HEMODYNAMIC FEATURES EXTRACTION FROM A NEW

ARTERIAL PRESSURE WAVEFORM PROBE

V. G. Almeida, P. Santos, E. Figueiras, E. Borges, T. Pereira, J. Cardoso, C. Correia

Centro de Instrumentação (GEI-CI), University of Coimbra, Coimbra, Portugal

H. C. Pereira

Centro de Instrumentação (GEI-CI), University of Coimbra, Coimbra, Portugal

ISA- Intelligent Sensing Anywhere, Coimbra, Portugal

Keywords: Arterial pressure waveform, First derivative, Pulse contour analysis, Systolic peak, Dicrotic notch, Dicrotic

peak and reflection point.

Abstract: In this work, we discuss an algorithm that reliable and accurately identifies the prominent points of the

cardiac cycle: the systolic peak (SP), reflection point (RP), dicrotic notch, (DN) and dicrotic peak (DP). The

prominent point’s identifier algorithm (PPIA) action is based on the analysis a number of features of the

arterial pressure waveform and its first derivative, and is part of the fundamental software analysis pack for

a new piezoelectric probe designed to reproduce the arterial pressure waveform from the pulsatile activity

taken non-invasively at the vicinity of a superficial artery. The output PPIA is the coordinates (in time and

amplitude) of the above referred points. To assess the accuracy of the algorithm, a reference database of 173

pulses from eight volunteers, was established and the values yielded by the PPIA were compared to

annotations from a human expert engineer (HEE). The quality of the results is statistically quantified either

in time as in amplitude. Average values of 4.20% for error, 99.09% for sensitivity and 96.77% for positive

predictive value were found to be associated to time information while amplitude yields averages of 2.68%,

99.08% and 98.22%, respectively, for the same parameters.

1 INTRODUCTION

Over the last years, increasing attention has been

paid to the effect of arterial stiffness, the measure of

rigidity of arteries (Mackenzie et al, 2002), on the

development of cardiovascular (CV) diseases

(Laurent et al, 2006).

The study of non-invasive methods that address

this problem, using devices capable of precisely

assessing the arterial pressure waveform (APW)

remains a capital issue that mobilizes the interest of

researchers. Pulse diagnosis has proved to be a

convenient, inexpensive and painless diagnosis

method; however experience has also shown that, in

order to obtain reliable results, it may require

practitioners with considerable training and

experience (Avolio et al, 2010).

Historically, the cuff sphygmomanometer,

universally used by clinicians since the beginning of

the 20th century, was the first device to quantify a

part of the information contained in the APW and

conquered a (still) irreplaceable role in general

clinical practice. The measurement of pulse

pressure, however, is the simplest surrogate measure

of arterial stiffness.

Nevertheless, the APW contains a vast amount of

pathophysiological information about the

cardiovascular condition that is concealed in its

morphology. Many factors can determine the

contour of the APW: the blood volume ejected from

the heart, the distensibility of the arterial vessel,

runoff of the blood to the periphery, rate of the

velocity of blood in the vessels and the vascular

properties of the vessels (Oppenheim et al, 1995).

Although different technologies have been put to

the service of this major endeavour, a short review

of recent literature leads us to the conclusion that

applanation tonometry (e.g. Sphygmocor® from

195

G. Almeida V., Santos P., Figueiras E., Borges E., Pereira T., Cardoso J., Correia C. and C. Pereira H..

HEMODYNAMIC FEATURES EXTRACTION FROM A NEW ARTERIAL PRESSURE WAVEFORM PROBE.

DOI: 10.5220/0003148901950200

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2011), pages 195-200

ISBN: 978-989-8425-35-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

AtCor, West Ryde, Australia and PulsePen® from

DiaTecne, Milan, Italy) remains the golden standard,

only second to the invasive catheterization.

As for the technologies based on

electromechanic sensors, PZ probes have been

studied either for assessing timing parameters of the

APW (McLaughlin et al 2003) or for monitoring

APW at the radial artery (Clemente et al 2010).

They also are in the basis of some commercial

instruments dedicated to PWV assessment (e.g.

Complior® from Artech-Medical, Pantin, France).

