See discussions, stats, and author profiles for this publication at: [Link]

net/publication/14935832

Noninvasive acoustical detection of coronary artery disease: A comparative

study of signal processing methods

Article in IEEE transactions on bio-medical engineering · July 1993

DOI: 10.1109/10.237677 · Source: PubMed

CITATIONS READS

130 753

5 authors, including:

Yasemin M Akay Johb L Semmlow

University of Houston 265 PUBLICATIONS 4,965 CITATIONS
244 PUBLICATIONS 5,978 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Johb L Semmlow on 14 January 2017.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 40, NO.6, JUNE 1993 571

Noninvasive Acoustical Detection of

Coronary Artery Disease: A Comparative
Study of Signal Processing Methods
Yasemin M. Akay, Student Member, IEEE, Metin Akay, Senior Member, IEEE, Walter
Welkowitz, Fellow, IEEE, John L. Semmlow, Senior Member, IEEE, and John B. Kostis

Abstract- Previous studies have indicated heart sounds may stenosis was increased [21], [23], [24], [17]. However, for
contain information useful in the detection of occluded coro- severe obstructions (above 95\% occlusion) sounds may not be
nary arteries [1]-[12]. During diastole, coronary blood flow is
produced due to low blood flow. At the lower end, occlusions
maximum, and the sounds associated with turbulent blood flow
through partially occluded coronary arteries should be detectable producing as little as 25\% narrowing have been detected
[1]-[12]. In order to detect such sounds, recordings of diastolic and used as a nonvasive measure to assess arterial narrowing
heart sound segments were analyzed by using four signal process- in vessels of the neck, thorax, and abdomen [23,] .[24].
ing techniques; the Fast Fourier Transform (FFT), the Autore- The auditory component associated with coronary stenosis is
gressive (AR), the Autoregressive Moving Average (ARMA), and similar to that found in partially occluded carotid arteries, but
the Minimum-Norm (Eigenvector) methods. To further enhance
the diastolic heart sounds and reduce background noise, an is much attenuated by the intervening tissue [1]. It is also
Adaptive filter was used as a preprocessor [12]. masked by the comparatively loud valve sounds. These can be
The power ratios of the FFT method and the poles .of the eliminated by isolating diastolic portions of the acoustic signal
AR, ARMA, and Eigenvector methods were used to diagnose using a time window synchronized with the cardiac cycle [1].
patients as diseased or normal arteries using a blind protocol
The principal objective of this study is to compare different
without prior knowledge of the actual disease states of the patients
to guard against human bias. Results showed that normal and signal processing techniques which can be used to identify
abnormal records were correctly distinguished in 56 of 80 cases the additional signal components found in heart sounds of
using the Fast Fourier Transform (FFT), in 63 of 80 cases using patients with coronary artery disease. These added components
the AR, in 62 of 80 cases using the ARMA method, and in 67 of form the basis of our approach to noninvasive detection of
80 cases using the Eigenvector method. Among all four methods,
coronary artery disease. In earlier studies, the average power
the Eigenvector methods showed the best diagnostic performance
when compared with the FFT, AR, and ARMA methods. These spectrum of diastolic heart sounds was estimated by using
results confirm that high frequency acoustic energy between 300 traditional FFT methods [1]. Although some success was
and 800 Hz is associated with coronary stenosis [2]-[12]. achieved in distinguishing normal from abnormal patients, this
signal processing technique was sensitive to the effects of noise
I. INTRODUCTION which masked detection of the desired diastolic signal [1].
Since the application of parametric modeling methods to
signal identification problems can result in a better estimation
PREVIOUS
duce sounds studies
due to showed that coronary
the turbulent stenoses
blood flow pro-
in partially of spectral features, particularly for low signal-to-noise ratios
occluded arteries [1]-[12]. If these signals could be reliably (SNR)[37]-[ 45], such model-based methods were employed
detected, they would provide a simple, noninvasive approach to analyze the recordings of diastolic heart sounds and to
to the detection of coronary artery disease. detect features associated with coronary stenosis [2]-[10].
Studies involving turbulent blood flow have been carried From the many model-based methods, the adaptive AR method
out in many components of the cardiovascular system, and it was chosen to represent the diastolic signal source since
has been widely reported that turbulence produced by stenoses it does not require prior knowledge of the signal charac-
produces sounds due to the vibration of the surrounding teristics and it can track changes in signal characteristics
structures [15]-[17], [21], [23]-[25], [35]. These sounds have [33], [22]. Initial analyses were carried out on 15 patients
been detected and analyzed, and results generally showed (10 abnormal, 5 normal patients) using two fast adaptive
that the high frequency energy increased when the degree of AR methods, the Recursive Least Square Lattice (RLSL)
and the Gradient Adaptive Lattice (GAL) method [2], [4],
Manuscript received November 16, 1992; revised January 26, 1993. This [5], [9]. Spectra obtained from individual diastolic cycles
work was supported by a grant from Colin Medical Instrument Corporation, showed some variation. To obtain representative frequency
Japan.
Y. M. Akay, M. Akay, W. Welkowitz, and J. L. Semmlow are with information, averages were calculated from an ensemble of
the Biomedical Engineering Department, Rutgers University, Piscataway, NJ 20-30 individual diastolic spectra. Results showed encourag-
08855.
ing spectral differences between normal and diseased patients
J. Kostis is with the Department of Surgery (Bioengineering), Department
of Medicine, UMDNJ-Robert Wood Johnson MediCal School, NJ 08854. indicating a resonance process in diseased patients [2], [4],
IEEE Log Number 9208529. [5], [9]. The poles of the AR model were also calculated and
0018-9294/93$03.00 © 1993 IEEE
572 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 40. NO.6. JUNE 1993

