Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 6, No. 5, October 2016, pp. 2197 2204 ISSN: 2088-8708 2197 Empirical Mode Decomposition (EMD) Based Denoising Method f or Heart Sound Signal and Its P erf ormance Analysis Amy H. Salman , Nur Ahmadi , Richard Mengk o , Armein Z. R. Langi , and T ati L. R. Mengk o School of Electrical Engineering and Informatics, Institut T eknologi Bandung Article Inf o Article history: Recei v ed May 25, 2016 Re vised July 11, 2016 Accepted July 27, 2016 K eyw ord: Heart Sound Denoising Empirical Mode Decomposition ABSTRA CT In this paper , a denoising method for heart sound signal bas ed on empirical mode decompo- sition (EMD) is proposed. T o e v aluate the performance of the proposed method, e xtensi v e simulations are performed using syntheti c normal and abnormal heart sound data corrupted with white, colored, e xponential and alpha-stable noise under dif ferent SNR input v alues. The performance is e v aluated in terms of signal-to-noise ratio (SNR), root mean square er - ror (RMSE), and percent root mean square dif ference (PRD), and compared with w a v elet transform (WT) and total v ariation (TV) denoising methods. The simulation results sho w that the proposed method outperforms tw o other methods in r emo ving three types of noises. Copyright c 2016 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Amy H. Salman School of Electrical Engineering and Informatics Bandung Institute of T echnology Jl. Ganesha No 10, Bandung 40132, Indonesia Email: amy@stei.itb .ac.id 1. INTR ODUCTION Cardio v ascular disease (CVD) has long been the leading cause of death throughout the w orld with an estimate of 17.3 million people died in 2008 and is predicted to reach 23.3 million in 2030 [1]. According to WHO report, more than three quarters of the death tak es place in lo w- and middle-income countries [2]. A lo w-cost and non-in v asi v e diagnosis system based on heart sound can be used to minimize the risk of patients going into se v ere condition and reduce the financial b urden through an early accurate diagnosis follo wed by appropriate treatment. This electronic auscultation technique utilizes adv anced si gnal processing with f ast computation capability , thanks to the adv ancement of computer technology . Ho we v er , to produce an accurate diagnosis result is not an easy task since, in practice, heart sound signal is al w ays contaminated with noise and interference from v arious sources such as background noise, po wer interference, breathing or lung sounds, and skin mo v ements in the surrounding en vironment. Thus, signal denosing method is of paramount importance to remo v e all these unw anted noise. A poor signal denoising method can lead to catastrophic result. The most widely used method for denoising heart sound signal is based on w a v elet transform (WT) [3–5], a po werful signal analysis tool with the ability to repres ent a signal simultaneously in the time and frequenc y . Despite the f act that the w a v elet based denoising method has been pro v en to be able to pro vide good denoising performance, ho we v er , it suf fers from se v eral limita tions. It requires predefined basis function selection (from too man y choices) suited to signal under consideration, which limits the fle xibility of the method. In addition, the decomposition le v el and thresholding technique of w a v elet denosing also need to be carefully considered. F ailing to choose the right de- composition le v el and thresholding technique will result in bad denoising performance. V ar ghees and Ramachandran emplo yed another alternati v e method based on T otal V ariation (TV) [6]. TV method has been mostly used for image denoising due to its great benefit of preserving and enhancing important features such as edge in images. Ev en though it can be used for denoising 1D signal, ne v ertheless, there are v ery fe w literatures e xploiting TV method for deonising heart sound. The highly non-stationary property of heart sound signal is not suitable to the nature of TV method which performs best on piece wise constant signals [7]. In addition, Figueiredo et. al. mentioned that the performance of TV J ournal Homepage: http://iaesjournal.com/online/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
