• Author / Uploaded
• Bruno Leppe

Citation preview

2017 2nd IEEE International Conference On Recent Trends In Electronics Information & Communication Technology, May 19-20, 2017, India

Comparative Study of ECG Feature Extraction Methods Akanksha Agrawal, NBN Sinhgad School of Engineering, Savitribai Phule Pune University, Pune,India, [email protected]

Dhanashri H.Gawali, NBN Sinhgad School of Engineering, Savitribai Phule Pune University, Pune,India, [email protected]

Abstract—Electrocardiogram (ECG) signals are used for the detection of heart related diseases from more than a century now. Electrocardiogram (ECG) is an electrical activity of the heart that is recorded by placing many electrodes on the body. Hence, it is important to identify an appropriate method for extracting features in ECG signal. This paper provides insight into different efficient methods for this purpose. The prominent methods presented in this paper are Pan-Tompkins, Hilbert transform, Histogram approach, Wavelet transform, Auto-regression(AR), Independent Component Analysis(ICA), Linear prediction(LP), Adaptive threshold. A comparative analysis is done for these feature extraction methods based on various performance parameters like sensitivity, predictivity and accuracy. The above methods are reviewed on lines of these parameters, which can be used to identify suitable method for ECG feature extraction.

for the production of electrical signals that spreads through both the atria and makes them contract. The AV node, which is located on the opposite side of the right atria, serves as an electrical gateway to the ventricles. The signal spreads both the atria causing the muscle cells to depolarize and contracts to induce a phase known as atrial systole on the ECG this atrial depolarization is represented by P wave. The atrial systole starts after 100 millisecond's after the P wave begins. The periods of conduction that follows the atrial systole and follows the contraction of the ventricles is depicted on the ECG by the PR segment. A flat line following the P-wave when the signal leaves the atria, it entries the ventricles via the AV node located in the inter-atrial septum. It enters the bundle of His and spreads through the bundle branches and the large diameter Purkinje's fibers along the ventricle values. As the signal spreads through the ventricles, the contractile fiber depolarizes and contract very rapidly inducing ventricular systole. In ECG, the QRS complex represents these rapid ventricular depolarizations. Atrial repolarization also occurs at this time but any; atrial activity is hidden on the ECG by the QRS complex. Finally, as the signal passes out of the ventricles, the ventricular walls start to relax and recover a state described at ventricular diastole. The dome shape T wave on the ECG marks this ventricular repolarization. On the ECG, the ST segment depicts the period when the ventricles are depolarized.

Keywords—ECG; AR; Wavelet; Hilbert; Histogram; Adaptive threshold; ICA; LP; Pan-Tompkins;

I. INTRODUCTION The heart is a vital organ of the human body. It pumps oxygenated blood into different parts of the body with the help of arteries. The proper functioning of the heart is very important for the whole body to function smoothly. However, with time, a number of people are now suffering from different types 6% of men suffer from heart-related problems in the year 2016 by heart attacks and strokes mainly. Hence, to measure the performance rate of the heart, ECG signal is used extensively. It is the most reliable, safe and accurate method to estimate the working of the heart. The frequency of the ECG signal lies in the range of 0.05 to 100 Hz. and it has a dynamic range is of 1-10mV.The different clinical values of ECG signal and its associated peaks are mentioned in Table 1[1]. Table I: Clinical values of ECG signal Sr.No

Feature

Amplitude(mV)

Duration(ms)

1

P wave

0.25

60-80

2

PR interval

-

120-200

3

R wave

1.60

80-120

4

Q wave

25% of R wave

90-100

5

QT interval

-

360-440

6

ST segment

-

100-120

7

T wave

5-8

120-160

8

RR interval

-

(0.6-1.2)sec

9

PR segment

-

50-120

II. CARDIOVASCULAR SYSTEM OF HEART Each beat begins in the right atrium with an action potential signal from SA node. The SA node is responsible

Fig 1.(a) Structure of heart (b) ECG peaks (c) Peaks corresponding with the functioning of heart [2]

978-1-5090-3704-9/17/$31.00 © 2017 IEEE 246

2017 2nd IEEE International Conference On Recent Trends In Electronics Information & Communication Technology, May 19-20, 2017, India

