ECG Signal Compression using Optimum Wavelet Filter Bank ... - Core

Available online at www.sciencedirect.com

Procedia Engineering 38 (2012) 2889 – 2902

International Conference on Modelling, Optimisation and Computing (ICMOC 2012)

ECG Signal Compression using optimum wavelet Filter Bank based on Kaiser Window K. Ranjeet*, A. Kuamr*#, Rajesh K. Pandey* PDPM Indian Institute of Information Technology,Design & Manufacturing Jabalpur Airport Road-Jabalpur, (M.P.) 482005, INDIA

Abstract In this paper, an optimized wavelet filter bank based methodology is presented for compression of electrocardiogram (ECG) signal. The methodology employs new wavelet filter bank whose coefficients are derived with window techniques such as Kaiser Windows using simple linear optimization and run-length encoding (RLE). The Wavelet based compression techniques minimize the compression distortion, while RLE further increases the compression without any loss of relevant signal information. The developed technique employs a modified thresholding, which improves the compression of signal as compared to earlier existing thresholding technique. A comparative study of performance of different existing methods and the proposed wavelet filter is made in terms of compression ratio (CR), percent root mean square difference (PRD) and signal-to-noise ratio (SNR). When compared, the developed wavelet filter gives better compression ratio and also yields good fidelity parameters as compared to other methods. The simulation result included in this paper shows the clearly increased efficacy and performance in the field of biomedical signal processing.

© 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license. Keywords: ECG, Compression, DWT, DCT, Run-length encoding

1. Introduction An electrocardiogram (ECG) is the graphical representation of electrical impulses due to ionic activity in the cardiac muscles of human heart. It is an important physiological signal which is exploited to diagnose heart diseases because every arrhythmia in ECG signals can be relevant to a heart disease [1]. ECG signals are recorded from patients for both monitoring and diagnostic purposes. Therefore, the storage of computerized is become necessary. # Corresponding author. Tel.: +91-9425805412; fax: +91-761- 2632524. E-mail address: [email protected]

1877-7058 © 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license. doi:10.1016/j.proeng.2012.06.338

2890

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

However, the storage has limitation which has made ECG data compression as an important issue of research in biomedical signal processing. In addition, the transmission speed of real-time ECG signal is also enhanced and economical due to ECG signal compression. An ECG signal contains steep slopes QRS complexes and smoother P and T waves. It is recorded by applying electrodes to various locations on the body surface and connecting them to a recording apparatus. There are certain amounts of sample points in ECG signal which are redundant and replaceable. ECG data compression is achieved by elimination of such redundant data sample points. During the past few decades, many schemes for ECG signal compression have been proposed. Most of them are lossy compression techniques in which the reconstruction signal is not exact replica of the original input signal. Generally, these techniques are classified two categories: direct techniques and general techniques [2-4]. In early stage of research, several methods [4-6] such as the amplitude zone time epoch coding (AZTEC), and coordinate reduction time encoding system (CORTES) were developed based on direct scheme in which compression is achieved by eliminating redundancy between different ECG samples in the time domain. Some other examples are the turning-point (TP) data reduction algorithm, the scan-along polygonal approximation (SAPA) and differential pulse code modulation (DPCM). All such techniques involve simple signal processing and yield minimum distortion with good compression. A detailed review on these techniques is presented in [4-6] and the references there in. In the last two decades, a substantial progress has been made in the field of data compression. So far, several efficient ECG compression techniques have been reported in literature such as Linear Predictive Coding (LPC), Waveform coding and Subband coding. In these techniques, more sophisticated signal processing techniques are employed. Linear predictive coding is robust tool widely used for analyzing speech and ECG signal in various aspects such as spectral estimation, adaptive filtering and data compression [4]. Several efficient methods [7-9] have reported in literature based on linear prediction. While in subband decomposition, spectral information is divided into a set of signals that can then be encoded by using a variety of techniques. Based on subband decomposition, various techniques [8-13] have been devised for the ECG signal compression. In the past, marked researches have made in the many transformation methods such as Discrete Cosine Transform (DCT), Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT) which are extensively used in data compression. Here, compression is achieved by transforming original signal into another domain to compact much of the signal energy into a small number of transform coefficients. In this way, many small transform coefficients can be discarded in the hope of achieving better compression. The fast Fourier transform is a discrete Fourier transform (DFT) algorithm which reduces the number of computations required and is exploited for analyzing signal in frequency domain. Discrete cosine transform gives nearly optimal performance in the typical signal having high correlations in adjacent samples. The detailed discussion on different techniques based on FFT and DCT is given in [14-17]. During the last decade, the Wavelet Transform, more particularly Discrete Wavelet Transform has emerged as powerful and robust tool for analyzing and extracting information from non-stationary signal such as speech signal and ECG signal due to the time varying nature of these signals. Non-stationary signals are characterized by numerous abrupt changes, transitory drifts, and trends. Wavelet has localization feature along with its time-frequency resolution properties which makes it suitable for analyzing non-stationary signals such as speech and electrocardiogram (ECG) signals [18]. Recently, several other methods [19-24] have been developed based on wavelet or wavelet packets for compressing ECG signal. In above context, therefore, this paper presents a new optimized wavelet filter bank for ECG signal compression whose coefficients are derived from linear optimization using different windows. The paper is organized as follows. A brief introduction has been provided in this section on the existing compression techniques of ECG signal. Section 2 discusses overview of discrete wavelet transform (DWT). In Section 3, methodology for designing wavelet filter bank is presented. Section 4 reviews the compression