2198 ISSN: 2088-8708 method could produce better result than older w a v elet based methods, b ut it w as outperformed by recent state-the-art w a v elet based methods [8]. Empirical mode decomposition (EMD), a relati v ely ne w non-linear a n d non-stationary signal analysis method [9], of fers interesting feature of adapti v e and data-dri v en decomposition capability . Since its inception, EMD has attracted man y rese archers around the globe to utilize it as denoising method [10–12]. Ho we v er , the mechanism of discriminating the noise and useful information within decomposed signal and fitti n g it to heart sound signal to get good signal reconstruction performance remains challenging. In this paper , we propose an EMD based denoising method for heart sound signals. T o measure the per - formance of our proposed method, we perform repeated simulation o v er normal and abnormal synthetic heart sound signal b urried under dif ferent types of noise with SNR input v alues ranging from 0 to 15 dB. The qualitati v e e v alu- ation is performed by visual inspection while qu a ntitati v e e v aluation is carried out by using three standard metrics: signal-to-noise ratio (SNR), root mean square error (RMSE), and percent root mean square dif ference (PRD). The rest of this paper is or g anized as follo ws. Section 2. describes EMD denoising method wi th brief intro- duction to its theoritical background and mathematical notation. Section 3. e xplains the imulation setting and data. The simulation results and performance analysis of both qualitati v e and quantitati v e are gi v en in Section 4. Finally the conclusion is dra wn in Section 5. 2. EMD DENOISING METHOD 2.1. Empirical Mode Decomposition (EMD) Empirical mode decom p os ition (EMD), since firstly proposed by Huang in 1998 [9], has g ained popularity as data analys is method especiall y for non-stationary and non-linear signals such as biomedical (including heart sound) signals. EMD, in contrast to other methods such w a v elets and fourier whic h require predefined basis function, is fully data-dri v en method that does not require an y a priori kno wn basis. EM D adapti v ely decomposes a signal into a series of simple oscillatory AM-FM components called as intrinsic mode functions (IMFs) through itera ti v e procedure (kno wn as sifting ). An IMF is defined as a function that satisfies tw o conditions: the number of e xtrema (maxima and minima) and zero crossing must be equal or dif fer by at most 1; and the a v erage v alue of the en v elopes deri v ed from local maxima and minima is (approximately) zero. Despite EMD still being lack of a solid mathematical foundation which could be used for theoritical analysis and performance e v aluation, it has been pro v en to pro vide interesting and useful results. The sifting procedure of EMD for decomposing the signal x ( n ) into IMFs is systematically described as follo ws: 1. Specify all the local e xtrema (maxima and minima) of x ( n ) 2. Interpolate between local maxima using cubic spline line to form upper en v elope e max ( n ) and local minima to form lo wer en v elope e min ( n ) 3. Calculate the local mean based on formed upper and lo wer en v elopes, m ( n ) = ( e max ( n ) + e min ( n )) = 2 4. Substract this mean from the original signal to e xtract the detail d ( n ) = x ( n ) m ( n ) . If d ( n ) does not satisfy IMF conditions (stopping criteria), the procedure 1) to 4) are iterated with ne w input signal d ( n ) 5. If d ( n ) satisfies the criteria of an IMF , it is stored as an IMF , h i ( n ) = d ( n ) where i refers to i th IMF . Residue signal is obtained by substracting the IMF from the original signal, r ( n ) = x ( n ) h i ( n ) 6. Perform the same step from 1) with the ne w signal r ( n ) until the final residue signal is constant or monotonic function. After the completion of EMD proce ss, the original signal can be written in terms of its IMF and residue signal as follo ws: x ( n ) = L 1 X i =1 h i ( n ) + r L ( n ) (1) where L refers to decomposition le v el and i denotes IMF order . Lo wer -order of IMFs contains f ast osci llation modes (high frequenc y) while higher -order of IMFs represent slo w oscillation modes (lo w frequenc y). 