• Tidak ada hasil yang ditemukan

Automatic Detection of Premature Ventricular Contraction Based on Photoplethysmography Using Chaotic Features and High Order Statistics

N/A
N/A
Protected

Academic year: 2024

Membagikan "Automatic Detection of Premature Ventricular Contraction Based on Photoplethysmography Using Chaotic Features and High Order Statistics"

Copied!
5
0
0

Teks penuh

(1)

Automatic Detection of Premature Ventricular Contraction Based on Photoplethysmography Using

Chaotic Features and High Order Statistics

Mohammad Reza Yousefi*1, Mahdi Khezri2, Razieh Bagheri3, Reza Jafari4

1IEEE Senior Member, Microgrid Research Center, Najafabad Branch, Islamic Azad University, Najafabad, Iran

2 Digital Processing & Machine Vision Research Center, Najafabad Branch, Islamic Azad University, Najafabad, Iran

3Electrical Engineering Department, Najafabad Branch, Islamic Azad University, Najafabad, Iran

4Electrical Engineering Department, K. N. Toosi University of Technology, Tehran, Iran [email protected], [email protected], [email protected], [email protected]

Abstract—This paper deals with the analysis of photoplethysmography (PPG) signals for the recognition of premature ventricular contractions (PVC). PGG is an optical method used to measure blood volume changes in a non-invasive manner. As a diagnostic tool, PPG has recently been considered to evaluate the functioning of the cardiovascular system and identify its related disorders. PPG signals from 22 healthy and unhealthy subjects were used in this work. A number of chaotic and statistical features including Lyapunov exponent, skewness, kurtosis, fuzzy entropy and spectral entropy were extracted from the signals and selective features were identified by principle component analysis (PCA) to be used during data classification. Feature reduction method. k-nearest neighbors (kNN), support vector machine (SVM) and neural network were examined as classification algorithms. Results showed that the highest recognition accuracy of 95% and specificity of 90.4%

are obtained by the KNN classifier.

Keywords-photoplethysmography, Premature ventricular contraction, chaotic features, principle component analysis, KNN classifier.

I. INTRODUCTION

Cardiac arrhythmia detection is a prominent and essential issue in clinical centers, since long-term arrhythmia causes an increased risk of death for patients. Therefore, arrhythmia detection methods should have high accuracy and precision to track the arrhythmia levels. Traditionally, detection and classification of these arrhythmia have been discussed as a major concern among cardiologists [1].

Premature ventricular contraction (PVC) is one of the most common heart rate arrhythmia and abnormal activities, which occurs due to depolarized ventricles and then emergence of abnormal patterns in QRS complex of electrocardiogram (ECG). These repetitive contractions during physical activities increase the risk of death and can lead to a more serious cardiac arrhythmia and atrial fibrillation [2].

PVC in ECG can be often detected by discriminative waveform. To do this, specialists can use holter monitoring for signal morphology. However, long- term connection of the electrodes makes the patient’s chest sensitive and uncomfortable. To solve this problem, PVC detection based on low interference methods, e.g., photoplethysmography (PPG), has been considered by researchers. PPG is a noninvasive method for monitoring hemodynamic changes in a cardiovascular system by using infrared light. In contrast to ECG, photoplethysmographic sensors have simpler application and can even connect to fingers and ears. In fact, PPG is an easy and useful method for measurement of blood volume changes in peripheral circulation [3].

In PPG, some radiations are reflected from skin surface and other radiations are absorbed by oxy hemoglobin, as well as by some of proteins and glucoses. Each of the components absorbs some of radiations depending on their chemical characteristics.

The rest of radiations go through the tissue, and the radiation level can be detected by a receiver. Simple structure of the PPG system can be observed in Figure 1.

Figure 1. The PPG schematic

The amplitude of PPG signals needs to be amplified due to low SNR. The signals are denoised in the preprocessing step. After amplifying and This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the acceptance and publication.

978-1-5386-3392-2/18/$31.00 ©2018 IEEE

(2)

filtering, sampling is applied to the signals by using ADC and then, the signals are investigated for heart rate estimation and determination of arrhythmia.

In this paper chaotic features and high order statistics were used for automatic detection of PVC.