The QT interval represents the time it takes for both depolarization and repolarisation of the ventricles to occur. The sequence of the events just mentioned and the associated ECG trace repeats with every heartbeat. ECG is an amalgamation of many action potential that constitutes the electrical activity of the heart. In this paper, a few methods that are presented to extract the features of the PQRST peaks of the ECG signal. These methods are used in time domain, frequency domain and inboth time –frequency domain. The structure of the heart with its labeling and how each function of the heart corresponds to the different peaks in the ECG signal is depicted in fig.1 III.

x Shifting x Linearity. Hilbert transform has a few advantages over the other methods of detecting QRS. It can minimize the unwanted effect of the large peaks T and P by creating envelope around the R peak. Also, it is an odd function [5] impelling that Hilbert transformed signal will cross the X axis whenever the signal is disturbed by any means. When the ECG signal is transformed with the help of Hilbert transform, the waveform then obtained is shown in Fig 2.

METHODOLOGY

A. Time Domain 1) Pan Tompkins Algorithm The Pan-Tompkins algorithm is most widely used and highly acknowledged algorithm. It is a simple yet powerful algorithm for QRS detection. It is a real-time QRS detection algorithm based on slope, amplitude, and width of the QRS complex[3]. It is quiet robust and efficient in nature. The algorithm uses a special digital band-pass filter. It is used to reduce the noises and false detection caused by faulty signals that contaminate the ECG signal. In this case, the sensitivity increase as this method allows the use of low thresholds. The algorithm is simple because, it automatically adapts and adjusts the threshold value for next peak detection from the previous smallest peak. Boiling it down to, the steps of the procedure are stated below[4]: x x x x x x

Band-pass filtering Applying a derivative operator Square and integrate signal Fiducial mark determination by thresholding QRS wave detection Average R_R interval and rate limits adjustment

2) Hilbert Transform Bolton and Westphal proposed the Hilbert Transform method. This algorithm and the Hilbert transform utilize the first differentiation of the ECG signal. It is used to identify and detect the R-peaks in the ECG waveform. The Hilbert transform develops an envelope of ECG signals where the maximum peaks are obtained at the zero crossing point.Given that u(t) is a real time function. The Hilbert transform is [5] ȗ(t)= Ƕ[u(t)]

ȗ(t)= ∫

∞ ∞

( )

Scaling and time reversal Convolution Orthogonal

3) Linear Prediction Linear prediction method is used in audio signal and speech processing. It is used to reproduce the original signal accurately [8]. FIR filter of appropriate order is the basic block of the system. The forward coefficient of the signal and the solution viz the difference equation presented which expresses a linear combination of each sample with the previous sample has to be determined [9]. It is accurate in nature also; it has faster computation than others. The actual ECG sequence P(i) is approximated by another sequence Ṕ(i) which is further determined by a distinctive set of predictor coefficients and the past S samples as follows[8] Ṕ(i) = ∑

( )∗ ( − )+ ( )

(3)

Where Ṕ (i) is the predicted signal value, p(i-k) is the previous observed value and a is the predictor coefficient and e(i) is the residual error.

(1)

(2)

Here, ‘Ƕ’ is the Hilbert transform coefficient. Hilbert transform has a few properties that are stated below [6]: x x x

Fig. 2. Waveform after Hilbert transforms [7]

Fig 3. Linear prediction model [10]

4) Histrogram Method A histogram is a display of statistical information that uses rectangles to show the frequency of data items in successive numerical intervals of equal size[11]. This method is used as an estimator for the waves of ECG signal. This approach gives the measure of the orientation the 247

2017 2nd IEEE International Conference On Recent Trends In Electronics Information & Communication Technology, May 19-20, 2017, India

object points in a few quantized directions. The histogram can be divided into a horizontal and vertical histogram. Both these directional histograms are used for determination of separate aspects of the ECG signal. This technique does not employ the usage of any mathematical complexity. This is combined with an adaptive threshold value to calculate the maximum bin of the histogram for the detection of different peaks. In this method, first the determination of R peaks of the ECG signal by setting a threshold value and by computing the maximum value within the marked histogram. Subsequently, the T and the P peaks of the ECG signal is marked and thereby, the whole ECG signal is converted into an array of rectangular peaks corresponding to different peaks of the ECG. Amongst those rectangular peaks, the R peaks, the T wave, and the P peak are marked for the extraction of features present in ECG signal.

