A Comparison of Baseline Removal Algorithms for Electrocardiogram (ECG) based Automated Diagnosis of Coronory Heart Disease Fayyaz A. Afsar1, M. S. Riaz2 and M. Arif3 Department of Computer & Information Sciences Pakistan Institute of Engineering & Applied Sciences (PIEAS) P.O. Nilore, Islamabad, Pakistan 1 [email protected], [email protected], [email protected],

Keywords: ECG, Baseline Removal, Filtering, Wavelets, Coronary Heart Disease.

I.

T

INTRODUCTION

he ECG signal is one of the most commonly used diagnostic tools for detecting cardiac disorders. It offers the advantage of providing a sufficiently detailed view of the working of the heart non-invasively which can give useful insight for the diagnosis of a multitude of cardiac diseases especially coronary artery disorders such as Myocardial Ischemia, Injury and Infarction [1]. According to a survey by the American Health Association, one in every 12 adult Americans suffers from coronary artery disease. This percentage is significantly higher in Pakistan with one out of every 4 Pakistani citizens above the age of 35 suffering from coronary artery disorders [2]. These figures indicate the need for accurate and timely diagnosis of such disorders for initiation of proper treatment. With the development in computing technology, automatic cardiac disease diagnosis using ECG has emerged as one of the most promising areas of research. Unfortunately the ECG signal is contaminated by a variety of noise sources such as power line interference, electrode contact noise, electromyographic noise etc. [3] In the absence of noise removal techniques, the performance of automatic diagnosis algorithms is significantly reduced. A major type of noise in the ECG signal is baseline variation which stems from impedance variations between electrode and

skin resulting from movement of the electrode away from the contact area on the skin and respiration. An example of such baseline variation in the ECG signal is shown in figure 1. The Baseline signal within the ECG is a low frequency signal over the range of 0-0.8Hz. However different parts of the inherently non-stationary ECG signal such as the P-wave, T-wave and especially the ST-segment (see figure 2) have some of their corresponding frequency components over the same range. According to the findings by the American Health Association (AHA) the lowest frequency component in the ECG signal is around 0.05Hz [4]. Thus the baseline signal is a type of in-band noise. The simplest approach for removal of the baseline is to filter the ECG signal using high-pass digital filters with a cutoff of ~0.8Hz [5]. However such a filtering operation introduces distortions in the ST segment of the ECG which plays a vital role in the diagnosis of different life threatening cardiac disorders such as coronary artery heart disease (Myocardial Infarction and Ischemia). Deviation of the ST segment level point (usually defined at 60ms or 80ms after the J-point in the ECG) from the isoelectric level indicates the presence of coronary artery disease [1]. Therefore it is desirable that any method being used for removing baseline from the ECG introduces minimum distortions in the ECG so that effective diagnosis of these disorders can be done. ECG Signal with Baseline Variation 860 840 820 800 ADC Units

Abstract— This paper presents a comparison of different approaches for performing baseline removal in the electrocardiogram (ECG) signal for use in an ECG based decision support system for diagnosis of coronary heart disease. Our implementations of seven different algorithms for removal of baseline from the ECG signal have been compared which include methods based on use of linear Digital filters, Adaptive filters, Multiresolution analysis and Curve fitting or polynomial based approaches. The comparison was carried out using manual ST Segment level annotations in different ST segment deviation episodes from the European Society of Cardiology (ESC) ST-T database. Results indicate that the use of Wavelet Adaptive Filter for baseline removal produces ST segment levels which are the closest to those annotated by the human expert.

780 760 740 720 700 680 660

0

5

10

15

20 Time (s)

25

30

35

Figure 1. ECG Signal with slowly varying baseline

40

brief description of each of these methods is given in sectionIII.

Figure 2. Different components of a single beat ECG signal. The point at the junction of the QRS and the ST segment is called the J-point.

The literature gives a variety of methods for baseline removal from the ECG. The simplest approach, as discussed earlier, is to use high pass filtering. However the selection of an optimal cutoff frequency such that the filter introduces minimum distortion in the ECG is an issue. Some sophisticated filtering based approaches use adaptive filtering [6, 7], time varying digital filters using Short Time Fourier Transform (STFT) for estimation of cutoff filters [8, 9], wavelet transform [9-11]. Another paradigm is to use baseline estimation through cubic spline curve fitting [12] or through linear interpolation between isoelectric levels [13, 14]. The baseline estimate is then subtracted from the original ECG signal to produce the baseline removed (or corrected signal). This paper renders a novel comparison of seven different baseline removal techniques which have been implemented by the authors for use in conjunction with algorithms for automatic detection of Myocardial Infarction and Ischemia. The rest of the paper is organized as follows: Section-II presents briefly the different steps involved in this study, with Section-III describing the baseline removal techniques implemented in this investigation. Section-IV renders the results and concluding remarks. II.

