Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331824279

Non-invasive Calibration-Free Blood Pressure Estimation Based on Artificial


Neural Network

Chapter · January 2020


DOI: 10.1007/978-3-030-14118-9_69

CITATIONS READS

7 117

4 authors, including:

Nashat Maher Tamer Emara


Ain Shams University Ain Shams University
4 PUBLICATIONS 17 CITATIONS 54 PUBLICATIONS 486 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Nashat Maher on 20 June 2019.

The user has requested enhancement of the downloaded file.


None-Invasive Calibration-Free Blood Pressure
Estimation based on Artificial Neural Network

Nashat Maher1,G.A.Elsheikh2, WagdyR.Anis1and Tamer Emara3


1
ASU faculty of Engineering, Cairo, Egypt
2
PHI institute, Giza, Egypt
3
ASU faculty of medicine, Cairo, Egypt
nmr277@gmail.com ,gaelsheikh@gmail.com,
wagdy_anis@eng.asu.edu.eg,thmfe@yahoo.com

Abstract: This paper presents a non-invasive method for Blood Pressure (BP) estimation
based on extracted features from photoplethysmogram (PPG) and Electrocardiogram (ECG)
signals. The proposed method depends on a machine learning technique, namely Artificial
Neural Networks (ANN), to estimate blood pressure. The training is conducted on a real data
set (more than 2000 BP, ECG and PPG signals) recorded by patients' monitoring at various
hospitals between 2001 and 2008. In addition to the ten features that are usually used in litera-
ture, the proposed method uses the cross validation technique between features to provide more
robust estimation of the blood pressure. Furthermore, the proposed method provides accurate
and reliable blood pressure estimation while it is calibration-free. Compared to previous works,
we used half of the data and the results clarified that we achieved more accuracy in the systolic
pressure measurements. These results are expected to improve more by increasing the training
samples, which is planned in future work.

Keywords: Pulse Wave Velocity, BP Monitoring, Machine Learning.

1 Introduction

World Health Organization (WHO) in 2014‟s world health statistics reported that the hyper-
tension causes 9.4 million people death annually [1]. According to 2008‟s survey, 29.2 percent
of men and 24.8 percent of women suffer from the high blood pressure problem [2]. The hyper-
tension has been recognized as the second factor of cardiovascular disease after diabetes. It is
also called the silent killer, as many people are not aware of their hypertension and the way to
control it. The Blood Pressure (BP) is a periodic signal with the heart rate frequency. The upper
bound of the blood pressure is called the Systolic Pressure (SP) while its lower bound is called
the Diastolic Pressure (DP). The mean arterial pressure (MAP) is defined as the average of the
blood pressure in a cardiac cycle. If SP is above 140 mmHg or DP is above 90 mmHg, it is
called hypertension that can damage internal body organs. The normal range of MAP is be-
tween 70 mmHg and 110 mmHg. Patients with hypertension usually measure their blood pres-
sure occasionally. However, their blood pressure varies over time due to many factors such as
food taking, mental situations or stress. Therefore, continues blood pressure monitoring seems
necessary for accurate diagnosis and treatment of such patients. On the other hand continuous
„beat-to-beat‟ blood pressure monitoring is very useful in patients who are likely to display
sudden changes in blood pressure (e.g. vascular surgery), in whom close control of blood pres-
2

