A NEW METHOD FOR FETAL ELECTROCARDIOGRAM DENOISING USING BLIND SOURCE SEPARATION AND EMPIRICAL MODE DECOMPOSITION

Transcription

1 Rev. Roum. Sci. Techn. Électrotechn. et Énerg. Vol. 6,, pp , Bucarest, 206 A NEW METHOD FOR FETAL ELECTROCARDIOGRAM DENOISING USING BLIND SOURCE SEPARATION AND EMPIRICAL MODE DECOMPOSITION DRAGOS DANIEL TARALUNGA, ILINCA GUSSI 2, RODICA STRUNGARU Key words: Fetal electrocardiogram, Empirical mode decomposition (EMD), Independent component analysis. The abdominal fetal electrocardiogram (fecg), recorded noninvasively via surface electrodes placed on the maternal abdomen, has recently gained significant research attention from physicians due to its advantages over fetal monitoring techniques currently used in clinical practice. However, the fecg signal is often corrupted by different types of noise, generated by biological sources: maternal electrocardiogram (mecg), abdominal skeletal electromyogram (EMG), electrohysterogram (EHG) or non-biological sources: movement artefacts, electromagnetic and electronic noise. The presence of so many types of noise makes the direct fecg analysis impossible and, in addition, the spectrum of the noise is overlapping with the one of the signal of interest, making simple filtering techniques ineffective. In this paper, we propose a new method for abdominal fecg denoising based on a blind source separation method, i.e the independent component analysis (ICA), and empirical mode decomposition (EMD). The performance of the proposed algorithm, called ICA_EMD, was evaluated on both simulated and real data and compared with the performance obtained by ICA, showing a better removal of artefacts without disturbing the fecg signal.. INTRODUCTION Fetal monitoring is based mainly on the analysis of the fetal heart rate (fhr) and more recently on the analysis of the morphology of the electrical signal generated by the fetus heart, i.e. fetal electrocardiogram (fecg) []. In clinical practice, the fhr is obtained using a method based on Doppler ultrasound, called cardiotocography (CTG). The fecg obtained by placing surface electrodes on the maternal abdomen overcomes most of the present drawbacks of the methods used in clinical practice [2]: it can be used for long term monitoring, it is non-invasive, and it offers both the beat to beat fhr, by R-peak detection of the ventiruclar contraction which is represented in the fecg by three waves, Q, R and S waves (QRS) [3] and the morphology of the fecg signal. However, despite these clear advantages, abdominal recorded fecg has not made its way in clinical practice mainly due to the fact that it is corrupted by different powerful types of noise which makes impossible the direct analysis of the fecg. Biological sources generating disturbing noise include: maternal electrocardiogram (mecg), abdominal skeletal electro-myogram (EMG), electrohysterogram (EHG), electro-gastrogram (EGG), maternal and fetal respiration, etc. Other types of noise are present in the abdominal recordings: electronic noise generated by the electronic components, power-line interference (PLI) generated by power-line grid [4], and movement artefacts. The maority of these types of noise have spectra overlapping with the spectra of the signal of interest, and thus simple filtering methods are unsuccessful. There are different approaches based on different concepts for denoising the abdominal fecg. Wavelet transform (WT) has been used to remove artefacts from abdominal recorded signals [5]. However, the maority of WT-based methods require that the noise components have low amplitude and that they are stationary when compared with the abdominal fecg. This is not the case for the types of noise present in abdominal fecg recordings; they have a much higher amplitude than the signal of interest, i.e. up to mv for the mecg, 0 mv for the EMG, mv for the EHG etc. [6, 7], while the average amplitude of the abdominal fecg is 0 µv [6]. Moreover, they are not stationary and usually have frequency components close the frequency of the signal of interest. Blind source separation (BSS) methods such as principal component analysis (PCA) and independent component analysis (ICA) are more frequently used in biomedical signal processing of multivariate time series. Accurate fecg denoising is reported in literature using ICA algorithms [8, 9]. In the present paper, the aim is to design and introduce a new signal processing tool, called ICA_EMD, based on the ICA and EMD methods in order to remove the noise present in abdominal fecg recordings. The ICA algorithm is used to separate the fecg component from high disturbing noises like mecg and PLI. However, the ICA algorithm is not able to fully separate the fecg from all artefacts, especially the ones having a Gaussian distribution, impairing the analysis of the fecg. The second step of the proposed algorithm is to use empirical mode decomposition (EMD) to further denoise the fecg signal obtained with ICA. EMD was introduced by Huang et al. [0] and has become an important tool for non-stationary signal analysis. The performance of the proposed method is evaluated on simulated and real signals, the latter recorded on women during pregnancy. Moreover, the performance is compared with that obtained by the ICA algorithm in order to emphasize the important contribution of the proposed method. 2. MATERIAL AND METHODS 2.. DATA DESCRIPTION The data used for evaluating the proposed method are classified in simulated data and real recordings. In the next subsections, the modality of simulating realistic abdominal fecg and the procedure of recording real data are described. Politehnica University of Bucharest, Applied Electronics and Information Engineering Department, dragos.taralunga@upb.ro, rodica. strungaru@upb.ro. 2 Carol Davila University of Medicine and Pharmacy, Obstretics and Gynecology Department, Bucharest, Romania, ilinca. gussi@gmail.com.

