AN ENHANCED EMD ALGORITHM FOR ECG SIGNAL PROCESSING Foteini Agraoti, Student Member, IEEE and Dimitrios Hatzinakos,

Senior Member, IEEE

The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, 10 Kings College Road, Toronto, ON, Canada, M5S 3G4 {foteini,dimitris}@comm.utoronto.ca

ABSTRACT The Empirical Mode Decomposition (EMD) is becoming increasingly popular for the multi-scale analysis of signals. However, the data-driven and adaptive nature of the EMD raises concerns regarding the uniqueness of the decomposition as well as the extend to which oscillatory modes can be mixed across different IMFs. This paper proposes a solution to this problem for the analysis of ECG signals. The bivariate extension of the decomposition (BEMD) is used as the basis of an analysis in which a synthetic ECG signal of idealized waveform guides the decomposition of an input ECG segment. Essentially, this work provides the necessary ground for the deployment of signal processing algorithms on the ECG signal using a more robust EMD analysis. Index Terms Electrocardiogram, bivariate empirical mode decomposition, intrinsic mode function 1. INTRODUCTION The purpose of pattern recognition is to extract information that is originally hidden in the data. Multi-scale analysis is one way to address this problem. For one dimensional physiological signals such as the electrocardiogram (ECG), phonocardiogram (PPG), blood volume pressure (BVP), photoplethysmogram (PPG) and other that are widely studied in diagnostics, typical approaches include the Fourier transform, the spectogram, the cosine transform, the Wigner-Ville distribution and wavelet analysis. Although all these approaches are well established both theoretically and in practice, they are all employed on linearity and stationarity assumptions. However, this is not the case with most biosignals as the physical process that generates them is neither stationary nor linear. This gap in digital signal processing was emphasized by Huang et al.s work [1], and the Empirical Mode Decomposition (EMD) was proposed as an alternative solution. Since then, a signicant amount of work has been reported on the applications of EMD along with several variations of the original algorithm [2, 3, 4, 5, 6]. The essence of EMD is that

it decomposes signals in a data-driven and adaptive manner without linearity or stationarity restrictions. For EMD, every signal is the superposition of fast oscillations over slow oscillations [2]. Despite the fact that many real life signals do not exhibit oscillations naturally (for example images), rapidly oscillatory components can be extracted to describe the nest high frequency characteristics of the signal. Similarly, slow oscillations can be extracted to describe low frequency underlying phenomena that may exist in the signal. In EMD terminology, the rst are low order Intrinsic Mode Functions (IMFs) while the latter are higher order IMFs. It is important to note, that EMD can detect and extract these components without prior information on the morphological properties of the signal. However, adaptivity is a mixed blessing. The benet, as mentioned before, is that any signal can be processed with a decomposition that explores its intrinsic properties. The drawback of EMD is that it cannot be analytically dened. This causes uniqueness and mode mixing problems that limit its applicability. Since the decomposition depends on the signal itself, the number of IMFs in which it will result in, is unpredictable. In addition, similar oscillatory modes may be presented across different IMFs. This uncertainty poses great threats to the automatic deployment of EMD in signal processing. For instance, it is difcult to establish corresponding IMFs across different instances of the same signal (for example consecutive PPG recordings). This problem may be practically addressed by forcing the decomposition to stop once a predened number of IMFs is reached, however this does not solve the mode mixing problem nor does it guarantee that the resulting IMFs will have physical meaning. We advocate that this problem can be addressed by taking into account the particular morphology of the signal in-hand. Although it is difcult to predict the appearance of a stochastic signal, one can take explore its general structure to enhance the empirical decomposition. This work is interested in the ECG signal, as it is widely examined by both the medical [7, 8, 9] and lately by the biometrics [10, 11, 12] communities. A solution based on the Bivariate EMD (BEMD) [2] extension of the algorithm is presented.

978-1-4577-0274-7/11/$26.00 2011 IEEE

DSP2011

RR

Voltage (m V)

T P Q
0 0.5 1

T P

-2

S
1.5 2

Q S
2.5 3 3.5 4

Tim e (sec)

EMD seeks for oscillations through a sifting process. Once an IMF is found, it is removed from the signal and the algorithm iterates on the residual in order to nd more oscillatory modes. Fast oscillations (high frequency) are detected rst. Given a signal (), EMD operates as follows: 1. Detects local maxima () and minima () of (). 2. Interpolates among () to get an upper envelope (), and () for minima respectively. 3. Computes the average of the two envelopes () = ()+ () . 2 4. Subtracts from the original signal () = () (). 5. Iterates for the residual: () = (). This process is terminated when () meets the IMF criteria. If it does, () describes an underlying oscillation of (), refereed herein as (). EMD continues with sifting on the residual () = () () i.e., on a signal stripped of a fast oscillation. The original signal can then be expressed as: () =
1 =1

