Detection of S1 and S2 Heart Sounds by High Frequency Signatures
D. Kumar, P. Carvalho, M. Antunes†, J. Henriques, L. Eug´ enio†, R. Schmidt‡, J. Habetha‡
Centre for Informatics and Systems, University of Coimbra, Portugal
†Cardiothoracic Surgery Centre, University Hospital of Coimbra, Portugal
‡Philips Research Laboratories, Aachen, Germany
A new unsupervised and low complexity method for detec-
tion of S1 and S2 components of heart sound without the
ECG reference is described. The most reliable and invari-
ant feature applied in current state-of-the-art of unsuper-
vised heart sound segmentation algorithms is implicitly or
explicitly the S1-S2 interval regularity. However, this crite-
rion is inherently prone to noise influence and does not ap-
propriately tackle the heart sound segmentation of arrhyth-
mic cases. A solution based upon a high frequency marker,
which is extracted from heart sound using the fast wavelet
decomposition, is proposed in order to estimate instanta-
neous heart rate. This marker is physiologically motivated
by the accentuated pressure differences found across heart
valves, both in native and prosthetic valves, which leads
to distinct high frequency signatures of the valve closing
sounds. The algorithm has been validated with heart sound
samples collected from patients with mechanical and bio-
prosthetic heart valve implants in different locations, as
well as with patients with native valves. This approach ex-
hibits high sensitivity and specificity without being depen-
dent on the valve type nor their implant position. Further-
more, it exhibits invariance with respect to normal sinus
rhythm (NSR) arrhythmias and sound recording location.
Many heart disorders can be effectively diagnosed using
auscultation techniques. In potentially deadly heart dis-
eases, such as natural and prosthetic heart valve dysfunc-
tion or even in heart failure, heart sound auscultation is one
of the most reliable, cheap and successful tools for early di-
agnosis. Auscultation is the preferred method for the detec-
tion of prosthetic valve dysfunction; it exhibits 92% sensi-
tivity over echophonocardiography and cinefluroscopy .
To develop automatic heart disorder diagnosis tools based
on phonocardiogram, it is important to first segment the
heart sound into clinically meaningful segments or lobes,
such as the S1 and the S2 sound components associated
with closing valves during systole and diastole. Once these
are detected, diagnostic features may be subsequently ex-
tracted for each type of sound. However, S1 and S2 sound
detection is one of the major and most difficult problems in
heart sound analysis.
Heart sound segmentation algorithms found in literature
may be broadly classified into two main approaches: those
require an ECG reference to synchronize the segmentation
and those that do not. The latter may be further classified
into supervised and unsupervised methods. In ECG ref-
erence based approach, first QRS complexes and T-waves
are detected in order to locate the S1 and S2 segments, re-
spectively . In low quality ECG signals, T-waves are not
always clearly visible. In such cases, S2 sounds may be
classified by an unsupervised classifier . To avoid extra
hardware requirements and clumsy wiring arrangement for
data acquisition, many researchers tried to identify S1 and
S2 sounds by several means of signal processing and sta-
tistics without using ECG as a reference. In this context
several supervised techniques have been suggested, such
as artificial neural network  and decision trees . An-
other class of approaches involves unsupervised techniques
such as envelogram , spectrogram quantization method
, and self organizing map using Mel frequency cepstrum
coefficients . In practice, it is observed that these meth-
ods do not perform well for all type of heart sounds (e.g.
arrhythmic cases). Regarding the segmentation of heart
sounds produced by prosthetic valves, it is well known that
these sounds are dependent on several factors such as sur-
gical techniques, location of implantation and type of pros-
thetic valves. In practice, the aforementioned methodolo-
gies do not provide the necessary invariance in order to be
applicable for all kinds of heart valve implant patients.
The most reliable and invariant feature applied in cur-
Table 1. Result for the S1, S2 detection
S1 - S2
Mechanical 5893/6097122108 97.80 98.20%
Bio-prosthetic 729/7792619 96.77 97.62%
Native 642/65446 99.38 99.07%
7264/7530152 133 97.95% 98.20%
imum duration of 2 minutes. All collected heart sounds
were first preprocessed using a 4th order Butterworth high
pass filter with cut-off frequency of 40Hz in order to elim-
inate low frequencies produced by muscle and stethoscope
The proposed algorithm was tested for the collected
heart sounds. The achieved results are summarized in ta-
ble 1. In the worst case heart sound sample, 95.67% sensi-
tivity and 96.12% specificity were obtained, while in best
case 100% sensitivity and 100% specificity are found. It
has been noticed that wrong detection are normally caused
by noise removal; when noise such as speech, cough, sud-
den movements are removed adjacent relevant sound seg-
ments are also affected. The achieved sensitivity of noise
detection by jitter approach is 92.20%. A total of 46 noisy
segments out of 50 were detected in all tested sound sam-
ples. High frequency noisy segments which have duration
less than 50 ms were not detected by the jitter approach.