2 2 A new method for fetal electrocardiogram denoising Simulated data The special feature of simulated data, controllable signal-to noise ratio (SNR), makes them very useful and necessary for obectively evaluating the performances of a specific method or to compare the performances of different methods which have the same purpose. The simulated abdominal fecg signals are derived using the method proposed by Sameni et al. [, 2]. The electrical activity of the fetal and maternal hearts is modeled by a dipole, thus, each moving cardiac vector, either maternal or fetal, is described by three orthogonal components which represent the proection of the moving cardiac dipole on the recording electrode axes. The morphology and the quasi-periodicity of the ECG is reproduced within the dynamic model, in which each beat corresponds to one revolution around a limit circle. Time-varying autoregressive models are used to simulate realistic noise and the parameters are derived by fitting the model to real baseline wander, muscle artifacts, i.e. EMG, and electrode movement signals. A detailed description of the model can be found in [, 2]. The SNR between the fecg, the signal of interest, and the mecg, as noise component, is set to 20 db. By varying the SNR between the fecg and the contribution of all the other noise sources (abdominal muscle activity, electrode movement and baseline wander), different data sets can be created in order to evaluate the algorithms for fecg extraction: P SNR 0log fecg m = 0, PmECG P SNR 0log fecg n = 0, Pn where P mecg, P fecg, P n are the power of the mecg, the fecg and the other types of noise, respectively Real data The real data used for the performance evaluation of the proposed method contains data available free on Physiobank and a set of real data recorded by the authors. The latter is acquired with the MP 50 acquisition system from Biopac, using a matrix of electrodes at the Ioan Cantacuzino Maternity, Bucharest. The sampling frequency used is 000 Hz and the conversion from analog to digital is made on 8 bits. This set of data will be referred as set b. The other real data sets are from the set a available for the Physiobank challenge 203. Data for the challenge consist of a collection of one-minute fetal ECG recordings. Each recording includes four noninvasive abdominal signals sampled at 000 Hz. From set a the records a03, a05, a06, a08 are used to evaluate ICA_EMD METHOD Independent component analysis The aim of ICA is the decomposition of a set of multichannel data in an a priori unknown linear mixture of a priori unknown source signals, relying on the assumption that the source signals are mutually statistically independent [3]. Thus, an observation vector, y, can be defined as in (2): () y = Mx + n, (2) where x is the vector containing the independent sources, n is the additive noise and M is the mixing matrix. An element of the matrix M, m i indicates to what extent the th source component contributes to the i th observation channel, i.e. they determine how the sources are mixed in the observations [4]. ICA is used to estimate the mixing matrix, M, and the sources vector x, knowing ust the observation vector, y under the following assumptions: i) the mixing vectors are linearly independent; ii) the components of x are mutually statistically independent, as well as independent from the noise components. Thus, the source signals are estimated from the observations by using a simple matrix multiplication: x = My. (3) There are different approaches for the estimation of the mixing matrix and the source signals. In the present paper, the solution proposed in [4], called oint approximate diagonalization of eigen-matrices, JADE, is used, which exploits the oint diagonalization of a set of fourth order cumulant matrices Empirical mode decomposition The EMD is a signal processing technique used for nonstationary signal decomposition. It has the advantage of being fully data driven, adaptive, and thus no a priori knowledge is necessary [0]. Based on its principle, EMD decomposes the signal in a set of oscillations called intrinsic mode functions (IMF). A function is considered to be an IMF if it satisfies the following conditions: i) the number of extrema and zero-crossings must be equal or to differ at most by one and ii) it has to be symmetric with respect to local zero mean, i.e. the mean value of the envelope