Printer Friendly

A novel ECG signal processing and abnormal heart beat classification algorithm based on particle swarm optimization (PSO) model and the neural network.

1. Introduction

The ECG analysis is the important means of current examination heart disease. The bio-electrical signal of human body is low frequency faintness signal under the strong noise background (Alickovic, 2015; Bortolan, 2015). In the ECG collection process, how to take the measure to filter the industrial frequency interferences of 60 Hz, the baseline wander caused the signal distortion disturbing wave to decide the final results analysis fit and unfit quality (Civicioglu, 2013). At the same time, in order to record enough useful signals, the ECG signal must be taken to a certain degree of compression processing, in the limited storage space to store more useful information to meet the actual clinical needs and reduce hardware costs (Chen, 2013; Oster, 2015). Briefly, the electrocardiogram is the comprehensive electricity activity of the myocardial cell in the one reflection of body surface (Mukhopadhyay, 2013). Small electric current that the cardiac muscle activity produces may conduct the body surface after the tissue, making the different parts of the body surface have the different electric potentials (Gradl, 2012; Huang, 2015), records in the electric potentials body surface two points which will form a continual curve according to the heart excited time sequence, this is the electrocardiogram (the basis of the ECG). ECG signal is very weak it is very easy to be submerged in a strong noise signal. The traditional method of using the various filtering methods to process the ECG signal (Gawali, 2014; Ren, 2014), but this method in the short term low energy transient signal, after smoothing filter, not only the signal-to-noise ratio is not improved, but the signal position has also been blurred, these are fuzzy location information off transient it is very important for the analysis of ECG.

As shown in the figure one, we demonstrate the systematic architecture of the ECG signal processing extraction model. In order to enhance the signal-to-noise ratio, often uses the correlation method. The correlation method not only can examine the QRS wave, but can also withdraw QRS wave under the arrythmia but it floats to the high frequency noise and base is very sensitive (Garg, 2013). If the form changes, it will not be best, but the ECG has the obvious time-variable and non-linear response that is very big regarding time-varying signal its profile and template difference with the possible omission, and therefore it has not grasped fully to treat examines signal certain priori information under the foundations, this method lacks the validity (Wang, 2014; Lino et al., 2016).

To deal with this drawback, in this paper, we conduct research on the novel ECG signal processing and the abnormal heart beat classification algorithm based on the particle swarm optimization model and the neural network.

2. The Particle Swarm Optimization (PSO) Model

Regarding the biological population, the individual can gains the information in the process of search food from the experiences and experiences of other community members, when search unknown, the uncertain fragmentary distributed food and the superiority of this cooperation behavior is much greater than the pressure of the competition. Particle swarm optimization is a kind of evolutionary computation technology based on swarm intelligence. Its idea comes from the theory of artificial life and evolutionary computation (Rai, 2013; Duan, 2013). The PSO algorithm is one of colony intelligence optimal algorithm main representatives, because its principle and mechanism are simple and does not need the gradient information, the adjustable parameter are few, the algorithm is easy programming implementation, and operating efficiency is high, the convergence rate is fast (Yu, 2013). Therefore the PSO algorithm since publication has caught wide attention. After several years of research and the development, the PSO algorithm had been widely used in the function optimization, multi-objective programming and the neural network training, the fuzzy system control and other domains with satisfactory and result and has put forwarded enough modifications.

The basic concept of PSO stems from the study of bird predation behavior, people from the bird prey model was enlightened, and as used to solve the optimization problem. In PSO, the solution to each optimization problem is a bird in the search space which is called a particle. All particles have a fitness value that is determined by the optimized function, and each particle has the velocity that determines the direction in which they fly and the distance. PSO indicates each possible solution for a particle in group, each particle has own position vector and velocity vector, with one by the sufficiency of objective function decision, makes various granules swim to the sufficiency function value highest direction group through the similar gradient descent algorithm. Each particle evaluates the merits of current position of the particle by an optimization function related to the objective function, and then iteratively finds the optimal solution or approximate optimal solution and in each iteration the particle updates its speed and position by tracking two "extremes".

The PSO algorithm can be described as follows. Suppose we are in the n dimension searching space, the group X = {[x.sub.i]} is consists of m particle, the i - th particle's location and velocity can be defined as formula 1 and 2, respectively.

