Sie sind auf Seite 1von 6

Proceeding of the 2014 IEEE Students' Technology Symposium

An Automated Method for Detecting Systolic Peaks


from Arterial Blood Pressure Signals
Dandu Sriram Raju, M. Sabarimalai Manikandan and Barathram.Ramkumar
School of Electrical Sciences
Indian Institute of Technology Bhubaneswar
Bhubaneswar, Odisha-751013, India.
Email:{sriram.iitbbs, msm.sabari, and barathram24}@gmail.com

AbstractIn this paper, we present an automatic method


for determining time-location of systolic peak in arterial blood
pressure (ABP) signals. The method consists of four major
steps: Gaussian derivative filtering, nonlinear peak amplification,
Gaussian derivative based peak finding scheme, and peak position
adjustment procedure. The method is tested and validated using
the standard MIT-BIH Polysomnographic database containing a
wide range of ABP signals, artifacts and high-frequency noises.
Our results demonstrate that the proposed method can achieve
better peak detection performance while maintaining very small
detection error rates for both clean and noisy ABP signals.
The method achieves an average sensitivity of 99.89% and
positive predictivity of 99.59% on test ABP datasets consisting
of 67,125 beats. Unlike other existing methods, our method is
quite straightforward and simple in the sense that it does not
use search-back algorithms with secondary thresholds.

I. I NTRODUCTION
Arterial blood pressure (ABP) signal contains vital clinical
information about the cardiovascular system, including heart
rate, systolic and diastolic arterial pressures. These clinical
information are widely used to assess properties of the arterial
vessel wall [1]-[6]. Furthermore, the morphological characteristics of an ABP waveform are closely related with the
hemodynamic behaviors of blood circulation [3]. The ABP
signal consists of a systolic peak, dichrotic notch, and dichrotic
peak. The onset and steep upstroke points reflect the aortic
valve opening for blood ejection. The systolic peak of the
ABP waveform indicates the integrated behaviors of cardiac
blood ejection and arterial wave reflection. The dicrotic notch
indicates the closure of aortic valve [3]. The detection of
those pressure components is essential for assessing different
abnormal functions of structural components of heart [4][6]. Thus, an automatic detection method has become very
important tools for pulse oximetry, cardiac arrhythmia detection, and pulse contour analysis [6]. There are numerous
current and potential applications for pressure components
detection methods. Although many detection methods have
been reported for ECG signals [7], [8], there are only a few
methods to detect essential peaks in ABP signals [6].
In [11], X. Liu and J. Liu (2012) presented a method
for detecting the systolic peak of the atrial blood pressure
signals. The method consists of four components: a low-pass
filter to reduce high-frequency noise, a slope sum function
(SSF) to enhance the morphological up-slope features of the

TS14SPSAA02 355

pressure signal and to suppress other components of the


ABP signal, modified mean shift algorithm to detect local
maxima and peak position calibration strategy to find true
peak of the ABP signals. The standard MIT-BIH database was
used for validation. In [3], B. N. Li et al. (2010) presented
an automatic delineator for characterizing the fiducial points
(namely, onsets, systolic peaks and dichrotic notches) of
arterial blood pressure waveforms. The delineation method
is based on the critical point detection in the derivative of
filtered ABP waveform that detects systolic peaks by using
zero-crossing points after the maximal inflection in waveform
derivative. The amplitude and interval thresholds are used in
the detection process. The delineator had an average error rate
1.14%, sensitivity 99.43% and positive predictivity 99.45%. In
[1], W. Zong et al. (2003) presented an open-source algorithm
to detect onset of arterial blood pressure pulses. The method
is based on a windowed and weighted slope sum function to
extract the onsets of ABP waveforms, and claimed its accuracy
99.31% on 368,364 beats with reference to ECG reference
and annotations and 96.41% on 39,848 beats with reference
to manual annotation.
In [13], Z. T. Beattie presented an algorithm to detect
pressure peak beats in ABP, intracranial pressure (ICP), and
pulse oximetry (SpO2) signals. The beat detection algorithm
includes decimation process, 10th order high pass elliptical filter, 6th order autoregressive model, spectral density estimator,
peak detection algorithm and decision logic. The algorithm
had an average sensitivity of 99.6% 0.27 and an average
positive predictivity of 98.6% 1.1. In [6], Mateo Aboy et al.
(2005) designed an automatic detection algorithm for pressure
signals that locates the first peak following each heart beat.
This is called the percussion peak in intracranial pressure,
the systolic peak in ABP and the pulse oximetry (SpO2)
signals. The proposed algorithm incorporates a filter bank with
variable cutoff frequencies, spectral estimates of the heart rate,
rank-order nonlinear filters, and decision logic and achieved
a sensitivity of 99.36% and positive predictivity of 98.43%.
In [15], F. Scholkmann et al. (2012) presented a new method
for automatic detection of peaks in noisy periodic and quasiperiodic signals. The method is based on the calculation and
analysis of the local maxima scalogram, a matrix comprising
the scale-dependent occurrences of local maxima.
Most existing detection methods use similar decision rule