defined by the local maxima and that defined by the local minima should be zero. The decomposition process of the EMD, called sifting process, is as follows: a) determine all the local maxima and minima and construct the upper and lower envelopes by cubic spline interpolation; b) determine the mean of the upper and lower envelops, m (t), and then subtract it from the original signal x(t) as in (4): ( ) x( t) ( t) h t m =. (4) If h (t) satisfies the conditions of the IMF, then it is considered to be the first IMF obtained in the sifting process. If h (t) doesn t satisfy the conditions to be an IMF, then steps a) and b) are reiterated for h (t): h ( ) h ( t) ( t), t m, =. (5) After k cycles of the iterative steps, the IMF is obtained: h ( t) h ( t) ( t), k =, k m, k. (6) Thus, the first IMF is obtained, c = h,k- (t). c) subtract the obtained IMF from the original data, obtaining the residue r (t). r ( ) x( t) ( t) t c =. (7)

3 96 Dragoş Daniel Ţarălungă, Ilinca Gussi, Rodica Strungaru 3 d) r (t) is considered the new signal and all the steps above are reiterated in order to obtain the second IMF, c 2 (t) and the second residue r 2 (t). Thus, the residue can be expressed as following: r () t r () t c () t =. (8) When the residue r (t) becomes a constant or monotonic function, the sifting process is terminated. The data can be reconstructed by summing up all obtained IMFs: L () t = c () t + r () t i x, (9) where L is the number of the IMFs obtained after the sifting process Independent component analysis empirical mode decomposition The BSS method JADE is used as a first step in extracting the fecg from the abdominal recorded signals based on the hypothesis that the sources are independent. Hence, each recording channel is considered an element of the observation vector x. With the help of JADE, the fecg is separated from the maority of noise, including mecg, EHG, PLI etc. However, the fecg component obtained with the BSS method is not entirely free of noise, i.e. a perfect separation is not achieved. It still contains noise like electrode artefacts, baseline wander, muscle artefacts which can still impair the fecg analysis. The EMD is chosen to further denoise the signal, because of its advantage to be entirely data driven and applicable on nonstationary signals Comparative study The performance of the proposed method is compared with the one obtained by JADE with no further denoising step. The identification of fecg component is made by visual inspection. For a quantitative comparison two performance indices are defined for the simulated and real signals. Thus, for the simulated data sets, the performance of the fecg denoising is measured as in (0): L The fecg component is further denoised by applying the EMD to obtain the IMFs. Next, based on visual inspection, the IMFs which contain only noise are eliminated, and the rest of the IMFs are used to reconstruct the fecg signal. In Fig. a is depicted one of the ten channels containing the original simulated abdominal signal which is the mixture of fecg, mecg and other types of noise; in Fig. b the fecg is obtained after JADE is applied, and in the last subplot, the result of fecg denoising with ICA_EMD is depicted. JADE and the proposed method, ICA_EMD, are applied on all simulated data sets, i.e. when the SNR of the other types of noise except mecg varies with the following values: 20 db, 5 db, 0 db, 5 db, 0 db. The error calculated with (0) for each data set when JADE and ICA-EMD are applied, is illustrated in Fig. 2. a) b) ε = M 2 ( () i fecg () i ) fecg () i orig est M fecg, (0) orig 2 where fecg orig is the original fecg signal introduced in the simulation and fecg est is the fecg signal after denoising with JADE and ICA-EMD, respectively. For the case of the real data set, because there is no access to the clean reference fecg, the performance of denoising is quantified using the SNR computed as following: the energy of the fetal QRS complex over the energy of the base line present before and after a fetal QRS complex. c) Fig. a) Original simulated abdominal signal; b) fecg obtained after JADE is applied; fecg estimated with ICA_EMD. 3. RESULTS 3.. SIMULATED SIGNALS The first step is to apply the chosen BSS method in order to decompose the simulated signal and to obtain a fecg component, which is identified based on visual inspection. Fig. 2 The error computed for each simulated data set.

