Determining Mean Square Error and Standard Deviation Error For Measurement of Non-Invasive Blood Pressure Using ANN

European Journal of Scientific Research
ISSN 1450-216X / 1450-202X Vol. 141 No 2 August, 2016, pp.209-222

http://www.europeanjournalofscientificresearch.com
Determining Mean Square Error and Standard Deviation Error for

Measurement of Non-Invasive Blood Pressure Using ANN
M. El-Sayed Waheed
Department of Computer Science
Faculty of Computer and Information
Suez Canal University, Egypt
Mewahed@yahoo.com.
G. M. Behery
Department of Mathematics
Faculty of Science
Damietta University, Egypt
Behery2911961@gmail.com.
Ahmed M. Elshewey
Department of Mathematics
Faculty of Science
Damietta University, Egypt
Elshewey_2008@yahoo.com.
Abstract
The measurements of blood pressure techniques are based on measuring the

pressure of the cuff and on sensing the variation of the amplitude of pulsatile. Patient
movement is sensitively effected by these measurements. The movements that are
slightest unexpected could offset the readings of the automatic blood pressure meter
by a large amount or render the readings totally meaningless. In this paper,
windkessel models (two element windkessel model, and three element windkessel
model) were applied to generate blood pressure measurement. Blood pressure
measurement for healthy person varies in the range 80mmHg/120mmHg. Artificial
neural network classification is one of the artificial intelligence techniques used for
the statistical analysis and classification. Artificial neural network (ANN) is tools for
pattern classification computational that have been the matter of research that is
renewed. ANN uses several learning algorithms and formats are being used in
medical applications, industrial, and academic research artificial neural network were
also used in previous studies to estimate blood pressure. In this paper, artificial neural
network designs have been used for determining systolic pressure and diastolic
pressure, where the input values in the neural network are the values estimated from
windkessel models (two element windkessel model, and three element windkessel
model). Gradient descent and an algorithm of adaptive learning rate back propagation
were used for the training. The result showed that the artificial neural network was
more accurate method. However a huge database is required for this method. Both
under training and over training will affect the measurement accuracy.
Keywords: Mean square error (MSE), Standard deviation (SD), Artificial neural
network (ANN), Multilayer perceptron (MLP), Blood pressure (BP), Systolic pressure
(SP), Diastolic pressure (DP).
Determining Mean Square Error and Standard Deviation Error for Measurement of Non-Invasive
Blood Pressure Using ANN 210
1. Introduction
Blood pressure is a very important measurement for estimation the health of human [1, 2].
Respectable analysis has been dedicated towards effective blood pressure measurement in
many situations. Hence, a variety of methods are available for estimating and monitoring
the pressure of blood. These methods can be classified as invasive and non-invasive.
Invasive method involves in inserting blood pressure measuring cannula (thin flexible
tubes) into the body [3] and making incisions. Non-invasive method involves some form of
a cuff wrapped around a persons arm that is inflated and deflated while monitoring the
pulse oscillations [4, 5, 6]. Non-invasive techniques are better because the painless nature is
inherent and also have been automated to free up a doctors attention for more significant
tasks [6, 7]. In this paper, windkessel models (two element windkessel model, and three
element windkessel model) was designed to generate blood pressure measurement. The
models of windkessel was created to describe the systemic arterial system and the heart as a
hydraulic circuit in closed shape [8]. Here, the circuit contained a pump of water connected
to a chamber, filled with water except for a pocket of air. As its pumped up, the water
compresses the air that successively pushes the water out of the chamber. This analogy
shows the mechanics of the heart. The values estimated from windkessel models are trained
through ANN training algorithm. An ANN is a mathematical computing paradigm that
models the operations of biological neural systems [9]. Artificial Neural Network (ANN)
are tools for pattern classification computational that have been the matter of research that
is renewed and interest during 15 years past. ANN, uses several learning algorithms and
formats, are being used in medical applications, industrial, and academic research [10, 11,
12, 13, 14].
The model of feed forward multilayered neural network played a significant role in
neural networks applications [10]. The algorithm of back propagation is a mostly used as
training transaction that modifies the relation weights of multilayer perceptron (MLP) [15].
It is an algorithm of gradient descent that decreases the mean square error (MSE) between
the desired response signals and perceptron output signals in an manner that is iterative [10,
11, 12, 13, 14].
On analyzing recent expansions, it becomes clear that the tendency is to create new
methods for decision making of computer in medicine and to estimate critically these
methods in practice of clinical. Diseases diagnosis may be gazed as a task of pattern
classification [10, 16]. ANN applications in the field of medical are great and implicate
myocardial infarction diagnosis [17], analysis of electrocardiogram [18, 19], analysis of
electrogastrogram [20], analysis of electroencephalogram [21], artery disease coronary [22,
23, 24, 25], warning forecasting for patients with failure of heart, analysis of
electromyogram [20], differential of varied data that are pathological [13].
The estimated blood pressure values from windkessel models are used as inputs
through ANN training algorithm. Mean square error (MSE) and standard deviation error
(SD) were calculated between the inputs and outputs of artificial neural network training
algorithm.
2. Materials and methods