used as a discriminant criterion. In most cases, the second and We selected the delay d = 5 and for ALE [5], [12] based
usually the third poles were fmther from the unit circle in on the self correlation lengths of the diastolic heart sounds
normal patients than in diseased patients [2], [4], [5], [9]. This components [12], [33]. The ALE filter order was taken as
indicates that the diastolic heart sounds of diseased patients m = 10 based on our previous findings [7], [12].
contain more signal within a limited spectral range which is
indicative of resonance processes. Indeed, a theoretical model
B. Autoregressive (AR) Method:
has indicated that coronary stenosis produces well defined
acoustic resonances [11]. The AR model is the most widely used modeling method to
A number of other studies including several of patients be- estimate the power spectral density (PSD) function associated
fore and after angioplasty have shown that parametric model- with some biological signals. The AR model is called the
ing methods can be used to detect sounds associated with coro- all-pole method [32], [28]. Each sample of a signal can
nary stenosis [2]-[12]. The purpose of this paper is to compare be expressed as a linear combination of previous samples
the effectiveness of the FFT, AR, ARMA, and Minimum- and an error signal e(n). The error signal is assumed to be
Norm (Eigenvector) methods in detecting acoustic signals independent of the previous saIllples [2], [5], [7], [9], [10],
associated with occluded coronary arteries. Recently, we uti- [12], [32].
lized adaptive filters to reduce background noise in the input m
signal [12]. The FFT, MODIFIED YULE-WALKER (MYW)
AR, ARMA, and Minimum-Norm (Eigenvector) methods were y(n) =- L
p=l
apy(n - p) + e(n) (6)

applied to the filtered signal. The MYW AR and ARMA

methods were chosen since they show better performance than
the maximum-likelihood realizations of the AR and ARMA where y( n) represents the signal to be modeled, ap represents
the AR coefficients of the AR process at the pth stage, e( n)
methods [26]-[28], [19],[20] when the poles of the AR and
represents the estimated error signal, and m represents the AR
ARMA models are very close to the unit circle. The Minimum-
model order.
Norm method was chosen because of its higher resolution,
The estimation of the PSD of the AR method was carried
particularly in the presence of noise [29], [33]. The analyses
out by using the MYW AR method [28], [22], [33].
were carried out on 80 recordings (32 normal and 48 abnormal)
from 73 normal or diseased patients using a blind protocol. A
blind protocol was followed so that the actual disease states C. Autoregressive Moving Average (ARMA) Method:
of the patients were not disclosed until the acoustical results Some biomedical signals can be modeled using the ARMA
were completed. method when the signal is corrupted by heavy observation
noise. The output sequence y( n) can be modeled by assuming
II. METHOD
an input driving sequence v( n) as follows [8], [10], [12], [22]:

A. Adaptive Line Enhancer (ALE) Method: m q

The Adaptive Line Enhancer (ALE) [13], [18] method y(n) =- L

p=l
apy(n - p) + L
p=O
bpv(n - p) (7)
which is a special case of adaptive noise canceling method has
been widely applied in many fields including biomedical signal
processing [12], [14], [31], [33]. A particular advantage of this where ap represents the AR coefficients of the AR process at
method is that it does not require a reference (noise) signal the pth stage, bp represents the MA coefficients of the MA
to extract the desired biomedical signal from the background process at the pth stage, m represents the AR model order,
noise. In this study, we utilized the least mean square (LMS) and q represents the MA model order.
implementation of the line enhancement method [12]. The The PSD of the ARMA model can be calculated accurately
computational algorithm of the ALE can be summarized as and efficiently using the MYW ARMA method [7]-[10], [12],
follows: [28], [19], [20]. Previous studies showed that for a filter order,
m (the AR), greater than 10 the predicted error power as
x(n) = y(n) + v(n) (1) a function of filter order was relatively stable (see fig. l(a)
r(n) =x(n-d) =y(n-d)+v(n-d) (2) in [5]). Based on the data of Fig. 1 (a), (b) in [5] and our
m
initial empirical findings, filter orders between 10 and 15 are
y(n) = L
p=O
hp(n)r(n - p) (3) sufficient to represent the signal records. Filter orders m = 10
(the AR order) (see Fig. 1 in [5]), m = 10 and q = 3
e(n) = x(n) - y(n) (4) (the ARMA orders), were chosen as adequate to produce an
estimate of the white noise process after filtering y (n) with
hp(n + 1) = hp(n) + 2JLe(n)x(n - p - d) (5)
the estimated inverse transfer function of the ARMA method
where x (n) represents the primary signal, y (n) represents [7], [8], [22], [32], [33]. For the initial estimation, one chooses
the desired diastolic heart sound signal, r( n) represents the a(O) = b(O) = 1. The AR filter poles are distributed between
reference signal, v( n) represents the primary noise signal, five complex conjugate pole pairs. The ARMA filter poles and
hp (n) represents the adaptive filter weights, m represents the zeros are distributed between five complex conjugate poles and
adaptive filter order, and d represents the delay. between one complex conjugate zero pair and one real zero.
AKAY, et al.: NONINVASIVE ACOUSTICAL DETECTION OF CORONARY ARTERY DISEASE 573

The autocorrelation function (ACF) order was chosen as 15.

The number of signal components were between 3 and 5 for
normal and abnormal patients, respectively.
Details of the FFT, MYW AR, ARMA, and Minimum-
Norm (Eigenvector) methods have been described elsewhere
[1]-[10], [12].