For this purpose, high order statistics features, e.g., skewness and kurtosis, as well as chaotic features, e.g., Lyapunov exponent, spectral entropy, and phase entropy, in order to use by several classification methods, e.g. support vector machine (SVM), K nearest neighbors (KNN) and multi-layer perceptron (MLP) for detection of PVC.

In recent years, several studies have been carried out in modeling fields and detecting PVC. Several researchers have verified the feature extraction methods and discriminative features. The use of efficient processing techniques, especially in the selection of relevant features and classification methods, can lead to an accurate detection of PVCs in patients with unexpected conditions. Studies on PVC systems started from 1996 by Hamm and Han [4]. In their study, two linear predictors of coding coefficients were used by the QRS complex square mean. After that, Millet et al. [5] represented an algorithm for PVC detection based on morphological characteristics of the QRS complex. In another similar study, a set of 4 features has been used based on QRS.

Moreover, Chikh et al. [6] proposed three methods of feature extraction. These proposed methods were based on signal spectral components, linear predictor coding and principle component analysis. Gill et al.

[7] classified PVCs by a linear classifier. In this research, parameters such as a turbulence set point and its slope were discussed for turbulence determination. Also, the system function on 4131 PVCs gathered from 27 subjects was verified.

Overall, the results showed that PPG can play role of ECG in turbulence analysis. Yet in another study, Paradkar et al. [8] represented a PVC method based on phase entropy and heart rate estimation from finger PPG. In their study, the Capnobase dataset was used [9] and accuracy of heart rate estimation was obtained as 99.33%. Solosenko [10] represented another method for PVC detection based on six features extracted from PPG signals. This method was implemented based on six features, including pulse power and pick to pick time interval. After that, artificial neural network was implemented on the PhysioNet dataset, including MMIC and MMIC II

[11]. In [12] seven temporal features from PPG signals were extracted and a Bayesian classifier was implemented. Ribeiro et al. [13] proposed an algorithm based on three structures including SVM, MLP and RVM.

II. WAVEFORM AND DATASET A. Waveforms

The variation of PPG is relevant to the peripheral blood volume changes [14]. Premature contractions lead to a reduction of ventricular filling and lessen the peripheral pulse amplitude. Therefore, PPG pulses during PVC may become rarely detectable, or may still have a sufficient amplitude for peak detection [15]. The premature pulses in PPG are denoted as P.

A sample of PVC pluses in PPG and ECG signals has been shown in figure 2.

B. Dataset

Twenty-two 30 minutes PPGs sampled at 500 Hz (the MIMIC dataset [17]) were used; 18 PPGs as a training set and 4 PPGs for testing. Some of the signals in the MIMIC database were omitted from the study as they contained severe signal corruptions or various pathologies. The list of considered signals is presented in Table 1.

Table 1. Test PPGs obtained from the MIMIC database (No 1-22) No. Signal

name #P No. Signal

name #P

1 039m 0 12 252m 0

2 041m 0 13 253m 0

3 055m 0 14 404m 208

4 211m 0 15 408m 2

5 212m 30 16 439m 3

6 218m 0 17 442m 366

7 221m 0 18 444m 10

8 224m 0 19 449m 2

9 225m 11 20 466m 4

10 230m 4 21 471m 0

11 237m 14 22 474m 4

Total: 59 Total: 649

(3)

Figure 2. Example of PVC pulse types in PPG together with reference ECG [2]

III. METHODS

The proposed method for PVC detection and classification extracts chaotic and high-order statistics features for each PPG pulse. The method is composed of 3 major parts:

A. Preprocessing and feature extraction

PPGs are preprocessed by using a band-pass finite impulse response (FIR) filter with 0.5-4 Hz sub-band frequency respectively, to minimize high frequency noise and omit the baseline. Then, the signals were smoothed by baseline alteration. In this paper, kurtosis, skewness, fuzzy entropy, spectral entropy and Lyapunov exponent were extracted as the chaotic and high-order statistics features.

Kurtosis (K) is a measurement of the combined sizes of two tails. It measures the amount of probability in these tails:

K = [( ) ] (1) where X is the i X value, X is the average and s is the sample standard deviation.

Skewness (SK) in a measure of a dataset’s symmetry or lack of it. A perfectly symmetrical dataset will have a skewness of 0. Skewness is defined as:

SK = ∑(( )) (2) where n is the sample size, X is the i X value, X is the average and s is the sample standard deviation.