4 4 A new method for fetal electrocardiogram denoising REAL SIGNALS The ICA_EMD method is applied also on real data. The steps of the proposed method are presented for set b, i.e. the signals recorded by the authors. In Fig. 3a a channel from the original real data from set b is depicted. It can be observed that the PLI has a strong burying effect even on the mecg, which usually contributes the most interference in abdominal signals. After the ICA algorithm is applied, the fecg component is identified in IC 9. The fecg component, i.e. IC9, is further decomposed using EMD in order to improve the SNR. Based on visual inspection, the IMFs containing noise are reected and the remaining IMFs are used to reconstruct the signal and obtain a fecg estimate. In Fig. 3b the fecg estimated with JADE is illustrated, while in Fig. 3c the fecg estimated with ICA_EMD can be observed. JADE and ICA_EMD are also applied on the recordings in set a from Physiobank, i.e. on a03, a05, a06 and a08. The SNR computed as described in Section 2, is depicted in Fig. 4. Quantitative differences can be observed between the SNR obtained with JADE and with ICA_EMD. The results show that ICA_EMD has the highest SNR, especially for the data in set b, where the SNR for JADE is db, indicating the presence of high noise in the fecg IC, as confirmed by Fig. 3b, and the SNR obtained with ICA_EMD is 24.7 db, an improvement of 5.32 db. a) Fig. 4 The SNR computed for each data set, set a, with a03, a05, a06, a08 and set b. 4. CONCLUSIONS In this paper, a novel combination of two important signal processing methods has been introduced in order to completely remove various sources of noise from abdominal signals in order to obtain a clean abdominal fecg. The ICA_EMD method consists of two steps: first the JADE algorithm is applied on the raw signals to obtain the ICs, and then the EMD is further applied on the IC containing the fecg for a final denoising. From Fig., Fig. 2, Fig. 3 and Fig. 4 it can be observed that the proposed ICA_EMD method has higher performance than the JADE algorithm on both simulated and real data. For the latter data, the quantitative analysis based on the computed error shows that ICA_EMD overcomes the JADE performance for all datasets. In addition, the method described is computationally economical and it can be used as preprocessing step for fecg segmentation in order to determine the fhr, the durations of specific waves, segments and intervals, etc. Finally we conclude that this study brings significant contribution in the domain of abdominal fecg denoising, confirming that this type of signal can be obtained with an improved SNR so that is reliable in clinical practice. ACKNOWLEDGMENTS The research is financed by the Sectoral Operational Programme Human Resources Development of the Ministry of European Funds through the Financial Agreement POSDRU/59/.5/S/ Received on July 23, 205 b) c) Fig. 3 a) Original real abdominal signal; b) fecg obtained after JADE is applied; c) fecg estimated with ICA_EMD. REFERENCES. H. Noren, A.K.Luttkus, J.H. Stupin, et al., Fetal scalp ph and ST analysis of the fetal ECG as an adunct to cardiotocography to predict fetal acidosis in labor: a multi-center, case controlled study, J. Perinat. Med., 35, pp , G. Clifford, R. Sameni, J. Ward, J. Robinson, A.J. Wolfberg, Clinically accurate fetal ECG parameters acquired from maternal abdominal sensors, Am. J. Obstet. Gynecol., 205,, pp. 47.e 5, M.A. Hasan, M.D. Mamun, M. Marufuzzaman, Hardware approach of a novel algorithm of r-peak detection for the simultaneous measurement of fetal and maternal heart rates during pregnancy, Rev. Roum. Sci. Techn. Électrotechn. et Énerg., 57, 4, pp , D. D. Taralunga, G. M. Ungureanu, I. Gussi, R. Strungaru, W. Wolf, Fetal ecg extraction from abdominal signals: a review on suppression of fundamental power line interference component and its harmonics, Comput. Math. Methods. Med. (239060), pp. 5, 204.

5 98 Dragoş Daniel Ţarălungă, Ilinca Gussi, Rodica Strungaru 5 5. S. Wu, Y. Shen, Z. Zhou, L. Lin, Y. Zeng, X. Gao, Research of fetal ECG extraction using wavelet analysis and adaptive filtering, Computers in Biology and Medicine, 43, 0, pp , M. Peters, J. Crowe, J. F. Piéri, H. Quartero, B. Hayes-Gill, D. James, J. Stinstra, S. Shakespeare Monitoring the fetal heart non-invasively: a review of methods, Journal of Perinatal Medicine, 29, 5, pp , D. Devedeux, C. Marque, S. Mansour, G. Germain, J. Duchêne Uterine electromyography: a critical review, Am. J. Obstet. Gynecol., 69, 6, pp , L. D. Lathauwer, B. D Moor, J. Vanderwalle Fetal electrocardiogram extraction by blind source subspace separation, IEEE Trans. Biomed. Eng., 47, pp , R. Sameni, G. D. Clifford, A Review of Fetal ECG Signal Processing; Issues and Promising Directions. The Open Pacing, Electro-physiology & Therapy Journal, 3, pp. 4 20, N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, E. H. Shih, Q. Zheng, N. C. Yen, C. C. Tung, H. H. Liu, The empirical mode decomposition method and the Hilbert spectrum for non-stationary time series analysis, Proc. R. Soc. Lond, A454, pp , R. Sameni, Extraction of Fetal Cardiac Signals from an Array of Maternal Abdominal Recordings, Ph.D. dissertation, Sharif University of Technology - Institut National Polytechnique de Grenoble, R Sameni, G. D. Clifford, C. Jutten, and M. B. Shamsollahi. Multichannel ECG and Noise Modeling: Application to Maternal and Fetal ECG Signals, EURASIP Journal on Advances in Signal Processing, Article ID 43407, L. De Lathauwer, B. De Moor, J. Vandewalle, An introduction to independent component analysis, J. Chemometr., 4, 2000, pp J. F. Cardoso, High-order contrasts for independent component analysis, Neural Computation,,, pp , 999.