978-1-4799-2608-4/14/$31.00 2014 IEEE

41

Proceeding of the 2014 IEEE Students' Technology Symposium

II. A LGORITHM D ESCRIPTION AND T HEORY


The proposed detection method consists of four major steps:
Gaussian derivative filtering, nonlinear peak amplification,
Gaussian derivative based peak finding scheme, and peak
position adjustment procedure. In the first stage, the ABP
signal is preprocessed for suppressing the baseline drifts and
enhancing the high-slope portions of ABP signals. The filtered
signal is obtained by convolution of the ABP signal with Gaussian derivative function. In the second stage, nonlinear signal
processing performs high-frequency noise suppression and
further amplifies the magnitude of the high-slope components
in the filtered signal. In the third stage, the signal envelope
from the nonlinear signal processing stage is processed to find
approximate locations of systolic peaks. The locations of local
peaks are automatically determined by processing the positive
zero-crossing points of the output waveform obtained from the
convolution of the signal envelope with Gaussian derivative
kernel. Finally, the detected locations of candidate peaks are
used as guides to find true time-locations of systolic peaks in
the ABP signal.

100point Gaussian
Kernel

Gaussian pulse peak


1
0.8
0.6
0.4
0.2

(a)
20

In practice, the raw ABP signals are often corrupted by


various noises and artifacts including baseline drift, amplitude modulation with respiration, power-line interference and
instrument noise [6]. To suppress noises and artifacts, many

40

60

80

100

0.1
0.05

ZeroCrossing Point

0
0.05

(b)

0.1

20

40
60
Sample Number

80

100

Fig. 1. (a) The 100-point Gaussian kernel window with spread = 6.25,
and (b) The Gaussian derivative kernel.

noise filtering techniques including derivatives, digital filters,


wavelets, empirical mode decomposition, adaptive filters and
filter banks have been used in the processing stage [1], [3], [6],
[10]. In this work, the first stage of the proposed preprocessor
employs a Gaussian derivative filter to remove baseline wander
and eliminates high-frequency noise. The output of this filter
contains only high-slope portions. The second preprocessing
stage employs a squarer and an adaptive energy-thresholding
rule for further attenuating other low-amplitude components
and enhancing the high-slope systolic peak portions of the
ABP waveform. The third Shannon energy computation stage
further magnifies the magnitude of a low-amplitude peak. The
smoothing filter provides smooth signal envelope, wherein
the smoothing filtering is determined from refractory period.
The preprocessing stage finally provides a signal envelope
containing smooth local peaks that correspond to the systolic
peaks contained in the ABP signal. Thus, the resulting smooth
signal envelope is further processed to find time-locations of
true systolic peaks in ABP signals.
1) Gaussian Derivative Filtering: The Gaussian derivative
filter is designed to emphasize high-slope systolic peak portions of the ABP waveform. The length and spread of the
Gaussian pulse kernel is determined empirically in such that it
can simultaneously emphasize the high-slope peak portion and
suppress the baseline wander and high-frequency noise. Thus,
the proposed preprocessing can reduce the computational load
as compared to most existing methods which use a set of
digital filters and derivatives. The M 1 coefficients vector
of a Gaussian kernel g[m] are computed as
M 2
1 (m 2 )
2