Fig. 1. Main components of an ECG heart beat. The remaining of this paper is organized as follows. Section 2 provides an overview of the ECG from a signal processing point of view while Section 3 is a brief introduction to the EMD algorithm. The proposed decomposition is presented in Section 4 and some experimental results are provided in Section 5. 2. THE ECG WAVEFORM The ECG reects the cardiac electrical potential over time. From a signal processing perspective, the ECG is a non periodic but highly repetitive signal that is composed of three main waves, as shown in Figure 1, typically referred to as the P wave, QRS complex and the T wave. The P wave has usually positive polarity and a duration of approximately 120 ms. This wave mainly reects the depolarization of the right and left atria. The QRS complex describes the depolarization of right and left ventricles. In normal rhythms, its duration varies between 70-110 ms. Finally, the T wave reects a depolarization of the ventricles and is usually observed about 300 ms after the QRS complex. However, its exact position relies on the heart rate and appears closer to the QRS complex at rapid rhythms [13]. The spectral characteristics of ECG waves are central to the application of signal processing algorithms. A healthy P wave is considered to contribute to the low frequency components at about 10-15 Hz. On the other hand, a QRS complex has a spectrum of comparably high frequencies due to its steep slopes. The spectral content of this complex is usually found in the 10-40 Hz band. 3. THE EMPIRICAL MODE DECOMPOSITION

() + ()

(1)

where () denotes the extracted IMF and () the nal residual. Note that by denition, () is not an IMF. BEMD was proposed by Rilling et al. [2] as a way to decompose complex signals with the EMD. The BEMD algorithm decomposes naturally bivariate signals in a consistent way i.e., by examining the real and imaginary parts simultaneously. Other complex solutions [5] risk a physically meaningless decomposition because the complex components are treated independently. In [2], the oscillatory rationale of the EMD is directly translated to the bivariate case. Because of the consistency in the analysis of the real and the imaginary parts, the BEMD has been suggested for signal separation (detrending) in ltering applications [14, 15]. 4. PROPOSED DECOMPOSITION

This section provides an overview of the EMD algorithm, proposed by Huang et al. [1], in order to then extend it to the bivariate case, and present the proposed solution. EMD decomposes a signal adaptively into a number of IMFs. Each IMF describes a disctint oscillation and has the following characteristics [1]: 1. The number of extrema and the number of zero crossings must be equal or differ at most by one. 2. The mean of the envelopes that are dened by the maxima and the minima is zero at every time instance.

In this work, the BEMD is used on ECG signals in order to overcome the drawbacks of the data-driven EMD. As explained in Section 1, the EMD suffers from uniqueness issues i.e., the number and type of IMFs at which the decomposition will result in, is uncertain, even for signals of similar statistics. Unless properly treated, this poses great restrictions to its usability, as it renders comparisons among different ECG segments meaningless. Furthermore, predetermining the number of IMFs (by forcing the decomposition to stop) beats the purpose of EMD as the analysis will no longer be adaptive.

a) Real ECG segment

3

Real ECG IMFs (Real Part)

1 0 1 0 0.5 1 2 3 4 5 2

Synthetic S h i ECG IMFs IMF (Imaginary Part)

IMF 1
0 2 0 1 1 2 3 4 5

V Voltage (mV)

2 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

b) Synthetic ECG segment

3

Time (sec)

Oscillation Magnitude n e

1 0

0 0.5 0 1 0 1 0 0.1 0 0.1 0 0.2 0.1 01 0 0 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

IMF 2

0 1 0 0.5 1 2 3 4 5

2 1 0 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

IMF 3

0 0.5 0 0.2 1 2 3 4 5

IMF 3

0 0.2 0 0.4 1 2 3 4 5

Time (sec)

