Cuff-less Calibration-free Blood Pressure Estimation under

Ambulatory Environment using Pulse Wave Velocity and

Photoplethysmogram Signals

Haruyuki Sanuki

, Rui Fukui

, Tsukasa Inajima

and Shin'ichi Warisawa

Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa-shi, Chiba 277-8563, Japan

The University of Tokyo Hospital, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8655, Japan

Keywords: Blood Pressure Monitoring, Pulse Wave Velocity, Photoplethysmogram, Electrocardiogram.

Abstract: This paper presents a blood pressure estimation method based on pulse wave velocity (PWV). Although

there are a variety of methods based on PWV to estimate blood pressure, most of them require calibration

per patient, and the patient has to remain still. The goal of our research is to develop a calibration-free blood

pressure estimation method that is applicable not only during rest but also during exercise. To accomplish

our goal, we extracted properties of blood vessels from photoplethysmogram (PPG) signals, and compared

several regression models, such as the deductive model based on blood vessel physics equation, and the

inductive model based on machine learning. Twenty-four participants performed exercise, measuring blood

pressure, electrocardiogram (ECG) and PPG. The best result showed that the mean error for the estimated

systolic blood pressure (SBP) against cuff-based blood pressure was 0.18 ± 8.68 mmHg. Although there

was not a big difference between the regression models, PWV and Augmentation Index are effective

features to estimate SBP. In addition to this, Heart Rate was effective only for the young men, and height

ratio of c-wave to a-wave of acceleration pulse wave might be effective for elderly men. These results

suggest that our proposed method has the potential for cuff-less calibration-free blood pressure estimation

which include measurements during rest and exercise．

1 INTRODUCTION

In recent years, the number of hypertension patients

has increased, and around 40% of adults aged 25 and

over were estimated to have hypertension (World

Health Organization, 2014). Hypertension can lead

to various diseases such as a life-threatening heart

disease, cardiovascular diseases (CVDs), and renal

insufficiency. Since most people are not aware of

their hypertension, they are not treated in time.

Monitoring one's blood pressure is required for the

prevention, early detection, and early recovery of

hypertension.

However, single blood pressure measurement is

the mainstream in hospitals or at home. It is difficult

to monitor the changes of blood pressure, especially

indicators like short-term changes and changes

during the day, which are important to diagnose a

patient’s body. Moreover, white-coat hypertension,

which leads to high blood pressure when measured

in the medical environment, could cause

misdiagnosis. To diagnose and treat such patients

properly, continuous blood pressure monitoring is

required.

Nowadays, Ambulatory Blood Pressure

Monitoring (ABPM) is used for continuous blood

pressure measurement. Figure 1 shows ABPM

equipment. ABPM measures blood pressure by a

cuff every 15 minutes or so. It is rather

uncomfortable and the patient has to remain still.

Figure 1: ABPM equipment.

The method based on pulse wave velocity (PWV)

has been intensively studied because of its potential

Sanuki H., Fukui R., Inajima T. and Warisawa S.

Cuff-less Calibration-free Blood Pressure Estimation under Ambulatory Environment using Pulse Wave Velocity and Photoplethysmogram Signals.

DOI: 10.5220/0006112500420048

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 42-48

ISBN: 978-989-758-212-7

for tracking blood pressure change continuously

without a cuff (Mukkamala et al., 2015). PWV is the

velocity of an arterial pulse propagating through the

arterial wall and can easily be calculated from Pulse

Transit Time (PTT). PTT is the time interval

between an R-wave peak of electrocardiogram

(ECG) and a particular point of

photoplethysmogram (PPG). PWV is obtained by

dividing distance from the heart to a particular

peripheral site by PTT.

Based on previous researches, a formula which

takes continuity equation and Navier-Stokes

equation estimates systolic blood pressure (SBP).

The formula is as follows (Lopez, 2010, Inajima,

2012).

  













(1)





is the pulse wave velocity, and coefficients b

and b

are parameters related to individual blood

vessel properties. Traditionally, the coefficients are

calibrated by measuring blood pressure and PWV

beforehand.

Gesche et al. established a model with PWV to

estimate systolic blood pressure (SBP) during

exercise with initial calibration, and the standard

deviation of estimation error (SD) was 10.1 mmHg

(Gesche, 2012). Ding et al. established a model with

PWV and photoplethysmogram intensity ratio to

estimate blood pressure during rest with initial

calibration, and SD was 5.21 mmHg for SBP and

4.06 mmHg for diastolic blood pressure (DBP) (Ding,

2015). Kauchuee et al. investigated the relationship

between PTT and blood pressure and found that non-

linear models are better than linear models.