E. Patient Analysis:
Patients were selected from those undergoing catheteriza-
tion and/or angioplasty at the Cardiodynamics Laboratory of
Robert Wood Johnson University Hospital. Diastolic heart
sounds were recorded from the 4th intercostal space on the
chest of patients using a specially designed high sensitivity
accelerometer [34]. These sounds were recorded while the
patients held their breath and were supine. The sounds were
normalized by the system gain.
Fig. 1. Averaged PSD functions' obtained from the FFT method, over 10 The objective of this study was to investigate the diagnostic
periods (1024 samples each), applied to the isolated diastolic heart sounds of ability of diastolic heart sounds to detect coronary artery
Normal Patient 4 and Abnormal Patient 1 after adaptive line enhancement. disease noninvasively in a typical patient environment. For
Solid line: Abnormal. Dashed line: Normal. Signals were windowed using the
Hamming window. each patient, 10 cardiac cycles were digitized (fs, sampling
frequency = 4 kHz) and 1 024 samples were analyzed for each
period. An average spectrum over 10 cardiac cycles was con-
D. Minimum-Norm (Eigenvector Method): structed to obtain representative frequency information since
The Minimum-Norm method produce unbiased spectral spectra obtained from each diastolic cycles showed some slight
estimates when the signal-to-noise ratio (SNR) is low. This variation. The spectra were obtained using the techniques
important property is not shared by the linear prediction, described above. Before the analysis, the dc component from
maximum-likelihood (minimum energy), and Bartlett methods each isolated diastolic heart sounds record was eliminated.
due to limits on resolution [4], [6], [7], [9], [10], [22], [33]. The These signals were multiplied by a Hamming window. As
Minimum-Norm method has been applied to several important detailed elsewhere [1]-[12], the diastolic heart sounds passed
problems such as direction finding, echo resolution [3], [6], through an anti-aliasing low-pass analog filter with a cutoff
[7], [9], [10], [12]. Initial studies showed that the diastolic frequency of 1 200 Hz and a highpass digital filter with a
components associated with occlusive coronary stenosis con- cutoff frequency of 180 Hz.
tain narrow band components. This was confirmed by the For this study, 80 (48 abnormal and 32 normal) recordings
other studies [11]. In this case, the diastolic heart sound signal of isolated heart sound complexes obtained from 73 patients
y(n), may be represented as a number of sinusoids along with were analyzed using the FFT, AR, ARMA, and Eigenvector
background noise [3], [6], [7], [9], [10], [12]: methods. Patients with occlusions of less than 30% were con-
sidered to be normal. All of these patients, except for 7 pseudo-
L normals (young, symptomless subjects, assumed to be normal)

y(n) = L
i=l
Ciejwn + v(n) (8) were proven by cardiac catheterization as normal or diseased.
The diagnostic effectiveness of the decision parameters
(the power ratios of the FFT and the second poles of the
AR, ARMA, and Eigenvector methods) was estimated by
where Ci is the amplitude of components in the diastolic heart constructing receiver operating curves (ROC) of sensitivity
sounds, Wi is the frequency of components in this signal, L is versus specificity for these parameters. Sensitivity was defined
the number of components, and v( n) is the noise signal. as the ratio of correctly diagnosed diseased patients over
Eigenvector methods can then be used to extract the signal the total number of the diseased patients in the data base.
buried in noise. In theory, these Eigenvector methods have in- Specificity was defined as the ratio of the correctly diagnosed
finite resolution and provide an accurate frequency estimation normal patients over the total normal patients in the data base.
regardless of the SNR [7], [22], [33].
In order to calculate the Minimum-Norm spectra, the di-
III. RESULTS AND DISCUSSION
mension of the noise subspace K and the number of signal
components were determined by a hypothesis testing proce- Tables I and II show a description of the stenoses in terms
dure based on the confidence interval of the noise Eigenvalue of area reduction of the coronary arteries for the patients used
[30], [48], in which a threshold value was assigned in a blind in this study.
fashion to guard against human bias since the performance Fig. 1 shows the PSD function obtained from the FFT
of the Akaike information and minimum description length method applied to the isolated heart sounds (10 periods) of
criteria yield an inconsistent estimate and tend to over- or Normal Patient Number 4 and Abnormal Patient Number 1.
underestimate the number of signal components [12], [22]. The most apparent spectral feature associated with the
574 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 40, NO.6, JUNE 1993