[x.sub.i] [location] = [{[[x.sub.i1], [x.sub.i2], ..., []}.sup.T] (1)

[v.sup.i] [vilocity] = [{[v.sub.i1], [v.sub.i2], ..., []}.sup.T] (2)

Each iteration of the particle update itself by updating the individual extremum and the global extrema. The update paradigms are defined as follows.

[v.sub.i] (t+1) = w x [v.sub.i] (t)+[c.sub.1][r.sub.1] ([p.sub.i](t) - [x.sub.i](t))+[c.sub.2][r.sub.2] x ([p.sub.g](t) - [x.sub.i] (t)) (3)

[x.sub.i] (t + 1) = [x.sub.i] (t) + [v.sub.i] (t + 1) (4)

As shown in the figure 2, we illustrate the training steps of the particle swarm optimization algorithm we can therefore summarize the features of the basic PSO as follows. (1) The convergence speed is fast. In the iterative evolution, only the best particles pass the information to other particles which belong to the unidirectional information flow. (2) Has the memory, the historical best position of granule and community to remember and transmit to other granules. The parameter that must adjust are few, the structure is simple and is easy to realize. (3) With the decimal code that is decided by the solution of issue directly, the variable number of issue solution directly as granule dimension.

We need to optimize the PSO to achieve for the better performance. The granule renewal algorithm can potentially be modified based on the listed guidelines.

1. For in the middle of the difference. In the thought of the organic evolution survival of the fittest according to the nature, with increase of population evolution algebra, the average adaptation value of population will increase at the big probability. Like this under the elite algorithm condition, there is a reason to suppose the adaptation value of population with increase of the evolution algebra, but increases monotonously as random selection bodies from neighboring two generations or several generations of the population, then the difference of the two individuals represent the direction that the population adaptation value increased at the certain probability that is the evolution direction that hopes.

2. Compiler principles. Variation in nature or breeding is one of the reasons for the emergence of new varieties, the variation may be a good direction to develop and it may be the opposite. The mutation operation can effectively guarantee the diversity of information in the population and avoid falling into the local optimum.

3. Diffractive optical element. From perspective of organic evolution, the final outcome of although evolving is makes the species to most adapt to the direction of environment develops, but inherits and intermediate result of evolution by no means always fine that might also appear in the process of evolution is worse than the improved seed individual, they cannot certainly exclude from the species, but should be in species whole and this point has been reflecting the thought of diffractive optical element.

In our modification strategy, we need to integrate the area information. The inertia weight w is an important parameter in the PSO algorithm and is used to adjust the degree to which the particles have affected the present. And a good inertia weight adjustment strategy is: the early stage of the evolution uses a larger inertia weight, while the latter uses a smaller inertia weight. Because the larger inertial weight can enhance the overall exploration ability, and the smaller the weight can improve the local excavation ability. To balance the weight, we update it as follows.

w = ([w.sub.max] - [w.sub.min]) x exp (-[beta][(t / [T.sub.max]).sup.2]) + [w.sub.min] (5)

3. The Neural Network based Signal Classification

3.1. The Traditional Signal Classification Methodologies

Traditional data classification algorithm has achieved success in basic applications, the traditional two-class SVM has been applied to a large number of the practical applications such as digital recognition, face recognition, etc. It has been proved to be a good classification machine learning algorithm for small sample classification problem. A class of support vector machines, originally proposed by Harry et al. that was used to estimate minimum area containing the samples, and probability of occurrence of the samples outside the determined region is very small. Specifically, original spatial data are implicitly mapped to the high-dimensional feature space through the kernel function, and then the hyperplane is constructed such that the sample data is separated from the origin of the feature space as much as possible, one class of problems in the original space is transformed into two kinds of SVMs in the feature space. The kernel classification function can be listed as below.

Kernel (x, x') = exp (-[[parallel]x - x'[parallel].sup.2] / 2[[sigma].sup.2]) (6)

[f.sub.classification] = [[summation].sub.i] [[alpha].sub.i]k([x.sub.i], x) - [[rho].sup.*] (7)

In KFDA, the choice of kernel function is important, designs whether directly affects the classified effect appropriately. So long as although satisfies the Mercer condition function to elect for the general kernel function theoretically, but the performance of some different kernel function sorter is completely different. When chooses Gauss RBF kernel function, so long as chooses the appropriate parameter, the limited sample can always be been linearly separable in the feature space, therefore this article chooses the Gauss RBF kernel function as the kernel function of sorter, the expression is shown as the follows.

k (x, x') = exp [-[[parallel]x - x'[parallel].sup.2]/[[sigma].sup.2]] (8)

For the KFDA in regularization coefficient C, it also needs to be chosen to achieve highest classification accuracy. The commonly used parameter selection method of K-fold crosses validation method, steepest descent method, genetic algorithm and ant colony algorithm. The K-fold cross validation method is an unbiased estimate of the generalization error, its essence is through the training model in the training set, which has the certain accuracy and has good generalization performance. The distinction number of OAR method is K, because each criterion function needs all samples to participate in the training, therefore the support vector number in each criterion function is biggest. But in the OAO many kinds of classifications each criterion function only needs two kinds of samples to participate in the training, the support vector number of each criterion function is smallest. Figure 4 illustrates the mentioned optimal function finding procedures.

3.2. The Neural Network based Signal Classification

The one most commonly used application of general neural network is the pattern classification. The correct characteristics of the neural network sorter possibly are because the generalization ability and excitation function of neural network have some saturation areas. However most neural networks are the neuron constitution by use two-level excitation function, the output of this excitation function only has two saturation areas. However, in the neuron can also use to have the excitation function of saturated rank, at this time this neuron was also called the multistage neurons. Each sub-network in the network is only trained only by the two modes corresponding to its row and its corresponding column. In the recognition, a sub-network is judged as a pattern A instead of the pattern B. We can only understand that the sub-A instead of Mode B. Thus each sub-network cannot determine the mode, but can exclude another mode and the corresponding pattern. Under this basic guidance, we propose the multi-layer pattern as the follows.

f (net) = 1/L [[summation].sup.L.sub.r=1] [f.sub.s] ([beta](net - [[theta].sup.r])) (9)

For the classifier, it is important that the output values of the output layer neurons lie in the saturation region. So when the input mode becomes fuzzy, the output of the neural network classifier will not change greatly, that is, the output is not very sensitive to the input, which is also very important for the classifier. For accuracy concern, we take the rough set theory into basic consideration. Minimal attribute reduction set B set to the original condition attribute, regardless of any property removed from the B, are likely to cause the whole universe U reduced reliance, and decrease indirectly reflects the important degree of attribute importance that can be any attribute according to the definition in B.


Uses rule that the random unearths to help to construct the BP neural network for decision-making. The regular antecedent refers to some condition correspondences the neural network input level and the consequent refers to conclusion some output levels of correspondence as each rule is equal to an implicit strata node. This time rule after the knowledge of straight run, eliminated the information of redundancy and disturbance, the dimension of input vector greatly reduced, making the neural network the training price reduce, and was helpful to reducing fitting phenomenon.

4. The Proposed ECG Signal Processing and the Abnormal Heart Beat Classification Algorithm

4.1. ECG Signal Processing Paradigm Demonstration

The traditional frequency range de-noising method uses the Fourier transformation then maps the entire quite territory generally the signal through a filter of low-pass or bandpass. But because the Fourier transformation does not have the topicality, therefore throws over after the filtration image enhancementation that not only the signal-to-noise ratio cannot be improved, the CI of signal was also blurred. Wavelet analysis is a new method of general signal frequency analysis appears when it has characteristic of multi-resolution can handle the non-stationary signals like ECG signals. The wavelet de-noising method of wavelet threshold denoising mainly and modulus maxima de-noising while the modulus maximum primary value method of computation and convergence of non-linear wide slow. Method of calculation is small that can effectively remove the noise in the signal singularity at the same time as it has wide adaptability and the general originality.

For the wavelet analysis, the basis selection is the essential issue. Most signals R peak strange, because the R wave has certain frequency range. R wave ridge strange characteristics and a high frequency noise peak irregularity differ from. Separate DYWT can describe the irregularity of signal singular point approximately, based on this can use it to differentiate the R wave and high frequency noise. There are many kinds of base of wavelet analysis. In the process of noise reduction, the selection criteria are usually from the following aspects: (1) self-similarity principle, if the selected wavelet has a certain similarity to the signal, the transformed energy will be compared; (2) support the length of the signal for the local analysis requires wavelet function in time domain with the tight support; (3) symmetry in image processing to avoid the significant phase shift; (4) regularity, the signal or image reconstruction to obtain smooth The effect is very useful.

4.2.The Abnormal Heart Beat Classification

Wavelet transform is the common method of classification, but it cannot effectively improve the accuracy of the partial classification, even the loss of some key special cylinder. The Gabor transformation is a time-frequency domain method, may show that at temporal frequency from near time-frequency characteristics that often was used in the compression and feature extraction of signal. Therefore, in this paper, we conduct research on the novel Gabor based system. The significance that Gabor launches is primary function can construct to causes it is easy to locate and gather to the time and frequency highly. From now on, it launches coefficient may the display signal fall frequency nearby time-frequency characteristics. In the formula 11, we show the transformation of the Gabor model.

[c.sub.m,n] = [[summation].sup.L-1.sub.k=o] x(k) [[chi].sup.m] (k - mN) [W.sup.-nK] (11)

Pats when to the heart the sample enters the young Gabor transformation needs to choose appropriate comprehensive window. The Gauss function had guaranteed the time domain and frequency range energy comparatively is quite centralized that may obtain the best time and best frequency resolution. In addition to the feature based on the Gabor transform, the continuous two QRS waves of the RR period is another important feature, which describes the location of many heart pumping R waves in many cardiac the arrhythmias. Gabor transform coefficient and RR period together constitute characteristics of the. Classification performance is evaluated based on sensitivity, specificity, and overall accuracy. Sensitivity is defined as the percentage of abnormal heart beat that is correctly classified. Specificity is defined as the percentage of normal beat that is correctly classified. Accuracy is defined as the percentage of the heart beat that is correctly classified.

In the figure 6, we show the simulation on the proposed methodology. First calculates the Gabor conversion coefficient that each heart pats, this kind of 155 point heart patted the sample to turn into the 1875 point Gabor transformation signal while because in the real part of Gabor conversion coefficient symmetrical distribution, the feature selection was limited in the first 955 coefficients searches. RR that chooses the coefficient and current heart shear after the special cylinder to ask that the time constitutes the characteristic vector as the input of general sorter together. Most close neighbor method is a powerful non-parameter sorter, does not need, no matter what serves the time-consuming training step to knock does not need to redesign the sorter in the character subset of each choice. The above results show that the Gabor transform can extract the features better, as but the Gabor transform coefficients have information redundancy, which is not conducive to classification. Therefore, in the order to improve the classification accuracy, Gabor transform coefficients must be selected.

5. Conclusion and Summary

This paper proposes the novel ECG signal processing and the abnormal heart beat classification algorithm based on particle swarm optimization (PSO) model and the neural network. ECG is a reflection of the cyclical activity of the heart an important indicator in the clinical widely used. The current ECG detection equipment has not only limited to the ECG graphics records, and the ECG digital acquisition and automatic analysis, thereby improving the analysis speed and accuracy, reduce the work intensity of the doctor. Ectopic origin of general motility, frequency changes, conduction abnormalities or pacemakers is likely to cause arrhythmia. Arrhythmia classification is one of the important contents of ECG computer analysis. Under this basis, this paper proposes the novel perspective on the ECG signal processing and that the abnormal heart beat classification algorithm. The proposed method combines the PSO and neural network to optimize the wavelet. The simulation result proves the effectiveness of the method. In the future research, we will integrate more machine learning algorithm to enhance the robustness.

Recebido/Submission: 27/06/2016

Aceitacao/Acceptance: 14/09/2016


The work is financially supported by the following projects. (1) Research and application system of the ECG analysis technology based on the WIT120 (NO. 2016XJCQCX17). (2) Research on the hierarchical support vector machine based beat classification methodology (NO. 2016A030313658). (3) Research and the application scenario building of the ECG analysis technology for mobile medical treatment (NO. 2015KTSCX175).


Alickovic, E., & Subasi, A. (2015). Effect of Multiscale PCA de-noising in ECG beat classification for diagnosis of cardiovascular diseases. Circuits, Systems, and Signal Processing, 34(2), 513-533.

Abo-Zahhad, M., Al-Ajlouni, A. F. (2013). A new algorithm for the compression of ECG signals based on mother wavelet parameterization and best-threshold levels selection. Digital Signal Processing, 23(3), 1002-1011.