2891

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

methodology based DWT. Finally, a comparison of results obtained with wavelet filter bank and other wavelet filters is carried out in Section 5, followed by concluding remarks in Section 6 2. Overview of Discrete Wavelet Transform The Wavelet Transform has emerged as a powerful mathematical tool in many areas of science and engineering, more so in the field of data compression. The concept of wavelets was first introduced by Grossman and Morlet in 1984 to analyze signal structures of very different scales, in the framework of seismic signals [18, 23]. Basic principal of wavelet transform is that it decompose the given signal in too many functions by using property of translation and dilation of a single prototype function, called a mother wavelet (ψ (t )), defined as −1 § t −b · ψ a ,b ( t ) = a 2 ψ ¨ (1) ¸ © a ¹ where, a, b ∈ R, a ≠ 0 defines the dilation factor applied to the mother wavelet and b is a translation factor. In other words, wavelets are oscillations localized in time, which means their energy is also localized in time. There are many different types of mother wavelets such as Mexican hat, Morlet, Meyer, Haar and Daubechies, which presents different forms and specificities. The choice of wavelet determines the final waveform shape. To have a unique reconstructed signal from wavelet transform, we need to select the orthogonal wavelets to perform the transforms. A generic wavelet must satisfy the admissibility criteria and also belong to the space of square integral functions. A continuous-time wavelet transform of a signal (f (t)) is defined as −1 ∞ §t −b · W f (b, a) = a 2 ³ f (t )ψ * ¨ (2) ¸dt © a ¹ −∞ where, the asterisk denotes a complex conjugate and multiplication of a

−1 2

is for energy normalization

purposes so that the transformed signal will have the same energy at every scale. Hence, wavelets has adaptive nature, present a large time base for analyzing the low frequency components, and have a better time resolution for analyzing phenomena that are more transitory. As a and b are continuous over R. (over the real number), there is often redundant in CWT representation of the signal. A more compact representation can be found with a special case of WT, called the Discrete Wavelet Transform (DWT) where only the required wavelet coefficients for the reconstruction of x(t) are kept. This is possible by sampling the dilation and translation variables a and b according to criterion given in [18, 24]: (3) a = 2 − m and b = n 2 − m where, m and n is the set of the positive integers (Z). Substituting Eq. (3) into (2), DWT of a signal f (t) is

DWTψ f (m, n) =

∞

³

−∞

f (t )ψ *( m.n ) ( t )dt

where, ψ m , n (t ) = 2 − mψ (2 m t − n)

(4)

(5)

However, for computing Eq. (4), an infinite number of terms are required. Therefore, to overcome this problem, a new family of basis functions called scaling functions (φm, n (t )) was introduced, which are derived just similar to the wavelets: φm , n (t ) = 2 − m φ (2m t − n)

2892

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

where, φ (t ) is the mother scaling function. These scaling functions are complementary basis for wavelet function basis. Due to this, the multiresolution analysis (MRA) of signal is possible. MRA means decomposition a signal into different frequency bands. For MRA, the average and the details features of the signal is found via scalar products with scaling signals and wavelets. The algorithm of wavelet signal decomposition is illustrated in Fig 1. x0 (n)

y1 Level-1 DWT coefficients h(n)

2

g(n)

2

h(n)

2

g(n)

2

y2 Level-2

(a) 2

h(n)