Kachuee’s model without calibration during rest

achieved 16.17 mmHg for SBP and 8.45 mmHg for

DBP (Kachuee, 2015).

Although there are a variety of methods based on

PWV to estimate blood pressure, the application of

the PWV-based method has several problems. First,

individual blood vessel properties differ from person

to person. Most of the methods, therefore, require

calibration per person. Secondly, current methods

still lack application during exercise. Thirdly, there

is not enough accuracy of blood pressure

measurement based on PWV for medical use.

The objective of our study is to establish a

calibration-free blood pressure estimation method

based on PWV during rest and exercise. To achieve

the objective, we extracted properties of blood

vessels from photoplethysmogram (PPG) signals and

compared several regression models, such as the

deductive model based on blood vessel physics

equation and the inductive model based on machine

learning.

In this research, we focused on SBP that is

superior to Diastolic blood pressure as a predictor of

CVDs (Mourad, 2008).

This paper is organized as follows. Chapter 2

shows an overview of our methodology including

peak detection method, feature extraction, regression

models and evaluation method. Chapter 3 explains

the experiment, and Chapter 4 describes the result.

Lastly, Chapter 5 is the conclusion of this research

and future perspectives.

2 METHOD

Our method estimates SBP by using ECG and PPG.

The method is demonstrated in Figure 2. While

extracting features from ECG and PPG signals, we

use peak detection to extract features automatically.

Therefore, we first explain the peak detection

method before feature extraction.

Figure 2: Overview of the method to estimate SBP.

2.1 Peak Detection

In order to extract features from ECG and PPG

signals automatically, we need to build a robust

pattern-matching model. Therefore, we applied

Continuous Wavelet Transform (CWT), which is

widely used for R spike detection (Legarreta, 2005)

and PPG waveform analysis (Fan, 2011). The

Mexican Hat wavelet was selected as the mother

wavelet, because of its similarity with the ECG and

PPG signals (Daubechies, 1992). We found optimal

scales for each signal using annotations provided on

small data. Figure 3 shows an R spike detection of

ECG signal, and Figure 4 shows a foot point

detection of PPG signal.

Cuff-less Calibration-free Blood Pressure Estimation under Ambulatory Environment using Pulse Wave Velocity and Photoplethysmogram

Signals

Figure 3: R spike detection of ECG signal.

Figure 4: Foot point detection of PPG signal.

2.2 Feature Extraction

We extract Heart Rate (HR) from ECG signal, PPG

features from PPG signal, and PWV from ECG and

PPG signals.

2.2.1 PWV and HR

PWV was calculated by dividing the participant's

height by time interval between R-wave peak of

ECG and three points of PPG, which are the steepest

slope of the corresponding upstroke (PWV

), the

maximum point (PWV

), and the minimum point

(PWV

), as shown in Figure 5. HR is calculated by

the time interval between the nearest R spikes.

Figure 5: Definition of each PWV.

2.2.2 PPG Features

The PPG signal reflects the blood volume of the

vessel, measured by red, green or infrared light,

which is irradiated into the tissue and is absorbed or

reflected. The features extracted from PPG signal

have relationships with blood vessel conditions

(Elgendi, 2012). Most researches extract features

from velocity pulse waves (first derivative) and

acceleration pulse waves (second derivative) of the

PPG signal to interpret the original PPG signal

(Takazawa, 1998). In this research, features are

extracted from volume pulse waves and acceleration

pulse waves.

In volume pulse waves, Inflection Point Area

Ratio (IPA), Augmentation Index (AI), Crest Time

(CT), and Large Artery Stiffness Index (LASI) are

extracted.

 Inflection Point Area Ratio (IPA): IPA is the

ratio of the four pulse areas between the

selected points, S1, S2, S3 and S4, which are

shown in Figure 6. IPA is used as an indicator

of the total peripheral resistance (Wang, 2009).

In this research, it is proposed to use the ratio

of S2, S3, and S4 to S1.

Figure 6: Definition of S1, S2, S3 and S4.

 Augmentation Index (AI): AI is the ratio of the

height of the diastolic peak to height of the

systolic peak (Figure 7). AI is a measure of the

wave reflection and arterial stiffness

(Takazawa, 1998).

 Crest Time (CT): CT is the time interval

between the foot point and the systolic peak

(Figure 7). CT is an important feature for

classifying cardiovascular diseases (Alty,

2007).

 Large Artery Stiffness Index (LASI): LASI is

the time interval between the systolic peak and

the diastolic peak (Figure 7). LASI is related

to large artery stiffness (Elgendi, 2012).

