuzzy c-means (FCM) clustering algorithm, The continuous wavelet transform (Continuous Wavelet
forming a neuro-fuzzy network. His method uses the Transform - CWT) of a signal f is the inner product of the
discrete wavelet transform (DWT) to extract the feature set. signal by a wavelet, in this case called mother wavelet ψ, as
Similarly, Güler e Übeyli [4] proposed a method that uses a it appears in the following formula:
set of small MLPs combined with a larger MLP to perform
The variance is defined like the average of the square of the
distance of each value from the mean, it shown the
difference of a value in relation to the mean value. Can be
Where: expressed by the following formula:
The mother wavelet ψ is given by the following formula:
The Discrete Wavelet Transform (DWT) of a signal From the variance we can calculate the standard deviation
expressed as a vector of samples X = (x0, x1,, …, xT)t can which is defined as the square root of the variance, or even
be expressed as: the average distance of each value to the average, which
shows how the sample is distributed to the extent that
deviates from the mean.
Another important measure is the auto correlation, which is
considered a measure of similarity between a sample and
However, in practice, this turns into the use of a pyramidal its shifted version. The auto correlation can be expressed as
algorithm, where a series of low pass and high-pass filters follows:
are applied to the signal. The Pyramid algorithm that
computes the wavelet transform was proposed by Stephane
G. Mallat [11], and is shown graphically in Figure 1.
where x(n) is a sample of size N. This calculation tells how
much a certain value influences the neighborhood values,
i.e., how related are the values of a sample.
We still have the relative amplitude, which is nothing more
Figure 1. Two levels of the Mallat’s DWT
than the ratio between the minimum and maximum value
The algorithm uses a low-pass filters l and h high-pass to for a given sample, which shows morphological
calculate the approximation and detail coefficients, characteristics of the sample.
respectively. On the first level, the approximation
coefficients are calculated by the convolution of the signal 2.2.3 Principal Components Analysis
with the low-pass filter, and the detail coefficients by the The Principal Component Analysis (PCA) reduces the
convolution with the high-pass filter. After this process, the dimension of matrices building a smaller matrix only with
resulting coefficients are downsampled, since the the principal information of the original one. For example,
convolution modifies the size of the resulting signals. For considering the training and testing sets that will be
every new level of coefficients, the approximation classified, it is possible to reduce a 10-dimensional feature
coefficients from the previous level are taken as the set to a 5-dimension, extracting the higher five principal
original signal, and the same process is applied. components. The process to find principal components is
The filters depend on the type of wavelet-mother used on the process to find eigenvectors of a covariance matrix R
the transform. The most commonly used is the Haar formed by the feature set. The covariance matrix R is given
wavelet, and its low-pass and high-pass filters are shown by:
below:
h = [0.70710678118654757, 0.70710678118654757]
l = [0.70710678118654757, 0.70710678118654757] The eigenvalues and eigenvectors are the obtained solving
Some authors use the à trous wavelet transform [12], the eigenvalue problem. For this problem, ten eigenvalues
whose main difference to the Mallat’s algorithm is the lack and ten eigenvectors are obtained, the five eigenvectors
of the downsampling after the application of the filters. corresponding to the highest five eigenvalues will be used
This difference helps to preserve the temporal resolution to build a new feature set.
and the time-invariant property of the signal on different The new feature set is obtained by the multiplication of
scales [13]. each sample by the five eigenvectors, yk:
2.2.2 Statistical Measurements
Several statistical measurements are used by the ECG beat 2.2.4 Higher-order Statistics - Cumulants
classification methods, including the variance, standard A cumulant is a Higher Order Statistics measure that is
deviation, auto correlation and relative amplitudes. commonly used in physics. They can be seen as a set of
quantities that provide an alternative to the moments of the
signal, providing useful information about the probabilities it depends on the presence of the training instances on the
distribution of the signal [14]. classification stage. The network’s topology is composed of
four layers: Input Layer, Pattern Layer, Summation Layer
and Output Layer. That topology can be seen in Figure 2.
Figure 4. Probabilistic Neural Network Topology
The input layer distributes the samples to neurons of
Figure 2. 200 samples of a NORMAL beat: (a) original pattern layer, and all entries are propagated to all the
sample, (b) 2nd order cumulants, (c) 3rd order cumulants neurons of layer pattern. Each sample of the training set
and (d) 4th order cumulants corresponds to a unit-layer pattern, thus, the hidden layer
will have the same number of neurons as the size of the
training set. Each hidden layer neuron has a multi-
dimensional Gaussian kernel, capable of estimating the
probability density function (PDF) of the input relatively to
that pattern. The summation layer has the same number of
neurons as the number of classes on the training set, and
each neuron in this layer responsible for accumulating the
PDFs of all neurons from the pattern layer that have the
same class.