Entropy (H ) is a major tool in information theory of a random variable (X ) with continuous probability law f (X) defined as:

H = − f (X) log f (X) dx (3) Fuzzy entropy (FE) is the entropy of a fuzzy set, loosely representing the information of uncertainty

[18]. Fuzzy entropy can be calculated using a multi- stage algorithm for time series.

Power spectral entropy (PSE) describes the complexity of a system:

P(w ) = |X(w )| (4) P =∑ (( )) (5) Equation (5) describes normalized factor of P(w ).

PSE = − ∑ P ln P (6) PSE in Equation (6) is calculated by using a standard formula for entropy [19].

Maximal lyapunov exponent is used to quantify the nonlinear chaotic dynamics of the signal [20]. For a discrete time series X = (X ), starting with X we have:

1 '

0

0

( )

lim

1n ln ( )i

n i

x f x

λ n

→∞ =

=

(7) B. Classification

Three commonly used classifiers namely, MLP, SVM and KNN were applied to the obtained features.

KNN is a supervised learning algorithm. In the KNN method, two important factors are involved. The first factor is the selection of the distance, weighting function and selecting the best neighborhoods [21].

An artificial neural network consists of three layers of input, output, and processing unit. The neuron is the smallest data processing unit that forms the basis of neural network performance. A neural network is a collection of neurons that, when placed in different layers, forms a specific architecture based on the relationships between neurons in different layers [22].

The SVM classifier generates a hyper-plane in a way the boundary between data in different classes is maximized. In addition to the linear SVM, a nonlinear SVM uses a kernel function to create the maximum- margin hyperplanes. There are different kernel functions such as Linear, polynomial, radial basis

(4)

function (RBF) and sigmoid that can be used for support vector classification. The KNN algorithm is the simplest of other algorithms to classify objects in machine learning. t. The performance of the methods was evaluated in terms of sensitivity (Se), specificity (Sp) and accuracy (Acc). The training process was repeated 200 times and then, the averaged performance values were taken as the overall performance measure.

IV. RESULTS

The results are presented in Table 2, in which N and P indicate healthy subjects and patients, respectively. Table 2 shows that the performance of the classifier depends on the methods.

Table 2. Classification results obtained by the classifiers Classifier SVM MLP KNN

Acc% 90.9%

± 1.8% %72.3

± 3% 95.5%

± 4%

Sp% 92.9%

± 5.5% 76.9%

± 1.5% 100%

± 0%

Se% 87.5%

± 1.8% 66.7%

± 3.5% 88.9%

± 1%

In MLP, the number of layers, neurons in a input layer and neurons in hidden layer of ANN was set to 3, 20 and 10, respectively. by using scaled conjugate gradient descent (SCGD). In KNN, K was set to 5 by Euclidean distance criterion. In SVM, linear kernel has better results rather than others. We used k-fold (k=5) as cross validation where 5% of dataset were used for evaluation. In Table 2, the results showed the ideal specificity and highest percent of accuracy in the KNN classifier, although the minimum tolerance of accuracy is obtained by SVM. The results also showed that the Entropy, Maximum Lyapunov exponent and Kurtosis features presented the best performance, respectively.

V. DISCUSSION AND CONCLUSION

The goal of this work was to develop a pattern recognition method for detection of PVC by relying on photoplethysmography signal analysis. The first attempt to detect premature contractions by using PPG was presented in an earlier study. Using chaotic and statistical features of the signal leads to obtaining remarkable results using all three classification methods which shows a significant improvement compared to the previous studies. The results showed that KNN has an overall better performance as compared to others

.

REFERENCES

[1] Bozkurt, Alper, Arye Rosen, Harel Rosen, and Banu Onaral.

"A portable near infrared spectroscopy system for bedside monitoring of newborn brain." Biomedical engineering online 4, no. 1 (2005): 29.

[2] A. Solosenko and V. Marozas,"Automatic premature ventricular contraction detection in photoplethysmographic signals," New England Journal of Medicine, vol. 312, pp.

193–197, 2011.

[3] J. Zhou,"Automatic Detection of Premature Ventricular Contraction Using Quantum Neural Networks", IEEE Symposium on Bio Informatics and Bio Engineering, vol.15, pp. 125-131. 2003.

[4] F. Hamand and S. Han, "Classification of cardiac arrhythmias using Fuzzy ARTMAP", IEEE Transactions on Biomedical Engineering, vol. 43, 425-430, 1996.