Fig. 2. Example of ECG synthesis. a) Real ECG segment, b) Generated synthetic ECG. The two signals are synchronized at the main waves. What is needed, is a way for the decomposition to focus on oscillations that are directly linked to the inherent characteristics of the signal in-hand, and to ignore random oscillations that may be presented due to noise. We address this problem by decomposing a real ECG signal together with a synthetic, in a framework that allows the latter to act as the rule of decomposition i.e., to determine which type of IMFs are important from on the input side. The proposed framework encompasses two steps, namely a) ECG synthesis and b) driven BEMD analysis. Given an ECG segment, (), the rst step is to synthesize a synchronous ECG segment, () in order to decompose them simultaneously at the second step. 4.1. ECG Synthesis The goal of this step is to design a synthetic ECG signal, (), that is synchronized at the main waves with the input ECG signal ( ()) that is to be decomposed. Synchronization is necessary because both the EMD and BEMD algorithms operate in the time domain. The purpose if to decompose the synthetic signal simultaneously with (). () is rst ltered, to remove major noise artifacts (baseline wander, powerline interferences etc). A Butterworth bandpass lter or order 4, with cutoffs at 0.5 and 40 is used based on empirical results. The signal is then delineated, i.e., ducial points such as the ones shown in Figure 1 are detected. The QRS complex is detected using the algorithm described in [16], and surrounding waves are localized with empirical rules (the P waves healthy duration is approximately 120, and T wave extends about 300 after the QRS complex [13]). The position of the ducial points on (), guide the generation of a synthetic ECG signal (). Synthesis is performed using a dynamic model of three coupled differential equations, proposed by McSharry et al. [17]. A heart beat is modeled as a a circular movement of a trajectory in a 3D space, where a cycle completion corresponds to one repeti-

0.2 IMF 4 0 2 I 0 0

Time (sec)

Time (sec)

Fig. 3. Simultaneous decomposition of real and a synthetic ECG signal using the BEMD. tion. Because of the circular movement, the locations of the ducial points , , , , , are transformed to the respective angles , , , , within the unit circle. The generated synthetic ECG, (), is an idealized, noise free, representation of an ECG signal. Figure 2 shows an example of an input ECG segment and the corresponding synthetic that was generated. 4.2. Driven BEMD In this step, the synthetic signal () is decomposed simultaneously with the input one, (), using the BEMD. In essence, the difference between EMD and BEMD is that where EMD sifting builds envelopes around (), BEMD builds 3D cubes that surround the complex function () [2]. Thus, the analysis is performed simultaneously for the real and imaginary components, and results in the same number of IMFs for both: () =
1 =1

() + ()

(2)

where () denotes a complex IMF and () the complex residual. A complex signal is formed using () and () for the real and imaginary parts respectively: () = () + () By applying BEMD on () we get () =
=1

(3)

{ ()} +

=1

{ ()}

(4)

where the residual has been included in the summation for simplicity. A BEMD decomposition example is depicted in Figure 3. Since the synthetic ECG has an idealized waveform, the presence of oscillatory activity on the imaginary

Oscillatio Magnitud on de

Vo oltage (mV)

Synhtetic ECG signal

2 0 2 0 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

IMF 1 Tricomponent function

Oscillation Magnitude

1 0 0.5 0 0.5 0 0.1 0 0.1 0 5 0 5 0 0.01 0 x 10

3

0.5

1.5

2.5

3.5

4.5

IMF 2 Bicomponent Function

0.5

1.5

2.5

3.5

4.5

IMF 3 Monocomponent Function

0.5

1.5

2.5

3.5

4.5

IMF 4

0.5

1.5

2.5

3.5

4.5

IMF 5

0.01 0

0.5

1.5

2.5

3.5

4.5

Time (sec)

