IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO.

5, MAY 2013

1345

Fetal ECG Extraction by Extended State Kalman Filtering Based on Single-Channel Recordings
Mohammad Niknazar , Bertrand Rivet, and Christian Jutten, Fellow, IEEE
AbstractIn this paper, we present an extended nonlinear Bayesian ltering framework for extracting electrocardiograms (ECGs) from a single channel as encountered in the fetal ECG extraction from abdominal sensor. The recorded signals are modeled as the summation of several ECGs. Each of them is described by a nonlinear dynamic model, previously presented for the generation of a highly realistic synthetic ECG. Consequently, each ECG has a corresponding term in this model and can thus be efciently discriminated even if the waves overlap in time. The parameter sensitivity analysis for different values of noise level, amplitude, and heart rate ratios between fetal and maternal ECGs shows its effectiveness for a large set of values of these parameters. This framework is also validated on the extractions of fetal ECG from actual abdominal recordings, as well as of actual twin magnetocardiograms. Index TermsExtended Kalman ltering (EKF), fetal electrocardiogram (fECG) extraction, model-based ltering, nonlinear Bayesian ltering, twin magnetocardiogram (MCG) extraction.

I. INTRODUCTION INCE the rst demonstration of the fetal electrocardiogram (fECG) carried out in 1906 by Cremer [1], various methods for fECG monitoring have been proposed to obtain information about the heart status. The fECG can be measured by placing electrodes on the mothers abdomen. However, it has very low power and is mixed with several sources of noise and interference. Nevertheless, the main contamination is the maternal electrocardiogram (mECG) [2]. As a result, the basic problem is to extract the fECG signal from the mixture of mECG and fECG signals, where the interfering mECG is a much stronger signal. According to the review [3], existing fECG extraction approaches in the literature can be categorized by their methodologies, which include linear or nonlinear decomposition and adaptive ltering. Linear or nonlinear decomposition methods are common approaches, in which single- or multichannel recordings are

Manuscript received February 20, 2012; revised May 7, 2012, October 1, 2012, and November 27, 2012; accepted November 29, 2012. Date of publication December 20, 2012; date of current version April 15, 2013. Asterisk indicates corresponding author. M. Niknazar is with the GIPSA-lab (UMR CNRS 5216), University of Grenoble, Grenoble 38402, France (e-mail: mohammad.niknazar@gipsa-lab. grenoble-inp.fr). B. Rivet is with the GIPSA-lab (UMR CNRS 5216), University of Grenoble, Grenoble 38402, France (e-mail: [email protected]). C. Jutten is with the GIPSA-lab (UMR CNRS 5216), University of Grenoble, Grenoble 38402, France, and also with the Institut Universitaire de France, 75005 Paris, France (e-mail: [email protected]). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TBME.2012.2234456

decomposed into different components using suitable basis functions. Linear decomposition methods use either xed basis functions (e.g., wavelets [4]), or data-driven basis functions (e.g., singular vectors [5]). This limits performance of decomposition in nonlinear or degenerate mixtures of signal and noise [3]. Blind or semiblind source separation, which are categorized as linear decomposition approach, have also been used for fECG extraction [6], [7]. These methods are based on the assumption of independent components (or more generally independent subspaces [8] or partitions [9]) for the maternal and fetal signals, or of the existence of some temporal structure for the desired signals [10][12]. In [13] and [14], wavelet decomposition was also combined with blind source separation for extracting and denoising fECG signals. In another recent work, a new technique was proposed to fasten the traditional independent component analysis (ICA) method [15]. In blind source separation methods, it is usually assumed that signals and noises are mixed in a stationary and linear manner. However, fECG and other interferences and noises are not always stationary mixed and linearly separable [16]. Nonlinear transforms have been also used for mECG cancelation and fECG extraction. In these methods, constructed phase space of noisy signal and of its delayed versions is smoothed using conventional or principal component analysis (PCA) smoothers [17]. The samples are then transferred back to the time-domain representation. Although these methods are interesting since they are applicable to as few as one single maternal abdominal channel, the selection of the required time lags for constructing phase space representation is empirical and the important interbeat variations of the cardiac signals can be wiped-out during the state-space smoothing. Moreover, they demand higher computational complexity in comparison to linear methods, and the correct embedding dimension can change as the noise statistics change [3]. Adaptive ltering is another common approach for mECG cancelation and fECG extraction [18]. The conventional adaptive ltering is based on training an adaptive lter for either removing the mECG using one or several maternal reference channels [18], [19], or directly training the lter for extracting the fetal QRS waves [20], [21]. However, existing adaptive ltering methods for mECG artifact removal either require a reference mECG channel that is morphologically similar to the contaminating waveform or require several linearly independent channels to roughly reconstruct any morphologic shape from the references [18]. Both of these approaches are practically inconvenient and with limiting performance, because the morphology of the mECG contaminants highly depends on the electrode locations and it is not always possible to reconstruct the complete

0018-9294/$31.00 2012 IEEE

1346

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 5, MAY 2013

mECG morphology from a linear combination of the reference electrodes [3]. Practically, it has been shown that for fECG extraction, blind source separation methods outperform adaptive lters [22]. An important advantage of spatial ltering over conventional adaptive lters is their ability to separate mECG and fECG with temporal overlap but it often requires more than two sensors. The Kalman ltering (KF) framework, which can be considered as a member of the general class of adaptive lters, is a promising approach for both mECG cancelation and fECG enhancement. In [16] and [23], a set of state-space equations was used to model the temporal dynamics of ECG signals, for designing a Bayesian lter for ECG denoising. This Bayesian lter framework was used in [16, p. 50] to extract fECG from singlechannel mixture of mECG and fECG. However, as mentioned in [16], the lter fails to discriminate between the maternal and fetal components when the mECG and fECG waves fully overlap in time. The reason is that when mECG is being estimated, fECG and other components are supposed to be Gaussian noises. However, this assumption is not true, especially when mECG and fECG waves fully overlap in time it is difcult for the lter to follow desired ECG. Clinical monitoring of fetal cardiac activity is usually based on a small number of electrodes located on mothers abdomen, and on a sound sensitive sensor. In such a context, in this study, we wonder what performance can be obtained with only one electrode, by using a rened model of the signal recorded on the unique electrode: the model will explicitly take into account that the signal is the superposition of a few ECG signals. The rest of this paper is organized as follows. In Section II, equations and theory supporting our proposed method including the Bayesian ltering theory and dynamic ECG model are described. In Section III, results of the proposed method applied on different data and discussion about the results are presented. Finally, our conclusion is stated in Section IV. II. METHODOLOGY A. Extended Kalman Filter Framework for ECG Extraction The goal of KF is to estimate the state of a discrete-time controlled process. Consider a state vector xk +1 governed by a nonlinear stochastic difference equation with measurement vector yk +1 at time instant k + 1: xk +1 = f (xk , wk , k + 1) yk +1 = h(xk +1 , vk +1 , k + 1) (1)

Fig. 1.

Illustration of the phase assignment approach on one ECG.

In this study, a synthetic dynamic ECG model [25] is used to extract fECG from mixture of an mECG, one (or more) fECG(s), and other signals considered as noises. In polar coordinates [23], one ECG signal can be expressed as the sum of ve Gaussian functions dened by their peak amplitude, width, and center, denoted i , bi , and i , respectively: z () = i W i exp(( i )2 /(2b2 i )). Each Gaussian function thus models one of the ve waves W = {P, Q, R, S, T } of a heart beat. The state vector in (1) is dened by the phase and the amplitude z of the ECG: xk = [k , zk ]T . Assuming a small sampling period , the state noise k , and dening wk as [0, k ]T , the state process f () is k +1 = (k + ) mod(2 ) zk +1 =
i W

(2)
2 i,k 2b2 i

i i,k exp b2 i

+ zk + k (3)

where the random variables wk and vk represent the process and measurement noises, with associated covariance maT T and Rk = E vk vk . The extended trices Qk = E wk wk Kalman lter (EKF) is an extension of the standard KF to nonlinear systems f () and h(), which linearizes about the current mean and covariance [24]. In order to improve the estimations, EKF can be followed by a backward recursive smoothing stage leading to the extended Kalman smoother (EKS). However, since EKS is a noncausal method, it cannot be applied online but it is useful if a small lag in the processing is allowed.

where is the phase increment and i,k = (k i ) mod(2 ). From the ECG, one can dene the observed phase k by a linear time wrapping of the RR time intervals into [0, 2 ) (see Fig. 1). The measurement process h() is nally dened as yk +1 = xk +1 + vk +1 , where yk +1 = [k +1 , sk +1 ]T . The ECGs composing the observed mixture can be estimated by recursively applying the described EKF: at each step, one ECG is extracted according to a deation procedure. In case of a mixture of mECG and one fECG, the rst step extracts, from the raw recording, the dominant ECG (often the mECG) considering the concurrent ECG (respectively, fECG) and other noises as a unique Gaussian noise. After subtracting the dominant ECG from the original signal, the second step is the extraction of fECG from the residual signal. This procedure is referred to as sequential EKF or EKS (seq-EKF or seq-EKS). In this recursive extraction, during the rst step, the concurrent ECG (i.e., fECG) and additional noise are modeled by Gaussian noises vk and wk , which is not a very relevant assumption. In fact, although this assumption may be acceptable when there are not strong artifacts interfering with the ECG, it is no longer accurate when other ECG artifacts are considerable (i.e., at the rst step) since the noise is no longer normally distributed. In addition, concurrent ECGs can be confused with dominant ECG when their waves (especially QRS complexes) fully overlap in time. Meanwhile, resultant inaccuracies, which are generated by the previous steps of the ECG extraction, will propagate to the next steps while the residuals are computed.

NIKNAZAR et al.: FETAL ECG EXTRACTION BY EXTENDED STATE KALMAN FILTERING BASED ON SINGLE-CHANNEL RECORDINGS

1347

B. Extension to Multiple ECGs: Extended State EKF In this paper, the dynamic equations (2) and (3) are extended for simultaneously modeling N ECGs mixed in a single observation. The related extended state vector xk = (1) (1) (N ) (N ) [k , zk , . . . , k , zk ]T is thus dened by (1) (1) k +1 = (k + (1) ) mod(2 ) (1) 2 (1) i,k i (1) (1) (1) zk +1 = i,k exp (1) 2 (1) 2 b 2bi i W i 1 (1) (1) +zk + k . . . ( N ) (N ) + ( N ) ) mod(2 ) k +1 = (k (N )2 (N ) i,k (N ) i ( N ) ( N ) i,k exp ( N ) 2 (N )2 zk +1 = b 2bi i W i N (N ) (N ) + zk + k where each [k , zk ]T is related to one of the ECGs. Finally, the measurement process leads to express the measurement vector (1) (N ) yk +1 = [k +1 , . . . , k +1 , sk +1 ]T as (n ) (n ) (n ) = k +1 + vk +1 , n {1, . . . , N } k +1 N (4) (n ) ( N +1) zk +1 + vk +1 . sk +1 =
n =1 (i) (i)

fECG is obtained by using the seq-EKF algorithm, which now allows us to detect easily the fetal R-peaks.1 Then, for each ECG, each beat (dened by the signals between two consecutive Rpeaks) is time wrapped into [0, 2 ). The average of the ECG waveform is obtained by the mean of all time-wrapped beats, for all phases between 0 and 2 . Finally, by using a nonlinear least-squares approach [26], the best estimate of the parameters in the minimum mean square error (MMSE) sense is found. III. RESULTS AND DISCUSSIONS Both synthetic and actual data have been used to study performance of the proposed method. In Subsection III-A, quantitative results coming from simulations and inuence of the main parameters of mixed ECGs on performance of the method has been studied. They will present the conditions in which the proposed method is efcient. In Subsection III-B, the effectiveness of the method on actual data has been examined. A. Experimental Performance Analysis on Synthetic Data Since there is neither ground truth nor golden standard on single-channel recording, it is important to provide quantitative performance with simulations to validate the behavior of the proposed method. In order to do so, realistic synthetic mixtures of mECG and fECG with white Gaussian noise have been generated for different situations, and the proposed method has been applied on them to extract mECG and fECG. Synthetic mECG and fECG used in this study are based on a 3-D canonical model of the single dipole vector of the heart, proposed in [27] and inspired by the single-channel ECG dynamic model presented in [25]. Sampling frequency is set to 500 Hz and signals include 20 000 samples. The main parameters that can affect the mixtures are input noise power, ratio between amplitudes of fECG and mECG, and ratio between fetal and maternal heart rates. In order to investigate the performance of the proposed method, 100 trials were carried out under each value of these parameters. In the output, estimated mECG and f , are assumed to be the sum of mECG, fECG signals, s m and s fECG, and noise, such that s m = 1 sm + 2 sf + 3 n s f = 1 sm + 2 sf + 3 n (5)

This extended state Kalman ltering procedure is referred to as parallel EKF or EKS (par-EKF, or par-EKS, respectively). As shown in Section III, this par-EKF or par-EKS is more accurate to extract fECG from abdominal sensors than the seq-EKF or seq-EKF. Indeed, in the proposed method, all ECGs are jointly modeled by dynamic states so that only the state and measurement noise vectors are assumed to be normally distributed. Moreover, the extended state par-EKF fully models overlapping waves of several ECGs. Finally, the state and observation noises, n n and vk , respectively, allow the lter to t some variabilities k of the ECG shapes. Although the model does not t too large variations (for example, due to arrythmia), an inspection of the residue will reveal these abnormal beats. C. Model Parameters Estimation The proposed par-EKF and par-EKS lie on several state pa(n ) (n ) (n ) rameters {i , bi , i }i Wn , n {1, . . . , N }. The procedure described below is an extension of the single ECG parameter estimation [23]. The parameters estimation procedure rst needs the R-peak detection for all ECGs to perform the time wrapping of the RR (n ) intervals into [0, 2 ) to dene k . The R-peaks are found from a peak search in windows of length T , where T corresponds to the R-peak period calculated from approximate ECG beat-rate. R-peaks with periods smaller than T 2 or larger than T are not detected. Although maternal R-peaks are easily detectable from the mixture, fetal R-peak detection is more complex due to its lower amplitude than mECG. Therefore, a rough estimation of

where coefcients 1 , 2 , 3 , 1 , 2 , and 3 have to be estimated and sm , sf , and n denote mECG, fECG, and noise, respectively. In order to estimate the coefcients, sm , sf , and n are assumed to be orthogonal, i.e., decorrelated. The orthogonality principle states that an estimator s achieves MMSE if and = 0. Satisfaction of this criterion leads to only if E ( s s)T s 1 = 1 = E ( sT E ( sT E ( sT m sm ) m sf ) m n) , , = = 2 3 T s ) T n) E (sT s ) E ( s E ( s m m m f m E ( sT f sm ) E (sT f sm ) 2 = , E ( sT f sf ) E (sT f sf ) 3 = , E ( sT f n) E (sT f n) . (6)

1 In practice, one could also use a sound sensor to have a reliable R-peak detector. In this case, even if there exists a delay, it does not impact the method, since it can be synchronized.

1348

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 5, MAY 2013

Fig. 2. Mean SNR improvement results of the EKF and EKS against input noise power (bold lines). Upper and lower borders (thin lines) present maximum and minimum, respectively.

Fig. 3. Mean SIR improvement results of the EKF and EKS against amplitude ratio (bold lines). Upper and lower borders (thin lines) present maximum and minimum, respectively.

In a successful estimation, contribution of desired ECG in output should be much more than contribution of undesired ECG and noise. In other words, in extraction of fECG, the power of 2 sf should be much larger than the power of 1 sm + 3 n, which means the contribution of mECG and noise is very low in the fECG estimate. In the same manner, the power of 1 sm should be much larger than the power of 2 sf + 3 n in mECG extraction. In order to quantize contribution of the desired ECG in the output, output signal-to-noise ratio (SNR) for maternal and fetal ECGs are dened as SNRm o u t = SNRf o u t =
2 Ps m 1 2 2P 2 Ps f + 3 n

2 Ps 2 f 2 Pn 2 Ps +
1
m

(7)

where Ps m , Ps f , and Pn denote power of mECG, fECG, and noise, respectively. Output SNR is now compared to input SNR to investigate performance of desired ECG extraction. Input SNRs are dened as Ps f Ps m and SNRf i n = . (8) SNRm i n = Ps f + Pn Ps m + Pn Input signal-to-interference ratio (SIR) and output SIR are also dened as Ps f Ps SIRm i n = m , SIRf i n = Ps f Ps m SIRm o u t =
2 2 Ps Ps m 1 2 f , SIR = . fo u t 2P 2 Ps 2 sf m 1

(9)

1) SNR Analysis: Fig. 2 shows SNR improvement results of EKF and EKS over a wide range of input noise power. The SNR improvement in decibels is dened as the output SNR of the lter minus the input SNR. In all trials, power of mECG signals is normalized to 1 (0 dB) and the ratio of amplitudes of fECG and mECG is 0.3. Maternal and fetal heart rates are set to 1.1

and 2 Hz, respectively. Moreover, in order to have more realistic signals, mECG and fECG are allowed to have slight random uctuations (5%) in amplitude and duration at each beat. Moreover, initial phases of ECGs are random. As can be seen in Fig. 2, both EKF and EKS successfully improved the SNR for all ranges of the input SNRs. When the mixture is rather noise free (noise power 30 dB), the minimum SNR improvement of fECG is 40 dB, which means efcient cancelation of mECG. Nevertheless, even for very noisy mixtures (noise power 20 dB), the SNR improvement of fECG remains over 20 dB. According to this gure, EKF is more effective when a rather clean signal is available. On the contrary, as power of noise increases, EKS signicantly outperforms EKF. As it has been explained in the previous section, the EKS algorithm consists of a forward EKF stage followed by a backward recursive smoothing stage. Therefore, if a rather clean signal is available, the recursive smoothing stage will deteriorate EKF output, because the output is smooth enough and recursive smoothing leads to overltering. Conversely, if the signal is very noisy, EKF output is not denoised enough yet. Therefore, recursive smoothing stage can be successfully used to cancel more noise from the signal. 2) Amplitude Ratio Analysis: The basic problem of fECG monitoring is to extract the fECG signal from the mixture of mECG and fECG signals, where the interfering mECG is a stronger signal. Therefore, it is necessary to evaluate the performance of the method for different ratios of fECG and mECG amplitudes. For this purpose, SIR improvement of output signals have been calculated in the range of 0.11 of amplitude ratio of fECG and mECG. Fig. 3 shows SIR improvement results of the EKF and EKS for different values of amplitude ratios. Power of mECG signals are normalized to 1 (0 dB) with 5% random uctuation, input SNR with respect to (w.r.t.) mECG is 10 dB, and average maternal and fetal heart rates are 1.1 and 2 Hz, respectively. As is seen in Fig. 3, although the fetal SIR improvements of both EKF and EKS remain over 30 dB for all ranges of the amplitude ratios, results of EKS are slightly better.

NIKNAZAR et al.: FETAL ECG EXTRACTION BY EXTENDED STATE KALMAN FILTERING BASED ON SINGLE-CHANNEL RECORDINGS

1349

Fig. 4. Mean SIR improvement results of the EKF and EKS against heart rate ratio (bold lines). Upper and lower borders (thin lines) present maximum and minimum, respectively.

3) Heart Rate Ratio Analysis: Since fetal heart rate may vary in a wide range [28], the performance of the method was studied on a wide range of 0.33.6 Hz of fetal heart rate. Fig. 4 shows SIR improvement results of EKF and EKS. Power of mECG signals are normalized to 1 (0 dB) with 5% random uctuation and the ratio of amplitudes of fECG and mECG is 0.3. Input SNR w.r.t. mECG is 10 dB, and maternal heart rate is set to 1.1 Hz. In this section, heart rate uctuations are slighter (1%) to study harmonic issues more accurately. As expected, SIR improvement diagram has three deep local minima at ratios 1, 2, and 3. The reason is that when main frequencies of mECG and fECG are proportional, the signals overlap more closely in the frequency domain. Therefore, discrimination of mECG and fECG is more difcult for these ratios. Nevertheless, these situations are unlikely happening because the heart rate ratio is usually more than 1 and less than 2. Even in these cases, fetal SIR improvement remains over 20 dB. Here again, EKS slightly outperforms EKF. B. Fetal ECG Extraction on Actual Data In Subsection III-A, efciency of the proposed method in fECG extraction for a wide range of possible congurations has been examined using synthetic data. In this section, the results of application of the proposed method on actual data are presented. 1) DaISy Database: The DaISy fetal ECG database [29] consists of a single dataset of cutaneous potential recording of a pregnant woman. A total of eight channels (ve abdominal and three thoracic) are available, sampled at 250 Hz and lasting 10 s. Fig. 5 presents the results of par-EKS and seq-EKS using the rst channel of the dataset. Moreover, the periodic component analysis ( CA) [8] using the eight channels, which is a multichannel method, is also included as the golden standard. Results of CA method are then postprocessed via EKS on the best ECG estimate [23]. As already mentioned, unlike seq-EKS, par-EKS does not fail when mECG and fECG fully overlap in time. This

Fig. 5. Comparison of fECG extraction by par-EKS, seq-EKS and CA on the rst channel of DaISy data.

is particularly noticed between t = 6 s and t = 7 s in Fig. 5, in which some parts of fECG signal have been deteriorated during mECG extraction by the seq-EKS method. On the contrary, the proposed par-EKS jointly models the fECG and mECG, resulting in a better estimate of fECG than seq-EKS. Since par-EKS estimates a single component while CA can estimate several components (typically one or two), the cosine between subspaces is used and is equal to 0.92 in this experiment. With a value close to 1, these estimates are quite similar. Finally, Fig. 6 shows the results of fECG extraction using par-EKS applied on the other abdominal channels of the DaISy dataset. It experimentally proves that par-EKS is able to extract fECG even in ill-conditioned mixtures, such as channels 4 or 5. 2) Noninvasive fECG Database: This database consists of a series of 55 multichannel abdominal fECG recordings, taken from a single subject between 21 and 40 weeks of pregnancy. The recordings include two thoracic signals and three or four abdominal signals. The signals were recorded at 1 kHz, 16-bit resolution with a bandpass lter (0.01100 Hz) and a main notch lter (50 Hz) [30]. Fig. 7 shows results of seq-EKS and par-EKS using channel 3, and CA using all channels of the rst 20s of namely the ecgca771 dataset. To show the effectiveness of the proposed method in extraction of the fECG at different periods of pregnancy, and from different channel locations, the rst 20s of the mixtures and fetal par-EKS outputs of the datasets ecgca274 channel 5, ecgca748 channel 4, and ecgca997 channel 3 are plotted in Fig. 8. 3) Twin Magnetocardiograms Extraction: The proposed method has been principally designed for ECG signals. Nevertheless, due to the morphological similarity of the ECG and the magnetocardiogram (MCG), it is also directly applicable to MCG recordings. In this section, twin fetal cardiac mag-

1350

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 5, MAY 2013

Fig. 8. ECG mixtures of the datasets ecgca274 channel 5, ecgca748 channel 4, and ecgca997 channel 3, and their fetal par-EKS outputs.

Fig. 6. Results of fECG extraction using par-EKS applied on channels 25 of the DaISy dataset (up to down). Note differences of scales, according to the channels and the fetal estimates.

Fig. 7. Comparison of fECG extraction by par-EKS, seq-EKS, and CA on ecgca771 of the PhysioNet database. Fig. 9. Results of the seq-EKS, par-EKS, and CA on twin MCG data.

netic signals recorded by a SQUID Biomagnetometer system are extracted. The dataset has been recorded in the Biomagnetic Center of the Department of Neurology (Friedrich Schiller University, Jena, Germany) and it consists of 208 channels sampled at 1025 Hz over 30 min.

Fig. 9 presents the results of the proposed par-EKS to extract the two fetal MCG signals from a single sensor. A typical channel (indexed 92) of namely the q00002252 dataset has been selected. Even though the multichannel CA method provides better results in this case than single-channel methods (par-EKS

NIKNAZAR et al.: FETAL ECG EXTRACTION BY EXTENDED STATE KALMAN FILTERING BASED ON SINGLE-CHANNEL RECORDINGS

1351

Fig. 10. MCG mixtures of the channels 126, 152, and 160, and their fetal par-EKS outputs.

Performance of the proposed method on extraction of fECG from one mixture of mECG and fECG was examined according to noise level, amplitude ratio, and heart rate ratio parameters: results show that the proposed method can be successfully employed in many scenarios. According to the obtained results, as long as R-peaks are correctly detected, the proposed model achieves good results. Although a reliable R-peak detection is a straight forward procedure in a single fetal pregnancy (which most likely happens) even with a single sensor, it is much more difcult in multiple fetal pregnancy (twin or more). Nonetheless, in these situations, the R-peak detection can be provided by other modalities such as echocardiography. Finally, the proposed method compares favorably with efcient multisensor methods such as CA (which also requires reliable R-peak detection), while it requires only one sensor. The latter criterion is of high interest, since using a single channel does not only mean less electronic components (such as analog to digital converters or ampliers) and thus a cheaper device, but also a more convenient and portable device for a long-term monitoring system or at home since only a single electrode has to be placed on the mothers abdomen. Perspectives include extension of the proposed method to apply on multichannel (but with a small number of channels, e.g., up to 3 or 4) mixtures of mECG and fECG. Moreover, synchronous echocardiography data can also be used in future works, especially for a reliable R-peak detection. ACKNOWLEDGMENT The authors would like to thank Dr. R. Sameni for his support and assistance in this study, and Dr. D. Hoyer for providing the twin MCG dataset. REFERENCES
[1] H. M. Jenkins, Technical progress in fetal electrocardiographyA review, J. Perinat. Med., vol. 14, no. 6, pp. 365370, 1986. [2] G. Camps, M. Martinez, and E. Soria, Fetal ECG extraction using an r neural network, in Proc. Comput. Cardiol., 2001, pp. 249252. [3] R. Sameni and G. D. Clifford, A review of fetal ECG signal processing; Issues and promising directions, Open Pacing, Electrophysiol. Ther. J., vol. 3, pp. 420, 2010. [4] A. Khamene and S. Negahdaripour, A new method for the extraction of fetal ECG from the composite abdominal signal, IEEE Trans. Biomed. Eng., vol. 47, no. 4, pp. 507516, Apr. 2000. [5] P.P. Kanjilal, S. Palit, and G. Saha, Fetal ECG extraction from singlechannel maternal ECG using singular value decomposition, IEEE Trans. Biomed. Eng., vol. 44, no. 1, pp. 5159, Jan. 1997. [6] J.-F. Cardoso, Multidimensional independent component analysis, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., May 1998, vol. 4, pp. 19411944. [7] L. de Lathauwer, B. de Moor, and J. Vandewalle, Fetal electrocardiogram extraction by blind source subspace separation, IEEE Trans. Biomed. Eng., vol. 47, no. 5, pp. 567572, May 2000. [8] R. Sameni, C. Jutten, and M. B. Shamsollahi, Multichannel electrocardiogram decomposition using periodic component analysis, IEEE Trans. Biomed. Eng., vol. 55, no. 8, pp. 19351940, Aug. 2008. [9] J. L. Camargo-Olivares, R. Marti-Clemente, S. Hornillo-Mellado, M. M. Elena, and I. Roman, The maternal abdominal ECG as input to MICA in the fetal ECG extraction problem, IEEE Signal Process. Lett., vol. 18, no. 3, pp. 161164, Mar. 2011. [10] A. K. Barros and A. Cichocki, Extraction of specic signals with temporal structure, Neural Comput., vol. 13, no. 9, pp. 19952003, Sep. 2001.

or seq-EKS), the proposed par-EKS succeeds to extract the two fetal MCG (fMCG), while seq-EKS fails to discriminate correctly the two fMCGs when they overlap (see highlighted signal parts in Fig. 9). In order to show the good behavior of par-EKS in several congurations, par-EKS is applied on other sensors (see Fig. 10). One can note that the proposed par-EKS succeeds to extract the two fetal MCGs. Finally, it is worth noting that the crucial part of the proposed par-EKS is the R-peak detection. Although this detection is quite direct when a single fetus is present (see Subsection II-C), some words should be added on twin data. Indeed, on such data, the detection of the mothers R-peaks is still direct since it is the dominant signal. On the contrary, the discrimination between the two fetal R-peaks is much more difcult. Even though in this study, the oracle is obtained using several sensors and applying an ICA algorithm (here, we used Fast-ICA), it can be replaced in practice by a sound sensor located on the mothers abdomen. IV. CONCLUSION In this paper, a synthetic dynamic ECG model within a KF framework has been extended to jointly model several ECGs to extract desired ECGs from a unique mixture (i.e., one channel recording) of maternal and fetal ECGs and noise. Although the proposed method only uses a single channel to separate different ECGs, since each ECG has a corresponding term in the model, the proposed model can efciently discriminate ECGs even if desired and undesired ECG waves overlap in time. As proved on synthetic data and illustrated on actual data (single and multiple fetal pregnancy), the main merit of the proposed algorithm relies on its performance in a large class of situations.

1352

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 5, MAY 2013

[11] Z. Yi and Z. Zhang, Extraction of temporally correlated sources with its application to non-invasive fetal electrocardiogram extraction, Neurocomputing, vol. 69, no. 79, pp. 894899, 2006. [12] Y. Li and Z. Yi, An algorithm for extracting fetal electrocardiogram, Neurocomputing, vol. 71, no. 79, pp. 15381542, 2008. [13] V. Vigneron, A. Paraschiv-Ionescu, A. Azancot, O. Sibony, and C. Jutten, Fetal electrocardiogram extraction based on non-stationary ICA and wavelet denoising, in Proc. 7th Int. Signal Process. Appl. Symp., 2003, vol. 2, pp. 6972. [14] M. G. Jafari and J. A. Chambers, Fetal electrocardiogram extraction by sequential source separation in the wavelet domain, IEEE Trans. Biomed. Eng., vol. 52, no. 3, pp. 390400, Mar. 2005. [15] R. Martin-Clemente, J. L. Camargo-Olivares, S. Hornillo-Mellado, M. Elena, and I. Roman, Fast technique for noninvasive fetal ECG extraction, IEEE Trans. Biomed. Eng., vol. 58, no. 2, pp. 227230, Feb. 2011. [16] R. Sameni. (2008, Jul.) Extraction of fetal cardiac signals from an array of maternal abdominal recordings, Ph.D. thesis, Sharif University of TechnologyInstitut National Polytechnique de Grenoble, [Online]. Available:http://www.sameni.info/Publications/Thesis/PhDThesis.pdf [17] M. Kotas, Projective ltering of time-aligned ECG beats, IEEE Trans. Biomed. Eng., vol. 51, no. 7, pp. 11291139, Jul. 2004. [18] B. Widrow, Jr, J. R. Glover, J. M. McCool, J. Kaunitz, C. S. Williams, R. H. Hearn, J. R. Zeidler, Jr, E. Dong, and R. C. Goodlin, Adaptive noise cancelling: Principles and applications, Proc. IEEE, vol. 63, no. 12, pp. 16921716, Dec. 1975. [19] N. J. Outram, E. C. Ifeachor, P. W. J. Van Eetvelt, and J. S. H. Curnow, Techniques for optimal enhancement and feature extraction of fetal electrocardiogram, IEE Proc.Sci., Meas. Technol., vol. 142, no. 6, pp. 482 489, 1995. [20] A. G. Favret, Computer matched lter location of fetal R-waves, Med. Biol. Eng., vol. 6, no. 5, pp. 467475, Sep. 1968. [21] Y. C. Park, K. Y. Lee, D. H. Youn, N. H. Kim, W. K. Kim, and S. H. Park, On detecting the presence of fetal R-wave using the moving averaged magnitude difference algorithm, IEEE Trans. Biomed. Eng., vol. 39, no. 8, pp. 868871, Aug. 1992. [22] V. Zarzoso and A. K. Nandi, Noninvasive fetal electrocardiogram extraction: blind separation versus adaptive noise cancellation, IEEE Trans. Biomed. Eng., vol. 48, no. 1, pp. 1218, Jan. 2001. [23] R. Sameni, M. B. Shamsollahi, C. Jutten, and G. D. Clifford, A nonlinear Bayesian ltering framework for ECG denoising, IEEE Trans. Biomed. Eng., vol. 54, no. 12, pp. 21722185, Dec. 2007. [24] G. Welch and G. Bishop, An introduction to the Kalman lter, Dept. Comput. Sci., Univ. North Carolina, Chapel Hill, Tech. Rep. TR 95-041, 1995. [25] P. E. McSharry, G. D. Clifford, L. Tarassenko, and L. A. Smith, A dynamical model for generating synthetic electrocardiogram signals, IEEE Trans. Biomed. Eng., vol. 50, no. 3, pp. 289294, Mar. 2003. [26] G. D. Clifford, A. Shoeb, P. E. McSharry, and B. A. Janz, Model-based ltering, compression and classication of the ECG, Int. J. Bioelectromag., vol. 7, no. 1, pp. 158161, 2005. [27] R. Sameni, G. D. Clifford, C. Jutten, and M. B. Shamsollahi, Multichannel ECG and noise modeling: Application to maternal and fetal ECG signals, EURASIP J. Adv. Signal Process., vol. 2007, p. 43407, 2007. [28] T. J. DuBose, J. A. Cunyus, and L. F. Johnson, Embryonic heart rate and age, J. Diagnost. Med. Sonogr., vol. 6, no. 3, pp. 151157, 1990. [29] B. De Moor, P. De Gersem, B. De Schutter, and W. Favoreel, DAISY: A database for identication of systems, J. A, Spec. Issue Comput. Aided Control Systems Design, vol. 38, no. 3, pp. 45, Sep. 1997. [30] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals, Circulation, vol. 101, no. 23, pp. e215e220, Jun. 13, 2000, Circulation Electronic Pages, [Online]. Available:http://circ.ahajournals.org/cgi/content/full/101/23/ e215

Mohammad Niknazar was born in Tehran, Iran, in 1985. He received his B.Sc. degree in electrical electronic engineering from the University of Semnan, Semnan, Iran, in 2007, and the M.Sc. degree in electrical biomedical engineering from the Sharif University of Technology, Tehran, Iran, in 2010. As a member of GIPSA-lab, he is currently working toward the Ph.D. degree in signal, image and speech processing and telecommunications at Joseph Fourier University, Grenoble, France. His research interests include signal and image processing, nonlinear (chaos) analysis, blind source separation, and pattern recognition. Current areas of application include fetal ECG extraction and ECG processing. Previous areas of application include epileptic seizure detection and prediction.

Bertrand Rivet was graduated from the Ecole Normale Sup erieure de Cachan, France. He received the Agr egation de Physique Appliqu ee in 2002, the Masters degree from the University of Paris-XI, France, in 2003, and the Ph.D. degree from Grenoble Institute of Technology (GIT), Grenoble, France, in 2006. He is currently an Associate Professor in signal processing with PHELMA and a member of GIPSAlab, GIT. His research interests include biomedical signal processing, audiovisual speech processing, and blind source separation.

Christian Jutten (F08) received the Ph.D. degree ` s Sciences degree from the Instiand the Docteur e tut National Polytechnique of Grenoble, Grenoble, France, in 1981 and 1987, respectively. From 1982 to 1989, he was an Associate Professor in the Department of Electrical Engineering, Institut National Polytechnique de Grenoble. Since 1989, he has been a full Professor at University Joseph Fourier, Grenoble. He was a Visiting Professor in the Swiss Federal Polytechnic Institute in Lausanne in 1989 and in Campinas University (Brazil) in 2010. He has been the Deputy Director of the Grenoble images, speech, signal, and control laboratory (GIPSA, 300 people) and Director of the Department Images-Signal (DIS, 100 people) from 2007 to 2010. For 30 years, his research interests are blind source separation, independent component analysis, and learning in neural networks, including theoretical aspects (separability, source separation in nonlinear mixtures, sparsity) and applications in signal processing (biomedical, seismic, hyperspectral imaging, speech). He is an author or co-author of more than 65 papers in international journals, four books, 22 invited plenary talks, and 150 communications in international conferences. Dr. Jutten was an Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS in 19941995, and co-organizer the 1st International Conference on Blind Signal Separation and Independent Component Analysis (Aussois, France, January 1999). He has been a scientic advisor for signal and images processing at the French Ministry of Research from 1996 to 1998 and for the French National Research Center (CNRS) from 2003 to 2006. He is currently Deputy Director of the Institute for Information Sciences and Technologies of CNRS. He received the Medal Blondel in 1997 from SEE (French Electrical Engineering Society) for his contributions in source separation and independent component analysis, and became a Senior Member of the Institut Universitaire de France in 2008.