A. Preprocessing Stages

TS14SPSAA02 355

1.2

Gaussian Derivative
Kernel

with different amplitude-dependent and duration-dependent


thresholds for locating onset and peaks of the ABP waveforms. Many of the methods used in QRS detection that take
advantage of the impulsive shape of the QRS complex that do
not work well on pressure signals. In practise, ABP signals
are often contaminated by various noises and artifacts. Under
this scenario, most of these semi-automatic methods had a
large number of false positive and negative detections [6].
Furthermore, the morphology of the ABP waveform can vary
in response to different pathologic or physiologic stresses.
Therefore, reliable and accurate detection method is essential
for cardiac health monitoring applications.
In this paper, we propose an automatic peak detection
method based on Gaussian derivative filters and Shannon
energy envelogram. The method is much less sensitive to
parameter choices than conventional methods. The produces a
very robust and accurate detection performance on both clean
and noisy ABP signals. The key feature in our method is a
Gaussian derivative filtering to enhance the high-slope of APB
signal and to suppress baseline artifacts, a nonlinear signal processing to suppress the high-frequency noises and emphasize
magnitude of the low systolic peaks and a straightforward and
simple peak finding scheme to automatically determine timelocation of systolic peaks contained in the ABP signal.
The rest of this paper is organized as follows. Section
II describes the proposed systolic peak detection algorithm.
Section III provides the experimental results for a wide variety
of ABP signals from the standard MIT-BIH Polysomnographic
database. Finally, conclusions are drawn in Section IV.

g[m] = e 2

m = 1, 2, 3, ........, M

(1)

where M denotes the length of Gaussian kernel and denotes


the width spread (or standard deviation) of the Gaussian
kernel. The width of the Gaussian kernel is directly related

978-1-4799-2608-4/14/$31.00 2014 IEEE

42

Proceeding of the 2014 IEEE Students' Technology Symposium

Amplitude

(a)

Amplitude

(b)

0
1

10

12

14

10

12

14

Thresholded
Energy
Shannon
Energy (SE)

Energy

0.8
0.6
0.4
0.2

(e)

Smooth SE
Envelope

0.5

(d)

(f)

0.5

(c)

of the low-frequency artifacts meanwhile it emphasizes the


magnitude of the components of the systolic peaks. In this
work, adaptive thresholding is applied on the energy envelope
in order to remove the low-amplitudes of other components of
the ABP signal. The adaptive thresholding rule is defined as

0, e[n] <
(5)
eth [n] =
e[n], e[n] >

0.5

0
1
0.8
0.6
0.4
0.2

10

12

14

10

12

14

10

12

14

10

12

14

where the threshold is computed for each block. In the


thresholding process, the energy values e[n] smaller than the
threshold parameter are set to zero and other values are
retained. Here, the adaptive-threshold parameter for each
ABP segment is computed as


N
N
1 
1 
=
(e[n] e )2 and e =
e[n].
N n=1
N n=1

0.3
0.2
0.1
0

0.2
0.15
0.1
0.05
0

8
Time (second)

Fig. 2. Illustrates the outputs of different stages of the preprocessing scheme.