[5] J. Millet, M. Perez, G. Joseph, A. Mocholi and J. Chorro, ''Previous Identification of QRS Onset and Offset is not Essential for Classifying QRS Complexes in a Single Lead", IEEE Computers in Cardiology, vol.24, 299-302, 1997.

[6] M.A Chikh, N. Belgacem, Az. Chikh and F. Reguig, "The Use of Artificial Neural Network to Detect the Premature Ventricular Contraction (PVC) Beats", Electronic Journal Technical Acoustics, vol.2, 2004.

[7] E. Gil, P. Laguna, J. Pablo Martınez and O. Barquero-Perez,

"Heart Rate Turbulence Analysis Based on Photoplethysmography", IEEE transactions on biomedical engineering, vol. 60, pp. 23-30. 2013.

[8] N. Sudhir Paradkar and S. R. Chowdhury, " Fuzzy Entropy based Detection of Tachycardia and Estimation of Pulse Rate through Fingertip Photoplethysmography", Journal of Medical and Bioengineering, vol. 4, pp. 151-160. 2015.

[9] W. Karlen, M. Turner, E. Cooke, G. Dumont, and J. M.

Ansermino, "Capnobase: Signal database and tools to collect share and annotate respiratory signals", In Annual Meeting of the Society for Technology in Anesthesia (STA), vol. 15, pp.432-436, 2010.

[10] A. Sološenko, P. Andrius, and M. Vaidotas,

"Photoplethysmography-based method for automatic detection of premature ventricular contractions", IEEE transactions on biomedical circuits and systems, vol.23, pp. 234-239. 2015.

[11] A. L. Goldberger, L. A. N. Amaral, L. Glass, "PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals", Circulation, vol.

101, pp. 215–220, 2000.

[12] A. Solosenko and V. Marozas," Smartphone Based Application for Online Premature Ventricular Contraction Detection", conference of Biomedical Enginnering, 2010.

[13] E. Peper, R. Harvey, I.-M. Lin, H. Tylova, and D. Moss, “Is there more to blood volume pulse than heart rate variability respiratory sinus arrhythmia, and cardiorespiratory synchrony?,” Biofeedback, vol. 35, no. 2, pp. 54–61, 2007.

[14] D. Zheng, J. Allen, and A. Murray,“Determination of aortic valve opening time and left ventricular peak filling rate from the peripheral pulse amplitude in patients with ectopic beats,”

Phys. Meas., vol. 29, no. 12, pp. 1411–1419, 2008.

[15] Sološenko, Andrius, Andrius Petrėnas, and Vaidotas Marozas.

"Photoplethysmography-based method for automatic detection of premature ventricular contractions." IEEE transactions on biomedical circuits and systems 9, no. 5 (2015): 662-669.

[16] M. Saeed, M. Villarroel, A. T. Reisner, G. Clifford, L.-W.

Lehman, G. Moody, T. Heldt, T. H. Kyaw, B. Moody, and R.

G. Mark, “Multiparameter intelligent monitoring in intensive care II (MIMIC II): A public-access intensive care unit database,” Crit. Care Med., vol. 39, pp. 952–960, May 2011.

[17] Kantz, Holger. "A robust method to estimate the maximal Lyapunov exponent of a time series." Physics letters A 185, no. 1 (1994): 77-87.

(5)

[18] Kullback, Solomon. Information theory and statistics. Courier Corporation, 1997.

[19] Zhang, Aihua, Bin Yang, and Ling Huang. "Feature extraction of EEG signals using power spectral entropy." In BioMedical Engineering and Informatics, 2008. BMEI 2008. International Conference on, vol. 2, pp. 435-439. IEEE, 2008.

[20] A. Solosenko and V. Marozas, “Automatic extrasystole detection using photoplethysmographic signals,” in Proc. 13th Mediterranean Conf.Medical and Biological Engineering and Computing, 2013, vol. 41, pp. 985–988.

[21] Tan, Songbo. "An effective refinement strategy for KNN text classifier." Expert Systems with Applications 30.2 , pp. 290- 298.2006.

[22] Windeatt, Terry. "Accuracy/diversity and ensemble MLP classifier design." IEEE Transactions on Neural Networks 17.5 ,pp.1194-1211.2006.

Referensi

Dokumen terkait