The achieved results were verified by manually inspecting
QRS-complexes and T-waves of the corresponding ECG.
The entire algorithm was simulated in MATLAB using a
pentium 4 (3GHz). It takes between 6 to 10 seconds for the
complete segmentation of each 2 min heart sound sample.
This paper proposes an algorithm for S1 and S2 heart
sound segmentation that does not rely on the ECG refer-
ence. The boundaries of the sounds were identified and
sound lobes were validated using physiological based crite-
ria. Next, heart sound (S1 and S2) were identified based on
a high frequency marker, which is identifiable from accen-
tuated pressure differences found across heart valves, both
in native and prosthetic valves, that leads to distinct fre-
quencysignatures of the valveclosing sounds. Unlikeother
methods of segmentation, the proposed algorithm exhibits
excellent performance in case of SNR arrhythmic heart
sounds. Besides this, it is completely automatic, therefore,
is highly appropriate for eHealth.
Despite of the excellent performance for the correct
detection of S1 and S2 sounds in phonocardiogram, this
method still fails to segment heart murmurs which are the
major signs of heart diseases and prosthetic valve dysfunc-
tion. Precise boundary detection of heart murmurs present
between S1 and S2 or S2 and S1 is the future challenge in
this course of research.
This work was performed under the IST FP6 project
MyHeart ( IST-2002-507816) supported by the European
 G. Mintz, E. Carlson, M. Kolter, ”Comparison of noninva-
sive techniques in evalution of the nontissue cardiac valve
prothesis”, J. Med. Eng. Technol., 15(6), 1991, page. 222-
 M. El-Segaier, O. Lilja, S. Lukkarinen, L. Srnmo, R. Sep-
ponen and E. Pesonen, ”Computer-Based Detection and
Analysis of Heart Sound Murmur”, Annals of Biomedical
Engineering, vol. 33, no. 7, 07-2005, pages. 937-942.
 P. Carvalho, P. Gil, J. Henriques, M. Antunes and L.
Eug´ enio, ”LowComplexityAlgorithmfor HeartSoundSeg-
mentation using the Variance Fractal Dimension”, Proc. of
the Int. Sym. on Intelligent Signal Processing, 2005, page.
 T.OlmezandZ.Dukar, ”Classificationofheartsoundsusing
an artificial neural network”,Journal of Pattern Recognition
Letters, 08-2003, page. 617-629.
 J. E. Hebden and I. N. Torry, ”Neural Network and Conven-
tional Classifiers to Distinguish between First and Second
Heart Sound”, in IEE Colloquium (Digest), 08-1996, page.
 A. Ch. Stasis, E. N. Loukis, S. A. Pavlopoulos, D. Kout-
souris, ”Using decision tree algorithm as a basis for a heart
sound diagnosis desicion support system”, in IEE Collo-
quium (Digest) of Artificial Intelligence Methods for Bio-
medical Data Processing, 08-1996, page. 1-6.
 H. Liang, S. Lukkarinen and I. Hartimo, ”Heart Sound Seg-
mentation Algorithm Based On Heart Sound Envelogram”,
Proc. of IEEE Computers in Cardiology, 1997, page. 105-
 H. Liang, S. Lukkarinen and I. Hartimo, ”A boundary modi-
fication method for heart sound segmentation algorithm”, in
Proc. of IEEE Computers in Cardiology, 1998, page. 593-
 D. Kumar, P. Carvalho, P. Gil, J. Henriques, M. Antunes and
L. Eug´ enio, ”A new algorithm for detection of S1 and S2
heart sounds”, in Int. Conf. of Acoustic and Speech Signal
Processing (ICASSP), 2006, page. 1180-1183.
 K. O. Lim, Y. C. Liew and C. H. Oh, ”Analysis of mitral
and aortic valve vibrations and their role in the production
of the first and second heart sounds”, in Journal of Physics
in Medicne and Biology, 1980, page. 727-733.
 J. Pan, W. J. Tompkins, ”A real-time QRS Detection Al-
gorithm”, in IEEE Transaction on Biomedical Engineering,
vol. BME-32, No. 3, March 1985, page. 230-236.
 J. D. Bronzino, ”The Biomedical Engineering Handbook”,
CRC Press, 1995, NewYork.