Transcript Detection of A2-P2 Interval using S

Detection of A2-P2 Interval using S-Transform and Gaussian Modeling
2
Gavriely ,
1
Intrator
Guy Amit , Noam
Nathan
1School of Computer Science, Tel-Aviv University
2Rappaport Faculty of Medicine, Technion
Methods – Cont.
We have found that the analysis of S2
using the S-Transform leads to best
separation of the components: A2-P2
f
S ( , f )   s(t )
2

e
( -t )2 f 2
2
e
- i 2 f t
Conclusion
Two recordings from 5 healthy subjects
containing one minute of respiration
were inspected.
S-Transform
S-Transform of S2
150
A2 Component
100
P2 Component
50
0
dt
S-Transform analysis showed two
distinct clusters. Gaussian parameter
estimation enabled peak detection where
clusters were close in the time-frequency
plane.
0
0.05
0.1
0.15
0.2
time (sec)
0.25
0.3
0.35
Wigner-Ville Distribution of S2
150
Cross Component
100
A2 and P2 Components
50
0
0
0.05
0.1
0.15
0.2
time (sec)
0.25
0.3
0.35
S-Transform of second heart sound
Compariosn between Wigner-Ville
(Used by Duran) and S-Transform
0.02
160
0.01
140
amp
0
120
-0.01
freq (Hz)
Accurate automatic identification of the
A2 and P2 components is difficult due to
their temporal overlap and spectral
similarity.
Results
freq (Hz)
Abstract
The temporal difference A2-P2 is
correlated with pulmonary arterial
pressure, thus can be used to detect
pulmonary hypertension. Detection of
the precise timing of A2 and P2 events
can also help in triggering devices such
as left ventricular assist devices.
1
freq (Hz)
Eyal
1
Balla ,
100
-0.02
80
-0.03
• Phonocardiogram (PCG) from
pulmonary area
• Electrocardiogram (EKG)
• Respiration pressure signals
0.2
time (sec)
0.25
0.3
0.35
0.4
0.05
0.1
0.15
0.2
time (sec)
0.25
0.3
100
0.35
50
Gaussian Estimation
0
1
x



 i  i  x  i 
T
S ( x)  
i 1
1
 2  i
e
1/ 2

0.05
0.1
0.15
Low frequency peak
0.2
time (sec)
0.25
0.3
0.35
0.4
Change of Gaussian mean time interval while breathing
50
Gaussian Mean Time interval
Normalized Pressure
40
30
20
10
0
2
-10
0
10
20
30
40
50
60
70
time (sec)
Change of Gaussian mean frequency interval while breathing
70
Estimation of the parameters in the
Time-Frequency plane is done using the
EM algorithm
The algorithm estimates parameters of
two Gaussians (corresponding to the
subcomponents of S2) from a scatter
which can be created by the different
time-frequency distributions.
Scatter of time-frequency distribution
100
90
0
The interval times of the two clusters
varied between 5ms-40ms
Interval time (ms)
Peak detection is done using Gaussian
estimation for ensuring robustness. The
signal is modeled as a sum of two
Gaussians in the time-frequency plane:
1 Std from mean
Generated by covariance
matrix
80
Gaussian Mean Frequency interval
Normalized Pressure
60
50
40
30
20
10
0
10
20
30
40
50
60
70
time (sec)
Gaussian mean peak frequency while breathing
100
• The robustness is achieved using a
detailed T/F representation such as the
S-Transform.
• The robustness is increased by the
modeling of the two time/frequency
energy peaks as two Gaussian
distributions, thus, achieving a stable
estimation -- the Gaussian Mean, rather
then the energy peak.
High Freq Peak
Low Freq Peak
80
60
40
20
• The time interval between the two
components of S2 can be robustly
measured.
0
10
20
30
40
50
60
70
time (sec)
Plot of time intervals between A2 and P2
together with breathing pressure. Intervals
during inhalation were smaller than those
during exhalation and had smaller
variance.
• Mild agreement with literature which
states that the A2-P2 interval is
expanded during inspiration.
References
70
freq (Hz)
Digital recording of healthy volunteers
of the following channels:
0.15
freq (Hz)
0
2
Data Acquisition
0.1
High frequency peak
20
Frequency Difference (Hz)
Methods
0.05
150
40
Frequency (Hz)
A novel robust time-frequency method
for measuring the distance between the
two components A2 and P2 of the
second heart sound S2 is introduced.
0
60
60
50
Mean of
Gaussians
40
30
20
0.12
0.13
0.14
0.15
0.16
0.17
time (sec)
0.18
0.19
0.2
0.21
0.22
Both mean and full covariance matrix are
estimated
Amit, G., Gavriely, N., Lessick, J., Intrator, N.
Automatic Extraction of Physiological Features from
Vibro-Acoustic Heart Signals: Correlation with
Echo-Doppler. Computers in Cariology , 299-302,
2005