TABLE I TABLE IT
NORMAL/ABNORMAL PATIENTS DATA BASE NORMAL/ABNORMAL PATIENTS DATA BASE

# Patient ID Condition as % # of occlusion # Patient ID ___ C_ondition # of Occlusion

01 2703* LAD 55, CFX 20,40 3 41 14701 NORMAL
02 4803*- LAD 25,80, RCA 30,35,80, 8 42 6303 NORMAL
CFX 20,20,50 43 6403 RCA 90
03 4903* NORMAL 44 7201 NORMAL
04 6201 NORMAL 45 7301* NORMAL
05 3301+- LAD 10,50 2 46 7501 NORMAL
06 6301+-x LAD 60,90 2 47 7601*+- NORMAL
07 7903 LAD 35,90, RCA 100, CFX 5 48 7701*+- LAD 50 1
80,85 49 7901x LAD 35,90, RCA 100, CFX 5
08 10501 LAD 60,90,75, CFX 100 4 80,85
09 10301 LAD 75, RCA 68, CFX 99 3 50 4703+- NORMAL
10 10201 LAD 40,40, RCA 90, CFX 4 51 7503* NORMAL
45 52 8001-x NORMAL
11 10901 * LAD 90,RCA 90 2 53 14003+- NORMAL
12 11201 * NORMAL 54 14901 NORMAL
13 12001 NORMAL 55 15101 NORMAL
14 12101 NORMAL 56 5203+- CFX 20,95 2
15 12301 NORMAL 57 5703* LAD 25, CFX 100 2
16 12501 NORMAL 58 5903 LAD 25,50,RCA 25 3
17 10401 * NORMAL 59 3901 NORMAL
18 10201 *x LAD 40,40, RCA 90, CFX 4 60 4701 * NORMAL
45 61 8501 NORMAL
19 11101 *+x LAD 99 -1 62 8601 NORMAL
20 11303x NORMAL 63 3303x 3
LAD 40,40,CFX 85
21 11601*- LAD 70, RCA 100 2 64 3701* NORMAL
22 11501*+- LAD 45, RCA 45,68, CFX 4 65 7601+ NORMAL
48 3
66 11001+ LAD 40, RAC 40, CFX 40
23 12601 CFX 70,90 2
67 11303- RCA 90, CFX 30 2
24 12603* NORMAL 1
68 11403+- RCA 90
25 12701 LAD 30,40, RCA 100,30, 5
69 11503+- LAD 45, RCA 45,20, CFX 4
CFX30
48
26 12801 RCA 50 1
70 11703 CFX 60, RCA 100, CFX 100 3
27 12901*x LAD 50,95 2
71 12201 NORMAL
28 13101 LAD 75,RCA 40,65 3
72 13301 LAD 25,80,50, RCA 50, 5
29 13302*+- LAD 70,9l}, RCA 100 3 CFX 100
30 13303+-x LAD 100, RCA 100 2 73 13201 4
LAD 99,100, CFX 65, RCA
31 13901 LAD 70 1 35
32 13603 LAD 50,30 2 74 13203* NORMAL
33 13701 LAD 25, CFX 35,80 3 75 13401 CFX 60 1
34 13503x LAD 50 1 76 12703+-x RCA 100, CFX 30 2
35 14001 CFX 90 1 77 13901 LAD 90 1
36 14801*x NORMAL 78 14103 R CA 40,50 2
37 14201 LAD 90,100, RCA 80 3 79 15203* NORMAL
38 14301 LAD 95, RCA 40, CFX 4 80 12703+-x RCA 100, CFX 30 2
65,65 LAD is the left anterior descending artery. RCA is the right coronary
39 14501 CFX 70,90 2 artery. CFX is the circumflex artery. *Incorrectly Diagnosed Patients using
40 14601* LAD 80, CFX 80 2 the FFT method. +Incorrectly Diagnosed Patients using the AR method.
LAD is the left anterior descending artery. RCA is the right coronary -Incorrectly Diagnosed Patients using the ARMA method. xlncorrectly
artery. CFX is the circumflex artery. *Incorrectly Diagnosed Patients using Diagnosed Patients using the Eigenvector method.
the FFT method. +Incorrectly Diagnosed Patients using the AR method.
- Incorrectly Diagnosed Patients using the ARMA method. xlncorrectly
Diagnosed Patients using the Eigenvector method.
6 and 7 show the PSD functions and the poles obtained from
the Eigenvector methods applied to the isolated heart sounds
coronary artery diseased patients is the higher energy below (10 periods) of Abnormal Patient Number I as well as Normal
500 Hz when compared to normal patients after adaptive Patient Number 4, respectively. The most apparent spectral
filtering of the diastolic heart sounds using the FFT method. feature associated with' the coronary artery diseased patients
Figs. 2 to 5 show the PSD functions and the poles obtained is the higher energy between 300 and 500 Hz when compared
from the AR, and ARMA methods applied to the isolated heart to normal patients after adaptive filtering of the diastolic
sounds (10 periods) of Abnormal Patient Number I as well as heart sounds using the Eigenvector methods. Please note that
Normal Patient Number 4, respectively. The most apparent the spectral difference between normal and diseased patients
spectral feature associated with the coronary artery diseased seems to be a function of the signal processing techniques
patients is the higher energy between 400 and 800 Hz when since the statistical characterization and asymptotic behavior
compared to normal patients after adaptive filtering of the of these signal processing techniques are different (especially
diastolic heart sounds using the AR and ARMA methods. Figs. for finite data records) [22], [28], [32], [33].
AKAY, et al.: NONINVASIVE ACOUSTICAL DETECTION OF CORONARY ARTERY DISEASE 575

NORMAL '. "

25
10
15 ,, " 30 AR SPECTRUM ---- bd T
U
I
EA L
35
H
D
P ARKA SPECTRUM
20
10
25
16
20

ABNORMAL
NORMAL

o
4l2ll2l Bl2ll2l 12l2ll2l 16l2ll2l 2l2ll2ll2l o
Bl2ll2l 12l2ll2l 16l2ll2l 2l2ll2ll2l
o FRE:OUE:NCY (Hz) o FRE:OUE:NCY (Hz)

Fig. 2. Averaged PSD functions obtained from the AR model, over 10 Fig. 4. Average PSD functions obtained from the ARMA model, over 10
periods (I 024 samples each) applied to the isolated diastolic heart sounds periods (1024 samples each), applied to the isolated diastolic heart sounds of
of Normal Patient 4 and Abnormal Patient 1 after adaptive line enhancement. Normal Patient 4 and Abnormal Patient 1 after adaptive line enhancement.
Solid line: Abnormal. Dashed line: Normal. Solid line: Abnormal. Dashed line: Normal.

AR ARMA

1r/2 m ABNORMAL m ABNORMAL

• NORMAL • NORMAL
•
• m

o a
Fig. 3. Five complex conjugate pole pairs obtained from the AR model, over Fig. 5. Five complex conjugate pole pairs obtained from the ARMA model
10 periods (I 024 samples each) patients in Fig. 2 are shown in the upper half over 10 periods (1 024 samples each) applied to the same patients in Fig. 4
of the complex plane (phase in 1r radians with respect to /s/2. = 2kHz and are shown in the upper half of the complex plane (phase in 7r radians with
complex conjugates not shown). respect to /s /2. = 2 kHz and complex conjugates not shown).