sure is required (e.g. head injured patients), or in patients receiving drugs to maintain the blood
pressure. It also relies in the improvement of patient comfort, especially for those who are
likely to need close blood pressure monitoring for a long period of time e.g. ICU patients. The
most accurate and common blood pressure measurement devices are sphygmomanometers,
which must inflate a cuff around the arm so that BP can be measured with the height of a col-
umn of mercury [3]. This method requires inflatable cuff, which is inconvenient and prevents
continues measurements due to physiological limitations. Invasive (intra-arterial) blood pres-
sure (IBP) monitoring methods are commonly used to achieve this purpose but these tech-
niques generally used in the Intensive Care Unit (ICU) and are also often used in the operating
theatre .This technique involves direct measurement of arterial pressure by inserting a cannula
needle in a suitable artery. The cannula must be connected to a sterile, fluid-filled system,
which is connected to an electronic patient monitor, which is known to carry a risk, albeit a
small one and its measurement is complex and time consuming. Risks of these kinds could be
avoided, however, if there was a noninvasive method offering a high degree of accuracy and
real time operation in a continuous, beat-to-beat, mode. Further, the method should be insensi-
tive to the patient‟s movement (artifacts) and respond rapidly to cardiovascular changes, such
as a sudden drop in blood pressure. Recent researches suggest new cuff-less blood pressure
estimation methods. Although, there have been some attempts on estimating SP and DP based
on the Photoplethysmograph (PPG) signal shape, no clear relation between PPG and BP has
been found yet [4]. Another cuff-less method is based on the wave propagation theory for flu-
ids, which is founded on the natural relationship between the fluid pressure and wave propaga-
tion velocity. The theory implies that the blood pressure can be calculated from the heart pulse
wave velocity PWV. One of much simpler way to get PWV is by looking for arterial properties
such as arterial stiffness.

There are a number of disadvantages associated with this method, such as the need for cali-
bration for each person and the expiration of this calibration in short time intervals [5]. A novel
method is proposed for calibration free and accurate estimation of the blood pressure. The
proposed scheme is accomplished by the extraction of a number of physiological parameters
from Electrocardiogram (ECG) and PPG signals along with machine learning theories. The
main features of PPG signal are shown in Fig.1.
3

Fig.1. Features Extraction of PPG signal

2 Background

The main idea that motivate this work is that the velocity of the pressure pulse, which is ini-
tiated by the heart beat and propagates through arteries, is highly correlated with the elastic
properties of arteries, similar to a pipe with elastic walls. The velocity of displaced fluid in a
pipe is a function of its tension and elasticity. The relation between the Pulse Wave Velocity
(PWV), vessel parameters and blood properties can be represented as [6]:

Et (1)
PWV 
2 R
Where R is the inner radius of vessels, ρ represents the blood density, t is the vessel thickness
and E is Young‟s modulus, which is related to the vessels elasticity. For an elastic vessel, the
relation between the blood pressure and E is given by:
E  E  e  ( PP0 )
0
(2)

Where E0 and P0 are some constants and P can be interpreted as the blood pressure in arter-
ies. In fact (2) indicates that there exists an exponential relation between E and the blood pres-
sure. The formulations in (1) and (2) show that there is a relation between the pulse wave veloc-
4

ity and BP [7]. There are several methods for calculation of the pulse wave velocity, among
which one of the most well-known methods is the Pulse Transit Time (PTT) [8]. PTT is defined
as the time it takes for the heart beat pulse to propagate from heart to the body peripherals. The
pulse wave velocity can be estimated by dividing the distance from heart to a specific peripher-
al (dh,p) by the measured PTT through the following equation:

d h, p (3)
PWV 
PTT

The calculation of the blood pressure from Equation (3) incurs several challenges. One of them
is that arterial properties differ from person to another and are highly dependent to an individu-
al‟s age. Moreover, according to Equation (3), it is required to have the distance between the
peripheral and the heart, which is related to the person‟s height. PTT can be estimated as the
time interval between the R peak of the ECG signal as shown in Fig.2, which indicates the
electrical activity of the heart, and certain points in the finger PPG waveform.

Fig.2. Calculation of PTT from the heart beat pulse. a) ECG signal. b) PPG signal
.
3 Analysis Stepwise Procedure
The proposed method estimates the BP by extracting features from ECG and PPG signals by
building a regression model as shown in Fig.3.

Features extraction
ECG HR Regression
PWV model

PPG PPG features

Fig.3. Overview of the method to estimate BP


5

The work consists of the following steps:


1) Database: collection of a database with adequate sample size.
2) Preprocessing: smooth and remove invalid signals.
3) Feature extraction: extract useful features from signals.
4) Partition the samples into three subsets i.e. train, validation and test samples.
5) train the regression models using Machine Learning (ANN algorithm)
6) Evaluate the trained models‟ performance.

3.1 Database

Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC) II online waveform da-


tabase [9] provided by PhysioNet organization is used in this paper as the reference data-
base. It consists of thousands signals recorded by patient monitors at various hospitals be-
tween 2001 and 2008. Waveform signals were sampled at the frequency of 125 Hz with at
least 8 bit accuracy. We extracted PPG, ECG and arterial blood pressure waveform signals
from this database. Figure 4 shows one sample of a measured data from a patient, the first
is the ECG signal, second is the PPG and the last is continuous Arterial Blood Pressure
(ABP). All of these three signals are measured simultaneously.

1
ECG amplitude

0.5

0
0 200 400 600 800 1000 1200
ECG signal
ECG signals
3
PPG amplitude

0
0 200 400 600 800 1000 1200
PPG signal

200
PPG signal
BP amplitude

150

100

50
0 200 400 600 800 1000 1200
continous BP
ABP signal
Fig.4. Sample of ECG, PPG and ABP signals

3.2 Preprocessing

In order to make data ready for feature extraction, removing distorted and unreliable signals is a
vital task. Preprocessing is performed by dividing samples into fixed size signal blocks as fol-
lows:
 Step-1: Smoothing all signals with a simple averaging filter.
 Step-2: Removing signal blocks with irregular and unacceptable human blood pressure values.
6

 Step-3: Removing signal blocks with unacceptable heart rates.


 Step-4: Removing signal blocks with severe discontinuities, which was not resolved with the help of
smoothing filter in step 1.

3.3 Feature Extraction


Some useful features of the PPG signal are added to PTT features to improve the BP estima-
tion. These features are shown as follow:

1. PTT features: obtained by calculating the time distance between the ECG peak (R-peak) and
three points on the PPG signal. These features including (a) The maximum PPG peak (PTTp),
(b) The PPG minimum (PTTf), and (c) The point of maximum slope of the PPG waveform
(PTTd) (Refer to Fig. 2).
2. Heart rate: The heart rate is obtained by calculation of the peak-to-peak time interval of the PPG
or ECG signals.
3. The PPG features: in order to achieve a calibration free method various features related to the
blood pressure are selected as follow:
- Augmentation Index (AI): it is a measure of the wave reflection on the arteries [4],
which is calculated as the ratio between the diastolic peak and the systolic peak as
clarified in Fig.1.
- Large Artery Stiffness Index (LASI): it is a measure of the arterial stiffness and is re-
lated to the time interval between the systolic peak and the diastolic peak as shown
in Fig.1.
- Inflection Point Area ratio (IPA): it is defined as a function of the areas under the PPG
curve between selected points, denoted by S1, S2, S3 and S4 as shown in Fig.1.
4. Feature crosses: In this work cross validation between features is used to enhance the regression
model as shown in Fig.5.

Fig.5. Features cross validation


7

3.4 Data Partitioning

After processing of the database, more than 2000 records were obtained. The BP distribution
histograms are shown in Fig.6. The new database is then randomly partitioned into three sets:
70% as training, 15% as validation and the remaining as test samples.
450 500

400 (a) 450 (b)


350 400

300 350

300
250

frequency
250
200
200
150
150
100
100
50
50
0
60 80 100 120 140 160 180 200 0
systolic BP (mmHg)
50 60 70 80 90 100 110
Diastolic Pressure (mmHg)

Fig.6 Samples histogram (a) Systolic BP (b) Diastolic BP

3.5 Model based on Machine Learning

As mentioned before, there is an exponential relationship between the blood pressure and the
PWV, which causes a huge non-linearity near the high blood pressure values [10]. Consequent-
ly, the estimation of the high BP values becomes erroneous using the selected features. To
overcome this problem, two approaches can be used. First approach, polynomial terms of the
extracted features are added as extra features and then the training is performed via the linear
regression algorithm [11]. In this approach, Regularized Linear Regression (RLR) with the
Mean Squared Error (MSE) cost function used as a basic regression algorithm with a low com-
putational cost and fast training [11]. Second approach, non-linear regression algorithms, which
are proven to have a better capability to handle this non-linearity issue compared to the linear
regression approach at the cost of an extra computational complexity. ANN, which is one set of
algorithms used in machine learning for modeling data, is among these non-linear methods,
described in the following:
 ANN: Levenberg-Marquardt back propagation algorithm is used as the learning function
and MSE is used as the cost function. The network topology, trained with this algo-
rithm, consists of one input layer with the size of the feature vector, one hidden layer
with sizes between 5 to 15 and one neuron in the output layer as shown in Fig.7.
 Choosing more than 2000 samples from the same data used in literature work to make the
comparison more reliable.
 The correlation between the crossed features and the outputs is checked such that highest
relation maintained. This work proved that crosses feature between LASI and all ten
features used before in literature are enhanced as shown in Table-1 within next section.

+
10 cross features
Fig.7. Simplified block diagram for Artificial Neural Networks
8

3.6 Model Evaluation


The Mean Absolute Error (MAE) and Standard Deviation (STD) of estimation errors are
used for the model evaluation, which are calculated as,

STD 

d2 (5)

MAE 
d (6)

Where d is the error, n is the number of samples and σ is the standard deviation. These estima-
tion errors are calculated and shown in (Fig.8) via the histogram of systolic and diastolic error.

4 Experimental Results

Figure 8 shows the histogram of the BP estimation error using the ANN regression, which is
more accurate and reliable method in the model evaluation. Most of estimated error is ±20. the
results of ANN algorithm are presented in Table-1. It is clear that MAE in both DP and SP
enhanced in proposed work although the number of samples used in literature work is approxi-
mately double. In proposed work MAE of DP reduced to 6.8132 while it was 6.86 i.e. enhanced
by 0.6822%, and of SP reduced to 13.2493 while it was 13.78 in the previous work, i.e. en-
hanced by 3.85%. STD of DP enhanced by 1.943% from 8.96 to 8.7859 while STD of SP
increase by 0.369% from 17.46 to 17.5247.

600 800

(a) (b)
700
500

600

400
500

300 400

300
200

200

100
100

0 0
-80 -60 -40 -20 0 20 40 60 80 -60 -50 -40 -30 -20 -10 0 10 20 30
Systolic Error Diastolic Error

Fig.8. The proposed work results (a) Systolic BP error histogram (b) Diastolic BP error histogram
9

Table-1: Performance of Proposed & Literature Work

Algorithm DP SP Number of
samples
MAE STD MAE STD
Literature 6.86 8.96 13.78 17.46 4254
Work 6.8132 8.7859 13.2493 17.5247 2126
To evaluate the accuracy, cumulative percentage of readings falling within 5, 10 and 15 mmHg,
respectively is calculated for diastolic and systolic blood pressure. The proposed work en-
hanced result in systolic BP but still the literature result better in DP as shown in Table-2.

Table-2: Cumulative Percentage of Readings Comparison


≤5 mmHg ≤10 mmHg ≤15 mmHg
Absolute difference
Literature work Literature work Literature work
diastolic 51.2% 27.33 % 78.9% 52.59% 93.6% 74.13%
results
systolic 28.8% 34.01% 51.5% 59.92% 69.5% 75.96%

Although number of samples is one half of that used in literature, the proposed model results
are more accurate. Table-2 shows that in systolic pressure, the number of samples that has
absolute error less than 5 mmHg is increased by 5.21%, the number of samples that has abso-
lute error less than 10 mmHg is increased by 8.42% and the number of samples that has abso-
lute error less than 15 mmHg is increased by 6.46%.

Other cuff-less BP estimation designs in literature suffer from serious drawbacks such as the
need of calibration per each patient [5] or utilizing relatively small sample size database [10].
The proposed method in this work resolves all these issues while providing enough capability
for reliable and calibration-free, blood pressure estimation.

5 Conclusions
Continuous BP monitoring is vital; nevertheless it is unreachable through conventional cuff-
based BP measurement devices which require cuff inflation and deflation that and prevent
continuous measurement due to physiological limitation, moreover it is slow and uncomforta-
ble for many patients. The proposed method establishes a cuff-less BP estimation system with-
out calibration using ANN and cross validation of input features. The used cross feature valida-
tion method enhances the results of the output model compared to results achieved in Literature
work. Although number of samples is one-half of that used in the previous work, the results
obtained from the proposed model are more accurate in systolic pressure measurements. This
model clarified that the MAE in both systolic and diastolic are decreased. In addition, the re-
sults are promising and expected to increase accuracy of the model via increasing the number
of samples, which is planned for further work.

References

1. World Health Organization, “World Health Statistics 2014”, 2014.


2. Haruyuki Sanuki1 , Rui Fukui, Tsukasa Inajima and Shin'ichi Warisawa"Cuff-less Calibra-
tion-free Blood Pressure Estimation under Ambulatory Environment using Pulse Wave Ve-
locity and Photoplethysmogram Signals"In Proceedings of the 10th International Joint Con-
ference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 42-48
10

3. G. Van Montfrans, “Oscillometric blood pressure measurement: progress and problems,”


Blood pressure monitoring, vol. 6, no. 6, pp. 287–290, 2001.
4. M. Elgendi, “On the analysis of fingertip photoplethysmogram signals,” Current cardiology
reviews, vol. 8, no. 1, p. 14, 2012.
5. M. Wong, C. Poon and Y. Zhang, “An evaluation of the cuffless blood pressure estimation
based on pulse transit time technique: a half year study on normotensive subjects,” Cardio-
vascular Engineering, vol. 9, no. 1, pp. 32–38, 2009.
6. P. Obrist, K. Light, J. McCubbin, J. Hutcheson and J. Hoffer, “Pulse transit time: Relation-
ship to blood pressure,” Behavior Research Methods and Instrumentation, vol. 10, no. 5,
pp. 623–626, 1978.
7. D. J. Hughes, L. A. Geddes, C. F. Babbs, and J. D. Bourland, “Measurements of youngs
modulus of the canine aorta in-vivo with10 mhz ultrasound,” in Proc. Ultrasonics Symposi-
um, Sep 1978,pp. 326 326.
8. L. Peter, N. Noury and M. Cerny, “A review of methods for non-invasive and continuous
blood pressure monitoring: Pulse transit time method is promising?,” IRBM, 2014.
9. A. Goldberger, L. Amaral, L. Glass, J. Hausdorff, P. Ivanov, R. Mark, J. Mietus, G. Moody,
C. Peng and H. Stanley, “Physiobank, physiotoolkit, and physionet components of a new
research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. 215–
220, 2000.
10. H. Gesche, D. Grosskurth, G. Kuchler and A. Patzak, “Continuous blood pressure meas-
urement by using the pulse transit time: comparison to a cuff-based method,” European
journal of applied physiology, vol. 112, no. 1, pp. 309–315, 2012.
11. Mohamad Kachuee, Mohammad Mahdi Kiani, Hoda Mohammadzade, Mahdi Shabany
“Cuff-Less High-Accuracy Calibration-Free Blood Pressure Estimation Using Pulse Transit
Time”2015 IEEE International Symposium on Circuits and Systems (ISCAS).

View publication stats

You might also like