MATERIALS & METHODS

The basic approach implemented in this paper for performance analysis of different baseline removal algorithms comprises of the following steps: A. QRS onset and offset detection The QRS complex in the ECG is the most significant component of the ECG signal and is used for beat detection and the determination of heart rate through R-R interval estimation. For the detection of ST level deviation we need the isoelectric level and the J-point for each beat and their estimation requires the detection of the onset and offset of the QRS complex respectively. For the purpose of delineation of the QRS complex we have used the approach proposed by Martínez et al. [11] which is based on the wavelet transform and it offers an excellent accuracy in terms of both QRS detection and delineation. B. Baseline removal In this step the baseline in the ECG signal is removed through different techniques that we compare in this work. A

C. Detection of isoelectric level In this stage the isoelectric level point for a beat in the ECG signal obtained after baseline removal is detected by searching backwards from the QRS onset point for the flattest (having minimum absolute value of slope) 20ms segment within 80ms. The middle sample of this flattest waveform segment is used to define the position of the isoelectric reference point Ik for the corresponding kth beat in the ECG. The mean of the flattest 20ms segment is taken as the isoelectric level, zk. D. Detection of ST Deviation The end of the QRS complex for each beat is taken as the J-point (indicated by Jk for the kth beat). Measurements of ST segment level STk are taken 80 milliseconds after the J point if the heart rate does not exceed 120 bpm, and 60 milliseconds after the J point otherwise. The ST segment deviation for the kth beat is given by

D k = ST k − z k

(1)

E. Comparison with manual database annotations The European Society of Cardiology (ESC) ST-T database available with the Physiobank [15] repository has been used during this analysis. This database contains the annotated recordings of two channel ECG data for two hours at a sampling rate of 250Hz. The database contains 90 annotated excerpts of ambulatory ECG recordings from 79 subjects having or suspected to have Myocardial Ischemia. 48 of these 90 complete records are available freely on Physiobank and have been employed in this work. The annotations contain information about the onset and offset of different ST segment deviation episodes along with the peak deviation during an ST segment deviation episode. These manual peak ST deviation annotations were compared with the corresponding values of deviation obtained through different baseline removal algorithms using equation (1). The absolute error, Ep in micro-volts between the pth manually annotated peak deviation level and the one calculated from a baseline removal algorithm is taken. The mean and the median of these errors over all the manually annotated peak deviations in the database for different algorithms are reported. III.

IMPLEMENTED TECHNIQUES

In this section we present a brief introduction of different techniques that we have implemented for this comparison. A. Use of Cubic Spline Curve Fitting In this approach [12] isoelectric fiducial points are found in the ECG signal with baseline variation for each beat using an approach identical to the one discussed earlier and a third order cubic spline is fitted on these points to obtain an estimate of the baseline which is then subtracted from the original ECG signal. This method is among the most commonly used approaches for removal of ECG baseline variation. Cubic spline interpolation based baseline removal and other

interpolation based techniques adapt themselves automatically to the heart rate as more reference points become available with increase in heart rate. However, in the absence of any baseline variation in an ECG segment, an error in the calculation of the isoelectric reference point or the corresponding level causes undesired distortion in the ECG. Therefore accurate definition of the isoelectric reference point is mandatory for proper functioning which can become difficult in the presence of noise in the ECG signal. B. Use of Linear Spline Curve Fitting Papaloukas et al. [14] have proposed a simple and effective approach for the removal of baseline from the ECG signal. This method takes the ECG signal s[n] for a single cardiac cycle starting 60ms before the P-wave and ending 60ms after the T-wave and subtracts its mean from it to give y[n]. Next a first order polynomial p[n] is fitted on y[n]. The sample values of the QRS complex for each cardiac cycle in y[n] are replaced by the corresponding values of p[n] to give y*[n]. This replacement removes for the shift in p[n] towards main QRS polarity due to the high peaks in the QRS complex. Thereafter, a first order polynomial curve is fitted to y*[n] which is then subtracted from the corresponding region of y[n] to obtain the baseline removed signal. This method is very well suited for use in diagnosis procedures involving ST segment analysis because it does not affect the ST segment when no baseline variation is present. However, this method may produce discontinuities in the resulting signal at the end points of a cardiac cycle therefore reliable detection of the start and end points of the cardiac cycle along with accurate QRS delineation is required. C. Median Filtering Chouhan et al. [13] give a technique for baseline removal using median filtering on the electrocardiogram. In this procedure, firstly the median of the ECG signal is subtracted from the ECG signal. Then a fifth order polynomial is fitted to this shifted waveform to obtain a baseline estimate which is then subtracted from the ECG signal. The baseline drift is further removed by applying median correction, one by one, in each RR interval. This approach also offers the advantage that the signal is not distorted in the absence of baseline variation and is computationally efficient. D. FIR High Pass Filtering (FIR HPF) Since the baseline signal is a low frequency signal therefore we have used Finite Impulse Response (FIR) high pass zero phase forward-backward filtering [16] with a cut-off frequency of 0.5Hz to estimate and remove the baseline in the ECG signal. The order of the high pass filter is taken to be 700. This approach does not require the detection of any reference points in the ECG signal. However, due to the high filter order, this approach presents a high computational load. Moreover to nullify any phase distortions, zero-phase forwardbackward filtering has to be used which cannot be implemented in real-time.

