Application of the Empirical Mode Decomposition in the study of murmurs from Arteriovenous Fistula Stenosis

Pablo V squez1,2 , Marco Mungua M.1,2, , Members IEEE, Elisabeth Mattsson3 , and Bengt Mandersson2 a
Abstract The Empirical Mode Decomposition (EMD) is a method to decompose non linear, non stationary time series into a sum of different modes, named Intrinsical Mode Functions each one having a characteristic frequency. In the present work we used the EMD to investigate the properties of the recorded sounds from the Arteriovenous stula on hemodialysis patients. Phonoangiographic signals coming from two different vessel conditions, stenotic and non-stenotic, were analyzed by using EMD, the mean energy and mean instantaneous frequency per IMF proved to be good features for classication. Three types of classication schemes were tested on data from the rst IMf features achieving good results.

1. INTRODUCTION Patients suffering from end-stage renal disease are frequently treated using hemodialysis. During the treatment the blood is cleaned from waste and excess uid through a dialyzer. To reach the blood, a vascular access, usually placed on one of the patients forearms is used to insert the needles coming from the dialyzer. A very common type of vascular access is the so-called arteriovenous (AV) stula . The stula is made through a surgical operation during which an artery and a vein in the arm are connected together. The connection point is referred to as an anastomosis and is often located near the patients wrist or elbow. An AV stula is essentially a shortcut between the high pressure arterial side and the low pressure venous side that is aimed to achieve sufcient blood ow in the vein for cannulation. In well-developed stulas, ow rates may ultimately reach values of 600 to 1200ml/min. Problems may arise in the course of time however,the most frequent complication in AV stulas is that of stenosis. A stenosis is an abnormal narrowing of a bodily canal which can be caused by calcication or when the vessels wall is exposed to abnormal physical stress like turbulence or high blood pressure. When the dialysis becomes inadequate as a result of too low blood ow caused by the lumen narrowing the stula must be revised and remedied. An early detection of stenosis is desirable since it permits their correction prior to total occlusion and thereby prolongs the life of the stula. It has been reported that turbulent blood ow related to stenosis of vessels create audible sounds due to the vibrations
This work was supported by the Swedish International Development Cooperation Agency 1 UNI-Asdi-FEC Group, Faculty of Electrical Engineering and Computer Science, National University of Engineering, Managua, Nicaragua 2 Signal Processing Group, Department of Electrical and Information Technology, Lund University, Lund, Sweden {mmm, pvo,

on the surrounding structures, that can be analyzed to provide information about the severity of the blockage. Clinical practice guidelines [1] for auscultation suggest that physical examination of the stula (monitoring) should be performed weekly and should include inspection and palpation for pulse and thrill at different sites of the stula. The clinician can also use a stethoscope to listen for sounds originating from the stula, called bruits. The analysis of the sounds from vessels using signal processing techniques (phonoangiography) constitutes a non invasive, low cost technique that can be used with success for monitoring the vessels functioning. The technique may be used as a rst step diagnosis tool to assist clinician in the auscultation process and to decrease evaluation subjectivity of the stula state. It can also be a cost-effective technique as the rate of patient referral to a more sophisticate (and more expensive) diagnosis procedures like Doppler Ultrasound may be decreased. Previous research results have shown that stenosis has two basic acoustical effects: a general increase in the sound level and an introduction of new high frequency components in the power spectra. The changes in frequency are dependent on the distance from the stenosis and its severity. Empirical Mode Decomposition (EMD) is a technique used to decompose a signal into a set of individual characteristic oscillations called Intrinsical Mode Functions (IMF) [2]. This decomposition method is adaptive and local which make it appropriate to analyze non-linear and non-stationary signals such as the biological signals. Since its invention EMD when it was initially applied to the study of ocean waves it has found a broad range of applications, the rst application of this method to biological signals was in the study of blood pressure in rats [3]. Recently, it has been used to study also EMG [4] , EEG [5], and ECG [6] signals. Here EMD was applied to the recorded sounds from AV stulas of a group of patients an a subset of the resulting components were analyzed to extract the features used for classication. 2. METHODOLOGY For this study we used recordings taken form 5 patients labeled A to E which are currently undergoing hemodialysis treatment at the Department of Clinical Physiology, Lund University Hospital. The recordings were made with a BIOPAC TM , BP-150 system connected to a notebook PC. Input lter settings were 0.05Hz for low pass ltering and 5kHz for high pass ltering. A sample frequency of 10000 was selected in all measurements. A microphone attached to a stethoscope head was used as a measuring device.

bm}@eit.lth.se
3 Dialysis Unit, Department of Nephrology, Skane University Hospital, Lund, Sweden Elisabeth.Mattsson@skane.se Corresponding author

TABLE I VASCULAR ACCESS TYPE , CONDITION AND DURATION OF THE SOUND

RECORDINGS PER PATIENT

Subject

Type

graft

stula

C D E

stula stula stula

Signal A1 A2 A3 A4 A5 A6 B1 B2 B3 B4 C1 C2 D1 D2 E1 E2

Condition bf. angioplasty bf. angioplasty bf. angioplasty af. angioplasty af. angioplasty af. angioplasty bf. angioplasty bf. angioplasty af. angioplasty af. angioplasty bf. angioplasty aft. angioplasty no stenosis no stenosis no stenosis no stenosis

Duration 60 s 60 s 60 s 60 s 60 s 60 s 60 s 60 s 60 s 60 s 20 s 20 s 60 s 60 s 60 s 60 s

Fig. 1. EMD of a phonoangiographic signal. Only the rst 4 IMfs are shown. The rst row represent the signal.

Note: The patient C was selected from the database available in [10].

C. Feature Extraction During the recording sessions, patients rested their arms on a cushion an the stethoscope head was gently placed on the skin. For our analysis several sites were selected, common sites were at the anastosmosis, some centimeters downstream anastosmosis and if stenotic lesions were present additional points downstream stenosis were also considered. A. Data A summary of the patients involved in the study and the recorded signals is shown in table I. Patients A, B and C from the group were diagnosed with venous stenosis and referred to a surgery procedure for alleviating the problem. These patients offered a unique possibility for studying changes in the sound produced by the blood ow in the stula before and after the correcting procedure. Patients labeled D and E have their stulas working properly and were included as reference patients. B. Pre-processing A high pass lter of 40Hz was applied to the signals under study for base line wander and other artifacts removal. Segmentation of the signals into periods was achieved using the following procedure: homomorphic ltering was rst applied to calculate the signal envelope, then we performed peak detection on a smoothed version of the envelope (low pass ltered to 5Hz), these peaks were later used to divide the signal into segments. The segments were dened as the signal portion comprising 0.1 seconds before and 0.4 seconds after one peak. The reason for that selection is because in that portion of the signal (systolic) is where the high frequency content is concentrated. Incomplete of bad formed segments were not taken into account for calculations. Additionally signals were downsampled to 2.5kHz and normalized in amplitude for fair comparison since signal had varying degrees of amplitude. The algorithm of Empirical Mode Decomposition, EMD proposed by Huang [2] decomposes non-linear and not stationary signals in a sum of components which have frequencies and amplitudes that vary with time, called Intrinsic Mode Functions IMF. The major advantage of the EMD is that the IMFs are derived directly from the signal itself and does not require any a priori known basis. Hence the analysis is adaptive, in contrast to Fourier or Wavelet analysis, where a signal is decomposed in a linear combination of predened basis functions. ,gure 1 depicts the decomposition of a phonoangiographic signal. After the decomposition the original signal s(t) can be expressed as,
n

s(t) =
j=1

cj (t) + rn (t)

(1)

Where n is the number of components obtained after the sifting process, when a predetermined stoping criterion is satised. The algorithm used here for the EMD is the proposed in [7]. In equation (1), c1 (t), c2 (t), ..., cn (t) are the IMFs and rn (t) is termed the residue and can be regarded as a trend or a constant. These components represent the oscillatory modes embedded in the signal. The lower order IMFs capture fast oscillation modes of the signal, while the higher order IMFs capture the slow oscillation modes. We can apply the Hilbert Transform to each of the cj (t) to form an analytical signal zj (t) dened as z(t) = cj (t) + j (t) where j (t) is the Hilbert Transform of cj (t). The result of desribed process allow us express the decomposition in equation (1) in such a way that the timed dependency of amplitude and frequency are more evident.
n

s(t) =
j=1

Aj (t) cos (t) + rn (t)

(2)

with,

Fig. 3.

Selected features on a energy frequency plane.

modes. For this purpose we analyzed IMFs features like mean frequency fj , calculated from the marginal Hilbert spectrum, and energy content of the signal Ej . Expressions for the calculation on both features are as follows,
+ f h()df + h()df N cj (n) 2 n=1

fj Ej

(6)

(7)

Fig. 2. Hilbert Energy Spectra of recordings (3 periods) from patient B, taken before angioplasty (upper) and after angioplasty (lower).The spectra have been smoothed by a 7x7 gaussian lter

During the feature extraction step, every single period xi obtained from the signals in table I after the segmentation process, was normalized in energy (equation 8) and decomposed by EMD. Mean frequencies and energy per component were calculated and stored as relevant features for the classication step. xiN = xi
L

Aj (t) j (t)

= =

cj (t)2 + vj (t)2 arctan( vj (t) ) cj (t)

(3) (4)

, x2 i

(8)

i=1

The instantaneous frequency (t) is calculated as the time derivative of the instantaneous phase angle (t). The instantaneous amplitudes and frequencies are the inputs for the calculation of the Hilbert Spectrum H(, t) which is a high resolution time-frequency representation of the signal, and has been used in previous research for analysis of cardiac murmurs [8], [9]. The Hilbert Spectra of the phonoangiographic sounds recorded from patient B before and after the angioplasty procedure are shown in gure 2. With the Hilbert spectrum dened, the marginal spectrum h() can also be dened as,
T

Decomposition of these signals produced up to 12 components. After several initial tests it was concluded that IMFs beyond the second do not contributed to stenosis detection, so they were discarded. Only the IMF1 of all signals were used in further analysis. D. Classication For classication purposes we trained 3 classiers: An articial neural network (ANN) with one hidden layer of 5 units, a Gaussian Mixture Classier with 4 components, which are parametric representations of a probability density function (PDF) and a Parzen classier. We performed a 5fold and 12-fold cross-validation for the sound classication. In a N-fold cross-validation, data from the 5 patients are randomly divided into N data sub-group. Each sub-group is tested against data from the remaining sub-groups. The tests results are then averaged over the N validation outcomes.

h() =
0

H(, t)dt

(5)

It is expected that the additional high frequency components present in the sound recorded from turbulent blood ow will be reected in one or several of the signal oscillation

TABLE II C LASSIFIER P ERFORMANCE 5-Fold Error 0.12345 0.00960 0.19447 0.00687 0.15481 0.00908 12-Fold Error 0.12322 0.00784 0.19399 0.00799 0.15394 0.00941

Classier BP-ANN Mix.Gauss Parzen

tools when analyzing non stationary and non linear signals. The Hilbert Spectrum allow us to visually determine the high frequency content of the sounds produced by stenosis. In this paper, as a possible way to automate the diagnosis we proposed to use the features extracted from the rst IMF. Three different schemes were used for classications, the Neural Network classier obtained the best performance in this case. It is possible that the other classiers based on probability distribution need more data to improve their results. 5. ACKNOWLEDGEMENTS The authors gratefully acknowledge the patients and the staff of the Dialysis Unit, Department of Nephrology at Skane University Hospital for being very collaborative during the data acquisition. R EFERENCES
[1] G. Beathard, A practitioners Guide to Physical Examination of Dialysis Vascular Access, Fistula First Project, 2004. [2] N. E Huang,et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc Royal Society A, 454:903-995, 1998. [3] E.P. Souza Neto, P. Abry, P. Loiseau, J.C. Cejka, M.A. Custaud, J. Frutoso, C. Gharib, P. Flandrin, 2007: Empirical Mode Decomposition to Assess Cardiovascular Autonomic Control in Rats, Fundamental & Clinical Pharmacology, Vol. 21, pp. 481-496. [4] O. Andrade, A. Nasuto, P. Kyberd, C. M. Sweeney- R, F. R. V. Kanijn. EMG signal ltering based on Empirical mode descomposition. Biomed. Signal Proc. Control. 1, pp. 44-45, 2006. [5] Lorena Orosco, Eric Laciar An epileptic seizures detection algorithm based on Empirical Mode Decomposition of EEG. Proceedings of the 31st Annual International Conference of the IEEE EMBS 2009 [6] B. Weng, M. Blanco-Velasco, and K. E. Barner, ECG Denoising Based on the Empirical Mode Decomposition, 28th Ann. Conf. IEEEEMBS, 2006 [7] Rilling et al, On Empirical Mode Decomposition and Its Algorithms,IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing,2003. [8] Zhi-Dong Zhao, Yang Wang, Analysis of diastolic murmurs for coronary artery disease based on Hilbert Huang Transform. Proceedings of the6th International Conference on Machine Learning and Cybernetics, August 2007 [9] L.Sun, M.Shen, F.H.Y.Chan, P.J.Beadle, Instantaneous Frequency Estimate of Non Stationary phonocardiograph signals using Hilbert Spectrum.Proceedings of the 27th Annual International Conference of the IEEE EMBS 2005 [10] C. Nilsson, A. Larsson, A Pilot Study of the Acoustical Properties of the Arteriovenous Fistula using Digital Signal Processing Department of Electroscience, Lund University, 2005.

Fig. 4.

ROC curves for the tested classiers

3. RESULTS The trained classiers were applied to the patients data. Signals were classied in 2 classes: stenotic and non-stenotic. Since we trained the classier on periods rather than whole signals, we applied an empirical threshold; if at least the 75% of the periods were classied into one of the categories then the signal is classied according to that category, this was done to take into account temporary variations in some portions of the signal introduced by environmental noise, hand movements, variations in the applied pressure of the stethoscope head on the skin, etc. Under that criteria we obtained the classication results are summarized in table III. A percentage of correct classication obtained according 2 to the presented data is 16 85%. Interestingly recordings from patient A after the angioplasty were labeled as stenotic and were wrongly classied by the algorithm however we were informed by the hospital that the procedure had not been successful what was in accordance with our previous results. 4. DISCUSSION Murmurs produced by AV stulas can provide valuable clinical information on the functional state of the vascular access.The Hilbert Huang Transform and EMD are valuable
TABLE III C LASSIFICATION S UMMARY Signals 6 10 16 Real Condition Stenosis No-Stenosis Correctly Classied 6 8 14 Misclassied 0 2 2