2

g(n)

X(n) 2

h(n)

2

g(n)

(b) Fig. 1(a). Filter bank representation of DWT decomposition, (b) reconstruction of DWT decomposition. At each step of DWT decomposition, there are two outputs: scaling coefficients xj+1(n) and the wavelet coefficients yj+1(n). These coefficients are given as 2n

x j +1 (n) = ¦ g (2n − i ) x j (n)

(6)

i =1

and 2n

y j +1 (n) = ¦ h(2n − i ) x j ( n)

(7)

i =1

where, the original signal is represented by x0(n) and j shows the scaling number. Here g (n) and h (n) represent the low pass and high pass filters, respectively. The output of scaling function is input of next level of decomposition, known as the approximation coefficients. The approximation coefficients are low-pass filter coefficients, and high-pass filter coefficients are detail coefficients of any decomposed signal. The relation between the low-pass and high-pass filter and the scalar function and the wavelet can be states as: φ (t ) = ¦ h(k )φ (2t − k ) (8) k

ψ (t ) = ¦ h(k )ψ (2t − k )

(9)

k

These low-pass h (n) and high filters g (n) has mirror image at quadrature frequency, therefore filters satisfying this condition are known as the Quadrature Mirror Filters (QMF), commonly used in many engineering signal processing applications. In this paper, a two-channel QMF bank is designed and used

2893

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

as low-pass and high pass filters of wavelet for ECG signal processing. A Detailed discussion on wavelet and filter bank is given in [18, 24] and the references therein.

3. Design of Wavelet Filter Bank Among the various multirate systems, a quadrature mirror filter (QMF) bank was the first type of filter bank used in signal processing applications for separating signals into subbands and reconstructing them from individual subbands. The multirate filter banks suffer from phase distortion, amplitude distortion and aliasing distortion. Phase distortion is eliminated by use of the linear phase FIR filters. Amplitude distortion can be minimized using computer aided techniques. In case of a two-channel QMF bank shown in Fig. 2, the aliasing distortion is eliminated by

H1(z) = H0 (−z), G0 (z) = H1(−z) and G1(z) = −H0 (−z)

(10)

where, H 0 ( z ) and H1 ( z ) are the analysis filters, while G0 ( z ) and G1 ( z ) are the synthesis filters. If H 0 ( z ) is a finite impulse response (FIR) filter of even length having frequency response

H 0 (e jω ) = e− jω N /2 H 0 (ω ) (11) where, H 0 (ω ) is the amplitude response of a filter. From equations (9) and (10), the overall transfer function of QMF bank is reduced to

( )

T e jω =

e − jω N ­ jω ® H0 e 2 ¯

( )

2

(

+ H 0 e j (ω −π )

)

2½

(12)

¾ ¿

For perfect reconstruction, Eq. (13) must be satisfy

( )

H 0 e jω

2

(

+ H 0 e j (ω −π )

)

2

=1

(13)

In a two-channel QMF bank, perfect reconstruction (PR) is possible if Eq. (13) is satisfied. If this Eq. is evaluated at Ȧ = ʌ/2 , then, it leads to

H 0 (e jπ /2 ) = 0.707

(14)

Therefore, in this paper, new wavelet filter banks are designed with Kaiser Window based on algorithm given in [24-26]. In this algorithm, cut-off frequency is varied in each iteration so that Eq. (14) is approximately satisfied. A summary of steps for designing wavelet filter are: H0 (z)

2

2

G0(z)

H1(z)

2

2

G1(z)

y(n)

x(n)

Analysis section

Synthesis section

Fig.2 Two-channel QMF bank.

2894

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

Step 1: Specify design specifications such as filter taps (N), cutoff frequency ( ωc ), pass-band ripple and stop-band ripple. Step 2: Initialize Tolerance (Tol) and Step size (Step). Step 3: Evaluate a prototype filter coefficient using Kaiser and Blackman windows with N and

ωc before optimization start and the magnitude response of designed filter (MRD) at ω = π / 2. Also compute error or deviation of magnitude response of designed (MRD) filter from the ideal magnitude response (MR) given by Eq. (14). error = MR − MRD (15) Step 4(a): If tolerance is not satisfied, then ωc is varied in two ways: a. If MR >MRD, increase ωc = ωc + step b. Otherwise ωc = ωc − step.

Step 4(b): If Tol is satisfied. Then, design the other filters using this filter Redesign the prototype filter using new ωc and same order. Calculate MRD and also

Step 5: Error. Step 6:

Step = step /2. Go to step 4 till tolerance is not satisfied.

This method has been exploited for deriving wavelet filters with filter taps 16, 18 and 20 with Kaiser Window. The filter coefficients obtained are listed in Table I and corresponding wavelet are plotted in Fig. 3.

Kaiser 8

Kaiser 9

Kaiser 10

0.6

0.6

0.6

0.4

0.4

0.4

0.2

0.2

0.2

0

0

0

-0.2

0

10 N=16

20

-0.2

0

10 N=18

20

-0.2

0

10 N=20

Fig. 3 The plot of developed wavelet filters using Kaiser Window.

20

2895

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

Table 1 Wavelet Filter coefficients based on Kaiser Filter Coefficients

Filter Tap (N) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.000078243837 -0.001173173344 -0.001956998875 0.016338773143 -0.002485431166 -0.080722825679 0.08385299777 0.486068414306 0.486068414306 0.083852997777 -0.080722825679 -0.002485431166 0.016338773143 -0.001956998875 -0.001173173344 0.000078243837

0.0000725366460 0.0005407984977 -0.0034795868851 -0.0021973248409 0.0233983104728 -0.0070558086067 -0.0870987578319 0.0927756691044 0.4830441634436 0.4830441634436 0.0927756691044 -0.0870987578319 -0.0070558086067 0.0233983104728 -0.0021973248409 -0.0034795868851 0.0005407984977 0.0000725366460

-0.0000617450772 0.0006149264049 0.0011124070339 -0.0066294793195 -0.0016528844604 0.0295534665919 -0.0116352472194 -0.0913779903113 0.0997220984067 0.4803544479505 0.4803544479505 0.0997220984067 -0.0913779903113 -0.0116352472194 0.0295534665919 -0.0016528844604 -0.0066294793195 0.0011124070339 0.00061492640490.0000617450772

4. Compression based on Multiresolution analysis with Lossless encoding In this paper, the beta wavelet filters are used for the ECG signal compression. For ECG compression, three most commonly used steps are DWT decomposition, thresholding and Entropy encoding [27-29]. Using described methodology, obtained the compressed ECG signal data and its reconstructed signal. Further, QRS detection is apply for the evaluation of relevant signal clinical information at the reconstruction of signal [30-31]. 4.1. Compression methodology The compression methodology described in Fig. 4 proceeds with following three steps: Step I. In this step, the mother wavelet is chosen, and then DWT decomposition is performed on the ECG signal. Several different criteria can be used for selecting the optimal wavelet filter. For example, the optimal wavelet filter must minimize the reconstructed error variance and also maximize signal to noise ratio (SNR). In general, the mother wavelets are selected based on the energy conservation properties in the approximation part of the wavelet coefficients. Then, the decomposition level for DWT is selected which usually depends on the type of signal being analyzed or some suitable such as entropy. Here, the beta wavelets are used as the mother wavelet and 4-level decomposition of DWT is applied on the ECG signal. Step II. After computing the wavelet transform of the ECG signal, compression involves truncating wavelet coefficients below a threshold which make a fixed percentage of coefficients equal to zero. Two

2896

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

different approaches are available for calculating thresholds. The first type is known as global thresholding which involves taking the wavelet decomposition of the signal and keeping the largest absolute value coefficients. In this, the threshold value is set manually, this value is chosen from DWT coefficient (0….xmaxj), where xmaxj is the maximum value of coefficient. The second approach is known as level thresholding in which the threshold value is calculated using Birge-Massart strategy [32] and is well suitable in case of signal compression. In this paper, level thresholding is applied at the different decomposition levels.This strategy keeps all the approximation coefficients at the decomposition level N, the detail coefficients to be preserving at level i starting from 1 to N decomposition level are described as: M (16) ni = α ( N + 2 − i) where, Į is a compression parameter and its value is typically 1.5. The value of M denotes the distribution of the wavelet coefficients in the transform vector, its value depend on the length of approximation coefficients. In this paper, the Į is modified to α new which is given by

α new = eα / 2

for α =1.5

(17)

and threshold value is computed using this new value of Į. Step III. In this step, signal compression is further achieved by efficiently encoding the truncated smallvalued coefficients. The resulting signal data contains same redundant data which is waste of space. In this paper, run-length encoding is applied on the redundant data to overcome the redundancy problem without any loss of signal data Run length coding is a simple form of data compression in which runs of data are stored as a single data value and count, rather than as the original run [26-28]. In Fig. 4, it is shown that how RLE encode the data and give compressed signal data. Finally, the compressed ECG signal is obtained.

Signal Decomposition based on Analysis filter

Thresholding

ECG Signal

aaaabbbccaaadddbbaaccd

a4b3c2a3d3b2a2c2d1

Reconstructed ECG signal

Run-length Encoding

Compressed data value

Compressed Data

Synthesis Filter

Run-length Decoding

Fig. 4 Lossless compression methodology based on Kaiser Wavelet filter and run-length encoding

3.2. QRS Detection The QRS complex is an important waveform of electrocardiogram signal. It is having different shapes in different cases and represents the cardiac cycle and information about heart rate. In the ECG signal analysis, QRS detection is very useful, and it shows the accuracy of algorithm in case of signal compression and validates the reconstructed signal. There are several methods available for the QRS detection: Thresholding, transform methods, neural network approaches, morphology, matched filter and

2897

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

the methods described in the reference [29-30]. In this paper, QRS detection is done based on energy of decomposed signal coefficients and thresholding.

5. Results and Discussions In this section, a wavelet filter bank based methodology has been used for ECG signal compression. ECG records have been obtained from MIT-BIH Arrhythmia Database [33]. Here, wavelet filter bank with filter taps 16, 18 and 20 designed with Kaiser Window is exploited for ECG signal compression. Several examples are included to illustrate the effectiveness of the developed wavelet filters. The performance of the developed wavelet filters in the field of ECG signal compression can be evaluated by considering the fidelity of the reconstructed signal to the original signal. For this, following fidelity assessment parameters are considered [4-5]:

•

Percent root mean square difference (PRD): 1 /2

§ Reconstructed noise energy · PRD = ¨ ¸ © Origional signal energy ¹ •

½ ° 2¾ x ( n) − y ( n) ° ¿

¦ x ( n) 2

¦

(19)

Compression ratio (CR):

CR = •

(18)

Signal to noise ratio (SNR):

­ ° SNR = 10 log10 ® °¯ •

x 100 %

Number of significant wavelet coefficients Total number of wavelet coefficients

(20)

Retained energy (RE)

RE=

x ( n)

2

x ( n ) − y ( n)

2

× 100

(21)

Table 2 and Table 3 lists the simulation results obtained with beta wavelet filters (unmodified and new modified threshold) at level thresholding. Fig. 5 shows the plot of the original ECG signals (MIT/BIH100), its reconstructed version and preserved QRS impulse also with wavelet filters based on Kaiser Window. It can be seen from Table 1 that a significant compression ratio is achieved with the wavelets filters based on beta function at 4th level of decomposition. The average compression ratio obtained with filter tap N=16 Kaiser Wavelet filter is 5.40%, for N=18, compression ratio is 5.32% and in case of N=20, compression ratio is 5.45%. While, in case of the modified threshold from Table 3 compression ratio is 7.08% for N=16, at N=18, compression ratio is 7.06% and in case of N=20, compression ratio is 6.98%. It is also evident that the good fidelity measures can be achieved with the beta wavelet within acceptable range [5]. The average PRD obtained with Kaiser Wavelet filters is 1.91% and 2.64% for N=16 filter tap, 2.22% and 4.18% for N=18 and for N=20, 3.10% and 3.11% at both threshold approaches. From Table 2 and Table 3 the compression results illustrate that the new modified proposed threshold approach is give approximate 30% higher compression with small fidelity (PRD and SNR) degradation.

2898

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

TABLE 2 Fidelity assessment parameters obtained with Kaiser Wavelet filters at decomposition level 4

Signal

Wavelet Filter

RE

CR

PRD

SNR

QRS Detection (%)

MIT-BIH Rec. 100: M, 69

Kaiser8 Kaiser8

99.7772

5.57

5.46

25.24

99.90

MIT-BIH Rec. 112: M, 54

99.9711

5.28

1.18

38.50

99.00

MIT-BIH Rec. 117: M, 69

Kaiser8

99.9269

5.39

0.01

84.60

99.99

MIT-BIH Rec. 124: M, 77

Kaiser8

99.9891

5.34

1.02

38.81

99.99

MIT-BIH Rec. 100: M, 69

Kaiser9

99.7830

5.42

5.29

25.51

98.00

MIT-BIH Rec. 112: M, 54

Kaiser9

99.9690

5.36

1.74

35.16

97.50

MIT-BIH Rec. 117: M, 69

Kaiser9

99.9257

5.30

0.13

57.08

99.99

MIT-BIH Rec. 124: M, 77

Kaiser9

99.9895

5.20

0.96

40.35

99.90

MIT-BIH Rec. 100: M, 69

Kaiser10

99.7776

5.54

5.40

25.34

98.00

MIT-BIH Rec. 112: M, 54

Kaiser10

99.9693

5.25

0.57

44.82

99.90

MIT-BIH Rec. 117: M, 69

Kaiser10

99.9326

5.65

5.47

25.23

99.30

MIT-BIH Rec. 124: M, 77

Kaiser10

99.9886

5.37

0.96

40.30

99.80

TABLE 3 Fidelity assessment parameters obtained with Kaiser wavelet filters at Decomposition level 4 with modified thresholding

Signal

Wavelet Filter

RE

CR

PRD

SNR

QRS Detection (%)

MIT-BIH Rec. 100: M, 69

Kaiser8 Kaiser8

99.3499

7.04

5.46

25.24

99.90

MIT-BIH Rec. 112: M, 54

99.9600

6.87

0.78

42.07

99.90

MIT-BIH Rec. 117: M, 69

Kaiser8

99.7812

7.26

3.37

29.43

99.99

MIT-BIH Rec. 124: M, 77

Kaiser8

99.9798

7.14

0.98

40.14

99.90

MIT-BIH Rec. 100: M, 69

Kaiser9

99.2924

6.94

5.49

25.20

95.50

MIT-BIH Rec. 112: M, 54

Kaiser9

99.9558

7.11

1.29

37.77

98.00

MIT-BIH Rec. 117: M, 69

Kaiser9

99.7334

6.92

8.98

20.92

99.90

MIT-BIH Rec. 124: M, 77

Kaiser9

99.9820

7.29

0.96

40.35

99.90

MIT-BIH Rec. 100: M, 69

Kaiser10

99.3410

7.09

5.40

25.34

97.00

MIT-BIH Rec. 112: M, 54

Kaiser10

99.9534

6.87

1.51

36.40

98.00

MIT-BIH Rec. 117: M, 69

Kaiser10

99.7850

6.87

4.59

26.75

99.80

MIT-BIH Rec. 124: M, 77

Kaiser10

99.9772

7.09

0.96

40.30

99.90

When Kaiser Wavelet filters are compared, the Kaiser Wavelet filter based on filter tap N=20 is performed better in term of compression ratio and PRD as compare to lower filter tap at both threshold approaches. Therefore, Kaiser Wavelet can be effectively used for the ECG signal compression and modified threshold approach is suitable for selecting optimum threshold value in case of signal compression.

2899

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

ECG Signal (MIT-BIH Rec.100: M,69) 1 0.5 0 -0.5 -1

0

100

200

300

400

500

600

700

800

900

1000

Reconstructed ECG Signal (MIT-BIH Rec.100: M,69) Based on kaiser window 1 0.5 0 -0.5 -1

0

100 200 300 400 500 600 700 800 900 1000 QRS Complex in reconstructed ECG signal (MIT- BIH Rec. 100: M,69) Based on kaiser window

1 0.5 0 -0.5 -1

100

200

300

400

500

600

700

800

900

Fig. 5. Original ECG signal, reconstructed version and QRS Complex detection in Reconstructed ECG signal using Kaiser Wavelet filter TABLE 4 Comparison of performance of the developed method with other existing methods

ECG Signal MIT-BIH Rec. 100: MIT-BIH Rec. 117: MIT-BIH Rec. 100: MIT-BIH Rec. 117: MIT-BIH Rec. 100: MIT-BIH Rec. 117: MIT-BIH Rec. 100: MIT-BIH Rec. 117: MIT-BIH Rec. 117:

M, 69 M, 69 M, 69 M, 69 M, 69 M, 69 M, 69 M, 69 M, 69

Methods

CR

PRD

Kaiser wavelet [N=16] Kaiser wavelet [N=16] Kaiser wavelet [N=18] Kaiser wavelet [N=18] Kaiser wavelet [N=20] Kaiser wavelet [N=20] Miaou et al. [34] Manikandan et al. [35] Blanco et al. [36]

7.04 7.26 6.94 6.92 7.09 6.87 4.60 4.55 6.52

5.46 3.37 5.49 8.98 5.40 4.59 3.30 5.60 5.75

1000

2900

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

A comparison of the developed wavelet filters with other existing wavelet filters is carried out. For this, ECG records (MIT-BIH Rec. 100: M 69, MIT-BIH Rec. 112: M 54, MIT-BIH Rec. 117: M 69, MITBIH Rec. 124: M 77) have been taken from MIT-BIH Database and the performance measures obtained in each wavelet filter using same methodology and at level threshold value are listed in Table 2 and Table 3. As, it can be seen that the performances of the Kaiser wavelet filters as: Average retained energy (RE) for the filter tap N=16, 18 and 20 respectively are 99.92%, 99.91% and 99.92%. 100 CR

80

PRD SNR

60 40 20 0

1

2 3 4 No. of Decomposition level

5

(a)

100 80 60 40 20 0

1

2 3 4 No. of Decomposition level (b)

5

1

2 3 4 No. of Decomposition level (c)

5

100 80 60 40 20 0

Fig. 6 Performance of optimized wavelet based on Kaiser Window with filter tap (a) N= 16, (b) N= 18 and (c) N=20 for different decomposition levels

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

Table 4 illustrates that the performance of the methodology based on Kaiser Wavelet is superior as compared to other methods. Fig. 6 depicts the plots of performance obtained with Kaiser Wavelet at different decomposition levels using level thresholding. As it can be observed from the simulation results included in the figures that the compression ratio (CR) is increased at higher threshold values, while the quality of reconstructed signal is degraded. Thus, by choosing appropriate threshold value, good compression ratio can be achieved with preserving all clinical information.

6. Conclusions A wavelet based methodology is presented for ECG signal compression. In this methodology, new wavelet filters are derived using Kaiser Window techniques. Simulation results included in this paper clearly show the key advantageous features of the developed wavelet filters over others in the field of biomedical signal processing. It is found that the developed wavelet filter banks yields more compression with preserving all clinical information. All fidelity measuring parameters are improved. Therefore, it is concluded that it can be very effectively used in ECG signal compression.

Reference [1]. Afonso VX, Tompkins WJ, Nguyen TQ, Luo S. ECG beat detection using filter banks. IEEE Trans. Biomed. Eng. 1999; 46: 192–202. [2]. Wang X, Meng J. A 2-D ECG compression algorithms based on wavelet transform and vector quantization. Digital Signal Processing, 2008; 18:179-188. [3]. Blanco-Velasco M, Cruz-Roldan F, Lopez-Ferreras F, Bravo-Santos A, Martinez-Murioz D. A low computational complexity algorithm for ECG signal compression. Medical Engineering & Physics, 2004; 26: 553-568. [4]. Addison PS. Wavelet transform and the ECG: a review. Institute of Physics Publishing- Physiological Measurement, 2005; 26:155-199. [5]. Jalaleddine SMS, Hutchens CG, Strattan RD, Coberly WA. ECG data compression techniques: a unified approach.. IEEE Trans. Biomed. Eng. 1990; 37( 4): 329-343. [6]. Ole-Aase S, Nygaard R, Husoy JH. A comparative study of some novel ECG data compression technique. IEEE Nordic Signal Processing Symposium (NORSIG’98),1998; 3: 273-276. [7]. Nave G, Cohen A. ECG compression using long-term prediction. IEEE Trans. Biomed. Eng. 1993; 40: 877–885. [8]. Ramakrishna AG, Saha S. ECG coding by wavelet-based linear prediction. IEEE Trans. Biomed. Eng. 1997; 44(12): 1253–1261. [9]. Al-Shroud A, Abo-Zahhad M, Ahmed SM. A novel compression algorithm for electrocardiogram signals based on the linear prediction of the wavelet coefficients. Digital Signal Processing, 2003; 13:604–622. [10]. Aydin MC. ECG data compression by sub-band coding. IEEE electronics letters, 1991; 27:359-360. [11]. Husoy JH, Gjerde T. Computationally efficient sub-band coding for ECG signals. Medical Engineering and Physics, 1996; 18( 2):132–142. [12]. Zigel Y, Cohen A, Katz A. Signal Compression Using Analysis by Synthesis Coding. IEEE Trans. on Biomed. Eng. 2000; 47(10):1308-1316. [13]. Blanco-Velasco M, Cruz-Roldán F, Moreno-Martínez E, Godino JI, Barner KE. Embedded filter bank-based algorithm for ECG compression. Signal Processing, 2008; 88(6):1402–1412. [14]. Al-Nashash HAM. A dynamic Fourier series for the compression of ECG using FFT and adaptive coefficient estimation. Med. Eng. Physics, 1995; 17: 197-203. [15]. Al-Hinai N, Neville K, Sadik AZ, Hussain ZM. Compressed Image Transmission over FFT-OFDM: A Comparative Study. Australasian Telecommunication Networks and Applications Conference, 2007; 465-469. [16]. Batista LV, Melcher ElUM, Carvalho LC. Compression of ECG signals by optimized quantization of discrete cosine transform coefficients. Medical Engineering & Physics, 2001; 23:127-134. [17]. Aggarwal V, Patterh MS. Quality Controlled ECG Compression using Discrete Cosine Transform (DCT) and Laplacian Pyramid (LP). Multimedia, Signal Processing and Communication Technology, IMPACT 09, 2009; 12-15.

2901

2902

K. Ranjeet et al. / Procedia Engineering 38 (2012) 2889 – 2902

[18]. Mallat StG. A Theory for Multiresolution Signal Decomposition: The Wavelet Representation. IEEE Transaction on Pattern Analysis and Machine Intelligence, 1989; 11:674 - 693. [19]. Ahmeda SM, Abo-Zahhad M. A new hybrid algorithm for ECG signal compression based on the wavelet transformation of the linearly predicted error. Medical Engineering & Physics, 2001; 23: 117-126. [20]. Rajoub BA. An Efficient Coding Algorithm for the Compression of ECG Signals Using the Wavelet Transform. IEEE transactions on biomedical engineering, 2002; 49: 355-362. [21]. Kim BS, Yoo SK, Lee MH. Wavelet-Based Low-Delay ECG Compression Algorithm for Continuous ECG Transmission. IEEE transaction on information technology in biomedicine, 2006; 10: 77-83. [22]. Kumari RSS, Sadasivam V. A novel algorithm for wavelet based ECG signal coding. Computers and Electrical Engineering. 2007; 33: 186-194. [23]. Vetterli M, Kovacevic J. Wavelets and subband coding. Prentice Hall, 1995. [24]. Ranjeet K, Kumar A, Rajesh KP. Electrocardiogram signal Compression Using Beta Wavelet. J. Math. Model. Algor. 2012; DOI: 10.1007/s10852-012-9181-9. [25]. Kumar A, Singh GK, Anand RS. Near perfect reconstruction quadrature mirror filter. Int. J. of Computer Science and Engineering, 2008; 2(3):121-123. [26]. Kumar A, Ranjeet K. ECG Signal Compression using optimized Wavelet Filter Bank. Int. J. Signal and Image System Eng. (Accepted). [27]. Akhter S, Haque MA. ECG Compression using Run Length Encoding. European signal processing (EUSIPCO-2010), 2010; 1645-1649. [28]. Hellebrand S, Wurtenberger A. Alternation Run-length Coding- Technique for Improved test Data Compression. Workshop on Test resource partition. 2002. [29]. Lih-jen K., Yuan-pei L. Least squares- adapted edge-look-ahead prediction with run-length encoding for lossless compression of images. Acoustics, speech and signal processing (ICASSP-2008), 2008; 1185-1188. [30]. Kohler BU, Hennig C, Orglmeister R. The Principles of Software QRS Detection. IEEE Engineering in Medicine and Biology, 2002; 42-57. [31]. Gutierrez D, Garcia-Nocetti F, Solano-Gonzalez J. Classification of multichannel EEG data using length/energy transforms. IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, 2005; 221224. [32]. Misiti M, Misiti, Y, Oppenheim G, Poggi J. Matlab Wavelet Tool Box. The Math Works Inc. 2002 [33]. http://www.physionet.org [34]. Miaou S, Yen H. Quality driven gold washing adaptive vector-quantization and its application to ECG data compression. IEEE Transactions on biomedical Engineering, 2000; 47(2):209–218. [35]. Manikandan MS, Dandapat S. Wavelet threshold based ECG compression using USZZQ and Huffman coding of DSM. Biomedical Signal Processing and Control, 2006; 1: 261-270. [36]. Blanco-Velasco M, Cruz-Roldán F, Moreno-Martínez E, Godino JI, Barner KE. ECG compression withretrieved quality guaranteed. IEEE ELECTRONICS LETTERS. 2004; 40(23).