Fetal Heart Rate Monitoring Based on Adaptive Noise Cancellation and Maternal QRS Removal Window

Transcription

1 Euroean Journal of Scientific Research ISSN X Vol.7 No.4 (009), EuroJournals Publishing, Inc. 009 htt:// Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window M. Sheikh M. Algunaidi Student Member IEEE Deartment of Electrical, Electronic & Systems Engineering Faculty of Engineering and Built Environment, University Kebangsaan Malaysia (UKM), Bangi, Malaysia M. A. Mohd Ali Member IEEE, Deartment of Electrical, Electronic & Systems Engineering Faculty of Engineering and Built Environment, University Kebangsaan Malaysia (UKM, Bangi, Malaysia K. B. Gan Member IEEE, Deartment of Electrical, Electronic & Systems Engineering Faculty of Engineering and Built Environment, University Kebangsaan Malaysia (UKM), Bangi, Malaysia E. Zahedi Member IEEE, School of Electrical Engineering, SHARIF University of Technology Tehran, Iran Abstract In this aer a new method to extract the fetal signal from the abdominal electrocardiogram (ECG) is resented. A three-stage method for fetal heart rate detection from abdominal ECG recordings is roosed. After rerocessing, adative noise cancellation (ANC) is used to extract the fetal ECG. Then in the third stage maternal QRS comlex removal window is alied to eliminate or scale down the maternal residual eaks. The method is validated using 30 recorded data and comared with another three stage method using indeendent comonent analysis (ICA) for the fetal ECG extraction. The average sensitivity and average ositive redictivity of the ANC based method is 85.8 % and 67.6 % resectivly comared to 74.4% and 64.1% of the ICA based method. These show that the ANC based method was more successful in detecting the FHR than ICA. Keywords: Adative noise canceller, indeendent comonent analysis, fetal heart rate monitoring and QRS Removal Window.

2 Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window Introduction Fetal heart rate (FHR) monitoring is one of the methodologies to test fetal well being and diagnose for ossible abnormalities. Fetal monitoring throughout the regnancy enables the clinician to diagnose and recognize the athologic condition esecially ashyxia [1]. Although Doler ultrasound device is currently used for FHR monitoring, it is not suitable for long term monitoring due to its sensitivity to movement and its safety for long term exosure has yet to be established []. Besides ultrasound, non-invasive electrocardiograhy has been used to obtain valuable clinical information about the fetal well being during regnancy. The extraction of the fetal electrocardiogram (ECG) can be carried out via skin electrodes attached to the maternal abdomen. However, the abdominal ECG (AECG) is always corruted with ower line interference, maternal ECG (MECG) and electromyogram where its variability is influenced by the gestational age, osition of the electrodes and the skin imedance [3]. Therefore, aroriate signal rocessing techniques are required to reveal the fetal ECG (FECG) from the AECG. Various research efforts have been roosed to extract the FECG from the AECG such as adative filtering [4], correlation techniques [5], blind source searation [6] and a combination of wavelet analysis and blind source searation methods [7]. FHR can be calculated by determining the R-R intervals from the extracted FECG. However, the extracted FECG is still corruted by the residual eaks of MECG (esecially its QRS comlexes) hence the FECG detection remains difficult. In this aer, an adative noise canceller (ANC) is roosed to extract the FECG from the AECG. A QRS removal window (a window for removing the maternal QRS (MQRS) comlex), algorithm is develoed to eliminate the MECG residual eaks in the extracted FECG. The erformance of the roosed algorithm is evaluated and comared with the well-known indeendent comonent analysis (ICA) algorithm by using recorded data from the Universiti Kebangsaan Malaysia Medical Center (PPUKM).. Methodology.1. Data Acquisition AECG signals were recorded from 30 healthy regnant women (at 35 to 38 weeks of gestation), most of which are corruted with different levels of noises, using the lead system as shown in Fig. 1. The exerimental rotocol was aroved by the PPUKM Research and Ethical Committee rior to commencement of the study and informed consents were obtained from all subjects. Figure 1: Locations of the abdominal electrodes. com P 1 P BIOPAC- MP 100A PC 4 P 3 P 5 The AECG signals, = X ( n ), X ( n ),..., X ( n ) ] T X(n) where n denotes a discrete-time index, [ 1 and T is the transose oerator, were simultaneously recorded from maternal abdomen using six

3 567 M. Sheikh M. Algunaidi, M. A. Mohd Ali, K. B. Gan and E. Zahedi electrodes (five electrodes, [l,, 3, 4, 5], with a single common) using high gain amlifiers (BIOPAC- MP 100A). The AECG signals were digitized at 1000 Hz with 1 bit resolution. The total recording time during each session was about one minute. Electrode 1 is located in such a way that only MECG signals are acquired while the electrodes, 3, 4 and 5 acquired the mixture of MECG and FECG. Therefore, ( n) reference inut and X ( n) and X 3 ( n) four of the acquired AECG signals, = [ X ( n), X 3( n),..., X ( n)] algorithm as X 1 ( n) contains only MECG signal X 1 is defined as the are the rimary inut signals to the adative filter. As for ICA, X(n) [, 3, 4, 5], are fed into the ICA.. Algorithms The block diagram of the roosed algorithm is shown in Figure (). It consists of the re-rocessing stage, FECG extraction using ANC or ICA and the MQRS removal window. Figure : The block diagram of the roosed algorithm. Recorded ECG signals Pre-rocessing functions FECG extraction using ANC or ICA MQRS Removal FECG SIGNAL..1. Prerocessing Stage The rerocessing stage consists of the removal of the DC signal, baseline wander and the ower line interference. Each observation signal is made zero mean by subtracting its mean as follows: X ( n) = X ( n) mean( X ( n)) (1) Baseline wander is caused by the atient's breathing or movements during recording. The frequency of the baseline wander due to breathing is in the range of 1 Hz and the EMG noise (artifacts of muscular contractions) is characterized by relatively high frequency noise, hence the recorded signals were filtered by a FIR band-ass filter with cut-off frequencies at 4 Hz and 90 Hz. The ower line interference consists of 50 Hz sine wave and its harmonics. A notch filter centered at 50 Hz is used to eliminate this interference. An examle of an AECG signal and the rerocessed signal are shown in Fig. 3.

4 Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window 568 Figure 3: (a) AECG signal and (b) re-rocessed signal (a) (b)... FECG Extraction Technique Two algorithms namely ANC and ICA have been imlemented in this work to evaluate their erformance for FECG extraction. A. Adative Noise Canceller ANC is a method of estimating a signal (FECG ), contaminated by additive noise, MECG, [,3] with the rimary inut to the ANC becoming X using a reference inut, MECG1 X 1 [8]. The noise MECG 1 is uncorrelated with the FECG but correlated in some unknown way with the noise MECG as shown in Figure (4). The X 1 and X are the signals acquired from the maternal abdomen. Figure 4: Adative noise canceller system MECG FECG X = MECG + FECG [,3] + + e - MECG1 X 1 Adative filter y The noise MECG 1 is filtered to roduce an outut y that is as close a relica as ossible of MECG. This outut is subtracted from the rimary inut X to roduce the system outut e = MECG + FECG y () Where y is the outut of the adative filter Squaring both sides of Equation, we obtain e = FECG + ( MECG y) + FECG ( MECG y) (3) Alying exectations on both sides of Equation 3, we get

5 569 M. Sheikh M. Algunaidi, M. A. Mohd Ali, K. B. Gan and E. Zahedi E[ e ] = E[ FECG ] + E[( MECG y) ] + E[ FECG ( MECG y)] (4) As FECG is uncorrelated neither with MECG nor with y then E[ FECG ( MECG y)] = 0. Finally, we obtain E[ e ] = E[ FECG ] + E[( MECG y) ] (5) The goal of the adative filter is to minimize the mean square error (MSE) of E[ MECG y] = 0 This can be obtained iteratively, to give the otimal solution when y = MECG. B. Indeendent Comonent Analysis ICA is a method to find underlying factors or comonents from multivariate (multidimensional) statistical data. It looks for comonents that are both statistically indeendent and non-gaussian. Although an excellent review has been given by Cichocki & Amari [9], a brief descrition is given here. Given a set of mixed signals X(n) = [X 1 (n), X (n),, X (n)] T which are linear mixed with q (> q) unknown mutually statistically indeendent, zero-mean source signals S(n) = [s 1 (n), s (n),,s q (n)] T and noise contaminated. This can be written as i q X ( n) = A s ( n) + g ( n), i = 1,,, (6) j= 1 ij j ci or in the matrix notation X = AS + g c (7) where X = X(n) is the vector of sensor signals, S = S(n) is the source signal vector, g c = [g c (n), g c (n),, g c (n)] T is the additive noise vector, A is an unknown q mixing matrix and n is the discretetime index. The noise vector, g c is assumed Gaussian and indeendent. The mixing matrix A is determined by the body geometry and conductivity, as well as the electrode-source relative ositions [1]. Criteria based on maximization of non-gaussianity [13], maximum likelihood, minimization of mutual information [14], tensorial methods [15] and non-linear decorrelation [16] may be used to estimate the mixing matrix A and the source signal vector S. In the noise-free model, g c = 0, the identification of the mixing matrix A and the sources signal, S can be estimated if the sources are indeendent and non-gaussian, and the number of sensors is equal or larger than the number of indeendent sources to be estimated. However, a noisy estimates of the -1 sources signal may obtain, S = A ( X g c ), if g c 0. Therefore, re-rocessing before alying ICA may imrove the erformance of the ICA. In this aer, Second Order Blind Identification (SOBI) algorithm has been utilized to extract the FECG from the AECG. SOBI is a robust algorithm to searate the noises esecially EMG and electrode artifacts...3. Maternal QRS Removal In the ost rocessing stage two stes are imlemented which are the MQRS removal to eliminate the maternal residual eaks, and finally a 1 Hz notch filter to attenuate the residual baseline wander in the FECG. MQRS signal is catured within a window which is defined by taking 50 samles before and after every eak found in the inut signal X 1 with the condition as shown Fig. 5.

6 Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window 570 Figure 5: Flow Chart of the MQRS Removal Window i = 0 h = i = i+1 h=i yes X 1 (i)> X 1 (i ± 1) i > h no no X 1 (i-s) = 0 no yes i > h + s yes The QRS comlex interval is given by k = (i s) : (i + s); (8) where i is the samle index and s =50 is the number of samles. All samles that do not fall within this window will be zero added as shown in Figure 6(b). The MQRS window of Figure6(b) is used to eliminate or scale down the maternal residual eaks from the extracted FECG. A small amount of baseline wander was observed at this extracted signal. Therefore, a notch filter centered at 1 Hz is used to attenuate this baseline wander.

7 571 M. Sheikh M. Algunaidi, M. A. Mohd Ali, K. B. Gan and E. Zahedi Figure 6(a): Pre-rocessed AECG signal X1 and (b) MQRS interval definition. (a) k (b).3. Evaluation The roosed algorithms have been imlemented in Matlab codes using Matlab-7.4 (The Math-works Inc.). The erformances of the algorithms were then evaluated based on their sensitivities and ositive redictivities [15], when alied to AECG signal acquired from the PPUKM. The sensitivity is the fraction of real events that are correctly detected and it is defined by, TP Se = (9) TP + FN The Positive Predictivity is the fraction of detections that are real events and it is defined by, TP + P = (10) TP + FP where FN (False Negatives) denotes the number of missed detections, FP (False Positives) reresents the number of extra detections and TP (True Positives) is the number of correctly detected QRS comlexes. 3. Results and Discussions 3.1. Adative noise canceller extraction technique Examles of the extracted FECG using ANC, maternal QRS and FECG signal after alying maternal QRS removal window are shown in Figure (7). It is noted that the maternal residual eaks are still observed after the ANC and only eliminated after alying the maternal QRS removal window. After maternal residual eaks have been eliminated from the extracted FECG signal, a small amount of baseline wander has been observed. Therefore, notch filter centered at 1 Hz is adequate to attenuate this baseline wander.

8 Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window 57 Figure 7: (a) Extracted FECG using ANC, (b) MQRS and (c) FECG signal after alying MQRS removal window. 3.. Indeendent Comonent Analysis Extraction Technique Figure (8)shows examles of the result using ICA. The maternal residual eaks are still observed in the extracted FECG signal using ICA. The QRS Removal Window was then alied to remove the maternal residual eak from the FECG. After maternal residual eaks have been eliminated from the extracted FECG signal, a small amount of baseline wander has been observed. Therefore, notch filter centered at 1 Hz is adequate to attenuate this baseline wander.

9 573 M. Sheikh M. Algunaidi, M. A. Mohd Ali, K. B. Gan and E. Zahedi Figure 8: (a) Extracted FECG using ICA, (b) MQRS window and (c) FECG signal after alying MQRS removal window (a) (b) (c) With this imrovement on the extracted signal, the erformance of the FECG extraction techniques (ANC and ICA) is comared in terms of FHR detection. The FHR is calculated from the RR interval after eak detection [16] Performance Evaluation In this section, the ANC and ICA methods are evaluated using sensitivity and ositive redictivity. The effect of the lead osition in the rimary inut of the ANC is also evaluated. Table 1 shows the erformance using ICA and ANC based methods at signal extraction stage. The average sensitivity of the ANC based method is 85.5 % (X as rimary signal) as comared to 74.4% of the ICA based method. The average ositive redictivity of the ANC based method is 67.6% (X as rimary signal) as comared with that of the ICA based method which is 64.1%. It shows that the ANC based method was more successful in detecting the FHR than ICA. The QRS Removal window was emloyed to imrove this detection.. Table 1: Performance of ICA and ANC based method ANC method Weeks No Signals X as rimary signal X 3 as rimary signal ICA method Se (%) + P (%) Se (%) + P (%) Se (%) + P (%) (%) 67.6(%) 8.5(%) 66.7(%) 74.4(%) 64.1(%)

10 Fetal Heart Rate Monitoring Based on Adative Noise Cancellation and Maternal QRS Removal Window 574 With the availability of the multi-lead system for the ICA used in this work, it is ossible to evaluate the lead osition for the rimary signal of the ANC that gives otimum results. Hence a comarison is made between X 3 as the rimary signal and X as also shown in Table 1. The average sensitivity of the roosed algorithm from the rimary signal X is 85.8 % as comared to 8.5 % with rimary signal X 3. Also the average ositive redictivity of the roosed algorithm is 67.6% with rimary signal X as comared to X 3 with 66.7%. The erformance of the algorithm was better than ICA for both locations, although electrode location (associated with X ) is better than 3. This shows that the location of the electrode lays an imortant role in FHR detection. 4. Conclusion The roosed algorithm (ANC with the QRS Removal Window) has been demonstrated to have better erformance to extract the fetal signal. This method can use only two leads and a common. By using the MQRS removal window it is shown that it is ossible to control the amlitude of the maternal QRS comlex in the extracted signal or eliminate it. This facilitates detection of the fetal eaks and therefore the determination of FHR. The limitation of the roosed algorithm is that only signals which acquired later than 35 gestation weeks are tested. Farther imrovement is required to imlement the algorithm on ECG signal earlier than 35 gestation week. Current work is in rogress towards realizing an online FHR detection using 4 bit high resolution multi-channel bio-amlifier and finally the roosed algorithm will be fully tested in the clinical environment. Acknowledgement The authors would like to thank the Ministry of Science, Technology and Innovation, Malaysia; for suorting this work under the Science Fund Grant SF055. The authors would like also to exress their gratitude to Dr. Farshid Soheili for roviding the clinical data for this aer and esecially Professor Dr. Muhamad Abdul Jamil M. Yassin and Associate Professor Dr. Shuhaila Ahmad for their assistance in collecting the clinical data.

11 575 M. Sheikh M. Algunaidi, M. A. Mohd Ali, K. B. Gan and E. Zahedi References [1] R.K. Freeman; T. J. Garite; M. P. Nageotte, "Fetal Heart Rate Monitoring", cht.1,. 1 4, 003. Liincott Williams & Wilkins ( 003). [] G. M. Friesen, et al. "A Comarison of the Noise Sensitivity of Nine QRS Detection Algorithms." IEEE Trans. Biomed. Engineering 37: (1990). [3] R.C. Goodlin, "History of fetal monitoring", Am.J.Obstet,Gynecol 133, (1989). [4] E. R. Ferrara and B. Widrow, "Fetal electrocardiogram enhancement by time-sequenced adative filtering", IEEE Trans. Biomed. Eng. 9, (198). [5] S. Abboud, A. Alaluf, S. Einav, and D. Sadeh, "Real time abdominal fetal ECG recording using hardware correlator", Comut. Biol. Med.,, (199). [6] L. De Lathauwer, B. De Moor, and J. Vandewalle, "Fetal electrocardiogram extraction by source subsace searation", in Proc. IEEE SP/ATHOS Worksho HOS, (1995). [7] J.G. Maria, and C.A. Jonathon, "Fetal Electrocardiogram Extraction by Sequential Source Searation in the Wavelet Domain", IEEE Trans Biomed Eng, 5, (005). B. Azzerboni, F.L. Foresta, N. Mammone, and F.C. Morabito, "A New Aroach Based on Wavelet-ICA Algorithms for Fetal Electrocardiogram Extraction", in Proc. 13th EuroeanSymosium of Artificial Neural Networks, 7-9 (005). [8] B.Widrow, J.R., Jr.Glover, J.M.McCool, J. Kaunitz, C.S. Williams, R.H. Hearn, J.R. Zeidler, Jr. E. Dong, R.C.Goodlin, "Adative noise cancelling: Princiles and alications", Proceedings of the IEEE 63, (1975). [9] A. Cichocki & S. Amari, "Adative Blind Signal and Image Processing", , Wiley (00). [10] J. Vanderschoot, D. Callaerts, W. Sansen, J. Vandewalle, G. Vantraen, J. Janssens. "Two methods for otimal mecg elimination and fecg detection from skin electrode signals". IEEE Trans Biomed Eng March 1987;34(3): [11] A. Hyvarinen 1999 "Fast and robust fixed oint algorithm for indeendent comonent analysis" IEEE Trans. Neural Netw [1] P. Comon 1994 "Indeendent comonent analysis: a new concet" Signal Process [13] J. F. Cardoso 1989 "Source searation using higher order moments" Proc. ICASSP [14] C. Jutten and Taleb "A 000 Source searation: from dusk till dawn" Proc. ICA [15] Geng Jun "Find eak value of datas". USTB, Beijing, China for Dr. Ma Zheng, [Online]. Available: E -mail: dr.gengjun@16.com [16] (ANSI/AAMI EC57): "Testing and reorting erformance results of cardiac rhythm and ST segment measurement algorithms", (AAMI Recommended Practice/American National Standard).