(a) Original ABP signal. (b) Output of the Gaussian derivative filtering stage.
(c) Output of the squaring operation stage. (d) Output adaptive thresholding
stage. (e) Output of Shannon energy computation stage. (f) Output of the
smoothing filtering stage.

to the value of . A larger value of produces a more larger


pulse width. Then, Gaussian derivative kernel is computed as
h[m] = g[m + 1] g[m]

m = 1, 2, 3, ......., M 1 (2)

The 101-point Gaussian kernel with spread = 12.5 is shown


in Fig. 1(a) and the corresponding Gaussian derivative function
is shown in Fig. 1(b). The filtered signal is obtained by taking
the convolution of the input ABP signal x[n] and Gaussian
derivative function h[n]. The filtered signal d[n] is computed
as


x[k] h[n k].
(3)
d[n] =
k=

The resulting filtered signal is shown in Fig. 2(b). By referring


the figure, it is observed that the Gaussian derivative filtering
approach can effectively remove the baseline wander and highfrequency noise contained in the input ABP signal. The filtered
signal has maximum magnitude around systolic peak regions
in the input ABP signal. To simplify the peak detection task,
the bipolar filtered signal is first converted into unipolar signal
before applying peak detection logic.
2) Squaring and Adaptive thresholding: The filtered signal
d[n] is first squared to obtain a positive-valued signal and then
adaptive thresholding is performed on the energy (or squarer)
values e[n]. The energy is computed as
e[n] = d2 [n].

(4)

The output of the squaring operation stage is shown in Fig.


2(c). The squaring operation further reduces the magnitude

TS14SPSAA02 355

where N is the number of samples. The duration of the test


ABP segment is 10 s. The output of the adaptive thresholding
stage is shown in Fig. 2(d). By comparing Figs. 2 (c) and
(d), we can observe that the thresholding process effectively
eliminates spurious noise spikes and tend to reduce the number
of false positive detections for noisy ABP signals.
3) Shannon energy computation and Smoothing: Since
squarer may bury the peaks of low-amplitude systolic peak
under the high-amplitude ones, we thus use Shannon energy
transformation which results in small deviations between the
successive R peaks. The effectiveness of this nonlinear transformation was well studied in the previous published work [7].
To compute Shannon energy, the thresholded energy signal
eth [n] is first normalized as
eth [n] =

eth [n]
maxN
n=1 (|eth [n]|)

(6)

and then the Shannon energy of the normalized signal is


computed as
(7)
s[n] =
e2th [n] log e2th [n].
The output of the Shannon energy stage is shown in Fig.
2(e). The multiple peaks in the Shannon energy envelope
may increase number of false positive detections. Therefore,
the envelope smoothing process is implemented in this work.
The Shannon energy values are smoothed using a linear zerophase filtering with a rectangular impulse response, h(k) of
length L. The smoothing filter is designed to provide smoothed
peaked-waves around systolic peak portions and to smooth
out the multiple peaks. The smoothness depends on the filter
length L, which is found empirically. In this work, window
length of 100 ms is chosen that is less than refractory period
of 300 ms, during which no new ABP pulse is found [1].
The envelope smoothing process can reduce number of false
positive detections. The output of the smoothing filter is shown
in Fig. 2(f). It can seen clearly that the locations of candidate
systolic peaks in the smooth Shannon energy envelope s[n]
correspond to approximate locations of the systolic peaks in
the ABP signal shown in Fig. 2(a). Therefore, in this proposed
method, those candidate peaks are first detected and then used

978-1-4799-2608-4/14/$31.00 2014 IEEE

43

(e)

(f)

(g)

Convolution
Output
Output
Preprocessor

10

15

10

15

0
1.5
1
0.5
0
0
0.3

10

15

10

15

10

15

10

15

10

15

0
0.1

Output ZC
Detector

(d)

0.1

Output of
Peak Finder

(c)

0
0
0.2
0.15
0.1
0.05

Approximate
Location

(b)

1
0.5

Final Output

(a)

Original

Proceeding of the 2014 IEEE Students' Technology Symposium

0.2
0.1

1
0.5
0
1
0.5
0

Fig. 3. Illustrates the outputs of different stages of the peak finding scheme.
(a) Original ABP signal. (b) Output of the preprocessing stage. (c) Output
of the convolution of smooth signal envelope and Gaussian derivative kernel.
(d) Detected negative zerocrossing points. (e) Output of peak finding stage.
(f) Detected approximate locations of peaks. (g) Detected peaks using peak
adjustment procedure.