2.2. EMD Denoising EMD, whose decomposition is based on elementary substractions, enables perfect reconstruction of a signal. The EMD denoising method sta rts by identifying which IMFs carry dominantly noise and which IMFs contain pri- marily useful information. This is done by comparing the actual ener gy density with the estimated ener gy density (to IJECE V ol. 6, No. 5, October 2016: 2197 2204 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2199 form noise-only model [10]) of IMFs. The actual ener gy density of IMFs is calculated as follo ws E i = 1 N N X n =1 h i ( n ) ; i = 1 ; 2 ; 3 ; L (2) with i corresponds to IMF order . The estimated ener gy density (v ariance) of IMFs can be approximated using the formula [13] belo w V 1 = median ( j h 1 median ( h 1 ) j ) 0 : 6745 2 (3) V i = V 1 j ; i = 2 ; 3 ; 4 ; L (4) where and equal to 0.719 and 2.01, respecti v ely [10]. The IMFs whose actual ener gy density e xceed the v alue of their estimated ener gy density defined by noise-only model are cate gorized as information-dominated signal and should be included in signal reconstruction step; otherwise those IMFs will be e xcluded. Due to the f act that the noise em b e dd e d in IMFs is colored (not Gaussian distrib uted), e v en the information- dominated IMFs still may contain noise ha ving dif ferent ener gy density . T o remo v e those colored noise, IMF- dependent threshold v alue is required. Consideri ng that IMFs ha v e zero mean and in an y interv al of zero crossing [ z i j z i j +1 ] the absolute am plitude of i th IMF is v ery small, the thresholding scheme will be based on the single e xtrema h i ( r i j ) , where r i j corresponds to the e xtrema’ s time instance on this inte rv al. The tresholding scheme which follo ws the hard thresholding is e xpressed as follo ws ~ h i ([ z i j z i j +1 ]) = h i ([ z i j z i j +1 ]) ; j h i ( r i j ) j > T i 0 ; j h i ( r i j ) j T i (5) where h i ([ z i j z i j +1 ]) represents the samples from time instant z i j to z i j +1 of t th IMF . The threshold v alue used in this scheme is e xpressed belo w T i = C p V i 2 ln N ; i = 1 ; 2 ; 3 ; L (6) where C is 0.1 found by empirical simulations and V i is estimat ed ener gy density (v ariance) of i th IMF . This thresh- olding scheme which is inspired and adapted from w a v elet [11] will set to zero all the samples from time instant [ z i j z i j +1 ] if the single e xtrema amplitude belo w the theshold v alue meaning that there is no useful information (only noise) in the specified time instant. Otherwise, all the samples will be retained. The final signal reconstruction can be obtained by summing up all the included IMFs (whose act ual ener gy density e xceeding its estimated estimated ener gy as described pre viously) using the follo wing formula ^ y = q X i = p ~ h ( n ) (7) where p and q indicates the lo west and highest inde x of included IMF . 3. SIMULA TION SETTING T o e v aluate the qualitati v e and quantitati v e performance of our proposed method, we performed repeated simulations using synthetic heart sound data obtained from Uni v ersity of Michig an’ s Heart Sound & Murmur Library [14]. In this simulations, three types of heart sound signals used for simulations and their respecti v e recording location are ‘Normal S1 S2’ (Ape x, Supine, Bell), ‘S3 Gallop’ (Ape x, Left Ducubitus, Bell) and ‘S4 Gallop’ (Ape x, Left Ducubitus, Bell). These data are encoded in 44 ,100 Hz sample rate, 16 bits/sample, and 1 minutes of data length. The data w as then do