2.1. Blood pressure flow and windkessel model
The aorta is the largest artery within the human circulatory system, originating from the left
ventricle of the heart and extending down to the abdominal cavity, where it branches into
arteries that are smaller [26, 27].
211 M. El-Sayed Waheed, G. M. Behery and Ahmed M. Elshewey
The cycle of the cardiac is a closed loop, its system is pulsatile, and the heart flows
blood throughout the system of the circulation in order to form a pulse wave as shown
in (Figure 1).
Figure 1: Cardiac cycle form.
The first phase of the cardiac form, diastole of ventricular happens when the ventricles
are relaxed and allow the oxygenated new blood to flow in from the atria [28]. The diastole
of ventricular is followed by systole, where the ventricles contract and allow the blood to
eject out to the body through the aorta. The pressure of blood rises when the ventricles
contract and pump the blood into the aorta, at its maximum is named systolic pressure. At
the beginning of following cardiac cycle, as the blood begins to flow into the ventricles, the
pressure of the blood is at its lowest, and it is known as diastolic pressure [27]. The model
of windkessel was created to describe the systemic arterial system and the heart as a
hydraulic circuit in closed shape [8]. Here, the circuit contained a pump of water connected
to a chamber, filled with water except for a pocket of air. As its pumped up, the water
compresses the air that successively pushes the water out of the chamber. This analogy
shows the mechanics of the heart. Windkessel models are used commonly to represent the
undertaken load by the heart during the cardiac cycle. It connects blood pressure and blood
flow inside the aorta, and distinguishes the compliance of the arterial, peripheral resistance
of the valves and the inertia of the blood flow. The windkessel model takes the following
parameters into consideration while modeling the cardiac cycle:
Arterial compliance: assign to the main artery extensibility during the cardiac cycle.
Peripheral resistance: assign to the flow of resistance encountered by the blood as it
flows through the systemic arterial system.
Inertia: assign the blood inertia as it is cycled throughout the heart.
The windkessel is similar to the law of poiseuilles for a hydraulic system model. It
characterizes the blood flow through the arteries as the flow of fluid through pipes. In this
paper, we concentrate on the electrical circuit equivalent, as shown in (Figure 2).
Figure 2: Fluid dynamics and electrical circuit equivalents.

2.1.1. Windkessel model description

Windkessel model suppose that:
Cardiac cycle begins at systole.
The systole period is 2/5th of the cardiac cycle period.
Arterial compliance, peripheral resistance, and inertia are represented as a capacitor,
a resistor, and an inductor respectively.
Windkessel models are explained in two different models as below [29, 30].
2.1.1.1. Two element windkessel model

The simplest model of windkessel, shows the hemodynamic state is the two element model.
While a cardiac cycle, it takes into consideration the effect of total peripheral
resistance and arterial compliance [29].
Figure 3: Diagrammatic representation of ejection of blood from ventricular and arterial circulation.
Figure 4: Electrical analog of the two element windkessel model.
In the analog of the electrical, arterial compliance ( in /mmHg) is performed

as a capacitor with properties of electric charge storage, the peripheral resistance of the
arterial system ( in mmHg s/ ) is performed as a resistor of energy dissipating. The
blood flow from the heart ( in /s) is similar to the flow of current in the circuit and
the pressure of blood in the aorta ( in mmHg) represents the variation of time electric
potential. (Figure 4) shows that during systole state, the blood ejects from the ventricles to
the aortic chamber compliant. The stored blood in the vessels of peripheral and the elastic
recoil of aorta during diastole is seen as solid and dashed lines respectively. The theoretical
model is shown in the electrical analog in (Figure 3) is given as in (Equation 1).
(1)
2.1.1.2. Three element windkessel model

Three element windkessel model demonstrate the distinguish resistance of the proximal
aorta [30]. A resistor is combined in the series to account for this resistance to the flow of
blood due to the valve of the aortic. The existing parallel combination of resistor capacitor
performs the total peripheral resistance and aortic compliance in the two element model as
discussed before. A hydraulic equivalent of the three element model is shown in (Figure 5).
Compliance of the aortic due to the variation of the pressure is seen by allowing a
bottle to afford volume isolation. The tube geometry represents the distinctive aortic
resistance; resistance to flow is varied by fractional opening and closing of needle valve
shown.
Figure 5: A hydraulic equivalent of the three element windkessel model.
The theoretical modeling as seen in the electrical analog (Figure 6) is given as in

(Equation 2).
( ) (2)
Figure 6: Electrical analog of the three element windkessel model.
2.1.2. Model of the blood flow to the aorta

The blood flow in the aorta from the ventricle during the cycle of cardiac is represented as
in our model. is represented as a sine wave with amplitude during systole and is
zero otherwise. This follows our learning of the cardiac physiology. At diastole, where the
ventricles are restful, there is no flow of blood into the aorta, and then, = 0.
However with ventricular contract during the systole, the flow of blood is ejected
into the aorta and can be modeled as a sinusoidal wave, therefore the blood flow is
calculated by (Equation 3).
( ) (3)
Where is time in seconds, is the period of the cardiac cycle in seconds, is the
systolic period, in seconds, and represents the remainder of divided by .
is assumed to be 2/5 , according to the cardiac cycle dynamics. According to literature,
the flow of blood in one cardiac cycle is 90 . We use that information to get the
constant as shown in (Equation 4).
( )
( ) (4)
2.2. Statistical analysis and classification

2.2.1. Artificial neural networks
ANN consists of a large numbers of processing neurons (elements) that are connected and
linked with each other. The connections strengths are called weights. For the physical
systems modeling, a multilayered feed forward neural network is used commonly. It
contains a layer of input elements, a layer of output elements and one or more hidden
layers. To deal with separable problems that are nonlinearly, layer(s) of elements are added
to be placed among the input layer (containing nodes of the input) and the output neuron
are required leading to the architecture of MLP, as shown in (Figure 7). Since the layers of
the intermediate do not deal with the external environment, these are called hidden layers
and their elements called hidden elements. The intermediate layers that are added revived
the perceptron by spreading its ability to fix the problems of nonlinear classification. The
acknowledgement in ANN lies in the weights of the interconnection between
elements [10, 11, 12, 13, 14].
Figure 7: MLP neural network.

2.2.2. MLP neural network using training algorithm of back propagation

In MLP most applications, the weights are specified by back propagation algorithm means,
that is based on finding the surface of an error (error as ANN weights function) that uses
points of gradient descent with minimizing the error [10, 11, 12, 13, 14]. During the phase
of training, the weights are adjusted successively depend on a group of inputs and the
corresponding group of desired targets outputs. Every iteration in back propagation
modeling two extensions: producing a solution by forward activation, and a propagation of
backward of the calculated error to adjust the weights. The forward and backward
extensions are proceeding frequently till the solution of ANN reconciles with the value of
desired within a tolerance that is pre specified. The algorithm of back propagation provides
the weight adjustments needed in the backward extension [10, 11, 12, 13, 14].
In MLP neural network, in the hidden layer each neuron sums its signals of input
after multiplying them by the respective connection strengths weights and calculates
its output as a sum function as shown in (Equation 5).
(5)
Such that is activation function, the activation function for hidden neurons was
the traditional sigmoidal function that lies in range between zero and one. The squared sum
differences between the actual values and the desired of the output neurons that is given
in (Equation 6).
(6)
Such that is the value of desired output neuron and is the value of actual
output neuron. Each weight is modified when an increment is added to it.
was chosen to decrease as quickly as possible. How is calculated depends on the
algorithm training adaptive. The algorithm of back propagation is then mentioned to
modify all the weights in the network and gives the change in the weight of the connection
at iteration between neurons and as in (Equation 7).
(7)
Such that is known as learning coefficient, the momentum coefficient,

is the change in weight in the preceding iteration immediately. For this paper,
learning coefficient and momentum coefficient were empirically determined where
was determined as 0.1 and was determined as 0.2.
3. Results and discussion

In this paper, for solving the problem of pattern classification MLP neural network using
back propagation algorithm for training was used. Training algorithm that is effective and
system behavior that is better understood are the merits of this sort of neural network.
Selection of input parameters of the network and execution of neural network are
very significant for determining Mean square error (MSE) and standard deviation error
(SD) between the inputs and outputs of the artificial neural network training algorithm.
A feed forward neural network is used to classify the data estimated from two
element windkessel model. 15 values of the data for training and 15 values for testing. The
feed forward neural network consists of 3 layers, input layer consists of 15 nodes, one
hidden layer consists of 1 node, and output layer consists of 15 nodes. The mean square
error (MSE) was 0.1409 and standard deviation error (SD) was 0.3809.
(Figure 8) shows the relationship between blood pressure and time for both inputs
and outputs of the feed forward neural network using 1 node in hidden layer.
Figure 8: Relation between blood pressure and time for inputs and outputs, where hidden
layer consists of 1 node using two element windkessel model.
hidden layer consists of 2 nodes, and output layer consists of 15 nodes. The mean square
and outputs of the feed forward neural network using 2 nodes in hidden layer.
Figure 9: Relation between blood pressure and time for inputs and outputs, where hidden layer
consists of 2 nodes using two element windkessel model.
layer consists of 3 nodes using two element windkessel model.
layer consists of 4 nodes using two element windkessel model.
(Table 1) shows the values of mean square error (MSE) and standard deviation error
(SD) for different numbers of nodes in hidden layer using two element windkessel model.
Table 1: Values of mean square error (MSE) and standard deviation error (SD) for different
number of nodes in hidden layer using two element windkessel model.
No. of nodes in hidden layer MSE SD
1 0.1409 0.3809
2 0.2200 0.4742
3 0.2475 0.4478
4 0.3300 0.3682
A feed forward neural network is used to classify the data estimated from three
hidden layer consists of 1 node, and output layer consists of 16 nodes. The mean square
and outputs of the feed forward neural network using 1 node in hidden layer.
layer consists of 1 node using three element windkessel model.
layer consists of 2 nodes using three element windkessel model.
(Table 2) shows the values of mean square error (MSE) and standard deviation error
(SD) for different numbers of nodes in hidden layer using three element windkessel model.
Table 2: Values of mean square error (MSE) and standard deviation error (SD) for different
number of nodes in hidden layer using three element windkessel model.
No. of nodes in hidden layer MSE SD

1 0.1431 0.1394
2 0.2714 0.4047
3 0.3549 0.3878
4 0.4413 0.4210
4. Conclusion
In this paper, windkessel models (two element windkessel model and three element
windkessel model) were applied to generate blood pressure measurement. Blood pressure
for healthy person varies between 80mmHg/120mmHg. The estimated values from
windkessel models (two element windkessel model and three element windkessel model)
were applied and used as inputs for an artificial neural network. The feed forward neural
network used in this work consists of 3 layers, one input layer, one hidden layer, and one
output layer. Mean square error (MSE) and standard deviation error (SD) were calculated
between the inputs and outputs of artificial neural network training algorithm.
The results demonstrated that when the number of nodes decreases in hidden layer,
the results of mean square error (MSE) and standard deviation error (SD) are better as
shown in (Table 1) and (Table 2).
Acknowledgment
The study was supported generously by Department of Mathematics, Faculty of Science,
Damietta University.
I would like to thank Prof. Dr. Mohammed Elsayed Waheed, Dean of Faculty of
Computer and Information, Suez Canal University, for his most support and
encouragement. He kindly read my paper and offered invaluable detailed advices on
grammar, organization, and the theme of the paper.
I thank Prof. Dr. Gamal Mohammed Behery, Professor of Computer Science,
Faculty of Science, Damietta University, for his comments that greatly improved the paper.
References
[1] Loek M, Horlkov M, Havlk J. Mechanical model of the cardiovascular system:
determination of cardiac output by dye dilution. Lka a technika. 2012;42(2):77-80.
[2] McMahon N, Hogg LA, Corfield AR, Exton AD. Comparison of noninvasive and
invasive blood pressure in aeromedical care. Anaesthesia. 2012;67(12):1343- 1347.
[3] Ghostine S, Caussin C, Daoud B, Habis M, Perrier E, Pesenti-Rossi D, Sigal-
Cinqualbre A, Angel CY, Lancelin B, Capderou A, Paul JF. Non-invasive detection of
coronary artery disease in patients with left bundle branch block using 64-slice computed
tomography. Journal of the American College of Cardiology. 2006;48(10):1929-1934.
[4] Muecke S, Bersten A, Plummer J. The mean machine; accurate non-invasive blood
pressure measurement in the critically ill patient. Journal of clinical monitoring and
computing. 2009;23(5):283-297.
[5] Araghi A, Bander JJ, Guzman JA. Arterial blood pressure monitoring in overweight
critically ill patients: invasive or noninvasive?. Critical Care. 2006;10(2):R64.
[6] Sende J, Jabre P, Leroux B, Penet C, Lecarpentier E, Khalid M, Margenet A, Marty J,
Combes X. Invasive arterial blood pressure monitoring in an out-of-hospital setting: an
observational study. Emergency Medicine Journal. 2009;26(3):210-212.
[7] Schmidt B, Klingelhfer J. Clinical applications of a non-invasive ICP monitoring
method. European journal of ultrasound. 2002;16(1):37-45.
[8] Molino P, Cerutti C, Julien C, Cuisinaud G, Gustin MP, Paultre C. Beat-to-beat
estimation of windkessel model parameters in conscious rats. American Journal of
Physiology-Heart and Circulatory Physiology. 1998;274(1):H171-177.
[9] Negnevitsky M. Artificial intelligence: a guide to intelligent systems. Pearson
Education; 2005.
[10] Basheer IA, Hajmeer M. Artificial neural networks: fundamentals, computing, design,
and application. Journal of microbiological methods. 2000;43(1):3-31.
[11] Dirgenali F, Kara S. Recognition of early phase of atherosclerosis using principles
component analysis and artificial neural networks from carotid artery Doppler signals.
Expert Systems with Applications. 2006;31(3):643-651.
[12] beyli ED. Implementing eigenvector methods/probabilistic neural networks for
analysis of EEG signals. Neural networks. 2008;21(9):1410-1417.
[13] Shi Z, He L, Suzuki K, Nakamura T, Itoh H. Survey on Neural Networks used for
Medical Image Processing. International journal of computational science. 2009;3(1):86.
[14] Pham DT, Sagiroglu S. Training multilayered perceptrons for pattern recognition: a
comparative study of four training algorithms. International Journal of Machine Tools and
Manufacture. 2001;41(3):419-430.
[15] Suzuki K, Shiraishi J, Abe H, MacMahon H, Doi K. False-positive reduction in
computer-aided diagnostic scheme for detecting nodules in chest radiographs by means of
massive training artificial neural network 1. Academic Radiology. 2005;12(2):191-201.
[16] Dybowski R, Gant V. Clinical applications of artificial neural networks. Cambridge
University Press; 2001.
[17] Amato F, Lpez A, Pea-Mndez EM, Vahara P, Hampl A, Havel J. Artificial neural
networks in medical diagnosis. Journal of applied biomedicine. 2013;11(2):47-58.
[18] Engin M, Demira S. Fuzzy-hybrid neural network based ECG beat recognition using
three different types of feature sets. Cardiovascular Engineering: An International Journal.
2003;3(2):71-80.
[19] Prasad GK, Sahambi JS. Classification of ECG arrhythmias using multi-resolution
analysis and neural networks. InTENCON 2003. Conference on Convergent Technologies
for the Asia-Pacific Region 2003 Oct 15 (Vol. 1, pp. 227-231). IEEE.
[20] Gupta A, Vivekanandan S. EMG Myopathy Signal Detection Using Wavelet
Transform and Neural Network Techniques. International Journal of Science and Advanced
Technology (ISSN 2221-8386) Volume. 2012;2.
[21] Pavlopoulos S, Kyriacou E, Koutsouris D, Blekas K, Stafylopatis A, Zoumpoulis P.
Fuzzy neural network-based texture analysis of ultrasonic images. Engineering in Medicine
and Biology Magazine, IEEE. 2000;19(1):39-47.
[22] Mahmoodabadi Z, Abadeh MS. CADICA: Diagnosis of Coronary Artery Disease
Using the Imperialist Competitive Algorithm. JCSE. 2014;8(2):87-93.
[23] Er O, Yumusak N, Temurtas F. Chest diseases diagnosis using artificial neural
networks. Expert Systems with Applications. 2010;37(12):7648-7655.
[24] Mobley BA, Schechter E, Moore WE, McKee PA, Eichner JE. Predictions of coronary
artery stenosis by artificial neural network. Artificial intelligence in medicine.
2000;18(3):187-203.
[25] Dhande JD, Gulhane SM. Design of Classifier Using Artificial Neural Network for
Patients Survival Analysis. training. 2012;1(2).
[26] Akaike H. A new look at the statistical model identification. Automatic Control, IEEE
Transactions on. 1974;19(6):716-723.
[27] Lambermont B, Grard P, Detry O, Kolh P, Potty P, Defraigne JO, D'Orio V, Marcelle
R. Comparison between three-and four-element windkessel models to characterize vascular
properties of pulmonary circulation. Archives of physiology and biochemistry.
1997;105(7):625-632.
[28] Truant R. Design of a Pulsatile Pumping System for Cardiovascular Flow PIV
Experimentation. Bachelors in Engineering, University of Victoria, Victoria, British
Columbia. 2007.
[29] Gologorsky E, Gologorsky A, Barron ME. Intraoperative blood pressure measurement
modalities are separate and not equal. The Journal of the American Society of
Anesthesiologists. 2012;116(6):1394-1394.
[30] Liu SH, Cheng DC, Wang JJ. Estimating the mean blood flow of arm based on
Windkessel model. Biomedical Engineering: Applications, Basis and Communications.
2011;23(05):349-356.

Determining Mean Square Error and Standard Deviation Error For Measurement of Non-Invasive Blood Pressure Using ANN

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Determining Mean Square Error and Standard Deviation Error For Measurement of Non-Invasive Blood Pressure Using ANN

Hochgeladen von

Copyright:

Verfügbare Formate

European Journal of Scientific Research

ISSN 1450-216X / 1450-202X Vol. 141 No 2 August, 2016, pp.209-222

Determining Mean Square Error and Standard Deviation Error for

The measurements of blood pressure techniques are based on measuring the

2. Materials and methods

Figure 1: Cardiac cycle form.

Figure 2: Fluid dynamics and electrical circuit equivalents.

2.1.1. Windkessel model description

2.1.1.1. Two element windkessel model

Figure 4: Electrical analog of the two element windkessel model.

In the analog of the electrical, arterial compliance ( in /mmHg) is performed

2.1.1.2. Three element windkessel model

Figure 5: A hydraulic equivalent of the three element windkessel model.

The theoretical modeling as seen in the electrical analog (Figure 6) is given as in

Figure 6: Electrical analog of the three element windkessel model.

2.1.2. Model of the blood flow to the aorta

2.2. Statistical analysis and classification

Figure 7: MLP neural network.

2.2.2. MLP neural network using training algorithm of back propagation

Such that is known as learning coefficient, the momentum coefficient,

3. Results and discussion

No. of nodes in hidden layer MSE SD

Das könnte Ihnen auch gefallen