This algorithm is simple, accurate, real-time performance and stability. It can also effectively solve the baseline drift problems. When the ECG signal is exposed to adaptive thresholding method, the waveform hence obtained is shown in Fig 5

Fig 5: Waveform after applying Adaptive Threshold [7]

C. Time-Frequency Domain 1) Wavelet Transform

Fig 4. Peak Detection using Histogram Approach [11]

This proposed method is accurate in nature for the detection of peaks purpose [11].The ECG signal is converted to the form of histogram and the P-R-T peaks are detected. This is shown in Fig 4

Wavelet transform method is a mixture of two types, continuous and discrete. Discrete wavelet is sort off called wavelet filter banks, as it uses 2 filters, a Low pass Filter (LPF) and a high pass filter(HPF) to decompose the signal into different frequency scales[14].These functions of mathematics are used for decomposing data into different components of frequency[1]. Wavelet transform is suitable for both high and low frequencies varying its wavelet band depending on the range of frequency value. With the help of this technique, the transform divides the whole wave i.e. the mother wavelet into different wavelet transform bases. This is displayed in Fig 6, where a wavelet is obtained from the mother wavelet.

B. Frequency Domain 1) Adaptive -Threshold method One of the significant methods of detection of the R wave peak and others is the usage of adaptive threshold. The main idea behind this algorithm is that the value of threshold changes with time and it gets updated time to time in correspondence to the changes in the QRS morphology and also the levels of noise and artifacts present in the ECG signal[12]. Hence, to detect the peaks and the baseline drift and power interference, this method is widely used, as the method is much simpler and more effective. It can also do the real-time processing. It is important to fix an initial threshold value to detect the peaks. Nevertheless, fixing of threshold value will not be able to adapt the changing situations and will lead to false positivity. As stated above, this method operates on real-time processing; hence, the threshold value would automatically be updated to the ampthr0 value and the value in the new crest value. The formula for the same is [13] Amp-thr1= 0.6*amp-thr0+0.3*V

Fig 6: Wavelet Signal [16]

An important role in analysis and synthesis is of the proper selection of the wavelet basis function. The wavelet transform can be represented by [15] W(a,b) = ∫

,

=

( )

,

( )

∗ √

(5)

(6)

Where * represents the complex conjugation and the indicates the mother wavelet of the signal.

(4)

2) Independent Component Analysis(ICA) 248

,

2017 2nd IEEE International Conference On Recent Trends In Electronics Information & Communication Technology, May 19-20, 2017, India

Independent component analysis (ICA) is an application method used in signal processing. It has its main usage in separation of multivariate signal into its subcomponents underlying it, where these signals are mutually independent signals in nature[17]. ICA is used for the removal of hidden layers of noise in the ECG signal, but it needs to be assumed that these hidden signals are non-Gaussian in nature. Assume that x1,x2,x3….xn be n linear mixtures of n independent signals s1,s2,s3…sn[11] .In vector-matrix notations, can be written as follows[18] X= A*S

B. Predictivity (+P), it gives the balance between the count of events that are correctly detected, TP and the total count of events that the analyzer detects, and its formulated as +P(%) =

Here FP (false positive) gives the count of events that detection contains flaws. C. Accuracy (%) is calculated by percentage of detected peaks among total number of peaks.

(7)

Where X denotes a column vector having random variables x1,x2,x3..xn be n as its elements.S denotes a column vector having random variables s1,s2,s3…sk as its elements and A denotes the mixing mixture with elements aij [11]. There is different ICA algorithms are Jade ICA, Fast ICA, constrained ICA, and Kernel ICA and extended ICA [17].

Accuracy (A) =

V.