Bortolan, G., Christov, I., Simova, I., & Dotsinsky, I. (2015). Noise processing in exercise ECG stress test for the analysis and the clinical characterization of QRS and T wave alternans. Biomedical Signal Processing and Control, 18, 378-385.

Civicioglu, P., & Besdok, E. (2013). A conceptual comparison of the Cuckoo-search, particle swarm optimization, differential evolution and artificial bee colony algorithms. Artificial Intelligence Review, 39(4), 315-346.

Chen, W., Zhang, J., Lin, Y. (2013). Particle swarm optimization with an aging leader and challengers. IEEE Transactions on Evolutionary Computation, 17(2), 241-258.

Duan, H., Luo, Q. (2013). Hybrid Particle Swarm Optimization and Genetic Algorithm for Multi-UAV Formation Reconfiguration. IEEE Computational Intelligence Magazine, 8(3), 16-27.

Garg, H., & Sharma, S. P. (2013). Multi-objective reliability-redundancy allocation problem using particle swarm optimization. Computers & Industrial Engineering, 64(1), 247-255.

Gawali, D., & Wadhai, V. M. (2014). Implementation of ECG sensor for real time signal processing applications. Advances in Electronics, Computers and Communications, 1-3.

Gradl, S., Kugler, P., Lohmuller, C., & Eskofier, B. (2012). Real-time ECG monitoring and arrhythmia detection using Android-based mobile devices.20t2 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2452-2455.

Huang, W., Cai, N. (2015). ECG Baseline Wander Correction Based on Ensemble Empirical Mode Decomposition with Complementary Adaptive Noise. Journal of Medical Imaging and Health Informatics, 5(8), 1796-1799.

Lino, A., Rocha, A., & Sizo, A. (2016). Virtual teaching and learning environments: Automatic evaluation with symbolic regression. Journal of Intelligent & Fuzzy Systems, 31(4), 2061-2072.

Mukhopadhyay, S. K., Mitra, S. (2013). ECG signal compression using ASCII character encoding and transmission via SMS.Biomedical Signal Processing and Control, 8(4), 354-363.

Oster, J., Behar, J., Sayadi, O. (2015). Semisupervised ECG ventricular beat classification with novelty detection based on switching Kalman filters. IEEE Transactions on Biomedical Engineering, 62(9), 2125-2134.

Pereira, C., Ferreira, C. (2015). Identification of IT Value Management Practices and Resources in COBIT 5. RISTI - Revista Iberica de Sistemas e Tecnologias de Informacao, (15), 17-33.

Ren, C., An, N., Wang, J., Li, L. (2014). Optimal parameters selection for BP neural network based on particle swarm optimization: A case study of wind speed forecasting. Knowledge-Based Systems, 56, 226-239.

Rai, H. M., Trivedi, A., & Shukla, S. (2013). ECG signal processing for abnormalities detection using multi-resolution wavelet transform and Artificial Neural Network classifier. Measurement, 46(9), 3238-3246.

Wang, G. G., HosseinGandomi, A., Yang, X. S., & HosseinAlavi, A. (2014). A novel improved accelerated particle swarm optimization algorithm for global numerical optimization. Engineering Computations, 31(7), 1198-1220.

Yu, X. M., Qin, A. L., Xue, J. (2013). Application of Synergistic Optimization Method by Maximin Fitness Function Strategy Based Multi-Objective Particle Swarm Optimization Algorithm in Metal Bellows Structural Design. Advanced Materials Research, 722, 550-556.

Lixun Liu (1,2), Yujuan Si (1,2) *, Yuemeng Wang (2)

* Yujuan Si,

(1) Zhuhai College of Jilin University, Zhuhai, Guangdong 519041, China

(2) College of Communication Engineering, Jilin University, Changchun, Jilin 130012, China
COPYRIGHT 2016 AISTI (Iberian Association for Information Systems and Technologies)
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2016 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Liu, Lixun; Si, Yujuan; Wang, Yuemeng
Publication:RISTI (Revista Iberica de Sistemas e Tecnologias de Informacao)
Date:Nov 1, 2016
Previous Article:Research on the data aided English literature and culture corpus based on decision tree and natural language processing models.
Next Article:Construction and simulation on stock price forecasting model based on fundamental analysis and technical analysis.

Terms of use | Privacy policy | Copyright © 2019 Farlex, Inc. | Feedback | For webmasters