In all patients, large frequency peaks were found at either disease as compared to normal patients, as shown in Figs.
end of the observed spectra. The source of the low frequency 3, 5, and 7. This indicates that the diastolic heart sounds of
(around 200 Hz) sounds may be due to flow fluctuations, diseased patients contain more energy between 300 and 500
aortic pumping, or ventricle filling which are shaped by the Hz than that of normal patients.
highpass filter effects of the data acquisition system. At the As noted previously, the parameter z2 which is the ampli-
high frequency end of the spectra the peak (around 1000 Hz) tude of the second pole of the AR, ARMA, and Eigenvector
is caused by the transducer resonance. The specially designed methods was most significant in differentiating normal from
transducer used in this study has a relatively low frequency diseased subjects. For this reason, we utilized the second poles
resonance (1000 Hz) [34]. The low and high frequency peaks of these methods to diagnose patients as normal and diseased
gave rise to complex conjugate pole pairs falling close to the for the AR, ARMA, and Eigenvector methods. For the FFT
unit circle. These were not considered relevant as decision method, the power ratio (r), the ratio of the power below
criteria since they were produced by system components 500 Hz to the whole observed power, was used to diagnose
and by flow fluct~ations and are essentially the same in patients as normal and diseased. All these findings are in
all patients. Previous angioplasty studies showed that the agreement with theoretical studies [11] which have shown
amplitudes of these peaks were almost the same before that coronary occlusions produce resonances with frequencies
and after angioplastic surgery. An alternate decision criterion ranging between 200 and 800 Hz.
utilized an estimation of the poles obtained with the AR, Before differentiating between diseased and normal patients,
ARMA, and Eigenvector methods for the normal/abnormal a t -test (to test means) [31] was employed to determine
patient groups. The second complex conjugate pole pairs fell whether the decision parameters have the same mean and
closer to the unit circle for patients with coronary artery whether the differences in the means were statistically signif-
576 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 40, NO.6, JUNE 1993

- 25
•10 II •_a
•. •• ae.
I
Vl
D -II.. 20
I
-0
II-
80 MINIMUM-NORM ••SPECTRUM 100
NORMAL a.
::=- '0
~
10
30

10 501
~j 20
60J

EI FIT
• AR
a ARMA
EI3ENVECTOO

o
400 800 1200 1600 2000 ol-~-~-~-~-~-~-~------~
100 90 80 70 GO 50 '-0 30 20 10
o FREOUENCY (Hz)
Sensilivity (%)

Fig. 6. Averaged PSD function obtained from the Eigenvector method,

over 10 periods (1 024 samples each), applied to the isolated diastolic heart Fig. 8. Diagnostic Performance of Parameters z2 obtained from the AR,
sounds of Norma] Patient 4 (the dimension of the signa] subspace was 3) ARMA, and Eigenvector methods and r obtained from the FFT.
and Abnorma] Patient ] (the dimension of the signa] subspace was 5) after
adaptive filtering. Solid line: Abnormal. Dashed line: Normal.
selecting a threshold of 0.86 or below for the amplitude of the
second pole z 2 of the normal patients using the Eigenvector
MINIMUM-NORM
method leads to sensitivity of 79.16\% and a specificity of
•• ABNORMAL
11:/2 90.62\%. Using this threshold value, 3 of 32 normal and 10
• NORMAL
of 48 abnormal subjects were incorrectly classified. Finally,
- using a threshold of 0.46 or below for the power ratio r of
the normal patients analyzed by the FFT method leads to a
sensitivity of 75.00\% and a specificity of 62.50\%. Using
this threshold value, 12 of 32 normal subjects and 12 of 48
abnormal subjects were incorrectly classified.
Considering the incorrectly diagnosed diseased patients,
four patients (19, 30, 76, 80) had 100\% LAD occlusions.
These blockages would permit no blood flow and thus can be
7[ expected to appear normal. There is no clear trend among the
Fig. 7. Seven complex conjugate pole pairs obtained from the Eigenvector other nine misdiagnosed patients, however all of them were
method, over ]0 periods (1024 samples each) applied to the same patients in post -angioplasty patients.
Fig. 6 are shown in the upper half of the complex plane (phase in 11: radians
with respect to f8/2. = 2 kHz and complex conjugates not shown).
IV. CONCLUSION

icant. The second pole amplitude parameter z 2 obtained with In this study, an adaptive filter was used to reduce the back-
the AR, ARMA, and Eigenvector methods and parameter r ground noise from diastolic heart sounds recordings. The di-
obtained with the FFT method, showed significance values of agnostic performance of four signal processing techniques was
less than 0.001 for all four estimation techniques. evaluated after filtering the diastolic heart sounds. The AR and
Fig. 8 shows curves constructed from the parameter z 2 ARMA methods showed essentially the same diagnostic per-
obtained with the AR, ARMA, and Eigenvector methods and formance after adaptive filtering of the diastolic heart sounds.
parameter r obtained with the FFT method after adaptive Results obtained when comparing groups showed that the
filtering. Each point on the curve represents the combination spectral energy distribution differed markedly between normal
of true positives versus false negatives estimated for a given and diseased patients, with the energy between 200 and 800
threshold value of the parameter. For example, in Fig. 8 Hz being greater for diseased patients. For normal subjects, it
selecting a threshold of 0.83 or below for the amplitude of was found that the second poles of the AR, ARMA, and Eigen-
the second pole z2 of the normal patients using the AR vector methods were farther from the unit circle than those of
method leads to a sensitivity of 72.34\% and a specificity of diseased patients. The curves of sensitivity versus specificity
87.50\%. Using this threshold value, 4 of 32 normal and 13 using the AR, ARMA, and Eigenvector second poles and the
of 48 abnormal subjects were incorrectly diagnosed. In Fig. FFT power ratio for decision criteria show that this method
8 selecting a threshold of 0.83 or below for the amplitude of can successfully detect coronary artery disease noninvasively.
the second pole z 2 of the normal patients using the ARMA Among all four methods, the Eigenvector methods showed
method leads to a sensitivity of 70.71\% and a specificity of the best diagnostic performance when compared with the FFT,
87.50\%. Using this threshold value, 4 of 32 normal and 14 AR, and ARMA methods. Since the assumed power spectral
of 48 abnormal subjects were incorrectly classified. In Fig. 8 density function form imposes certain constraints on the ability
AKAY, et al.: NONINVASIVE ACOUSTICAL DETECTION OF CORONARY ARTERY DISEASE 577

of the estimator to estimate the PSD for finite data records, the [2] M. Akay, M. Bauer, J. L. Semmlow, W. Welkowitz, and J. Kostis,
parametric spectral estimators such as the AR, ARMA, and "AR modeling of diastolic heart sounds," in Proc. IEEE Frontiers in
Medicine, New Orleans, LA, pp. 172-175, 1988.
Eigenvector methods offer the promise of higher resolution [3] M. [Link], M. Bauer, J. L. Semmlow, W. Welkowitz, and J. Kostis,
over the FFT [22], [32], [33]. This is not surprising, since "Analysis of diastolic heart sounds before and after angioplasty," in
Proc. IEEE Frontiers in Medicine, New Orleans, LA, pp. 257-260, 1988.
the Eigenvector methods have infinite resolution and provide [4] M. Akay, J. L. Semmlow, W. Welkowitz, and in J. Kostis, "Parametric
an accurate estimate frequency estimation of the signals [22], analysis of diastolic heart sounds before and after angioplasty," in Proc.
[29], [32], [33]. However, the principal shortcomings of the IEEE Frontiers in Medicine, Seattle, WA, pp. 51-53, 1989.
[5] M. Akay, J. L. Semmlow, W. Welkowitz, M. Bauer, and J. Kostis,
AR, ARMA, and Eigenvector methods are that in each case "Detection of coronary occlusions using AR modeling of diastolic heart
poor spectral estimates occur [22], [32], [33] if the assumed sounds," IEEE Trans. Biomed. Eng., vol. 37, 366-373, 1990.
model is inappropriate, if the model orders chosen are in- [6] M. Akay, J. L. Semmlow, W. Welkowitz, M. Bauer, and J. Kostis,
"Noninvasive detection of coronary occlusions using Eigenvector meth-
correct, or if the dimension of the autocorrelation matrix is ods before and after angioplasty," IEEE Trans. Biomed. Eng., vol. 37,
inappropriate. pp. 1095-104, Nov. 1990.
None of the patients had aortic regurgitation, mitral stenosis, [7] M. Akay, "Noninvasive detection of coronary artery disease using
advanced signal processing methods," Ph.D. dissertation, Rutgers Uni-
or other audible diastole murmurs. Further work will involve
versity, Piscataway, NJ, 1990.
the development of additional signal processing techniques to [8] M. Akay, W. Welkowitz, J. L. Semmlow, and J. Kostis, "Application
of the ARMA method to acoustic detection of coronary artery disease,"
permit evaluation of such patients.
Med. Bioi. Eng. Comput., vol. 29, pp. 365-372, 1991.
Our results obtained with these four methods compare [9] J. L. Semmlow, M. Akay, and W. Welkowitz, "Noninvasive detection
quite favorably with other noninvasive methods for detecting of CAD using parametric analysis methods," IEEE, EMBS vol. 9, pp.
33-37, 1990.
coronary artery disease. For example, the sensitivity of the [10] W. Welkowitz, M. Akay, J.-Z. Wang, J. Semmlow, and J. Kostis,
Cardiointegram (CIG) technique developed by Teichholtz et "A model for disturbed coronary artery flow with phonocardiographic
at. [36], was found to be 73\% with a specificity of 78\%. verification," in Activation, circulation and transport in the cardiac
muscle, S. Sideman and R. Beyer, Eds. Norwell, MA: Kluwer, 1990.
This approach was considered to be a moderately useful [II] J.-Z. Wang, B. Tie, W. Welkowitz, J. L. Semmlow, and J. Kostis,
noninvasive method to detect coronary artery disease. "Modeling sound generation in stenosed coronary arteries," IEEE Trans.
The approach described in this paper is based on the Biomed. Eng., vol. 37, pp. 1087-1094, Nov. 1990.
[12] M. Akay, W. Welkowitz, J. L. Semmlow, and J. Kostis, "Application of
diastolic heart sound measurements associated with turbulence
the adaptive filters to noninvasive detection of coronary artery disease,"
due to stenoses. It is important to point out that the flow IEEE Trans. Biomed. Eng., vol. 39, pp. 176-184, 1992.
[13] B. Widrow, J. Glover, J. McCool, and J. Treicher, "Adaptive noise can-
patterns in the coronary tree do not conform with the simplified
celling: principles and applications," in Proc. IEEE, vol. 63,1692-1716,
assumption of the flow model. First of all, the measured 1975.