Figure 7: Definition of AI, CT, and LASI.

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

As Figure 8 shows, an acceleration pulse wave

includes five component waves, namely a-wave, b-

wave, c-wave, d-wave and e-wave. The type of

acceleration pulse waveform varies depending on the

blood vessel conditions. The height ratios and the

time intervals of each wave are extracted.

 Height ratio b/a, c/a, d/a, e/a; each height ratio

reflects arterial stiffness. If arterial stiffness

increased, b/a would increase and c/a, d/a, e/a

would decrease (Takazawa, 1998).

 Time interval between a-wave and b-wave

(a_b), a-wave and c-wave (a_c), a-wave and d-

wave (a_d), a-wave and e-wave (a_e). Time

interval of each wave describes acceleration

pulse waveform.

Figure 8: Acceleration pulse wave waveform. Waveform

differs depending on the vascular status (Homma, 1992).

In this research, features are selected in each

regression model by greedy forward selection

(Caruana, 1994).

2.3 Regression Model

In this research, two main approaches are taken to

choose a better regression model. One is the

deductive model based on blood vessel physics

equation, which is represented as Eq. (1), while the

other is the inductive model based on machine

learning.

2.3.1 Model based on Physics Equation

As shown in Eq. (2), the extracted features determine

individual blood vessel condition parameters b

and

. We use PWV

as pulse wave velocity in Eq. (2).

 























⋯





























⋯)

(2)





is the partial regression coefficient and 



is the

extracted feature. We named this model as LR.

2.3.2 Model based on Machine Learning

We use the inductive model based on machine

learning, not using a hypothesis but learning only

from the data.

Three regression models are selected, K-Nearest

Neighbours (KNN), Random Forest (RF) and Linear

Support Vector Machine (SVM).

 K-Nearest Neighbours (KNN): KNN is the

simplest nonparametric decision procedure,

and predicts a sample data by using its K-

nearest neighbors (Cover, 1967).

 Random Forest (RF): RF is a combination of

tree predictors, such that each tree depends on

the values of random features sampled

independently and with the same distribution

for all trees in the forest (Breiman, 2001).

 Linear Support Vector Machine (SVM): SVM

is the algorithm that maximizes the margin

between the training patterns and the decision

boundary, and is widely used for classification

and regression problems (B. E. Boser, 1992).

Each hyper-parameter is optimized by cross

validation.

2.4 Evaluation

As shown in Figure 9, in order to evaluate the

accuracy without any individual dependency, each

participant's data is taken out as test data in turn, and

is evaluated with the data of remaining train data.

Figure 9: Cross-validation for independent validation.

Though it is better to split data into three sets, which

are training data, validation data and test data, we

will split data into two sets, train data and test data,

because of the small sample size.

3 EXPERIMENTS

We conducted experiments on 18 young men

(22.9±1.2 years) and six elderly men (43.3±9.3

Cuff-less Calibration-free Blood Pressure Estimation under Ambulatory Environment using Pulse Wave Velocity and Photoplethysmogram

Signals

years). All participants underwent an exercise test for

28 minutes on a bicycle ergometer, and four

minutes of rest before and after the test, acquiring

ECG, PPG and SBP by sphygmomanometer. The

timing of load increase and decrease is shown in

Figure 10. The load was adjusted corresponding to

their exercise capacity.

Participants wear the ECG sensor surrounding

the heart, PPG sensor at the right index finger, and a

sphygmomanometer with a cuff (Tango M2 from

SunTech Medical) on the left arm. SBP was

measured every two minutes by a cuff and the

sampling rate of ECG and PPG measurements were

both at 1 kHz. ECG and PPG signals are sampled at

the same time with same microcomputer that would

guarantee the synchronization. As a reference,

participants wear PPG sensor at the right earlobe and

finger cuff (ClearSight from Edwards Lifesciences)

on the right middle finger. Figure 11 shows a

schematic of the experimental set-up. All the

participants gave their informed consent prior to the

experiment.

Figure 10: Exercise weight transition.

Figure 11: Schematic of the experimental set-up.

4 RESULTS

The SBP distribution histogram is shown in Figure

12, including 341 SBP measurements. The mean

SBP measured by the sphygmomanometer was

128.05±16.90 mmHg.

Figure 12: Histogram of SBP measurements.

We defined two groups, Group A only contains

young men and Group B contains both, young and

elderly men. The reason for not grouping elderly

men is that the sample size was not large enough.

The results from various regression models are

shown in Table 1. Although KNN showed the best

regression model for estimating SBP, there was not a

big difference between the deductive model based on

blood vessel physics equation and the inductive

model based on machine learning.

Table 1: Standard Deviation of estimation error for each

regression model.

LR KNN SVM RF

Group A [mmHg] 8.68 8.65 8.74 9.28

Group B [mmHg] 8.79 8.68 8.75 9.20

Table 2 shows the feature subset that is selected by

each regression model and Eq. (3) shows the

deductive model based on the blood vessel physics

equation. While PWV is an important feature in each

group, as expected, AI also appeared to be an

important feature. HR is only effective for young

men, and the height ratio of c-wave to a-wave of

acceleration pulse wave might be effective for

elderly men.

Group A

 



13.0  1.83/1









3.3+125.9

Group B

 



0.3/  11.2









4.8+128.0

(3)

Hereinafter, KNN is selected as the most accurate in

the regression model evaluation, according to Table

2. Figure 13 shows the plot of cuff-based SBP and

estimated SBP of Group B, and the correlation

coefficient was r=0.86 (p-value<0.01). Figure 14

gives the Bland-Altman plot, comparing for the

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

Table 2: The feature subset that is selected by greedy forward selection in each regression model. Group A is the young

men group and Group B is the young and elderly men group. ○ is the selected feature.

PWV

IPA

(S2/S1, S3/S1, S4/S1)

AI CT LASI

Height ratio

(b/a, c/a, d/a, e/a)

Time interval

(a_b, a_c, a_d, a_e)

HR × PWV

S3/S1 × PWV

○

KNN

○

a_b

SVM

○

a_c, a_e

○

c/a × PWV

KNN

○

e/a

SVM

○

c/a a_e

○

c/a

performance of the proposed method with the cuff-

based measurements of Group B. A total of 94.73%

of the measurements lies in the limits of agreement

(1.96×SD).

Figure 13: Correlation plot of SBP.

Figure 14: Bland-Altman plot of SBP.

Our proposed model for Group B, which contains

young and elderly men, showed that the mean error

was 0.18±8.68 mmHg, and the mean absolute

difference (MAD) was 6.93 mmHg, achieving grade

C as IEEE standard requirement (IEEE Standards

Association, 2014).

5 DISCUSSIONS

Previous researches show an effective model with

initial calibration during rest or exercise. The present

study proposed the calibration-free blood pressure

estimation method based on PWV during rest and

exercise and the method achieved grade C as IEEE

standard requirement. Moreover, we showed that

PPG features, especially AI, are effective to estimate

SBP. Although we tried to find the cause of large

error lying out of the limits of the agreement in

particular subjects, we were not able to find it out

because of the small sample size.

Some limitations remain in this research. One is

that our proposed model are validated by cuff-based

blood pressure but should be validated by invasive

arterial blood pressure. Furthermore, the proposed

model applied to 18 young men (22.9±1.2 years) and

6 elderly men (43.3±9.3 years), which is not enough

to pass the standard requirement. Finally, the

situation is limited to rest and specific exercise

compared to ambulatory environment.

6 CONCLUSIONS

In this research, we presented a calibration-free

blood pressure estimation method under ambulatory

environment. Using PPG features, especially AI,

enhances the accuracy of blood pressure estimation.

HR is only effective to estimate SBP for young men,

and height ratio of c-wave to a-wave of acceleration

pulse wave might be effective in elderly men.

According to the IEEE standard, the proposed

method achieved grade C in the SBP estimation.

In order to apply our method to daily use, we

have to address some issues.

 Validation should be conducted with a larger

sample size, including female participants,

elderly participants and hypertensive patients,

Cuff-less Calibration-free Blood Pressure Estimation under Ambulatory Environment using Pulse Wave Velocity and Photoplethysmogram

Signals

to pass the standard requirement and to

investigate the difference between the

hypertensive participants and non-

hypertensive participants as well as the elderly

participants and the young participants.

 Although we considered the situation of rest

and exercise, other situations that could cause

blood pressure changes, such as stressful

situations, should be taken into account.

 Motion artifact can obscure the waveform of

PPG signals obtained from the hand for daily

use. Therefore, obtaining PPG signal from

different specific portions of a body that are

less affected by motion artifact should be

considered.

ACKNOWLEDGMENT

This research was supported by Pacific Medico Co.

for providing ECG and PPG measurement devices.

REFERENCES

Alty, S., Angarita, J. N., Millasseau, S., Chowienczyk P.,

2007. Predicting arterial stiffness from the digital

volume pulse waveform. IEEE Transactions on

Biomedical Engineering, 54(12), pp.2268–2275.

Boser, B. E., Guyon, I. M., Vapnik, V.N., 1992. A training

algorithm for optimal margin classifiers. Proceedings

of the fifth annual workshop on Computational

learning theory - COLT '92, pp.144-152.

Breiman, L., 2001. Random forests. Machine Learning,

45(1), pp. 5-32.

Caruana, R. Freitag, D., 1994. Greedy attribute selection.

Machine Learning Proceedings 1994, pp.28–36.

Cover, T. M., Hart, P. E., 1967. Nearest neighbor pattern

classification. IEEE Transactions on Information

Theory, 13(1), pp.21-27.

Daubechies, I., 1992. Ten lectures on wavelets. Society for

Industrial and Applied Mathematics Philadelphia, PA.

Ding, X. R., Zhang, Y. T., Liu, J., Dai, W. X., Tsang, H.

K., 2016. Continuous cuffless blood pressure

estimation using pulse transit time and

photoplethysmogram intensity ratio. IEEE

Transactions on Biomedical Engineering, 63(5),

pp.964–972.

Elgendi, M., 2012. On the analysis of fingertip

photoplethysmogram Signals. Current Cardiology

Reviews, 8(1), pp.14–25.

Fan, Z., Zhang, G., Liao, S., 2011.Pulse wave analysis.

Advanced Biomedical Engineering, pp.21-40.

Gesche, H., Grosskurth, D., Kuchler, G., Patzak, A., 2012.

Continuous blood pressure measurement by using the

pulse transit time: comparison to a cuff-based method.

European Journal of Applied Physiology, 112(1),

pp.309–315.

Homma, S., Ito, S., Koto, T., Ikegami, H., 1992.

Relationship between accelerated plethysmogram,

blood pressure and arteriolar elasticity. Japanese

Journal of Physical Fitness and Sports Medicine,

41(1), pp.98–107.

Inajima, T., Imai, Y., Shuzo, M., Lopez, G., Yanagimoto,

S., Iijima, Katsuya, Morita, H., Nagai, R., Yahagi, N.,

Yamada, I., 2012. Relation between blood pressure

estimated by pulse wave velocity and directly

measured arterial pressure. Journal of Robotics and

Mechatronics, 24(5), pp.811-819.

IEEE Standards Association, 2014. IEEE standard for

wearable, cuffless blood pressure measuring devices.

Kachuee, M., Kiani, M. M., Mohammadzade H., Shabany

M., 2015. Cuff-less high-accuracy calibration-free

blood pressure estimation using pulse transit time.

2015 IEEE International Symposium on Circuits and

Systems (ISCAS).

Legarreta, I.R., Addison, P. S., Reed, M. J., Grubb, N.,

Clegg, G. R., Robertson, C. E., Watson, J. N., 2005.

Continuous wavelet Transform modulus maxima

analysis of the electrocardiogram: beat characterisation

and beat-to-beat measurement. International Journal

of Wavelets, Multiresolution and Information

Processing, 03(01), pp.19–42.

Lopez, G., Shuzo, M., Ushida, H., Hidaka, K.,

Yanagimoto, S., Imai, Y., Kosaka, A., Delaunay, J. J.,

Yamada, I., 2010. Continuous blood pressure

monitoring in daily life. Journal of Advanced

Mechanical Design, Systems, and Manufacturing, 4(1),

pp.179–186.

Mourad, J.J., 2008. The evolution of systolic blood

pressure as a strong predictor of cardiovascular risk

and the effectiveness of fixed-dose ARB/CCB

combinations in lowering levels of this preferential

target. Vasc Health Risk Manag, 4(6), pp.1315–1325.

Mukkamala, R., Hahn, J., Inan, O. T., Mestha, L. K., Kim,

C., Toreyin, H., Kyal, S., 2015. Toward ubiquitous

blood pressure monitoring via pulse transit time:

theory and practice. IEEE Transactions on Biomedical

Engineering, 62(8), pp.1879–1901.

Takazawa, K., Tanaka, N., Fujita, M., Matsuoka, O., Saiki,

T., Aikawa, M., Tamura, S., Ibukiyama, C., 1998.

Assessment of vasoactive agents and vascular aging by

the second derivative of photoplethysmogram

waveform. Hypertension, 32(2), pp.365–370.

Wang, L., Pickwell, M. E., Liang, Y. P., Zhang, Y. T.,

2009. Noninvasive cardiac output estimation using a

novel photoplethysmogram index. 2009 Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society, pp.1746-1749.

World Health Organization, 2014. Global Health Statistics

2014.

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing