Heart Sound Segmentation Algorithm Based on Heart Sound Envelolgram

H Liang, S Lukkarinen, I Hartimo

Helsinki University of Technology, Espoo, Finland

Abstract

based on the signal itself without a reference to ECG

using a set of normal recordings.
The purpose of this study is to develop ;in algorithm for
heart sound segmentation which uses the heart sound
signal as the sole source. Based on the algorithm, every
cycle of the PCG signals is separated into four parts: the
first heart sound, the systolic period, the second heart
sound and the diastolic period. The locations and
intervals of the first heart sounds and the second heart
sounds are computed first. Then based on this
information, the intervals of the systolic and diastolic
period are obtained consequently. Both normal and
abnormal heart sound recordings are investigated.

The segmentation algorithm, which separates the heart

sound signal into four parts: the first heart sound, the
systole, the second heart sound and the diastole, is
developed. The segmentation of phonocardiogram signal
is the first step of analysis and the most important
procedure in the automatic diagnosis of heart sounds.
This algorithm is based on the normalized average
Shannon energy of PCG signal. The performance of the
algorithm has been evaluated using 515 periods of PCG
signal recording from 37 objects including normal and
abnormal. The algorithm has shown 93 percent correct
ratio.

2. Materials

1. Introduction
Noninvasive study (diagnosis) methods, such as
phonocardiogram (PCG) and electrocardiogram (ECG),
offer useful information of functioning heart. In
auscultation, the listener tries to analyze the heart sound
components separately and then synthesize the heard
features. Heart sound analysis by auscultation highly
depends on the skills and experience of the listener [l].
Therefore the recording of heart sounds and analyzing
them by a computerized and objective way would be most
desirable. Before any automatic analysis can be done, the
heart sound needs to be segmented into components and
then analyze those components separately. The main
components are the first heart sound (Sl), the systolic
period, the second heart sound (S2),the diastolic period
in this sequence in time.
Some attempts to segment PCG signals have been
reported in the literature. The majority of them depend on
the reference of ECG signal or/and carotid pulse, such as
[2],[3]and [4]. M.W. Groch and A. Iwata have shown a
solution where the segmentation is based on the timedomain characteristics [3] and the frequency-domain
characteristics [4]of the PCG signal, respectively. David
S. Gerbarg [5],thirty years ago, took advantage of the
time relations of the signal components to separate them

0276-6547197 $10.00 0 1997 IEEE

105

The sound material consists of recoirdings of heart

sounds recorded with a multimedia PC equipped with an
electronic stethoscope. The sounds are recorded with 16bit accuracy and 11025H;z sampling frequency. No ECG
equipment has been used. Totally 37 recordings including
14 pathological murmurs and 23 physiological murmurs
with total cycles of 515 have been used to evaluate the
algorithm. These recordings have been made from
children aged from 0.4 to 13.9 years and they are taken at
several auscultation locations with duration of 7-12
seconds. The patients have different types of heart
diseases. An experienced children cardiologist has
pointed out the correct positions of S1 andl S2.

3. Methods
The segmentation algorithm is based cm the envelope
calculated using the normalized average Shannon energy,
which attenuates the effect of low value noise and makes
the low intensity sounds easier to be found.

3.1. The normalized average Shannon

energy
At first, the original signal is decimaited by factor 5
from 11025Hz to 2205IEz sampling frequency using an
eighth order Chebyshev type I lowpass filter with cutoff

Computers io Cardiology 1997 Vol24

frequency at 882Hz. Here, after filtering in the forward

direction, the filtered sequence is then reverses and run
back through the filter. The resulting sequence has
precisely zero-phase distortion, so there is no group delay.
Secondly, the signal is normalized to absolute
maximum of the signal according to the equation (1)

where x~~~~is the decimated signal.

Then, the envelope of normalized decimated signal is
calculated. Fig.l shows different methods of calculating
the envelope of the normalized signal. Because of the
symmetry of the results as we can see from the following
definitions, only the positive part is shown here. The
figure is drawn based on the following definitions, here x
is the normalized signal, which has the real value from -1
to 1.
Shannon energy: E = --x .logx2
(4)
Shannon entropy:
absolute value:
energy(square):

E = 4x1 loglxl
E = 1x1
E =x2

(5)

(6)

(7)

-- Shannon energy

solid llne
dotted llne

p0.6
c

-- Shannon entropy

.
.

Figure 1. The comparison of different envelope methods.

The figure indicates that the energy (square) will bury
the low intensity sounds under the high intensity ones by
enlarging the high/low intensity ratio. The Shannon
entropy accentuates the effect of low value noise that
makes the envelope too noisy to read. The absolute value
gives the same weight to all the signal. The Shannon
energy emphasizes the medium intensity signal and
attenuates the effect of low intensity signal much more
than that of high intensity signal. So, the Shannon
energy is better than the absolute value in shortening the
difference of the envelope intensity between the low
intensity sounds and the high intensity sounds. This
shortening makes the finding of low intensity sounds
easier.
Thirdly, the average Shannon energy is calculated in

106

continuous 0.02-second segments throughout the signal

with 0.01-second segment overlapping. The average
Shannon energy is calculated as
N

E , = -1 / N

x:oTm

(i> log x i o r m (i>

*

(2)

i=l

where , the xnOmisdecimated and normalized sample

signal and N is signal length in 0.02-second segments,
here N = 44.
Then the normalized average Shannon energy versus
time axis is computed. The normalized average Shannon
energy is computed as follows,
(3)

,
s( E,( t)) is the standard deviation of E ,( t ).

where, M ( E , (t)) is the mean value of E ( t) ;

3.2.

Picking up the peaks

Fig.2 Shows an original PCG signal , which is one of

the recordings, and its normalized average Shannon
energy and a threshold. Based on the envelogram
calculated by the normalized average Shannon energy
curve, a threshold is set to eliminate the effect of noise
and the very low-intensity signal. The peaks of each part
whose levels exceed the threshold are picked up and
assumed temporarily to be the first or the second heart
sound. Here, only one peak for each overshoot is chosen
even though there are more than one peaks above the
threshold. The choice of the peak for each overshoot is
based on the following criteria : (1) one peak is always
picked up; (2)more than two peaks means the existence of
splitted first or second heart sound, so the first peak is
picked up in order to get the onset of each sounds.

Ti me(s)

Figure 2. The original signal and its average Shannon

energy.
The actual abnormal heart sound recordings are very
complicated and the patterns of heart sounds and
murmurs vary largely fiom recording to recording even

for the normal ones. To solve these problems, we made

several additions in the procedure.

3.2.1. Rejecting extra peaks

The first problem is that many extra peaks are picked
up. This situation is shown in Fig.3. In order to eliminate
the extra peaks, the intervals between each adjacent peaks
are calculated. The low-level time limit and high-level
time limit, which are used for deleting extra peaks and
finding lost sounds respectively, are computed for each
recording based on the mean value and standard
deviation of these intervals. When an interval between
two adjacent peaks is less than the low-level time limit,
there must be one extra peak which should be rejected.
The following criteria are used to decide which one
should be dropped: (1) when two peaks appear within
50ms, which is the largest splitted normal sound
interval[61, and the energy of the first peak is not too
small compared to that of the second one, we picked up
the first one. Otherwise, the second one is chosen. Here
we assumed that the two peaks that meet the above
conditions are actually two parts of a splitted heart sound.
Otherwise, the first peak may belong to some other sound
class or can be classified as noise. (2) when the interval
between two adjacent peaks exceeds 50 ms, their
energies are compared. If the energy of the first peak is
larger than that of the second one and the last interval
meets certain consistency of every second interval, we
pick up the first one, otherwise the second one is picked
UP.

amount. This reduction of the threshold will be iterated

until the losing peaks are picked up or the iteration limit
is reached. Here the above criteria to eliminate the extra
peaks are used again to delete all extra peaks picked up
in the procedure of finding lost ones. Fig.4 shows the
lost peaks are found again.
The third problem is caused by the artifacts, that look
like real peaks both in time interval anid amplitude and
cant be discarded using the methods created for extra
peaks. They will be rejected in the following process.
1

61

SZ

61

3.5

62

Sl

62

Time(s)

61

6:,L

4.5

Figure 4. Recovering the lost peaks.

3.3 Identifying the Sls and S2s

After all the Sls and S2s have been recognised, we
need to identify which one is S1 and which one is S2.
Here, the identification i s based on the following facts:
(1) the largest interval of a recording (within 20 seconds)
is the diastolic period; (2) the systolic period is relatively
constant compared to tRe diastolic one. According to
these facts, starting from the largest interval, we examine
the intervals both forwards and backwards to maintain
the relative consistency of systolic and diastolic intervals
by different tolerances. Here, parameters c l and c2 are
used to represent the allowable tolerances in percentage
of the duration of the systolic and the diastolic period
respectively. After we have identified which are the
systolic and which are the diastolic periods, we have also
,
the peaks not
implied which are Sls and S ~ S and
included to these sets are detected and discarded. Fig.5
shows the identified Sls and S2s and a discarded
artifacts.

Time(s)

Figure 3. Rejection of extra peaks.

7
6

3.2.2. Recovering weak peaks of S1 or S2

a5

.z

-4

Secondly, some peaks are so weak, usually the Sls,

that their energies are less than the threshold and they are
about to get lost. These weak peaks definitely need to be
found, too. To do this, we still examine the intervals
between the adjacent peaks. When the interval e x c d s

the high-level time limit, it is assumed that a peak has

been lost and the threshold will be decreased by a certain

107

e3

4:

1
0
0

Time(s)

Fig.5 The identified S1 and S2 and artifacts

3.4.

5. Discussion

Decision of durations of S1 and S2

The detected Sls and S2s indicate the approximate

location of these sounds. The actual duration of these
sounds are obtained by defrning another lower threshold,
which differs from the one for detecting S1 and S2, to
find the two boundaries for each sound. The detection,
however, may lead to inaccurate boundaries because of
noise, closed murmur, the third heart sound, the fourth
heart sound, etc. We modified the deviations by confining
the duration of heart sounds between 2Oms and 12Oms.
The durations of the systole and the diastole are
accordingly decided with a small transition period before
and after these sounds.

4. Results
The results of the experiment indicate 93 percent
correctness in the automatic identification of S1 and S2.
The misinterpretations are caused by large background
noise and serious murmurs. Table.l shows the results. If
we omit the missing cycles, so that they will not effect our
further analysis, higher percentage of correctness will be
obtained.
Table 1. The statistics of the results.
I Numberofcvcles
correct
I
479
missing
30
6
incorrect
5 15
total

I
I

Percentage(%)
93.01
5.83
1.17
100

93.79%

1.36%

4.85%

6. Conclusions
The automatic segmentation algorithm is found to be
effective to segment phonocardiogram signals into four
parts. The algorithm has shown 93 percent success in 37
recordings, which include 515 cycles of heart sounds.
This is a good basis for further analysis of the heart sound
signals.With this segmentation, we can extract the
features of each segment, such as the root mean square,
the peak intensity, the peak location, the duration, the
splitted interval of S2, etc. We can also do other processes
to each segment for the classification.

Acknowledgements
The authors are grateful to Anna-Leena Noponen and
Anna Angerla for recording the signals and giving
clinical comments.

References

The performances of this algorithm with different sets

of parameters have also been compared. Table.:! shows
the results. The ratios of correctness for these three cases
do not differ much. It means that our algorithm is
relatively robust to fluctuations of these parameters. In
our research, it is the incorrect identification but not the
missing sounds that will degrade the performance of the
later extraction of the features of each segment. So, even
though the third set of parameters can give a little better
ratio of correctness, we choose the first set of parameters
which gives lower ratio of incorrectness.

I 0.20 0.40 0.5 I

One reason of the incorrect identification is the high

level of interfering signals like speech, crying, or other
ambient noise. Sometimes the sudden release of the
stethoscope from the patients for a short time during the
recording of data has also resulted in incorrect detection.
This can be helped by improving the recording
techniques.
Large murmurs that overlap S1 or S2 will make the
correct identification impossible. However, because the
murmur is large, it cant be neglected by auscultation.

108

[I] Sapire DW. Understanding and diagnosing pediatric heart

disease: Heart sounds and murmurs. Norwalk, Connecticut,
Applcton & Langc 1992: 27-43.
[2] Lehner RJ,Rangayyan RM. A three-channel microcomputer
system for segmentation and characterization of the
phonocardiogram. IEEE Trans. on Biomedical Engineering
1987; 34: .
[3] Groch M W , Domnanovich JR, Erwin WD, A new heartsounds gating devices for medical imaging. IEEE Trans. on
Biomedical Engineering 1992; 39: 307-310.
[4]Iwata A, Ishii N, Suzumura N. Algorithm for detecting the
f i s t and the second heart sounds by spectral tracking. Med.
& Biol. Eng. & Comput Jan 1980: 19-26.
[SI Gerbarg DS, etc. Analysis of phonocardiogram by 3 digital
computer. Circulation Research 1962; 11: 569-576.
[6] Hurst JW.The heart arteries and veins, 7th ed. McGrawHill Information Services Company, New York 1990: Ch.
14:175-242.
Address for correspondence: [email protected]
Helsinki University of Technology, Signal Processing and
Computer Technology Laboratory
P.O.Box 3000, FIN-02015 HUT, FINLAND