Auto-regression method is linear in nature. AR is a twolead ECG model to improve the cardiac problems, extract the needed features from the acquired ECG signal, and classify them related to its disease.[19] The AR coefficients computed from the ECG signal are classified using a generalized linear model and a multi-layer feed forward neural network. The AR model of order M is given by [19]: 1( − ) 1( ) + 1( ) 2( − ) 2( ) + 2( )

(8) (9)

Method of feature extraction

Applicatio n Domain

Sensitivi -ty

Predictiv i-ty

Accura -cy

Auto regression[19]

TimeFrequency

97.28

97.3

96.6

Wavelet transform[1,14,1 5, 16]

Timefrequency

99.89

99.86

99.75

Linear prediction[8,9,10 ]

Time

96.9

80.4

93.2

ICA[17,18,19]

Timefrequency

97.8

99

90.13

Hilbert transform[5,6]

Time

99.93

99.91

99.83

Frequency

99.55

99.28

99.70

Adaptive threshold[12,13,7 ]

A. Sensitivity(Se), it is given by the ratio by count of events that are correctly detected ,TP(true positive) to the total count of events is formulated by (10)

Here FN(false negative) is the count of missed events.

PERFORMANCE OF METHODS

Table II: Performance analysis of different Feature Extraction Methods

IV. PERFORMANCE PARAMETERS All the above-discussed methods are compared with one another based on three statistical parameters to evaluate. These parameters will determine which methods give more accurate, more precise and more determined results while it detects the peaks of the ECG signal .They are[11]:

%

(12)

With the help of above mentioned parameters, deduction of which feature extraction method has better performance role in comparison to the other proposed methods can be found out. Each method has been evaluated for a set of data base from MIT-BIH Arrhythmia and its parameters are calculated. In table II, each method and its calculated parametric values are mentioned.

Where; HR1 (n) and HR2 (n) denote ECG time series [19],e1(n)represents values of unknown zero means [19],e2(n) represents incorrect random variables [19],a1 (j) and a2 (j) are the AR model coefficients[19]. AR analyses has a lead over the other methods in the matters of its simplicity and also how it is well suited for the real-time applications at hospitals especially in the ICU . AR modeling has greater accuracy however, as AR modeling is linear in nature.

Se(%)=

100%

These parameters are the factors that help us evaluate which feature extraction method is better than the others are.

3) Auto Regression(AR)

1( ) = ∑ 2( ) = ∑

(11)

%

Histogram Approach[11]

Time

99.86

99.76

99.8

Pan [3,4]

Time

99.83

98.65

86

Tompkins

Our methods are functioning all in 3 different domainstime, frequency and time-frequency domain both. 249

2017 2nd IEEE International Conference On Recent Trends In Electronics Information & Communication Technology, May 19-20, 2017, India

In the time domain, the Hilbert and histogram methods giving out the optimum required result in terms of all the 3 parametric values. Hilbert transform requires a tedious amount of mathematical computation work and its method is more complex than histogram method. The histogram method however, requires less computational time & the method is less mathematically complex than another method.

[12]

[13]

In the mixed domain, the wavelet transform method performs fairly well above the other stated methods in relation to the calculated parametric terms. The advantage of this method is it can be used even for very small waveforms. This helps to expand its horizon of workspace.

[14]

VI. CONCLUSION AR, LP, Hilbert transform, Histogram, Adaptive threshold, Pan-Tompkins, wavelet transform, ICA are discussed throughout this paper for the extraction of different features from the ECG signal. Accuracy is a very importance aspect of the ECG feature extraction system. The sensitivity and predictivity is also compared for these methods. It is found that amongst these methods, Hilbert and histogram methods give the best results in terms of accuracy to detect the PQRST peaks of the ECG signal. The histogram method has an upper hand on Hilbert transform in terms of mathematical complexity and it can be used for online analysis of it. These 2 methods are apt for detecting the peaks correctly and not giving any false determination of any peaks. These approaches provide proper modeling of the ECG feature extraction methods.

[16]

[15]

[17]

[18]

[19]

REFERENCES [1]

G.S. Kumari, S. K. Kumar, J. Anusha and M. P. Rao, "Electrocardiographic signal analysis using wavelet transforms," 2015 International Conference on Electrical, Electronics,Signals, Communication and Optimization (EESCO), Visakhapatnam, 2015, pp. 1-6. [2] R. Gupta, M. Mitra and J.N. Bera, ECG acquisition and automated remote processing, Springer, New Delhi, India, 2014, pp. 1-6. [3] J. Pan and W. J. Tompkins, "A Real-Time QRS Detection Algorithm," in IEEE Transactions on Biomedical Engineering, vol. BME-32, no. 3, pp. 230-236, March 1985. [4] I. L. b. Ahmad, M. b. Mohamed and N. A. b. Ab Ghani, "Development of a concept demonstrator for QRS complex detection using combined algorithms," 2012 IEEE-EMBS Conference on Biomedical Engineering and Sciences, Langkawi, 2012, pp. 689-693 [5] Hahn, Stefan L., Hilbert Transforms, in Transforms and Applications Handbook, Ed. Alexander D.Poularikas, CRC Press Inc., Boca Raton, FL, 1996. [6] P. Valluraiah and B. Biswal, "ECG signal analysis using Hilbert transform," 2015 IEEE Power, Communication and Information Technology Conference (PCITC), Bhubaneswar, 2015, pp. 465-469. [7] A. Farahabadi, E. Farahabadi, H. Rabbani and M. P. Mahjoub, "Detection of QRS complex in electrocardiogram signal based on a combination of hilbert transform, wavelet transform and adaptive thresholding," Proceedings of 2012 IEEE-EMBS International Conference on Biomedical and Health Informatics, Hong Kong, 2012, pp. 170-173 [8] E. K. Zaghouani, A. Jemai, A. Benzina and R. Attia, "ELPA: A new key agreement scheme based on linear prediction of ECG features for WBAN," 2015 23rd European Signal Processing Conference (EUSIPCO), Nice, 2015, pp. 81-85. [9] P. Sarma, S. R. Nirmala and K. K. Sarma, "Classification of ECG using some novel features," 2013 1st International Conference on Emerging Trends and Applications in Computer Science, Shillong, 2013, pp. 187-191. [10] C. C. Lin, W. Hu and C. M. Yang, "Automatic classification of arrhythmic heartbeats using the linear prediction model," Computing in Cardiology 2013, Zaragoza, 2013, pp. 971-974. [11] B. Halder, S. Mitra and M. Mitra, "Detection and identification of ECG waves by histogram approach," 2016 2nd International

250

Conference on Control, Instrumentation, Energy & Communication (CIEC), Kolkata, 2016, pp. 168-172 K. Akazawa, K. Motoda, A. Sasamori, T. Ishizawa and E. Harasawa, "Adaptive threshold QRS detection algorithm for ambulatory ECG," [1991] Proceedings Computers in Cardiology, Venice, 1991, pp. 445-448 G. Chen, X. Wang and W. Wan, "An ECG R-wave detection algorithm based on adaptive threshold," 2015 International Conference on Smart and Sustainable City and Big Data (ICSSC), Shanghai, 2015, pp. 145-149. A. Mukherjee and K. K. Ghosh, "An efficient wavelet analysis for ECG signal processing," 2012 International Conference on Informatics, Electronics & Vision (ICIEV), Dhaka, 2012, pp. 411-415. C. L. P. Lim, W. L. Woo and S. S. Dlay, "Enhanced wavelet transformation for feature extraction in highly variated ECG signal," 2nd IET International Conference on Intelligent Signal Processing 2015 (ISP), London, 2015, pp. 1-6. M. J. d. Silva, "Characterization of QRS Complex in ECG Signals Applying Wavelet Transform," 2015 International Conference on Mechatronics, Electronics and Automotive Engineering (ICMEAE), Prague, 2015, pp. 86-89 M. Phegade and P. Mukherji, "ICA based ECG signal denoising," 2013 International Conference on Advances in Computing, Communications and Informatics (ICACCI), Mysore, 2013, pp. 1675-1680. H. P. Kasturiwale and C. N. Deshmukh, "Quality Assessment of ICA for ECG Signal Analysis," 2009 Second International Conference on Emerging Trends in Engineering & Technology, Nagpur, 2009, pp. 73-75. F. M. Vaneghi, M. Oladazimi, F. Shiman, A. Kordi, M. J. Safari and F. Ibrahim, "A Comparative Approach to ECG Feature Extraction Methods," 2012 Third International Conference on Intelligent Systems Modelling and Simulation, Kota Kinabalu, 2012, pp. 252256.