A number of beat detectors have been reported

(Donelli et al, 2002 and Oppenheim et al, 1995) for

characterizing APW in the literature; however the

full delineation of APW remains a capital issue that

motivates our investigation.

In a previous study carried out by our group, a

piezoelectric probe for non-invasive arterial pressure

waveform reproduction was developed.

In this paper an automatic delineator of the

characteristic features of the APW, meant to works

in articulation with the probe, is proposed. Its action

relies on the coordinates (time vs. amplitude) of the

4 prominent points that can be identified in the

APW: systolic peak (SP), reflection point (RP),

dicrotic peak (DP) and dicrotic notch (DN).

2 PIEZOELECTRIC PROBE

In a previous study, a piezoelectric probe capable of

recovering the arterial pressure was developed.

This device delivers as output, a high fidelity

replica of the APW, free from baseline drift due to

the action of a BLR circuit that avoids the need of

offline removal algorithms for this purpose (Xu et

al, 2007).

The elimination of baseline drift by the BLR

circuit consists in forcing the foot of systolic pulses

to start close to zero, without affecting the shape of

the signal.

This action is triggered only at the end of the

cardiac pulse, just before the starting of a new pulse.

Figure1 illustrates the main components of our

pulse acquisition system: The PZ probe, left

photograph, is held by a collar and placed in the

carotid artery site for the in vivo data acquisition.

Signal conditioning and data acquisition (USB

6009 from National Instruments are enclosed in the

electronics box (mid panel).

A personal computer running a dedicated

software package with PPIA, shown in the right

panel, completes the system.

3 METHODOLOGY

Once the foundations of our delineator are defined

we assess the capability of the algorithm in

identifying various prominent points from which the

clinically relevant waveform features are computed.

To attain this goal, a small universe of eight

healthy volunteers was organized as the seed of a

larger data base of cardiac pulses to be build in the

follow-up of this work.

The study protocol was approved by the ethical

committee of Centro Cirúrgico de Coimbra,

Portugal. All subjects were volunteers that

previously granted a written informed consent.

A set of pulses was acquired from each volunteer

followed by a pulse by pulse segmentation, routine,

as depicted in figure 2. The onset of the waveform is

identified in the signal conditioning circuit

according the heart rate value.

3.1 Physiologic Foundation

Four prominent points are generally identifiable in a

typical APW profile: SP, RP, DN and DP. Their

physiological origin is briefly recalled in the

following paragraphs.

The arterial pressure wave propagates along the

arterial tree and, as its branches change in diameter

Figure 1: The main components of our non-invasive system for hemodynamic analysis. (a) PZ probe; (b) Acquisition box;

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

196

Figure 3: Flowchart of the PPIA.

0 2 4 6 8 10 12 14 16 1

Time (s)

0 0.2 0.4 0.6 0.8 1

0.5

Pulses: 10

Time (s )

A.U.

Figure 2: Set of carotid pulses and the corresponding pulse

segments.

and stiffness, reflections of a portion of the original

energy is sent back towards the heart.

The waves reflected from the periphery,

superimpose to the forward wave originating a

visible change in the APW profile which determines

the RP.

Further to RP, the identification of the SP,

simply defined as the point of maximum systolic

pressure, allows the determination of a very

important hemodynamic parameter that denotes the

increase of pressure imparted by the reflected wave -

the Augmentation Index (AI), (Almeida et al, 2010).

The aortic DN pulse is associated to the effect of

the aortic valve closure when a small portion of

ejected blood moves back to the left ventricle

(Donelli et al 2002). This feature must be identified

by the delineator in order to accurately determine the

end of ventricular ejection, which is necessary for

cardiac output determinations.

The pulse profile is characterized by two peaks,

SP and the DP, one valley, DN, and one inflection,

RP. The PPIA must deal with the specification of

each one of these shapes in order to produce an

accurate identification of all four prominent points.

3.2 The Algorithm

As mentioned before, the main purpose of the

algorithm is the identification (in time and relative

amplitude) of the prominent points of the cardiac

cycle and make them available for analysis in an

understandable, clear way.

Flowchart of figure 3 depicts the main sequence

of operations performed by our Prominent Points

Identifier Algorithm (PPIA). The delineator is based

on the combined analysis of APW and its first order

derivative.

After acquiring a few pulses, around 10 in the

example of figure 2, the pulses are subject to a

segmentation process and normalised to the

diastolic-systolic pressure interval.

The use of the first derivative curve justifies that,

immediately following acquisition, a low pass filter

is used to suppress high frequency noise that would,

otherwise, turn the first derivative unusable. The

50Hz turn point of the filter ensures the interesting

range of frequencies to ensure the visibility of the

points. Figure 4 shows one pulse with its prominent

points identified by the PPIA. As an example, the AI

of the corresponding volunteer is also derived.

3.2.1 Systolic Peak (SP)

The SP identification is carried out using a local

extreme identification routine (Billauer, 2008).

This routine detects a maximum only if the value

of the previous minimum differs by a selectable

minimum amount, referred to as delta.

HEMODYNAMIC FEATURES EXTRACTION FROM A NEW ARTERIAL PRESSURE WAVEFORM PROBE

197

0 0.1 0.2 0.3 0.4 0.5 0.6

0.2

0.4

0.6

0.8

A.U.

0 0.1 0.2 0.3 0.4 0.5 0.6

-1

Time (s )

A.U.





Figure 4: Example of type A pulse waveform and its

prominent points identified by PPIA.

Since the cardiac cycle shows a variable number

of local maxima (2 to 4 depending on wave type and

artefacts), the routine must run repeatedly, with

different values of delta, to ensure the correct

identification of the SP.

3.2.2 Dicrotic Notch (DN) and Dicrotic Peak

(DP)

The identification of the SP combined with the zero

crossing values of the first order derivative is used to

identify the dicrotic points (notch and peak) that

must be located between the SP and the end of the

pulse.

The oscillations that inevitably occur can mask

the true DN and DP values. To prevent this, zero

crossing information must be combined with the

amplitude values of the APW pulse.

3.2.3 Waveform Type

According to the criteria proposed by Murgo et al

(1980) the pressure waveform can be classified into

one of four types (A, B, C, D) depending on the

location of the reflected wave, as shown in table 1.

To achieve this, the algorithm analyses the

waveforms deriving the number of maximum peaks

of the 1st order derivative of the APW taking as

reference the localization of the SP.

3.2.4 Reflection Point (RP)

The RP identification is carried out in three steps:

the localization of candidates, the elimination of

oscillations and the comparison with the APW

amplitude.

Table 1: Classification proposed by Murgo et al (1980).

3.2.5 Augmentation Index (AI)

Augmentation index (AI) is one of the most widely

used indices to quantify the arterial stiffness, based

on the measurement of the strength of the reflected

wave relative to the total pressure waveform.

−

=±

−

(1)

Where P

is the APW peak pressure, P

its

pressure at the inflection point and P

is the diastolic

blood pressure.

Types A and B show positive values of AI

denoting high arterial stiffness situations, while in

type C waveforms the negative values of AI is

characteristic of a relatively elastic and healthier

artery condition.

The key point in estimating AI is the correct

identification of the RP, to allow the subsequent

assessment of the relative augmentation that the

reflected wave imparts to the pressure waveform.

Several methods have been described in the

literature to evaluate this parameter. The algorithm

developed in this work is compared with the

classification of a human expert engineer (HEE) to

understand the effect of wrong RP identifications in

AI calculation.

3.3 Evaluation Procedure

To acquire the pulses we used a previously

developed pulse probe and organised a small

universe of eight volunteers from which a reference

database was built up.

In a first step, we used the algorithm described

above to identify the prominent points - SP, DN, DP

and RP. A human expert engineer (HEE) carefully

inspected each APW pulse and manually annotates

the same points.

All points are classified in the three classic types:

true positive (TP) and, just for the discrepant ones,

false negative (FN) and false positive (FP). FNs

occur only when the point cannot be identified.

The classification takes the HEE results as

reference and uses the same 8 ms threshold adopted

by others (Li et al, 2010 and Zong et al, 2003).

Type A

The RP occurs before SP

Type B

The RP occurs shortly before of the SP (a

threshold must be defined)

Type C

The RP occurs after SP

Type D

The inflection point cannot be observed

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

198

y=0.736x+0.208

0.65

0.75

0.85

0.95

0.65 0.75 0.85 0.95

CSAAmplitude(V)

ManualAnnotationsAmplitude(V)

y=1.003x+1.398

110

130

150

170

190

210

230

250

70 120 170 220

CSATime(ms)

ManualAnnotationsTime(ms)

y=0.997x+0.533

130

180

230

280

330

80 130 180 230 280 330

CSA Time (ms)

ManualAnnotationsTime(ms)

0.6 0.7 0.8 0.9

-0.2

-0.1

0.1

0.2

Me an (HA and CSA) [Amp(V)]

HA-CSA [Amp(V)]

1.96 sd

-1.96 sd

mean

100 150 200 250 300

-20

-10

Mean (HA and CSA) [Time(s)]

HA-CSA [Time(s)]

1.96 sd

-1.96 s d

mean

0 50 100 150 200

-50

-30

-10

Mean (HA and CSA) [Time(s)]

HA-CSA [Time(s)]

1.96 sd

-1.96 sd

mean

Figure 5: Agreement between the SP and RP obtained via PPIA and HEE (a) time information of SP and RP; (b) relative

amplitude of RP. Regression plots (top) and Bland-Altman plots (bottom).

Sensitivity (S), positive predictive value (P

) and

error of the PPIA algorithm shown in table I are now

computed according to:

TP FN

(2)

TP FP

(3)

FP FN

error

TP FP

(4)

An identical procedure is carried out in the

amplitude sense, using a 1% of pulse maximum,

threshold.

4 RESULTS

Tables 2 and 3 show the sensitivity, positive

predictive value and error of the algorithm for the

time and amplitude information, respectively.

Statistical analysis was performed using

Microsoft Excel® 2007 and SPSS® version 18.

As expected, RP shows the worst performance

indices due to its inherent detection difficulty, either

for the HEE as for the PPIA.

The three last columns of table 2 yield average

values of 4.20%, 99.09% and 96.77 % for error, S

and P

, respectively.

The same study applied to the amplitude

information of table 3 yields averages of 2.68%,

99.08% and 98.22 %.

Table 2: Validation of PPIA performance (time

information) compared to HEE annotations.

Table 3: Validation of PPIA performance (amplitude

information) compared to HEE annotations.

Figure 5 a) shows the correlation between the

reference HEE and PPIA values (SP and RP

analysis), as well the corresponding Bland - Altman

plots.

The straight line fittings for SP and RP show an

excellent correlation, where R

is the

square of the

sample correlation coefficient, between both

estimates (R

=0.996 and R

=0.907, respectively).

The PPIA, in average, overestimates the SP by

0.009 ms and underestimates the RP by 0.99 ms. To

explain the 100 fold factor between the estimation

errors for SP and RP one should bear in mind that, in

the APW curve, SP is a peak while RP is an

inflection. This circumstance makes the SP

estimation much easier for the HEE as well as for

the PPIA. Should the data base be large enough and

Pulses TP FP FN Error (%) S (%)

(%)

173 171 2 0 1.15 100 98.84

173 165 8 0 4.62 100 95.37

173 169 4 0 2.31 100 97.69

173 159 8 6 8.38 96.36 95.20

Pulses TP FP FN Error (%) S (%)

(%)

173 173 0 0 0 100 100.00

173 171 2 0 1.15 100 98.84

173 172 1 0 0.58 100 99.42

173 158 9 6 8.98 96.34 94.61

HEMODYNAMIC FEATURES EXTRACTION FROM A NEW ARTERIAL PRESSURE WAVEFORM PROBE

199

the SP error would tend to zero while RP would

eventually converge to a minimum, non-zero, biased

level. This, however, remains a hypothesis to be

demonstrated in the follow-up of this work.

Another interesting consequence of the nature of

SP and RP is that amplitude errors summarized in

table 3 are much lower than the corresponding time

errors (table 2). In fact as they are both associated to

a peak and an inflection, the first derivative of the

APW curve shows close-to-zero values in their

vicinity, hence the small resulting estimation error.

This fact is visible in figure 5 b) that shows the

correlation between the reference HEE and PPIA

values for RP analysis, as well the corresponding

Bland - Altman plot.

In fact, as can be seen in figure 6, AI error,

measured by the difference between AI values from

HEE and PPIA, amounts to an average value of just

0.53 %.

Figure 6: The relationships between AI obtained from

PPIA and HEE.

Table 4: Statistics information of measurements depicted

in figure 6.

5 CONCLUSIONS

We described new automatic feature extraction

algorithm capable of detecting the prominent points

of the APW: SP, DN, DP and RP. This algorithm is

a fundamental part of the automatic analysis tool in

our non-invasive system for hemodynamic analysis.

The clinical use of our probe, however, will still

require a medical oriented, multicenter study

including comparison with standard methods, e.g.

applanation tonometry and catheter collected data.

The need for a larger data base has also emerged as

the only means of attaining the necessary levels of

confidence.

ACKNOWLEDGEMENTS

We acknowledge support from Fundação para a

Ciência e a Tecnologia for funding (PTDC/SAU-

BEB/100650/2008) and SFRH/BD/61356/2009) and

from ISA, Intelligent Sensing Anywhere.

REFERENCES

Almeida V., Pereira T., Borges E., Figueiras E., Cardoso

J., Correia C., Pereira H. C., Malaquias J. L. and

Simões J. B., 2009. Synthesized cardiac waveform in

the evaluation of augmentation index algorithms. In

IEEE EMB, Proceedings of the 3rd International Joint

Conference on Biomedical Engineering Systems and

Technologies (BIOSTEC 2010). Valencia, Spain 20-

23 January 2010.

Avolio A. P., Butlin M., and Walsh A., 2010. Arterial

blood pressure measurement and pulse wave analysis -

their role in enhancing cardiovascular assessment.

Physiol. Meas. 31, R1–R47.

Billauer E.,2008. Peak detection using MATLAB.

http://www.billauer.co.il/peakdet.html. [Accessed 13

July 2010]

Clemente F, Arpaia P and Cimmino P 2010 A piezo-film-

based measurement system for global haemodynamic

assessment Physiol. Meas. 31 697–714.

Laurent S., Cockcroft J., Van Bortel L., Boutouyrie P.,

Giannattasio C., Hayoz D., Pannier B., Vlachopoulos

C., Wilkinson I., and Struijker-Boudier H., 2006.

Expert consensus document on arterial stiffness:

methodological issues and clinical applications.

European Heart Journal, 27, 2588–2605.

Li, B. N., Dong, M. C., Vai, M. I., 2010. On an automatic

delineator for arterial blood pressure waveforms.

Biomedical Signal Processing and Control, 5, 76-81.

Oppenheim, M. J., Sittig, D. F., 1995. An innovative

dicrotic notch detection algorithm which combines

rule-based logic with digital signal processing

techniques. Computers and biomedical research, 28,

154-170.

Mackenzie I. S., Wilkinson I. B. and Cockroft J. R., 2002.

Assessment of arterial stiffness in clinical practice. Q J

Med., 95, 67–74.

Robinson L., 1961 Reduction of Baseline Shift in Pulse-

Amplitude Measurements. Rev. Sci. Instr. 32, 1057.

Xu, L., Zhang, D., Wang, K., Li, N., Wang, X., 2007.

Baseline wander correction in pulse waveforms using

wavelet-based cascaded adaptative filter. Computers in

Biology and medicine, 37, 716-731.

Zong, W., Heldt, T., Moody, G.B., Mark, R.G., 2003. An

open source algorithm to detect onset of arterial blood

pressure pulses. Computers in Cardiology, 30, 259-

262.

y=0.961x+0.962

R²=0.948

‐30

‐20

‐10

‐30 ‐20 ‐10 0 10203040

AI_PPIA (%)

AI_HEE (%)

Minimum

(%)

Maximum

(%)

Mean

(%)

Std.

Deviation (%)

Error 167 0.00 19.09 0.53 2.47

Descriptive Statistics

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

200