E. Adaptive Filtering (AF) Adaptive filtering has been used for baseline removal from the ECG in [7] using the architecture shown in figure 3.

Figure 3. Adaptive Filtering for ECG Baseline Removal [7] For adaptive filtering of baseline wandering, only one weight is needed and the reference input is a constant with a value of one. The optimal weight w is determined using the Least Mean Squares (LMS) algorithm as follows,

w ( k + 1) = w ( k ) + 2 μ e ( k ) x ( k )

(2)

This filter has a zero at 0Hz and consequently it creates a notch with a bandwidth of

(μ π )

f s where fs is the sampling

frequency. Because AHA recommends cutoff frequencies under 0.8Hz for the prevention of distortion of the STsegment, µ=0.0101 is taken (for fs =250Hz). This approach produces severe distortion in the ECG signal, especially in the ST segment area [17]. F. Wavelet Adaptive Filtering (WAF) Park et al. [17] (see fig. 4) have proposed a wavelet adaptive filter for baseline removal from the ECG to minimize distortion of the ST Segment. In this method the ECG signal with baseline is decomposed up to 7 levels using Wavelet Transform with Vaidyanathan-Hoang wavelet having orthogonal characteristics. The 7th level approximation coefficients have frequency components in the range of 01.4Hz. These coefficients are then subjected to the adaptive filter with a cutoff frequency of 0.8Hz. The filtered output and the details coefficients are used for reconstruction using inverse wavelet transform to produce the baseline removed signal.

Figure 4. Baseline Removal using WAF [17]

This approach presents a very effective approach for baseline removal as it does not require the calculation of any reference points and the use of wavelet transform for the analysis of the inherently non-stationary ECG signal. G. Use of Empirical Mode Decomposition (EMD) Empirical Mode Decomposition [18] is a fully data-driven technique for analysis of non-stationary signals in which no a priori known basis is required. In EMD the input signal is decomposed into a sum of intrinsic mode functions (IMFs). The use of EMD for removal of baseline from the ECG signal has been proposed by [19] in which partial reconstruction of the ECG signal from the IMF obtained by the decomposition of the input ECG signal is used. This is done in a way to remove low frequency components from the ECG signal which results

in the removal of baseline variation. For further details refer to [19]. This method also offers a promising approach for removal of baseline variations from the ECG signal. However it is computationally very demanding in comparison to other approaches. IV.

[5]

J. A. V. Alste and T. S. Schilder, "Removal of baseline wander and powe-line interference from the ECG by an efficient FIR filter with a reduced number of taps," IEEE Trans Biomed Eng., vol. 32, pp. 10521060, 1985.

[6]

H. Gholam-Hosseini, H. Nazeran, and K. J. Reynolds, "ECG noise cancellation using digital filters," presented at Bioelectromagnetism, 1998. Proceedings of the 2nd International Conference on, 1998.

RESULTS AND CONCLUSION

The absolute values of error between manually annotated peak ST deviation and the one determined by different algorithms discussed earlier are shown below. These results clearly show that the wavelet adaptive filter outperforms all other approaches with the absolute error median being 42.6 micro-volts. The clinically used threshold for the detection of myocardial ischemia using the ST-segment is a deviation of one small square (100 micro-volts) of the ECG. Therefore wavelet adaptive filter is the recommended approach for baseline removal for the analysis of ST segment deviations in the ECG. Figure 5 shows an example of baseline removal with WAF. Table 1 Mean and Median values of the errors (in micro-volts) Method Mean(Ep) Median(Ep) 85.3965 53.6057 Cubic Spline 77.9238 54.9557 Linear Spline 86.1120 55.9492 Median 73.3396 50.8942 FIR HPF 78.9087 56.8559 AF WAF 67.5886 42.5735 76.7244 54.0533 EMD Original ECG Signal with Baseline

[7]

P. Laguna, R. Jane, and P. Caminal, "Adaptive Filtering Of ECG Baseline Wander," presented at Engineering in Medicine and Biology Society, 1992. Vol.14. Proceedings of the Annual International Conference of the IEEE, 1992.

[8]

S. V. Pandit, "ECG baseline drift removal through STFT," presented at Engineering in Medicine and Biology Society, 1996. Bridging Disciplines for Biomedicine. Proceedings of the 18th Annual International Conference of the IEEE, 1996.

[9]

Z. Donghui, "Wavelet Approach for ECG Baseline Wander Correction and Noise Reduction," presented at Engineering in Medicine and Biology Society, 2005. IEEE-EMBS 2005. 27th Annual International Conference of the, 2005.

[10] O. Sayadi and M. B. Shamsollahi, "Multiadaptive Bionic Wavelet Transform: Application to ECG Denoising and Baseline Wandering Reduction," EURASIP Journal on Advances in Signal Processing, pp. Article ID 41274, 11 pages, 2007. [11] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna, "A wavelet-based ECG delineator: evaluation on standard databases," Biomedical Engineering, IEEE Transactions on, vol. 51, pp. 570-581, 2004. [12] C. R. Meyer and H. N. Keiser, " Electrocardiogram baseline noise

50

estimation and removal using cubic splines and state-space computation techniques," Comput. Biomed. Res. , vol. 10, pp. 459-470, 1977.

0

[13] V. S. Chouhan and S. S. Mehta, "Total Removal of Baseline Drift from ECG Signal," presented at Computing: Theory and Applications, 2007.

-50 3000

3500

4000

4500

5000

5500

6000

6500

7000

ICCTA '07. International Conference on, 2007.

Baseline Removed Signal using WAF 60

[14] C. Papaloukas, D. I. Fotiadis, A. Likas, A. Liavas, and L. Michalis, "A

40

Knowledge -based Technique for the Detection of Ischemic Episodes in

20 0

Long Duration ECGs," Medical and Biological Engineering and

-20

Computing, vol. 39, pp. 105-112, 2001.

-40 -60 3000

3500

4000

4500

5000

5500

6000

6500

7000

[15] A. Taddei, G. Distante, M. Emdin, P. Pisani, G. B. Moody, C. Zeelenberg, and C. Marchesi, "The European ST Database: Standard for

Figure 5. Baseline Removed Signal using WAF

REFERENCES [1]

A. R. Houghton and D. Gary, Making Sense of the ECG: A Hands-on Guide: Arnold Publishing Co., 2003.

[2]

T. H. Jafar, F. H.Jafary, S. Jessani, and N. Chaturvedi, "Heart disease epidemic in Pakistan: Women and men at equal risk," American Heart Journal, vol. 150, pp. 221-226, 2005.

[3]

G. D. Clifford, "ECG Statistics, Noise, Artifacts and Missing Data," in Advanced Methods and Tools for ECG Analysis, G. D. Clifford, F. Azuaje, and P. McSharry, Eds. London: Artech House Publishing, 2006.

[4]

L. T. Sheffield, C. A. Berson, D. B. Remschel, P. C. Gillette, R. E. Hermes, L. Hinkle, H. Kennedy, D. M. Mirvis, and C. Oliver, "Recommendations for standard of instrumentation and practice in the use of ambulatory electrocardiography," Circulation, pp. 626A-636A, 1985.

Evaluating Systems for the Analysis of ST-T Changes in Ambulatory Electrocardiography. Eur Heart J " European Heart Journal, vol. 13, pp. 1164-1172, 1992. [16] J. G. Proakis and M. D. G., Digital Signal Processing: Principles, Algorithms, and Applications: Prentice-Hall, 1999. [17] K. L. Park, K. J. Lee, and H. R. Yoon, "Application of a wavelet adaptive filter to minimize distortion of the ST-Segment," Medical and Biological Engineering and Computing, vol. 36, pp. 581-586, 1998. 18] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.C. Yen, C. C. Tung, and H. H. Liu, "The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis," Proc. R. Soc. Lond., vol. 454, pp. 903-995, 1998. [19] M. Blanco-Velasco, B. Weng, and K. E. Barner, "ECG signal denoising and baseline wander correction based on the empirical mode decomposition," Computers in Biology and Medicine, vol. 38, pp. 1-13, 2007.