wn-sampled into 2000 Hz to increase the computation process without violating Nyquist theorem, since the frequenc y content of heart sound data is maximum at around 700 Hz. The simulations were carried out 100 times in each case using MA TLAB 2015a which runs on Intel(R) Core(TM) i7-4790 CPU @3.6 GHz W indo ws 7 en vironment. T o measure the performance under v arious noises, we added four types of noises, which are white, colored (bro wn), e xponenti al and alpha-stable noise to the clean input signal y ( n ) to form noisy signal x ( n ) . In each noise case, we used dif ferent input SNR le v el of 0 ; 5 ; 10 , and 15 dB. The noisy heart sound signal, x ( n ) , is e xpressed as x ( n ) = y ( n ) + e ( n ) ; n = 0 ; 1 ; 2 ; N 1 (8) where y ( n ) and e ( n ) denotes the clean signal and noise, respecti v ely . EMD Based Denoising Method for Heart Sound Signal and Its P erformance Analysis (Amy . H. Salman) Evaluation Warning : The document was created with Spire.PDF for Python.
2200 ISSN: 2088-8708 F or the purpose of performance benchmark, we also performed simulations o v er the same data using W a v elet T ransform (WT) and T otal V ariation (TV) based denoising methods. In WT based simulation, Daubechies db10 w a v elet function w as used since it highly resembles the heart sound signal, which lead to yield better performance. In addition, db10 w a v elet has orthogonal property which enables perfect reconstruction of signal and ha v e been reported to produce best result among others [3]. The decomposition le v el, N = 5 , is chosen as recommended in [4]. Hard thresholding technique is selected as it pro vides better result compared to soft thresholding technique. MA TLAB has b uilt-in function wden for w a v elet denoising as described in [15]. Se v eral input parameters and their setting for this function are e xplained in T able 1 [16]. P arameter rigsure represents the selection using the principle of Steins Unbiased Risk Estimate (SURE), h means hard thresholding, and mln denotes threshold rescaling using a le v el-dependent estimation of the le v el noise. T able 1. Input parameters setting P arameter Description Chosen Setting s original signal x ( n ) tptr threshold selection rule rigsure sorh thresholding technique h scal threshold’ s rescaling method mln n decomposition le v el 5 wav (mother) w a v elet function db10 As for TV based denoising method, a Majorization-Minimzation (MM) algorithm [7] w as used to solv e a sequence of optimization problems. The parameter is set to 0.3 based on the e xperiment of and c h a racteristic heart sound signals. The algorithm is run for 50 iterations to find more accurate result. 4. RESUL TS AND AN AL YSIS Figure 1(a) and Figure 1(b) sho ws ‘Normal S1 S2’ heart sound signal x ( n ) decomposition into 11 IMFs along with its final residue signal and the IMFs’ ener gy density comparison under 0 dB le v el of white noise, respecti v ely . As sho wn in Figure 1(b), only information-dominated IMFs (number 3 ; 4 ; 5 ; 6 ; 7 ; and 11 ) will be processed for final reconstruction, which leads to a term “partial reconstruction” of a signal. The qualitati v e performance e v aluation of our proposed method compared to other denoising methods were performed by visual inspection and comparison. Figure 2 presents the input clean signal, noisy signal, and denoised (reconstructed) signal of ‘S4 Gallop’ using w a v elet, TV , and EMD denoising methods. In each noise case, only one simulation result under 5 dB input SNR le v el is sho wn. Based on Figure 2, it is sho wn that EMD denosing method performs better among ot h e rs in three types of noises: white (a), bro wn (b) and e xponential noise (c) as its denoised signal most resembles the original signal. If we look closely and zoom in the figure, we will kno w that the amplitude of denoised signal by TV method are slightly reduced. Ev en though the denoised signal by TV method k eeps the amplitude of its main components almost the same as original one, the amplitude outside the main components interv al x(n) h 1 h 2 h 3 h 4 h 5 h 6 h 7 h 8 h 9 h 10 h 11 r (a) EMD based signal decomposition 0 2 4 6 8 10 12 −18 −16 −14 −12 −10 −8 −6 −4 IMF log 2  (Energy)     Real Energy Estimated Energy (b) Real vs estimated ener gy density of IMFs Figure 1. Signal decomposition and ener gy density comparison of IMFs under 0 dB le v el of white noise IJECE V ol. 6, No. 5, October 2016: 2197 2204 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2201 0 0.5 1 1.5 -1 0 1 Input Signal 0 0.5 1 1.5 -1 0 1 Noisy Signal 0 0.5 1 1.5 Amplitude -1 0 1 Wavelet Denoising 0 0.5 1 1.5 -1 0 1 TV Denoising Time 0 0.5 1 1.5 -1 0 1 EMD Denoising (a) White noise with SNR 5 dB 0 0.5 1 1.5 -1 0 1 Input Signal 0 0.5 1 1.5 -1 0 1 Noisy Signal 0 0.5 1 1.5 Amplitude -1 0 1 Wavelet Denoising 0 0.5 1 1.5 -1 0 1 TV Denoising Time 0 0.5 1 1.5 -1 0 1 EMD Denoising (b) Bro wn noise with SNR 5 dB 0 0.5 1 1.5 -1 0 1 Input Signal 0 0.5 1 1.5 -1 0 1 Noisy Signal 0 0.5 1 1.5 Amplitude -1 0 1 Wavelet Denoising 0 0.5 1 1.5 -1 0 1 TV Denoising Time 0 0.5 1 1.5 -1 0 1 EMD Denoising (c) Exponential noise with SNR 5 dB 0 0.5 1 1.5 -1 0 1 Input Signal 0 0.5 1 1.5 -1 0 1 Noisy Signal 0 0.5 1 1.5 Amplitude -1 0 1 Wavelet Denoising 0 0.5 1 1.5 -1 0 1 TV Denoising Time 0 0.5 1 1.5 -1 0 1 EMD Denoising (d) Alpha-stable noise with SNR 5 dB Figure 2. V isual performance comparison of ‘S4 Gallop’ heart sound signal denoising methods is slightly changed compared to the original signal. As for alpha-stable noise as sho wn in Figure 2(d), WT and TV denoising methods pefrorms better than EMD method. In order to obtain more e xact comparison, a quantitati v e performance w as e v aluated based on three metrics namely signal-to-noise ratio (SNR), r o ot mean square error (RMSE), and percent root mean s q ua re dif ference (PRD), T able 2. Performance comparison of denoising methods for ‘S3 Gallop’ heart sound data Noise Input SNR (dB) RMSE PRD (%) T ype SNR WT TV EMD WT TV EMD WT TV EMD White 0 8.9463 7.6747 9.9222 0.0391 0.0452 0.0352 35.7720 41.3448 32.1339 5 12.7895 11.8554 13.2863 0.0251 0.0280 0.0240 22.9760 25.5490 21.9542 10 16.6698 13.6183 17.6169 0.0161 0.0228 0.0145 14.6864 20.8534 13.2708 15 20.5798 14.2784 20.9220 0.0102 0.0211 0.0101 9.3646 19.3253 9.3160 Bro wn 0 -1.3465 -1.3822 0.58753 0.1329 0.1334 0.1071 121.4947 121.9054 97.8917 5 3.4123 3.1738 5.0755 0.0768 0.0786 0.0641 70.1732 71.8188 58.5944 10 8.3727 7.5069 9.8494 0.0431 0.0471 0.0374 39.4378 43.0857 34.1929 15 13.1883 10.8689 15.1542 0.0246 0.0316 0.0205 22.499 28.8733 18.6951 Exponential 0 -0.9209 -0.8023 6.2413 0.1217 0.1200 0.0537 111.2174 109.7049 49.0467 5 4.0152 4.1953 10.1811 0.0689 0.0675 0.0349 63.0014 61.7036 31.8804 10 8.8459 8.4530 14.5646 0.0395 0.0413 0.0223 36.1247 37.7918 20.3669 15 13.6832 11.6200 18.3461 0.0226 0.0287 0.0157 20.6983 26.2448 14.3572 Alpha-stable 0 9.8326 10.1323 8.0622 0.0440 0.0423 0.0564 40.2600 38.6638 51.5371 5 15.7833 13.0523 11.3505 0.0220 0.0272 0.0392 20.1374 24.8517 35.8671 10 20.1803 13.8206 14.2772 0.0158 0.0259 0.0336 14.4564 23.6848 30.6898 15 25.2577 14.3523 18.3497 0.0073 0.0213 0.0262 6.6370 19.4375 23.9710 EMD Based Denoising Method for Heart Sound Signal and Its P erformance Analysis (Amy . H. Salman) Evaluation Warning : The document was created with Spire.PDF for Python.
2202 ISSN: 2088-8708 which are calculated as follo ws: S N R = 10 log 10 P N n =1 [ y ( n )] 2 P N n =1 [ y ( n ) ^ y ( n )] 2 (9) R M S E = s P N n =1 [ y ( n ) ^ y ( n )] 2 N (10) P R D = v u u t P N n =1 [ y ( n ) ^ y ( n )] 2 P N n =1 [ y ( n )] 2 100 (11) where y ( n ) denotes the clean original signal, ^ y ( n ) refers to t he denoised (reconstructed) signal, and N represents the length of the signal. SNR is defined as the ratio of the po wer of a signal (useful information) and the po wer of noise (irrele v ant signal). RMSE is used to measure the accurac y of denoising method in preserving the quality of information in the denoised signal by calculating the sample standard de viation of the dif fe rences between denoised signal and original signal. PRD is frequently used as a method of quantifying the distortion or the dif ference between the original and the reconstructed signal. The PRD indicates reconstruction fidelity by point wise comparison with the original data. A denoising method is said to perfom better if at a particular input SNR, the v alue of output SNR is lar ger while the v alue of RMSE and PRD are smaller . Comparati v e simulation results of three denoising methods (WT , TV , and EMD) o v er ‘S3 Gallop’ heart sound data on the basis of SNR, RMSE, and PRD are sho wn in T able 2. The simulation result v alues were rounded into 4 digits after comma. Highlighted (bold) v alues indicates the best performance among others. It is sho wn that for three cases of noises (white, bro wn, and e xponential) under dif ferent input SNR v alues ( 0 ; 5 ; 10 ; and 15 dB), EMD denoising method consistently yields lar gest SNR v alue, and smallest RMSE and PRD v alues (see bold v alues). F or instance in white noisy en vironment with 0 dB input SNR le v el, EMD method sho ws SNR v alue 9.9222 dB, RMSE 0.0352 and PRD 32.1339 % where as WT (TV) method sho ws 8.9463 (7.6747) dB SNR, 0.0391 (0.0452) RMSE, and 35.7720 % (41.3448 %) PRD. The performance of EMD method in these three types of noises for other heart sound signals (‘Normal S1 S2’ and ‘S4 g allop’) o v er input SNR le v el range ( 0 dB - 15 dB) is superior as well compared to WT and TV methods. Ho we v er , for heart sound signals contaminated with alpha-stable noise, EMD method does not perform well compared to its counterparts especially for input SNR le v el 0 - 10 dB. In this type of noise, on a v erage, WT method outperforms other tw o methods, e xcept for the case of 0 dB input SNR where TV method produces the best performance on all three metrics. Alpha-stable noise being used in this simulation represents the impulsi v e noise or disturbance characterized by high amplitude and short time duration within arbitrary location along the data. This impulsi v e disturbance usually occurs when there is quick mo v ement or friction between chest skin and stethoscope during recording heart sound data. This alpha-stable noise has four parameters: (characteristic e xponent), (sk e wness), (scale) and (location) [17]. P arameter indicates the tail of distrib ution while specifies whether the distrib ution is right- or left-sk e wed. In this simulation, we used = 1 : 6 , = 1 , = 0 : 1 and = 0 . Graphical visualization of comparati v e simulation results of ‘S4 Gallop’ heart sound signal under four types of noises is depicted in Figure 3. Figure 3(a-c) sho ws the comparati v e output SNR, RMSE and PRD v alue of three denoising methods with respect to dif ferent input SNR le v els in white n oi sy en vironment. It is sho wn that EMD method (blue line with triangle point) on a v are ge performs better than WT (black line with rectangle point) and TV (red line with circle point), indicated by lar ger output SNR v alue and smaller RMSE and PRD v alues. The same trend is also observ ed in simulation results o v er bro wn and e xponential noisy signal as sho wn in Figure 3(d-f) and Figure 3(g-i). EMD is equi v alent to dyadic filter structure which can ef fecti v ely decompose fractional Gaussian noise processes such as white and colored (bro wn) noises. This leads to ef fecti v e denoising method o v er dif ferent class of fractional Gaussian noises [18–20]. Moreo v er , EMD method does not require an y predefined basis function and is fully data-dri v en which of fers more fle xibility and adaptability to an y signal under consideration. Ho we v er , EMD method does not perform well compared to its counterparts under alpha-stable noise simulation as shw on in Figure 3(j-l). According to our observ ation during repeated simulations, we chose constant v alue C = 0 : 6 in threshold v alue calculation within EMD denoising mechanism to obtain good performance. This constant v alue appli es well on three types of noises (white, bro wn and e xponential). Ho we v er , based on our simulation, the performance of EMD denoising method under alpha-stable noise can be impro v ed by increasing the constant v alue C up to 1.5. In addition, to miti g ate this impulsi v e disturbance in heart sound analysis, an adapti v e selection algorithm based on Heron’ s formula can be emplo yed in the subsequent process [21]. IJECE V ol. 6, No. 5, October 2016: 2197 2204 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2203 - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 8 1 0 1 2 1 4 1 6 1 8 2 0 2 2 O u t p u t   S N R   ( d B ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (a) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 0 . 0 1 0 0 . 0 1 5 0 . 0 2 0 0 . 0 2 5 0 . 0 3 0 0 . 0 3 5 0 . 0 4 0 0 . 0 4 5 0 . 0 5 0 R M S E I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (b) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 5 1 0 1 5 2 0 2 5 3 0 3 5 4 0 4 5 P R D   ( % ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (c) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 - 4 - 2 0 2 4 6 8 1 0 1 2 1 4 O u t p u t   S N R   ( d B ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (d) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 0 . 0 2 0 . 0 4 0 . 0 6 0 . 0 8 0 . 1 0 0 . 1 2 0 . 1 4 0 . 1 6 R M S E I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (e) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0 P R D   ( % ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (f) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 - 5 0 5 1 0 1 5 2 0 O u t p u t   S N R   ( d B ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (g) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 0 . 0 2 0 . 0 4 0 . 0 6 0 . 0 8 0 . 1 0 0 . 1 2 0 . 1 4 R M S E I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (h) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 2 0 4 0 6 0 8 0 1 0 0 1 2 0 P R D   ( % ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (i) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 8 1 0 1 2 1 4 1 6 1 8 2 0 2 2 2 4 2 6 O u t p u t   S N R   ( d B ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (j) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 0 . 0 0 0 . 0 1 0 . 0 2 0 . 0 3 0 . 0 4 0 . 0 5 0 . 0 6 0 . 0 7 R M S E I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (k) - 2 0 2 4 6 8 1 0 1 2 1 4 1 6 0 1 0 2 0 3 0 4 0 5 0 6 0 P R D   ( % ) I n p u t   S N R   L e v e l   ( d B )   W T   T V   E M D (l) Figure 3. Performance comparison o v er ‘S4 Gallop’ heart sound signal under (a-c) white (d-f) bro wn (g-i) e xponential and (j-l) alpha-stable noise EMD Based Denoising Method for Heart Sound Signal and Its P erformance Analysis (Amy . H. Salman) Evaluation Warning : The document was created with Spire.PDF for Python.
2204 ISSN: 2088-8708 5. CONCLUSION Empirical Mode Decomposition (EMD) based denois ing method is proposed in this paper . Its performance and analysis compared to other tw o methods based on w a v elet transform (WT) and total v ariation (TV) are presented. F our types of noises with input SNR le v el 0 dB, 5 dB, 10 dB and 15 dB are artificially added to clean original nor - mal and abnormal heart sound signal s obtained from the Uni v ersity of Michig an Health System. Based on e xtensi v e simulations, our proposed EMD based denoising method consistently yields better performance in terms of three stan- dard metrics: signal-to-noise ratio (SNR), root mean square error (RMSE), and percent root mean square dif ference (PRD) under white, colored (bro wn) and e xponential noises. As for alpha-stable noise, on a v erage, WT and TV based denoising methods perform better than EMD method. REFERENCES [1] G. Redlarski, D. Gradole wski, and A. P alk o wski, A system for heart sounds classification, PloS one , v ol. 9, no. 11, p. e112673, 2014. [2] WHO, Global status r eport on noncommunicable diseases 2014 . W orld Health Or g anization, 2014. [3] L. H. Cherif, S. Debbal, and F . Bereksi-Re guig, “Choice of the w a v elet analyzing in the phonocardiogram signal analysis using the discrete and the pack et w a v elet transform, Expert Systems with Applications , v ol. 37, no. 2, pp. 913–918, 2010. [4] S. R. Messer , J. Agzarian, and D. Abbott, “Optimal w a v elet denoising for phonocardiograms, Micr oelectr onics J ournal , v ol. 32, no. 12, pp. 931–941, 2001. [5] F . Liu, Y . W ang, and Y . W ang, “Research and implementation of heart sound denoising, Physics Pr ocedia , v ol. 25, pp. 777–785, 2012. [6] V . N. V ar ghees and K. Ramachandran, A no v el heart sound acti vity detection frame w ork for automated heart sound analysis, Biomedical Signal Pr ocessing and Contr ol , v ol. 13, pp. 174–188, 2014. [7] “T otal v ariation denoising (an MM algorithm), T echnical Report, Polytechnic Uni v ersity , Mar . 2014. [8] M. A. Figueiredo, J. B. Dias, J. P . Oli v eira, and R. D. No w ak, “On total v ariation denoising: A ne w majorization- minimization algorithm and an e xperimental comparisonwith w a v alet denoising, in 2006 IEEE International Confer ence on Ima g e Pr ocessing . IEEE, 2006, pp. 2633–2636. [9] N. E. Huang, Z. Shen, S. R. Long et al. , “The empirical mode decomposition and the hilbert spectrum for non- linear and non-stationary time series analys is, in Pr oceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences , v ol. 454, no. 1971. The Ro yal Society , 1998, pp. 903–995. [10] G. Rilling, P . Flandrin, and P . Goncalv es, “Detrending and denoising with empirical mode decomposition, in Pr oceedings of the Eur opean signal pr ocessing confer ence (EUSIPCO04) , v ol. 2, 2004, pp. 1581–1584. [11] Y . K opsinis and S. McLaughlin, “De v elopment of EMD-based denoising methods inspired by w a v elet threshold- ing, IEEE T r ansactions on Signal Pr ocessing , v ol. 57, no. 4, pp. 1351–1362, 2009. [12] M. A. Kabi r and C. Shahnaz, “Denoising of ECG signals based on noise reduction algorithms in EMD and w a v elet domains, Biomedical Signal Pr ocessing and Contr ol , v ol. 7, no. 5, pp. 481–489, 2012. [13] W . W u and H. Peng, Application of EMD denoising approach in noisy blind source separation, J ournal of Communications , v ol. 9, no. 6, 2014. [14] Heart Sound & Murmur Libr ary , Uni v ersity of Michig an Health System, Ann Arbor , MI. [Online]. A v ailable: http://www .med.umich.edu/lrc/psb open/repo/primer heartsound/primer heartsound.html [15] M. Misiti, Y . Misiti, G. Oppenheim, and J.-M. Poggi, W avelet T oolbox User’ s Guide , Re v . 4.14 (Rel. 2014b), The MathW orks, Inc., Oct. 2014. [16] A. H. Salman, N. Ahmadi, R. Mengk o, A. Z. R. Langi, and T . L. R. Mengk o, “Performance comparison of denoising methods for heart sound signal, in 2015 International Symposium on Intellig ent Signal Pr ocessing and Communication Systems (ISP A CS) . IEEE, 2015, pp. 435–440. [17] M. V eillette, Alpha-Stable distrib utions , Boston Uni v ersity , Aug. 2015. [18] Z. W u and N. E. Huang, A study of the characteristics of white noise using the empirical mode decomposi- tion method, Pr oceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences , v ol. 460, no. 2046, pp. 1597–1611, 2004. [19] P . Flandrin and P . Goncalv es, “Empirical mode decompositions as data-dri v en w a v elet-lik e e xpansions, Inter - national J ournal of W avelets, Multir esolution and Information Pr ocessing , v ol. 2, no. 04, pp. 477–496, 2004. [20] P . Flandrin, P . Gonc ¸ alv es, and G. Rilling, “EMD equi v alent filter banks, from interpretation to applications, Hilbert-Huang T r ansform and Its Applications , pp. 57–74, 2005. [21] A. H. Salman, N. Ahmadi, R. Mengk o, A. Z. Langi, and T . L. Mengk o, Automatic se gme n t ation and detection of heart sound components S1, S2, S3 and S4, in 4th International Confer ence on Instrumentation, Communi- cations, Information T ec hnolo gy , and Biomedical Engineering (ICICI-BME 2015) . IEEE, 2015, pp. 103–107. IJECE V ol. 6, No. 5, October 2016: 2197 2204 Evaluation Warning : The document was created with Spire.PDF for Python.