The classification phase of this neural network is simple,
yet costly. Each input from the training set will have its
PDF estimated for every neuron from the Pattern Layer.
The next layer, the Summation Layer will compute the
probability of the sample of being of each class
accumulating the values from the pattern layer and
grouping them by class. Finally, the Output Layer will
select the neuron with a higher value from the Summation
Layer to be the class of the sample.
To estimate the probability density function of a given
Figure 3. 200 samples of a Left Bundle Branch Block sample is performed using the kernel density estimation
beat: (a) original sample, (b) 2nd order cumulants, (c) 3rd approach, also called Parzen window, which is basically a
order cumulants and (d) 4th order cumulants technique of data interpolation. Given an instance x, the
As it can be seen on Figure 2 and Figure 3, the employment Parzen window PDF estimates b fx(x) for each sample data
of the cumulants on the task of classifying ECG beats can set. As the following expression:
help to remove the variability between beats from the same
type, and to enhance the differences between beats of the
different types.
2.2.5 Probabilistic Neural Network where K is the kernel function, h is a smoothing parameter
and n is the number of samples in the data set.
The Probabilistic Neural Network (PNN) is a neural
network used primarily as a classifier, but it can also be The kernel function most used is a Gaussian that can be
seen as an alternative way to implement a statistical represented by:
algorithm called discriminant [5]. It is instance-based, since
Thus, the expression (1) is denoted by:
Finally, the PNN has fast training, the training set is used
directly in the kernel, and converges to the optimal
Bayesian decision surface.
2.2.6 Nearest Neighbor
The K Nearest Neighbors algorithm calculates the
euclidean distance of a testing sample with all the training
samples. The input sample is classified with the same class
of the mode class of the k nearest neighbor according to the Figure 6. Multi-Layer Perceptron Topology
euclidean distance [8]. The pseudo code is described in the The input layer propagates the input values to the hidden
figure 3. layer, multiplying the value of each neuron with the
respective weight. The final value on the hidden layer
neurons are obtained applying the activation function on
the summation of all weighted values. The value of the
output layer neurons are calculated with the same
procedure, however, this value is compared with the
desired value from the training set, and the difference (the
error) is backpropagated to correct the neuron weights for
all layers.
One efficient method to backpropagate the error,
minimizing the overall error of the network, was proposed
Figure 5. KNN pseudo algorithm
by Broyden-Fletcher-Goldfarb-Shanno (BFGS) as a generic
The Euclidean distance is defined as the distance between optimization algorithm. It is a quasi-Newton non-linear
two points, is based on the Pythagorean formula, can be method for optimization that transforms the error problem
expressed like the square root of the square of the on a minimization problem solving the hessian matrix
difference between the points: approximation problem [15]
3. THE IMPLEMENTATION OF THE
METHODS
The sort procedure only order the distances vector, from the Although in the previous sections we presented only a
smaller to the bigger distance. The mode procedure subset of the methods that have been proposed in the
chooses, between the k smaller distances, the mode class literature, we have chosen these methods relying on their
point. accuracy and the sensitivity rates. Thus, the most efficient
2.2.7 Feed forward Back propagation Neural methods are covered by this study.
Network (MLP) Method # beats Accuracy Lesser
The Multi-Layer Perceptron (MLP) is a neural network that (%) Sensitivity
is usually used to solve classification problems. A common Engin 4 98,0 95,3
type of MLP is the Feed forward Back propagation Neural
Güler e Übeyli 4 96,94 95,56
Network (FFBNN), whose name is given because the input
is forward propagated and the errors are back propagated to Yu and Chen 6 99,65 99,04
correct the neurons weights. Yu and Chou 6 99,51 98,24
To explain the behavior of the FFBNN, consider the Chen and Yu 6 99,70 98,88
topology described in figure.