Fig. 4. EMD decomposition of a synthetic ECG signal. The rst three IMFs are quasi periodic. side guarantees that the corresponding mode is present on the real side. When the input ECG is contaminated with high frequency noise, low order IMFs on the real part will exhibit strong (but physiologically random) oscillations while almost zero activity will exist on the imaginary side,. The last observation makes noisy IMFs easily detectable. 5. ILLUSTRATION It has been observed that when analyzing the ECG signal with the EMD, the rst three IMFs tend to preserve information from the QRS complex [18]. The same is true for the BEMD. In the absence of noise, the rst IMF typically exhibits three distinct oscillations (tricomponent). Once this oscillation is removed from the signal and a sifting process is completed, the subsequent IMF exhibits two oscillations (bicomponent) and similarly the third is monocomponent. Figure 4 shows an example of a synthetic ECGs IMFs that illustrate this property. The IMFs of order higher than three are not quasi-periodic and the oscillation magnitude is near zero. However, due to noise the clear oscillatory structure of Figure 4 is not always apparent when decomposing real ECG signals. This is because the decomposition is driven by the noise as well. In the proposed extension for the BEMD algorithm, since the a noise-free synthetic ECG (imaginary part) is decomposed together with the input segment (real part), oscillations due to noise on the real side will result in near zero IMF activity on the imaginary side, which renders these IMFs easily detectable. In addition, by driving the BEMD decomposition using an idealized waveform, the IMFs will conform better to the expected structure i.e., IMF 1 to be a tricomponent function, IMF 2 a bicomponent and so on. Figure 5 shows a comparison between the univariate EMD and the driven BEMD decompositions for the same ECG segment. Even though there is no theoretical guarantee for uniqueness, the proposed driven BEMD algorithm bypasses this inadequacy by ensuring that the three most substantial IMFs for ECG analysis will be present in the decomposition, without mode mixing. The purpose of this analysis is to examine the consistency of the IMFs across different ECG recordings. To quantify this performance we dene a measure of oscillatory activity as follows. For every IMF we measure the time between consecutive extrema interchange which is essentially a measure of local oscillation. The assumption is that that there will be three dominant oscillation types within the rst IMF of an ECG, two in the second and one in the third. Even though some variability is expected among different ECG recordings (ECG is affected by both physical and psychological activity), it is anticipated that with the proposed BEMD decomposition the dominant oscillations at every IMF level will exhibit small variance i.e., they will not be affected by random noise oscillations or by mode mixing. Therefore the standard deviation of the dominant oscillations within every IMF is used in this work as a measure of stability for the proposed decomposition. The driven BEMD was tested on ECG recordings from 44 individuals. The collection took place at the Affect and Cognition Laboratory at the University of Toronto. For every subject, 50 recordings are available (5 sec each) which makes a total of 2200 ECG segments. The BIOPAC MP 150 system was used for the collection, and the signals were digitized at 1KHz. Across all recordings, the standard deviations of the three most dominant oscillations in IMF 1, obtained with the proposed BEMD decomposition, are {0.0820, 0.0593, 0.0072}. When the same measure is esti-

Voltage (mV)

20 0 -20 0

0.5

1.5

2.5

3.5

Time (sec)

Univariate EMD
10 0 -10 0

BEMD w ith synthetic ECG

20

IMF 1
0.5 1 1.5 2 2.5 3 3.5

0 -20

Oscillation Ampltitude

10 0 0 0.5 1 1.5 2 2.5 3 3.5

IMF 2

0 -20 0 0.5 1 1.5 2 2.5 3 3.5

-10

10 0 -10 0 5 0 -5 0 5 0 -5 0

IMF 3
0.5 1 1.5 2 2.5 3 3.5

5 0 -5 5 0 -5

0.5

1.5

2.5

3.5

IMF 4
0.5 1 1.5 2 2.5 3 3.5

0.5

1.5

2.5

3.5

Time (sec)

IMF 5
0.5 1 1.5 2 2.5 3 3.5

Time (sec)

Fig. 5. Comparison of Univariate and driven Bivariate EMD decomposition on the same ECG signal. For the BEMD case the IMFs exhibit less mode mixing as well as the oscillation structure follows the properties of ECG decomposition in the absence of noise i.e., IMF 1 is tricomponent, IMF 2 is bicomponent and IMF 3 is monocomponent. mated for a decomposition using the typical EMD algorithm, the oscillations of the rst IMF are not as consistent, with a standard deviation of {0.1652, 0.1034, 0.0453}. Similarly, for the second IMF, the deviation of the two dominant IMFs of the proposed method is {0.0871, 0.0268} while for the univariate EMD it is {0.1320, 0.0569}. For the third IMF there should be only one dominant oscillation, the deviation of which using the driven BEMD and the univariate EMD are 0.0283 and 0.0835 respectively. These measurements indicate that the proposed driven BEMD algorithm manages to provide a more consistent decompositions for the ECG signal. 6. CONCLUSION 8. REFERENCES This work dealt with the problem of ECG analysis using the EMD algorithm. Despite the advantages of the data-driven nature of the decomposition, the EMD is considerably vulnerable to noise. This practically leads to uniqueness and mode mixing issues. We advocate that this problem can only be addressed by taking into consideration the morphology of the signal on which the method is applied. In particular, this work provides a standard for the deployment of the EMD algorithm on ECG signals. Instead of risking a univariate decomposition that may lead to incomparable IMFs across different ECG segments, the bivariate extension [1] N.E. Huang, Z. Shen, R.R. Long, M.L. Wu, Q. Zheng, N.C. Yen, and C.C Tung, The empirical mode decomposition and hilbert spectrum for nonlinear and nonstationary time series analysis, Proc. Roy. Soc. London, vol. 454, pp. 903995, 1998. [2] G. Rilling, P. Flandrin, P. Gonalves, and J.M. Lilly, Bivariate empirical mode decomposition, IEEE Signal Processing Letters, vol. 14, no. 12, pp. 936 939, Dec. 2007. of the algorithm is proposed as promising tool for the simultaneous decomposition of the ECG signal with a synthetic equivalent. The synthetic ECG segment has a noise-free idealized waveform that acts as the rule of the decomposition process. This driven BEMD decomposition results in robust mode functions, with respect to the oscillatory functions that are conveyed. 7. ACKNOWLEDGEMENTS This work has been supported by the Natural Sciences and Engineering Research Council of Canada (NSERC).

Oscillation Ampltitude

0 20

0.5

1.5

2.5

3.5

[3] D. Looney and D.P. Mandic, Fusion of visual and thermal images using complex extension of EMD, in 2nd Int. Conf. on Distributed Smart Cameras, sep. 2008, pp. 1 8. [4] S. M. A. Bhuiyan, R. R. Adhami, and J. F. Khan, Fast and adaptive bidimensional empirical mode decomposition using order-statistics lter based envelope estimation, EURASIP Journal on Advances in Signal Processing, 2008. [5] T. Tanaka and D. P. Mandic, Complex empirical mode decomposition, IEEE Signal Processing Letters, vol. 14, no. 2, pp. 101 104, Feb. 2007. [6] M.U. Bin Altaf, T. Gautama, T. Tanaka, and D.P. Mandic, Rotation invariant complex empirical mode decomposition, in IEEE International Conference on Acoustics, Speech and Signal Processing. ICASSP 2007., April 2007, vol. 3, pp. 10091012. [7] P.S. Addison, J.N. Watson, G.R. Clegg, M. Holzer, F. Sterz, and C.E. Robertson, Evaluating arrhythmias in ECG signals using wavelet transforms, IEEE Engineering in Medicine and Biology Magazine, vol. 19, no. 5, pp. 104109, Sep/Oct 2000. [8] C.J. Finelli, J.M. Jenkins, and L.A. DiCarlo, Detection and identication of cardiac arrhythmias using an adaptive, linear-predictive lter, Computers in Cardiology, pp. 177180, 5-8 Sep 1993. [9] Chuang Chien Chiu, Tong Hong Lin, and Ben Yi Liau, Using correlation coefcient in ECG waveform for arrhythmia detection, Biomedical Engineering applications, Basis and Communication, vol. 17, pp. 147152. [10] F. Agraoti, F.M. Bui, and D. Hatzinakos, Medical biometrics: The perils of ignoring time dependency, in IEEE 3rd Int. Conf. on Biometrics: Theory, Applications, and Systems, 2009. BTAS 09, Washington, Sept. 2009, pp. 1 6. [11] S. A. Israel, J. M. Irvine, A. Cheng, M. D. Wiederhold, and B. K. Wiederhold, ECG to identify individuals, Pattern Recognition, vol. 38, no. 1, pp. 133142, 2005. [12] F. Agraoti, F.M. Bui, and D. Hatzinakos, On supporting anonymity in a BAN biometric framework, in To appear in the Proc. of 16th International Conference on Digital Signal Processing (DSP 2009), Santorini, Greece, July 2009. [13] L. Sornmo and P. Laguna, Bioelectrical Signal Processing in Cardiac and Neurological Applications, Elsevier, 2005.

[14] P. Flandrin, P. Goncalves, and G. Rilling, Detrending and denoising with empirical mode decompositions, in XII EUSIPCO, Sept. 2004. [15] Md. K. L Molla, T. Tanaka, T. M. Rutkowski, and A. CiChocki, Separation of EOG artifacts from EEG signals using bivatiate EMD, in IEEE Int. Conf. on Acoustics, Speech and Signal Processing, 2010, March 2010, pp. 562565. [16] J. Pan and W. J. Tompkins, A real-time QRS detection algorithm, IEEE Trans. on Biomedical Eng.,, vol. 32, no. 3, pp. 230 236, March 1985. [17] P.E. McSharry, G.D. Clifford, L. Tarassenko, and L.A. Smith, A dynamical model for generating synthetic electrocardiogram signals, IEEE Trans. on Biomedical Eng.,, vol. 50, no. 3, pp. 289 294, March 2003. [18] A. Karagiannis and P. Constantinou, Noise-assisted data processing with empirical mode decomposition in biomedical signals, IEEE Transactions on Information Technology in Biomedicine, vol. 15, no. 1, pp. 11 18, Jan. 2011.