as guides to find locations of true systolic peaks in the input


ABP signal.
B. Peak Finding Logic and Peak Location Adjustment Rule
The peak-finding scheme employs the peak-finding logic
used in our previous work [8]. We study the performance
of peak-finding logic which uses Gaussian derivative kernel
to locate the peaks contained in the smooth Shannon energy
envelope. The 101-point Gaussian kernel with spread = 12.5
is shown in Fig. 1(a) and the its first order Gaussian derivative
is shown in Fig. 1(b). By referring figures, we can observe
that the Gaussian kernel is symmetric about  M
2  + 1 and
its first-order Gaussian derivative is an antisymmetric. The
odd order derivatives are odd functions (antisymmetric around
zero). Fig. 1(a) clearly shows that the Gaussian kernel function
has peak at m =  M
2  + 1. Fig. 1(b) shows that the slope of
Gaussian derivative is positive and negative for 1 m  M
2

+
1

M,
respectively
and
also
its
slope
is
and  M
2
zero at m = M
,
where

denotes
the
floor
function.
By
2
using this property of Gaussian kernel, the peaks of the signal
envelope are located by finding convolution of the smooth
signal envelope s[n] and the Gaussian derivative d[n]. The
convolution between these two signals is computed as
z[n] =

d[k] s[n k].

(8)

k=

The output of the convolution stage is shown in Fig. 3(c) for


the smooth signal envelope shown in Fig. 3(b). By referring

TS14SPSAA02 355

these figures, it can be seen that the convolution output


has both positive and negative zero-crossings because of the
anti-symmetric nature of the Gaussian derivative kernel [8].
Fig. 3(c) clearly demonstrate that the negative zero-crossing
points correspond to the locations of the local peaks (or local
maxima) of the smmoth Shannon energy envelope s[n] shown
in Fig. 3(b). Therefore, the negative zero-crossing points are
detected and used as guides to find locations of true systolic
peaks in the original ABP signal shown in Fig. 3(a). The
negative zero-crossing points are detected by checking the sign
of the convolution output z[n] at time t(n) and t(n + 1). The
output of the negative zero-crossing point finder is shown in
Fig. 3(d). By comparing Figs. 3(d) and (e), it can be observed
that the peak-finding technique accurately finds time-locations
of candidate peaks contained in the smooth signal envelope.
The output shown in Fig. 3(f) demonstrate that the detected
time-locations are slightly shifted from the locations of true
systolic peaks of the input ABP signal. Therefore, the final
peak location adjustment procedure is used to correct the errors
of peak position shift. Firstly, a ABP segment centered around
a detected location (p[m] w2 ) is extracted from the original
ABP signal. Secondly, a simple algorithm finds a location of
maximum peak in the extracted ABP segment. Then, it stores
the new time-locations of real systolic peaks. This correction
procedure is repeated for all the detected locations p[m]. The
window size is 30 samples. The output of the peak location
adjustment stage is shown in Fig. 3(g). The experimental
results clearly show that the time-locations of systolic peaks
are accurately detected regardless of varying amplitudes and
morphologies of APB waveforms.
III. R ESULTS AND D ISCUSSION
The standard MIT-BIH Polysomnographic database
(http://www.physionet.org/physiobank/database/slpdb/)
is
used for testing and validating the performance of the
proposed detection method. The database contains 18
ABP recordings of variable duration [1]. By following the
guidelines given by the Association for the Advancement of
Medical Instrumentation (AAMI), two benchmark parameters
are used to assess the detection performance: sensitivity
and positive predictivity [16]. Sensitivity (Se) and positive
predictivity (+P) are defined as
TP
100 %
TP + FN
TP
Positive Predictivity (+P) =
100 %
TP + FP
Sensitivity (Se) =

(9)
(10)

From the experimental results, we therefore compute three


quantitative results: true positive (TP) when a systolic peak
is correctly detected, false negative (FN) when a systolic is
not detected, and false positive (FP) when an noise peak is
detected as systolic peak. By using these quantitative results,
we also calculate the detection error rate (DER) and overall

978-1-4799-2608-4/14/$31.00 2014 IEEE

44

Proceeding of the 2014 IEEE Students' Technology Symposium

TABLE I
P ERFORMANCE OF THE PROPOSED DETECTION METHOD IN FINDING
TIME - LOCATIONS OF SYSTOLIC PEAKS IN ABP SIGNAL .

Amplitude

Original ABP Signal

1
0.5
0
0

10

15
20
Output of Preprocessor

25

30

10

15
20
Output of Peak Finding Logic

25

30

10
15
20
25
Detected Peaks from PeakLocation Adjustment Stage

30

Negative
Zerocrossings

Smooth Signal
Envelope

0.2
0.15
0.1
0.05

0.1
0
0.1

Detected Peaks

1.5
ABP signal

Detected Peaks

Manual Annotations

1
0.5
0
0

10

15
Time (second)

20

25

30

Original ABP Signal

Amplitude

Fig. 4. Evaluation of the proposed systolic peak detection in MIT-BIH


polysomnographic database. (a) ABP segment from the record SLP/slp02. (b)
Output of the preprocessing stage. (c) Output of the peak-finding logic. (d)
Output of peak adjustment stage. Performance evaluation with respect to both
manual annotations (black asterisk) and systolic peaks detected (red circle)
by the proposed method. The test ABP signal includes baseline wander and
variable peak interval.

ABP
TP
FN
FP
DER
Se
+P Accuracy
Record (peaks) (peaks) (peaks) (%)
(%)
(%)
(%)
slp01a
3332
0
0
0
100
100
100
slp01b
3150
3
27
0.95 99.90 99.15
99.06
slp02a
4653
0
6
0.13 100.00 99.87
99.87
slp02b
3937
0
5
0.13 100.00 99.87
99.87
slp03
3339
0
8
0.24 100.00 99.76
99.76
slp04
4118
0
11
0.27 100.00 99.73
99.73
slp14
3178
0
14
0.44 100.00 99.56
99.56
slp16
4325
0
11
0.25 100.00 99.75
99.75
slp32
3467
0
17
0.49 100.00 99.51
99.51
slp37
4377
0
11
0.25 100.00 99.75
99.75
slp41
3486
67
75
4.07 98.11 97.89
96.09
slp45
3854
0
12
0.31 100.00 99.69
99.69
slp48
3598
0
8
0.22 100.00 99.78
99.78
slp59
3811
0
11
0.29 100.00 99.71
99.71
slp60
3730
0
15
0.40 100.00 99.60
99.60
slp61
3728
0
10
0.27 100.00 99.73
99.73
slp66
3477
0
14
0.40 100.00 99.60
99.60
slp67x
3495
0
9
0.26 100.00 99.74
99.74
Overall 67055
70
264
0.52 99.89 99.59
99.49

0.5
0

detection accuracy as

10
15
Output of Preprocessor

20

25

10
15
Output of Peak Finding Logic

20

25

10
15
Output of PeakLocation Adjustment Stage

20

25

Negative
Zerocrossing

Although many peak detection algorithms were reported,


the detection parameters of those algorithms are not provided
for reproducing their experimental results reported in the previously works. Therefore, we compare the detection accuracy
of the proposed method with expert manual annotations of
ABP signals provided in the standard database for algorithm
validation. First, we used the proposed method to annotate
the systolic peaks of the ABP signals. Then, the detected
systolic peaks are manually inspected for each of test ABP
signal. By comparing the manual annotations and methodbased annotations, we find a number of true positives, a
number of false negatives, and a number of false positives
for each ABP signal. The detection results of the proposed
method are summarized in Table I for all 18 recordings.
From the results shown in the table, we can observe that the
proposed detection method achieves an average sensitivity of
99.89%, a positive predictivity of 99.59% and DER of 0.52%.
From the detection results shown in Figs. 4-6, we can see
that the the proposed method has excellent performance in
detecting systolic peaks contained in the ABP signal including
baseline wanders, time-varying peak intervals and different
ABP morphologies.
Our method consists of several stages. Most stages are
computationally efficient enough to implement in a nearly real-

Detected Peaks

FP + FN
100 %. (11)
TP
TP
100 %. (12)
Accuracy (Acc) =
(TP + FP + FN)

Smooth Signal
Envelope

Detection Error Rate (DER) =

TS14SPSAA02 355

0.3
0.2
0.1
0

0.1
0
0.1

1.5

ABP signal

Detected Peaks

Manual Annotations

1
0.5
0
0

10

15

20

25

Time (second)

Fig. 5. Evaluation of the proposed systolic peak detection in MIT-BIH


polysomnographic database. (a) ABP segment from the record SLP/slp04. (b)
Output of the preprocessing stage. (c) Output of the peak-finding logic. (d)
Output of peak adjustment stage. Performance evaluation with respect to both
manual annotations (black asterisk) and systolic peaks detected (red circle) by
the proposed method. The test ABP signal includes long pause and different
ABP morphologies.

time block processing architecture. The proposed peak detection method is implemented using MATLAB programming.
The maximum and minimum computation time for the block
with duration of 10 s are 0.8580 s and 1.061 s, respectively that
is less than test block duration. The computational efficiency
can be further improved if we implement detection method
using C/C++ programming. In future directions, we study

978-1-4799-2608-4/14/$31.00 2014 IEEE

45

Proceeding of the 2014 IEEE Students' Technology Symposium

Amplitude

Original ABP Signal

0.5

Smooth Signal
Envelope

Negative
Zerocrossing

R EFERENCES

10
15
Output of Preprocessor

20

25

10
15
Output of Peak Finding Logic

20

25

10
15
Output of PeakLocation Adjustment Stage

20

25

0.2
0.15
0.1
0.05

0.1
0
0.1

Detected
Peaks

1.5

ABP signal

Detected Peaks

Manual Annotations

1
0.5
0

10

15

20

25

Time (second)

Fig. 6. Evaluation of the proposed systolic peak detection in MIT-BIH


polysomnographic database. (a) ABP segment from the record SLP/slp06. (b)
Output of the preprocessing stage. (c) Output of the peak-finding logic. (d)
Output of peak adjustment stage. Performance evaluation with respect to both
manual annotations (black asterisk) and systolic peaks detected (red circle)
by the proposed method.

the effectiveness of the method in real-time signal processing


hardware.
IV. C ONCLUSION
This paper presents an automatic detection method for
determining time-locations of systolic peaks contained in the
ABP signal. The proposed method comprises the Gaussian
derivative filtering to simultaneously emphasize high-slope
systolic peak portion and removes baseline wanders and
high-frequency noise. The non-linear peak amplification stage
magnifies the medium-amplitude systolic peaks and further
suppresses amplitude of low-frequency artifacts. The Gaussian
derivative kernel based peak finding scheme automatically
finds locations of peaks contained in the smooth signal envelope. The proposed detection method is tested and validated
using the standard MIT-BIH Polysomnographic database. The
method achieves an average sensitivity of 99.89% and positive
predictivity of 99.59% on test ABP datasets consisting of
67,125 beats. Our results show that this method is effective
in detecting the systolic peak in continuous blood pressure
signals including different ABP morphologies and various
kinds of artifices and noise. Unlike previous peak detection
methods, the proposed method is quite straightforward and
simple in the sense that it does not use search-back algorithms
with secondary thresholds. Thus, the proposed method can
be suitable for many cardiac health monitoring applications
including pulse oximetry, cardiac arrhythmia detection, and
cardiac output monitoring devices.

TS14SPSAA02 355

[1] W. Zong, T. Heldt, G.B. Moody, R.G. Mark, An open-source algorithm to


detect onset of arterial blood pressure pulses, Computers in Cardiology,
2003 , pp. 259-262, Sept. 2003.
[2] M. A. Navakatikyan, C. J. Barrett, G. A. Head, J. H. Ricketts, S. C.
Malpas, A real-time algorithm for the quantification of blood pressure
waveforms, IEEE Trans. Biomedical Engineering, vol. 49, no. 7, pp.
662-670, July 2002.
[3] B. N. Li, M. C. Dong, Mang I. Vai, On an automatic delineator for
arterial blood pressure waveforms, Biomedical Signal Processing and
Control, vol. 5, pp. 76-81, 2010.
[4] M. Aboy, J. McNames, and B. Goldstein, Automatic detection algorithm
of intracranial pressure waveform components, Proc. 23rd Int. Conf.
IEEE Engineering in Medicine and Biology Society, vol. 3, pp. 22312234, 2001.
[5] M. Aboy, C. Crespo, J. McNames, J. Bassale, and B. Goldstein, Automatic detection algorithm for physiologic pressure signal components,
Proc. 24th Int. Conf. IEEE Engineering in Medicine and Biology Society,
vol. 1, pp. 196-197, 2002.
[6] M. Aboy, J. McNames, T. Thong, D. Tsunami, M. S. Ellenby, B.
Goldstein, An automatic beat detection algorithm for pressure signals,
IEEE Trans. Biomedical Engineering, vol. 52, no. 10, pp. 1662-1670,
Oct. 2005.
[7] M. Sabarimalai Manikandan, and K. P. Soman, A novel method for
detecting R-peaks in electrocardiogram (ECG) signal, Biomedical Signal
Processing and Control, vol. 7, no. 2, pp. 118-128, March 2012.
[8] P. Kathirvel, M. Sabarimalai Manikandan, S. R. M. Prasanna, and
K. P. Soman An efficient R-peak detection based on new nonlinear
transformation and first-order Gaussian differentiator, Cardiovascular
Engineering and Technology, vol. 2, no. 4, pp.408-425, Dec. 2012.
[9] A. Donelli, J.R.C. Jansen, B. Hoeksel, P. Pedeferri, R. Hanania, J.
Bovelander, et al., Performance of a real-time dicrotic notch detection
and predication algorithm in arrhythmic human aortic pressure signals,
Journal of Clinical Monitoring and Computing,vol. 17 182-185, 2002.
[10] L. Xu, D. Zhang, K. Wang, N. Li, X. Wang, Baseline wander correlation in pulse waveforms using wavelet-based cascaded adaptive filter,
Computers in Biology and Medicine, vol.37, pp. 716-731, 2007.
[11] X. Liu, J. Liu, A novel beat detection method for blood pressure signals
using slope sum mean shift technology, IEEE Int. Conf. Biomedical
Engineering and Biotechnology, pp. 670-673, 2012.
[12] B. N. Li, B. B. Fu, M. C. Dong, Development of a mobile pulse
waveform analyzer for cardiovascular health monitoring, Computers in
Biology and Medicine, vol. 38, pp. 438-445, 2008.
[13] Z. T. Beattie, Algorithm for automatic beat detection of cardiovascular
pressure signals, Proc. 30th IEEE Int. Conf. EMBS, pp. 2594-2597, 2008.
[14] B. Nenova, I. Iliev, An automated algorithm for fast pulse wave
detection, International Journal of Bioautomation, vol. 14, no. 3, pp.
203-216, 2010.
[15] F. Scholkmann, J. Boss, M. Wolf, An efficient algorithm for automatic
peak detection in noisy periodic and quasi-periodic signals, Algorithms,
vol.5 ,pp. 588-603, 2012.
[16] ANSI/AAMI CE57: Testing and Reporting Performance Results of
Cardiac Rhythm and ST Segment Measurement Algorithms. (AAMI)
Recommended Practice/American National Standard, 1998. [Online].
Available: http://www.aami.org

978-1-4799-2608-4/14/$31.00 2014 IEEE

46