Minhas and Arif 6 99,49 98,92
Table 1. Accuracy and sensitivity rates for the ECG
Beat classification methods [3][4][5][7][6][8]
Table 1 shows the accuracy of the methods, and the lowest
sensitivity rate between the sensitivities for all beats. The
latter statistical measurement is important because it
assures that all kinds of ECG beats will be properly
classified, avoiding the overestimation that the accuracy
measure can cause. Considering these facts, we decided to For each sample that needs to be calculated, the algorithm
choose three methods to compare and implement on a needs to compute its PDF relative to each instance from the
mobile device, they are: Chen and Yu method [6], Minhas training set. After this phase, which is made on the Pattern
and Arif method [8] and Yu and Chen method [5]. Layer, the Summation layer will sum the results of all
The selected methods were implemented in the Java neurons from the Pattern Layer, grouping by the class of
Language because it can be easily ported to the mobile the neuron and making a weighted average of the value.
phones. With little adaptations, the methods were capable Finally, the Decision Layer will classify the instance with
of executing on PCs and on a mobile phone with the the class that had the higher average on the Summation
Google’s Android Operational System, allowing us to Layer.
analyze the results on a real mobile phone. The mobile 3.2 Minhas & Arif method
phone used for the tests was a Motorola Milestone A853, The method developed by Minhas and Arif [8] is similar to
which has 256Mb RAM memory, an Arm Cortex A8 CPU, the Yu and Chen method, they use the same statistical
which has a frequency of 600MHz and runs the version 2.2 measurements to build the feature set, but there are some
of the Android system. differences on the DWT and on the classifier. The authors
The following subsections will review the selected also apply a resampling on the records from 360Hz to
methods. 250Hz.
3.1 Yu & Chen method 3.2.1 Feature Extraction
The method proposed by Yu and Chen [6] is based on a To perform the feature extraction the à trous DWT, using
Probabilistic Neural Network (PNN) to classify up to six the Quadratic Spline Wavelet as the mother wavelet, is
different ECG beat types – including Normal beat (N), left applied to the samples.
(L) and right (R) bundle branch block beats, premature The first and second level detail coefficients, and the
ventricular contraction (PVC), atrial premature beat (APB) second level coefficients are selected as a base for the
and paced beat (P). The features are extracted calculating feature set. From them, the same statistical features as the
statistical measurements from the DWT coefficients of the FS2 from the Yu and Chen method are calculated.
raw signal.
The authors also employed the Principal Component
3.1.1 Feature Extraction Analysis to generate a second feature set, with a reduced
The DWT is applied on the raw ECG signals to perform the dimensionality. However, this procedure also reduced the
extraction of the features. From the first and second level accuracy of the method, thus, we will focus only on the
wavelet (WT) detail coefficients (D1 and D2), and from the original feature set, which contains 11 features for each
second level approximation coefficients (A2) several sample.
statistical measurements are calculated, including the 3.2.2 Classification
variance of the coefficients, the variance of the auto Minhas and Arif employed a k-Nearest Neighbor algorithm
correlation function of the coefficients, the ratio between to classify the testing set. This classifier is extremely easy
the maximum and minimum values and the interval to implement, since it’s depends only on a ordered list of
between the current beat and the last one. distances between the sample that is being classified and all
The authors proposed two feature sets, namely: FS1 and other samples from the training set. The Euclidean Distance
FS2. FS1 includes the variance of the original signal, the was the selected distance metric, and the k parameter was
variance of D1, D2 and A2, the variance of the set to 3.
autocorrelation of D1, D2 and A2 and the ration between 3.3 Chen & Yu method
the maximum and minimum values of D1, D2 and D3, for a The method proposed by Chen and Yu uses Higher Order
total of 10 features. FS2 contains all the features of FS1 Statistics with the DWT to build a large feature set for each
plus the instantaneous RR interval, since it is an important ECG beat, what is classified by a Multilayer Perceptron.
information for arrhythmias. The classifier covers the same beat types set as the Yu &
After the feature extraction, each feature is normalized, Chen method.
using the hyperbolic tangent sigmoid function to map the 3.3.1 Feature Extraction
values of each feature to the range [-1, +1]. To be successful on the difficult task of finding a reliable
3.1.2 Classification feature set the authors employed Higher Order Statistics
Yu and Chen use a PNN to perform the classification of the allied to the DWT.
beats. All the training instances must be loaded on the Three higher order statistics features are selected to be help
network at the moment of the classification, thus, this to build the feature set, namely: second order cumulant,
method is very costly in terms of memory usage. The fact third order cumulant and fourth order cumulant. The
that the PNN is instance-based also implies that the method authors use the à trous DWT, with the ‘sym6’ as the
is costly in terms of time, because it needs to process the mother wavelet, to calculate five levels of coefficients. The
whole training set to classify each testing instance. 2nd, 3rd and 4th order cumulants are extracted from the detail
wavelet coefficients of level 3, 4 and 5. As a result, nine to achieve a good performance, we used 11,600 instances
cumulants are calculated for each beat, and from them on each set.
several features are calculated, including: Cumulant Our implementation of the Chen and Yu method is not fully
Variance and Normalized Summation for each cumulant, functional yet. We need to review the feature extraction
the Number of Zero-Crossings on the cumulants of the phase in order to achieve better results. In our experiments,
level 5 coefficients and the Symmetry of the 3rd and 4th it achieved at most 89.52%, while the authors reported an
order cumulants. Additionally, three RR-interval related accuracy of 99.70% in their work.
features are defined, including: current RR interval,
previous RR interval and the ratio between the current and Statistical Measurement Minhas Yu and
the previous RR interval. In summary, the feature set is and Arif Chen
formed by 30 features. Accuracy % 99.0 98.67
3.3.2 Classification
Specificity % 99.81 99.5
The classification stage of this method is performed by a
MLP, which is a neural network composed by, in this case, Sensitivity % LBBB 98.29 97.58
three layers of neurons. The first layer only propagates the Sensitivity % RBBB 99.04 97.79
inputs signal, while the hidden and output layers will
Sensitivity % PVC 97.91 98.87
combine the inputs with a set of neuron weights and apply
an activation function to compute the output. The output is Sensitivity % APB 98.00 98.82
compared to the desired output and an error value is Sensitivity % PB 99.67 99.83
produced. This error is used to adjust the previous neurons
weights in a process called Backpropagation training. Table 3. Comparative results of the Minhas and Arif
and Yu and Chen methods
The authors empirically defined that the hidden layer is
formed by sixty neurons, and the activation function of the Our experiments for the Minhas and Arif method and for
network is the hyperbolic tangent sigmoid. The the Yu and Chen methods were successful on the PC and
Backpropagation training uses the BFGS method to on the mobile phone, achieving an accuracy of 99.0% and
minimize the error. 98.67%, respectively.
Due to the heavy computations needed by the classification
4. RESULTS ANALYSIS phase of the PNN and the KNN, the mobile phone needed
To evaluate the behavior of the implemented methods we too much time to classify each beat. For the Yu and Chen
used 23 records from the MIT/BIH arrhythmia database method, the PNN classifier needs to compute the PDF for
[9], which consists of a large database of annotated each of the 11,600 neurons from the Pattern Layer, perform
electrocardiograms containing several types of arrhythmias. the summation and chose the desired class. The selected
We used the same dataset to test all the implemented mobile phone (Motorola Milestone A853) took an average
methods in order to avoid overestimating of any of them. of 623ms to perform this task for each beat. The classifier
Type MIT/BIH Record # Training # Testing used by the Minhas and Arif method is the KNN, a
classifier that is similar to the PNN, however, instead of
N 103, 113, 115, 123, 600 600 finding the PDF for each neuron, it needs to compute the
220, 234 distance of the beat to each other beat from the training set
LBBB 109, 111, 207, 214 600 600 and select the class of the closest class to be the target class
RBBB 118, 124, 212, 231 600 600 of the beat. The average time taken to classify each beat
with this method was 573ms.
PVC 119 200 200
On an implementation that targets the monitoring of real
221 150 150 patients, the CPU will also be busy acquiring and decoding
200, 233 400 400 the patient’s signal from the electrocardiographer, detecting
APB 209 150 150 the QRS complexes to find the bounds of each beat and
extracting the features from the beat. As a result, all these
222 100 100 operations, together with the classification phase, can take
232 600 600 more than 800ms to be completed for each beat. A
107, 217 600 600 traditional heart rate of 75bpm will generate 1 beat each
PB
800ms, in other words, on the selected phone, these
Total 11600 11600 methods would not be able to monitor patients with heart
Table 2. The records selected from the MIT/BIH rates higher than 75bpm, turning their use impracticable in
database, grouped by the beat type. a real system.
Table 2 shows details of the records from the MIT/BIH In contrast to them, the Chen and Yu method could be able
database that were used to form the training and testing to classify each beat in a small time. Due the nature of the
instance sets. Since neural networks need large training sets MLP classifier, that just needs to propagate the inputs
through the network to find the target class, it achieved an http://www.who.int/entity/healthinfo/statistics/bodgbd
average of 27ms to classify each beat. However, as we have deathdalyestimates.xls.
already mentioned, our implementation of the method still [2] Pan, J.; Tompkins, W. J. 1985. A Real-Time QRS
needs to be reviewed in order to achieve the accuracy Detection Algorithm. IEEE Trans. Biomed. Eng.,
reported on the author’s paper. BME-32 (3):230-236.