Reynolds number in the coronaries is well below the critical [14] J. Chen, J. Vanderwalle, W. Sansen, G. Vantrappen, and J. Janssens,
"Adaptive method for cancellation of respiratory artifact in electrogastric
value (around 2000) necessary for the transition to turbulence measurements," Med. Bioi. Eng. Comp., vol. 27, 57-63, 1989.
flow. Bayliss reported that transition to turbulence occurs in [15] T. O. Cheng, "Diagnostic Murmur Caused by Coronary Artery Steno-
the neighborhood of Reynolds number 200 [46], [47]. sis," Ann. Intern. Med., vol. 72, pp. 543-546, 1970.
[16] W. Dock, and S. Zoneraich, "A diastolic murmur arising in a stenosed
The importance of an effective procedure for noninvasively coronary artery," Am. J. Medicine, vol. 42, p. 617, 1967.
detecting coronary occlusions is self-evident. The advantages [17] G. W. Duncan, T. O. Gruber, C. F. Dewey, G. S. Myers, and R. S. Lees,
"Evaluation of Carotid Stenosis by Phonoanogiography," New England
of this approach over the other approaches are that it is
J. Med., vol. 27, pp. 1124-1128, 1975.
noninvasive, passive, quick, and inexpensive. The acoustic [18] E. R. Ferrara and B. Widrow, "Adaptive filtering for signal enhance-
approach for detecting coronary artery disease using signal ment," IEEE Trans. Acoust. Speech Signal Processing, vol. ASSP-29,
June 1981.
processing methods is based on using isolated diastolic heart [19] B. Friedlander, and B. Porat, "Modified Yule-Walker of ARMA spectral
sounds. Although the acoustic approach fails to diagnose estimator," IEEE Trans. Aerosp. Electron. Syst., vol. AES-20, pp.
patient with 100\% occlusion, since these blockages permit 158-172, 1984.
[20] D. Izraelevitz, and J. S. Lim, "Spectral characteristics of the overdeter-
no flow and thus no turbulent sounds, it differs from other
mined normal equation method for spectral estimation," in Proc. 2nd
noninvasive techniques which are based on the symptoms of ASSP Spectral Estimator Workshop, pp. 49-54, 1983.
coronary artery disease. The decision parameters used in this [21] D. P. Giddens, "Measurements of disordered flows distal to subtotal
vascular stenoses in the thoracic aortas of dogs," Circulation Research,
study to detect coronary artery disease noninvasively are in- vol. 39(1), pp. 112-119, 1976.
dependent parameters obtained with traditional methods. They [22] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice
Hall, 1986.
may be combined further to improve noninvasive diagnosis of
[23] R. S. Lees, et al., "Phonoangiography: A new noninvasive diagnostic
coronary artery disease as a mass screening procedure. method for studying arterial disease," in Proc. Nat!. Acad. Sci., vol. 67,
pp. 935-942, 1970.
[24] R. S. Lees and G. S. Myers, "Noninvasive diagnosis of arterial disease,"
ACKNOWLEDGMENT
Adv. Intern., vol. 23, pp. 475-509, 1982.
[25] M. M. Kartchner, L. P. McRae, F. D. Morrison, "Noninvasive detection
The authors wish to thank 1. Redling, D. Shen, A. Smith, and evaluation of carotid occlusive disease," Arch. Surg., vol. 106, pp.
and V. Padmanabhan for collecting and preprocessing the data 528-535, 1973.
used in this study. [26] M. Kaveh, "High resolution estimator for noisy signals," IEEE Trans.
Acoust. Speech Signal Processing, vol. ASSP-27, pp. 286-297, 1979.
[27] M. Kaveh and S. P. Bruzzone, "A comparative overview of ARMA
REFERENCES spectral estimators," in Proc. 1st. ASSP Spectral Estimation Workshop,
2.4.1-2.4.8., 1981.
[1] J. L. Semmlow, W. Welkowitz, J. Kostis, and J. W. Mackenzie, "Coro- [28] S. M. Kay, and S. L. Marple, "Spectral analysis: A modern perspective,"
nary artery disease-correlates between diastolic auditory characteristic in Proc. IEEE, vol. 69, pp. 1380-1419, 1981.
and coronary artery stenoses," IEEE Trans. Biomed. Eng., vol. BME-30, [29] R. Kumerasan and D. W. Tufts, "Estimating the angles of multiple plane
pp. 136-139, 1983. waves," IEEE Aerosp. Electron. Syst., vol. AES-19, pp. 134-139, 1983.
578 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 40, NO.6, JUNE 1993

[30] T. W. Anderson, "Asymptotic theory for principle component analysis," He acted as coauthor with Dr. Welkowitz and Dr. Deutch for the new edition
Ann. 1. Math. Stat., vol. 34, pp. 122-148, 1963. of the book entitled Theory and Design of Biomedical Instruments (Academic
[31] M. Kendall, and A. Stuart, The Advanced Theory of Statistics, Fourth Press, 1991) and is the author of Biomedical Signal Processing (Academic
Ed. London, UK: Griffin, 1977. Press, 1993 (in press)). His research areas of interest are adaptive signal
[32] J. Makhoul, "Spectral linear properties and applications," Trans. Acoust. processing, detection and estimation theory, and application to biomedical
Speech Signal Processing, vol. ASSP-29, pp. 282-296, 1981. signals.
[33] S. J. Orfanidis, Optimum Signal Processing. New York: Macmillan, Dr. Akay is a member of Eta Kappa, Sigma Xi, Tau Beta Pi, the American
1985. Heart Association, the New York Academy of Sicences, and the American
[34] V. Padmanabhan, R. Fisher, J. L. Semmlow, W. Welkowitz, and Association of the Advancement of Science.
J. Kostis, High sensitivity PCG transducer for extended frequency
applications, in Proc. IEEE Frontiers in Medicine, Seattle, WA, pp.
57-59, 1989.
[35] J. F. Sangster, et al., "Diastolic murmur of coronary artery stenosis,"
Br. Heart J., vol. 35, pp. 840-844, 1973. Walter Welkowitz (S'46-A'49-M'55-SM'74-
[36] L. E. Teichholtz, et al.,"Cardiointegram detection CAD using in normal F'76-LF'92) received the B.S. degree in 1948 from
resting electrocardiograms," J. Am. Cardio., vol. 3, pp. 598. Cooper Union, NY; the M.S. degree in 1949 and
[37] J. D. Markel and A. H. Gray, Linear Prediction of Speech. Berlin, Ph.D. degree in 1954 from the University of Illinois.
Germany: Springer-Verlag, 1976. He was a research associate at Columbia
[38] V. K. Iyer, P. A. Ramamoorthy, and Y. Ploysongsang, "Autoregressive University, New York, from 1954 to 1955. He
modeling of lung sounds: Characterization of source and transmission," joined Gulton Industries, Inc., in 1956, and worked
IEEE Trans. Biomed. Eng., vol. 36, pp. 1133-1136, 1989. in various phases of medical instrumentation until
[39] O. Paiss and G. F. Inbar, "Autoregressive modeling of surface EMG and 1964. He then joined Rutgers University, New
its spectrum with application to fatigue," IEEE Trans. Biomed. Eng., Brunswick, NJ, where he is currently Professor of
vol. 34, pp. 761-771, 1987. Biomedical Engineering, and Adjunct Professor in
[40] M. N. Ansourian, J. H. Dripps, G. J. Beattie, K. Boddy, "Autoregressive the Department of Surgery (Bioengineering), Robert Wood Johnson Medical
spectral estimation of fetal breathing movement," IEEE Trans. BME, School (UMDNJ).
vol. 36, pp. 1076-1084, 1989. Dr. We1kowitz is the author of Engineering Hemodynamics: Application to
[41] J.-Y. David, S. A. Jones, and D. P. Giddens, "Modern spectral analysis Cardiac Assist Devices, and coauthor with Drs. S. Deutsch and M. Akay of
techniques for blood flow velocity and spectral measurements with Biomedical Instruments: Theory and Design. His research concentrations are in
pulsed doppler u1trasounds," IEEE Trans. Biomed. Eng., vol. 38, pp. the areas of cardiovascular systems analysis, heart assist devices and diagnosis
589-596, 1991. of cardiac disorders. He has served as a member of the Administrative
[42] H. Sakai and H. Tokumaru, "Statistical analysis of a spectral estimator Committee of the IEEE Engineering in Medicine and Biology Society and
for ARMA process," IEEE Trans. Autom. Control, vol. AC-25, pp. was chairman of its annual conference in 1984. He actively participated in
122-124, 1980. the ABET accreditation process and was a member of the IEEE Committee
[43] S. V. Narasimhan, "Pole-Zero Spectral Modeling of EEG," Signal on Engineering Accreditation Activities (CEAA).
Processing, vol. 19, pp. 17-32, 1989. Dr. Welkowitz is a member of Tau Beta Pi, Eta Kappa Nu, Phi Kappa
[44] J. Chen, J. Vandewalle, and W. Sansen, "Adaptive spectral analysis of Phi, Sigma Xi, the American Heart Association, the New York Academy of
cutaneous electrogastric signals using autoregressive moving average Sciences, and the American Society for Artificial Internal Organs. He received
modeling," Med. Bioi. Eng. Comput., vol. 28, pp. 531-536, 1990. the IEEE Centennial Medal in 1984. He was awarded the IEEE EMBS Career
[45] 1. Gath, C. Feuerstein, D. T. Pham, and G. Rondouin, "Tracking ofrapid Achievement Award in 1991.
dynamic changes in seizure EEG," vol. 39, pp. 952-959, Sept. 1992.
[46] L. E. Bayliss, "The Rheology of Blood," Handbook of Physiology,
Section 2: Circulation, 1:137, 1962.
[47] S. E. Logan, "On the fluid mechanics of human coronary artery
stenosis," IEEE Trans. Biomed. Eng., vol. 22, pp. 327-334, July, 1975. John L. Semmlow (M'79-SM'89) was born in
[48] R. O. Schmidt, "A signal subspace approach to multiple emitter location Chicago, IL, in 1942. He received the B.S.E.E.
and spectral estimation," Ph.D. dissertation, Stanford University, CA, degree from the University of Illinois, Champaign,
1981. in 1964. Following several years as an R. F. design
engineer for Motorola, Inc., he entered the Bio-
engineering Program at the University of Illinois
Medical Center, Chicago, receiving the Ph.D. degree
in 1970.
Yasemin M. Akay (S'91) was born in Istan-
He has held faculty positions at the University of
bul, Turkey. She received the Phar.D. degree from
California, Berkeley, and the University of Illinois,
Hacettepe University, Turkey, in 1980. She received
Chicago, and currently holds a joint position as As-
the M.S. from Rutgers University in 1992. From
1988 to 1990 she was a non-matriculated student at sociate Professor of Surgery. UMDNJ-Robert Wood Johnson Medical School
and Associate Professor of Biomedical Engineering at Rutgers University, NJ.
Rutgers University, where she is a graduate student.
In 1985 he was a NSF/CNRS Fellow in the Sensorimotor Control Laboratory
From 1980 to 1988 she worked as a pharmacist.
of the University of Provence, Marseille, France. His active research areas
Her research areas involve application of the neural
include physiological motor control and medical instrumentation.
networks to biomedical signals.

John B. Kostis was born in Greece. He received the

M.D. degree in 1960 at the University of Salonica,
Metin Akay (S'87-M'91-SM'93) was born in Greece.
Sivas, Turkey. he received the B.S. degree and He is now Professor of Medicine and Pharmacol-
the M.S·. degree in electrical engineering from ogy and Chief of the Division of Cardiovascular
Bogazici University, Istanbul, Turkey, in 1981 and Diseases and Hypertension at the University of
1984, respectively. From 1984 to 1986 he was Medicine and Dentistry of New Jersey-Robert Wood
in the Ph.D. program, at the same University. He Johnson Medical School. He is involved in several
received his Ph.D. degree from Rutgers University large multicenter NIH clinical trials and has pub-
in 1990. lished extensively in the areas of clinical cardiology
In 1986, he joined the Department of Biomedical and hypertension, cardiovascular pharmacology, and
Engineering, Rutgers, University, Piscataway, NJ, biomedical engineering.
where he held a Fellowship and was a teaching Dr. Kostis was elected president of the American Heart Association, New
assistant. He View
is publication
currently stats a visiting assistant professor at Rutgers University. Jersey Affiliate, in November, 1988.