8th Seminar on Neural Network Applications in Electrical Engineering, NEUREL-2006

Faculty of Electrical Engineering, University of Belgrade, Serbia, September 25-27, 2006 http://neurel.etf.bg.ac.yu, http://www.ewh.ieee.org/reg/8/conferences.html

20 06

ST Segment Change Detection by Means of Wavelets

Neboja Milosavljevi, Aleksandar Petrovi
amplitude change between ST point and PR point values grater then 0.1mV. Holter signals are prone to baseline drift. However, proposed algorithm performs over each beat separately. In that sense global baseline drift hardly affects each single beat. That is to say that relative amplitude changes due to baseline floating, typically do not interfere with ST deviation measurements in each beat.
1.5

Abstract This research aims to contribute to the automatic interpretation of long sequences of electrocardiograms (ECG) typical for Holter monitoring. We developed a method that uses wavelets for extracting ECG patterns that are characteristic for myocardial ischemia. It was our intention to detect the beats in the simplest possible manner and generate a quantitative estimate of myocardial ischemia likelihood which would suit needs of cardiologists. Biorthogonal wavelets were applied in order to define ST segment properties at different scales. The new method was tested on data from the European STT change database. Results show that this method it effective for distinguishing normal from ischemic ECGs. The element that makes the distinction is the correlation of number of ST deviations with the time of consecutive appearances.

Amplitude [mV]

Keywords Biorthogonal wavelets, ECG parameters,

ST segment deviation.

0.5

PR point

ST point

I. INTRODUCTION

-0.5

he object of automation of ECG analysis is to reduce the time required for human interpretation and analysis of ECG recordings from the Holter monitoring equipment. The algorithms which strive to detect specific ECG deviations in a computationally simple manner seem to be very popular [1]. Some of the existing techniques use the series of bandpass filters to extract the QRS complexes from the ECG signal, which under severe baseline drift and other high frequency noises fail to detect the characteristic points to an acceptable accuracy [2]. Wavelet Transformation (WT) has shown to be substantially noise-proof in ECG segmentation [2], [3], and thus very appropriate for ST-T segment (Fig. 1) extraction. In Holter monitoring systems it is of great importance to automatically detect ischemic ST deviation episodes (ST segment elevation or depression) with good accuracy. However, there is no strict definition of an ischemic ST episode and it is quite challenging to unify very different approaches of detecting these episodes. Here we define a single ST deviation as an absolute

P wave
-1 0 20

PQ QRS segment complex

40 60

ST segment
100

T wave
120 140 160

80 Samples

Fig. 1. Normal (solid line) and ST deviant (dashed line) ECG waveforms are compared (extracted from file e0121.dat in [4]). Segmentation is performed on normal ECG where horizontal axis (time) is sampled at 250 samples/sec and vertical axis represents normalized amplitude of the ECG signal. Algorithm detects ST point and PR point values.

Extensive consultations with medical experts from the Military Medical Academy in Belgrade, Serbia led to the conclusion that it is of most practical importance for a cardiologist to have a simple estimate (binary or quantitative) of whether an ECG recording includes segments that are characteristic for ischemic heart. This result should trigger the attention, and automatically annotated ST change episodes should be examined manually by an expert. Our aim here was to develop an algorithm that uses WT for identification of myocardial ischemic episodes, as well as the presentation of the results in a manner which could suit medical praxis. II. THE METHOD For ECG parameters estimation, it is desirable that the basis functions (wavelets) be symmetric/antisymmetric as proposed in [2]. In order to have a computationally simple analysis it is required for peaks to be detected as

Aleksandar Petrovi is with the School of Electrical Engineering, University of Belgrade, Serbia (phone +381-63-7762-264; e-mail: byron@sezampro.yu). Neboja Milosavljevi is with the School of Electrical Engineering, University of Belgrade, Serbia (phone +381-64-1713-218; e-mail: nebojsamilosavljevic@gmail.com).

1-4244-0433-9/06/$20.00 2006 IEEE

zero crossings (which is provided by antisymmetric basis) or local extrema (symmetric basis). It is also desirable that basis have a minimum number of sign changes. In practice, QRS complex is usually considered to be symmetrical, while T wave is less so. However, it has been shown in [2] that PR point and ST point can be estimated using biorthogonal wavelets under the assumption of QRS complex and T wave symmetry. Moreover, having as an aim proposed quantitative analysis, it is quite plausible to suppose basic QRS complex and T wave symmetry. Biorthogonal wavelets satisfy both these properties, so they are used for DWT (discrete wavelet transform) decomposition. In this paper we have used reverse biorthogonal spline wavelets implemented in rbio3.1 MATLAB package (Table 1). Decomposition and reconstruction filters satisfy further equations:
H ( ) + G ( ) K ( ) = 1 , where
H ( ) = e i / 2 (cos( / 2)) p
2

over the desired number of scales (Fig. 2).

H3 H2 H1 H ECG Filter Bank G d1,k s1,k d2,k s2,k d3,k s3,k d4,k s4,k

G3 G2

G1

Fig. 2. Wavelet decomposition of ECG signal. Decomposition performed over 4 scales with pre-filtering.

In order to obtain ST point and PR point values each decomposed ECG is segmented (Fig. 1). At the first and the second levels, QRS starting, ending and peak point are extracted. Let us denote as n R - the R peak point,
nl1 and nl 2 the beginning and ending points of QRS complex, respectively. PR point is thus calculated as: n nl 1 . PRpoint = n R l 2 2 At the fourth level, ST-T characteristic points are obtained: nT - the T peak point, nt1 and nt 2 the beginning and ending points of T wave, respectively. ST point is calculated as: n nt1 . STpoint = nT t 2 2 Wavelet decomposition introduces scale-dependent phase delay into signals. For example, each zero crossing point which corresponds to the peak of a symmetric uniphase wave is delayed for exactly 2 j 1 1 points, where j represents the scale (level).

G ( ) = 4ie i / 2 sin( / 2)
K ( ) = 1 H ( ) G ( )
2

and

i = 1 , p=3.
TABLE I: FILTER (LP), (HP) AND THEIR RECONSTRUCTION FILTERS. G, decomposition high pass filter coefficients; H, decomposition low pass filter coefficients; K, reconstruction high pass filter coefficients; L, reconstruction low pass filter coefficients (taken from rbio3.1 MATLAB help).
COEFFICIENTS OF SYMMETRIC LOW PASS ANTISYMMETRIC HIGH PASS

H 0.1768 0.5303 0.5303 0.1768

G 0.3536 1.0607 -1.0607 -0.3536

K 0.1768 -0.5303 0.5303 -0.1768

L -0.3536 1.0607 1.0607 -0.3536

A. Parameters Estimation ECG parameters are estimated in a way similar as suggested in [3]. Firstly, ECG signal is filtered so that only high frequency noise is reduced with minimal alterations in ECG amplitudes. This is important, since the amplitudes of ST deviations are taken to be the measure suggesting an ischemic pulse. This initial filter is also based upon WT. The denoising procedure is performed in three steps (MATLAB wavelet toolbox, [9]): 1) decomposition with sym8 wavelet at level 5 2) detailed coefficients tresholding choosing a threshold for each level 3) reconstruction ECG signal is further decomposed through wavelet decomposition dyadic tree where different ECG patterns are obtained through different decomposition scales. Namely, the signal is initially decomposed into low pass (smooth) and high passes (detailed) constituents. The same procedure is then applied to the smooth component and further on in the same way. This process is repeated

B. ST Deviation Analysis In our analysis we have used European ST-T change database [4] with ischemic ECG signals sampled at 250 Hz, two hours in duration each. The most elementary differentiate/threshold algorithm has been applied to extract each beat. Obtained signals were decomposed as described in A and for each beat/separate signal, an ST deviation value was calculated as , where ECGvalue( x ) denotes an amplitude at the point x of a given ECG. Cases with specific ECG beats, that is to say with indistinguishable ECG features (thus impossible segmentation), were excluded from further analysis. In that way, an ST deviation value was obtained for each beat that could be analyzed. The final report might include the comment that some beats were excluded from analysis as well as the number of the beats omitted, and therefore suggest a manual observation where required. III. RESULTS An array of ST deviations was transformed into an array which correlates number of ST deviations with time of consecutive appearances. This is to say that the algorithm counts the number of consecutive appearances
STdeviation = ECGvalue ( PRpoint ) ECGvalue ( STpoint )

A. Interpreting the Data Two or more ST change episodes could be separated by insignificant time gaps or with an unreadable ECG segment. This could lead to the conclusion that these episodes are separate, which is scarcely the case. Since the duration of the episodes are of great importance for making conclusions about myocardial ischemia, it was necessary to alleviate such shortcomings. Thus we applied a low pass filter (Chebyshev Type I filter) to the latter signal; thereby, constructed an envelope which was informative enough.
80 70 60 50 40 30 20

No. of consecutive ST deviation appearances

of ST deviations greater than 0.1mV and uses that number as a new arrays element. The procedure is repeated for each ST change episode.

4 3.5 3 2.5 2 1.5 1 0.5 0 -0.5

2000

4000 6000 Time correlated axis

8000

10000

No. of consecutive ST deviation appearances

Fig. 4. Number of ST deviations correlated with time of consecutive appearances case with few ST-T deviations. (file e0115.dat in [4] has been analyzed)

In Fig. 4 we can clearly see that vertical axis values are small compared to those in Fig. 3 or Fig. 5. Moreover, there are less high-valued ripples, and they occur with significant time gaps. This suggests that analyzed ECG is near normal in the terms of ST change.
250

10
No. of consecutive ST deviation appearances

0 -10

200

150

1000

2000

3000

4000

5000

6000

7000

Time correlated axis

Fig. 3. Number of ST deviations correlated with time of consecutive appearances functions envelope. (file e0111.dat in [4] has been analyzed)

100

50

Fig. 3 shows typical results from the readings in a patient with myocardial ischemia (manifested through Tand ST-wave change). High values at vertical axis suggest that a long ischemic episode has occurred (A part, Fig. 3). Within normal ECGs we expect that these values gather near zero. B part at Fig. 3 is a direct consequence of final low pass filtering and suggests that a series of shorter, but still significant, ST episodes have occurred with unsubstantial time gaps among them. B parts are not expected with normal ECGs. Negative ordinate values are a co product of LP filtering and bare no practical significance. Beside ECG analyzed in Fig. 3 which we considered to be a mid-case in the terms of number of ST deviations, we selected two more examples in order to demonstrate the efficiency of the algorithm. Fig. 4 presents performance of the algorithm in the case where there were only few ST changes, while Fig. 5 shows the case with large number of ST changes.

-50

-100

1000

2000

3000

4000 5000 6000 Time correlated axis

7000

8000

9000

Fig. 5. Number of ST deviations correlated with time of consecutive appearances case with a considerable number of ST-T deviations. (file e0121.dat in [4] has been analyzed)

In Fig. 5 high vertical axis values suggest long ST change episodes, while high-valued and almost unified ripples describe relatively long ST change segments which occurred within short time period. B. Estimating Algorithms Efficiency Annotation files provided for ST-T change ECG signals in [4] contain information about ST and T-wave deviations as well as the number of heart beat in that particular ECG at which a change occurred. ST changes are annotated by values 18 and T-wave changes by 19. In order to demonstrate similarity which exists between the algorithms outcome and manually annotated ST-T changes, in Fig. 6 and Fig. 7 we present the manually annotated change/time dependence for the same ECGs analyzed in Fig. 4 and Fig. 5, respectively.

20

19.5

19 Deviation code

18.5

accuracy could be obtained by improving beat separation algorithm. We concluded that the developed method can be applied in Holter monitoring when it is necessary to decide whether a patient suffers from myocardial ischemia manifested by ST segment change, or not. The algorithm can also be used for annotating ST change episodes which could be manually examined if that was suggested by the algorithms outcome. ACKNOWLEDGMENT

18

17.5

17

2000

4000

6000 No. of heart beat

8000

10000

We would like to express our gratitude to professor Dejan Popovi for his support and helpful comments, and docent Dr Slobodan Obradovi for valuable discussions. REFERENCES
[1] [2] [3] [4] [5] Li CW, Zheng CX, et al., Detection of ECG characteristic points using the wavelet transforms, IEEE Trans. on BME 1995; Jan: 2128 Ranjith P, Baby PC, Joseph P, ECG analysis using wavelet transform: application to myocardial ischemia detection, ITBMRBM 2003; 24: 44-47. Sivannarayana N, Reddy DC, Biorthogonal wavelet transforms for ECG parameters estimation, Medical Engineering and Physics 1999; 21: 167-174. European ST-T change database, [Online] http://www.physionet.org/. Jager F, Moody GB, Taddei A, Mark RG, Performance measures for algorithms to detect transient ischemic ST segment changes, Computers in Cardiology, Los Alamitos, CA: IEEE Computer Society Press 1991; p. 372-396. Garcia J, Sornmo L, Olmos S, Laguna P, Automatic detection of ST-T complex changes on the ECG using filtered RMS difference series: application to ambulatory ischemia monitoring, IEEE Trans Biomed Eng 2000; 47(9): 1195-1201. Sakar TK, Su C, Adve R, Salazar-Palma M, Garcia-Castillo L, Boix RR, A tutorial on wavelets from an electrical engineering perspective, part 1: discrete wavelet techniques, IEEE Antennas and Propagation Magazine 1998; 40(5). Dash PK, Electrocardiogram monitoring, Indian Journal of Anaesthesia 2002; 46(4): 251-260. MATLAB 7.01 reference guide

Fig. 6. Manually annotated ST-T changes which correspond to Fig. 4. (file e0115.atr in [4] has been used)

By comparing Fig. 4 and Fig. 6, as well as Fig. 5 and Fig. 7 it can be concluded that a strong correlation exists between the automatic and manual ECG analysis for the purpose of ST-T change/ischemia detection. Direct relationship exists between the time axes, too, but the algorithms outcome compresses time axis due to the nature of the constructed array whose envelope is the actual final result.
20

19.5

[6]

19 Deviation code

[7]

18.5

[8]
18

[9]
17.5

17

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 No. of heart beat

Fig. 7. Manually annotated ST-T changes which correspond to Fig. 5. (file e0121.atr in [4] has been used)

After such analysis, a medical doctor can examine those parts of ECG which were annotated for ST changes, if necessary. Performance on annotating ST changes of the proposed algorithm is the same as of [3], although different wavelet base function had been used. Demonstrated performance has been verified on 10 ECGs from [4], and confirms already described relationship between ECGs automatically and manually analyzed for the stated purpose. IV. CONCLUSION The method presented is computationally not demanding. It can be applied with low SNR (signal to noise ratio) and moreover, presents the results in a way that corresponds to common medical praxis. Better