5. CONCLUSION AND FUTURE WORK [3] Engin, M. 2004. ECG beat classification using
In this paper, we presented a comparison study of ECG neuro-fuzzy network. Pattern Recognition Letters, v.
beat classification methods destined to detect arrhythmias 25.
running in a mobile phone. This study was motivated by [4] Güler, I.; Übeyli, E. D.2005. ECG beat classifier
the need for a rapid response for the analysis of the exams designed by combined neural network model.
in some critical situations. Thus, our main objective was to Pattern Recognition.
point one or more methods that can successfully classify [5] Yu, S.-N.; Chen, Y.-H. 2007. Electrocardiogram beat
ECG beats on a mobile phone, with high accuracy fast classification based on wavelet transformation and
response. probabilistic neural network. Pattern Recognition
Relying on the specialized literature, we selected three Letters.
methods from the authors that reported the higher [6] Chen, Y.-H.; Yu, S.-N. 2007. Subband Feature
accuracies together with high sensitivities, namely: Chen Based on Higher Order Statistics for ECG Beat
and Yu method, Minhas and Arif method and Yu and Chen Classification. Proceedings of the 29th Annual
method. International Conference of the IEEE EMBS. Lyon,
As we reported in Section 4, our implementation of the France.
Chen and Yu method is not fully completed yet. By now, it [7] Yu, S.-N.; Chou, K.-T. 2007. A switchable scheme
achieves at most an accuracy of 89.52%. On the other hand, for ecg beat classification based on independent
our implementation of both the Yu and Chen method, and component analysis. Expert Systems with
the Minhas and Arif method were successful for both the Applications.
PC and the mobile phone, achieving accuracies of 98.67% [8] Minhas, F.-U.-A. A.; Arif, M. 2008. Robust
and 99.00%. This performance could be observed in both electrocardiogram (ECG) beat classification using
implementations, for the PC and for the mobile phone. discrete wavelet transform. Physiological
However, these two methods, when running in the mobile Measurement.
phone were too slow to classify the beats. Both of them
need more than 500ms. This is indeed a problem, since [9] Mark, R., Moody, G. 1988. MIT-BIH Arrhythmia
normal heart beat rates generate more than one beat per Database Directory. (Cambridge, MA: MIT Press).
second, and fast heart rates will generate more than one [10] Figueredo, M. V. M.; Dias, J. S. 2004. Mobile
beat per 500ms. Thus, both algorithms would not be able to Telemedicine System for Home Care and Patient
process all beats when running in a cell phone, at least in Monitoring. Proceedings of the 26th Annual
our implementation. International Conference of the IEEE EMBS.
5.1 Future work [11] Mallat, S. G. 1989. A theory for multiresolution
To overcome the performance limitations found in our signal decomposition: The wavelet representation.
study so far, we will focus on the implementation of the IEEE Transactions on Pattern Analysis and Machine
Chen and Yu method in order to be able to test it in a Intelligence
mobile phone. Since the MultiLayer Perceptron’s [12] Cohen, A., Kovacevic, J. 1996. Wavelets: the
classification phase is not instance-based, it is not costly in mathematical background. Proc. IEEE 84 514–22.
terms of CPU and memory usage. If we fail to fix our [13] J. P. Martίnez, R. Almeida, S. Olmos, A. P. Rocha and
implementation of the Chen and Yu method, we’ll try to P. Laguna, 2004. A wavelet-based ECG delineator:
develop a novel ECG beat classification method to run with evaluation on standard databases. IEEE Trans.
acceptable performance in a cell phone implementation. Biomed. Eng.
The new method will have to combine a fast classification [14] Osowski, S.; Linh, T. H. 2001. ECG Beat
algorithm with a feature set that provides better results. Recognition Using Fuzzy Hybrid Neural Network.
IEEE Transactions on Biomedical Engineering, v. 48.
6. REFERENCES [15] Flecher, R. 1980. Practical methods of optimization.
[1] WHO - World Health Organization. 2004. Table 3. Section 3.2, p38.
Estimated deaths per 100,000 population by cause,
and Member State. World Health Organization.
Available on: