A HEART CELL GROUP MODEL FOR THE IDENTIFICATION

OF MYOCARDIAL ISCHEMIA

Mohamed A. Mneimneh, Micheal T. Johnson and Richard J. Povinelli

Electrical and Computer Engineering, Marquette University, 1515 Wisconsin ave, Milwaukee, Wisconsin,USA

Keywords: Inverse problem, Ischemia, Decision Tree.

Abstract: Due to the increasing prices of medical care, and especially due to cardiovascular injury; scientists are

looking for inexpensive and less invasive ways to diagnose myocardial ischemia. Many studies have shown

that the variations of the ST-segment in the ECG signal are an indicator for ischemia. For this purpose, this

work proposes an approach based on a heart cell group model and principle component analysis, using a

decision tree classifier to differentiate between the ischemic and healthy beats. The cardiac based model is

based on a physiological model of the electrical cycle of depolarization and repolarization of the atria and

ventricles. The model parameters are estimated by minimizing the squared error between the generated

signal and the recorded ECG. The approach is applied to beats from the Long-Term ST database, which

consists of 86 subjects and more than 20,000 beats in which 80% of the beats are ischemic and 20% are

healthy. A 10-fold cross validation test is performed over the dataset. The accuracy of this approach is

91.62%, with sensitivity of 95.09% and specificity of 75.66%.

1 INTRODUCTION

Ischemic heart disease is the leading cause of death

in the world with almost 14% of all deaths (AHA

2005) Moreover, the average number of individuals

who undergo a heart attack as a result of myocardial

ischemia in the United States is approximately 1.5

million cases, of which 500,000 are fatal (AHA

2005). Myocardial ischemia is defined as the

deprivation of oxygen in some portions of the

cardiac tissue due to a blockage in the coronary

artery. If the deprivation continues for an extended

period, the effected cardiac tissue will die; thus,

leading to a heart attack. Tissue that has died is no

longer functional and diminishes the mechanical

pumping function of the heart (Pardee 1920).

Early detection of ischemia is crucial because, in

most cases, the effects of myocardial ischemia are

reversible if detected early enough (Long 1980).

General screening of patients is vital to preventing

myocardial infarction, since ischemia can be present

without exhibiting symptoms.

This work proposes a cardiac based model,

Principle Component Analysis (PCA) and a C4.5

decision tree classifier for the detection of

myocardial ischemia. The cardiac model is based on

a physiological model of the electrical cycle of

depolarization and repolarization of the atria and

ventricles. The Sinoatrial (SA) node, the

Atrioventricular (AV) node, bundle branches,

Purkinje fibers, and left and right ventricles are

modelled as signal generators. The ECG is generated

by the difference in signal amplitudes arriving at the

positive and negative terminals of an ECG lead. The

model parameters are estimated through the

minimization of the squared error between the

generated signal and the recorded ECG. In addition

to the obtained model parameters, 50 of the

components from applying PCA to the signal are

used in the diagnosis. A C4.5 decision tree is then

used as a classifier to determine if a beat is healthy

or ischemic.

The purpose of using electrocardiogram signals

for the diagnoses of myocardial ischemia is because

it is one of the least expensive techniques available

to physicians. Figure 1 shows a labelled ECG signal

showing the P, Q, R, S, T waves, the ST segment

and the J point. The use of the ST level in the

detection of myocardial ischemia was hypothesized

in 1920 (Pardee 1920). Examples of low cost

methods are ST event alerts ($250 cost) and easy to

administer) with sensitivity of 46% and specificity

of 91% and exercise stress testing ($200-$300 cost)

with 68% accuracy of 68% (R. Gianrossi 1989).

A. Mneimneh M., T. Johnson M. and J. Povinelli R. (2008).

A HEART CELL GROUP MODEL FOR THE IDENTIFICATION OF MYOCARDIAL ISCHEMIA.

In Proceedings of the First International Conference on Health Informatics, pages 51-58

 SciTePress

Figure 1: Labelled (ECG) signal.(Moody 2001).

Significant research has been undertaken to

develop a more accurate, less invasive, and less

expensive method for detecting myocardial

ischemia. Much of this research focuses on the use

of ECG signals. These methods build models or use

thresholds of the ST deviation to determine if a

patient’s ECG signal might indicate ischemia.

Previous techniques that dealt with ischemia

classification and detection when monitoring ECG

signals started with low accuracy that increased

significantly over time. These techniques are based

on the hypothesis that myocardial ischemia can be

detected by monitoring the ST variations.

Maglaveras et al. (N. Maglaveras 1994) have

investigated a method for ischemia detection that

uses supervised neural networks. The accuracy of

this approach is of sensitivity of 73.0% and positive

predictive accuracy of 69.5%.

RV Andreao et al. (R.V. Andreao 2004)

employed a Hidden Markov Model for beat

segmentation with the application of ischemia

detection. The accuracy of this model is of

sensitivity of 83.0% and positive predictive accuracy

of 85%.

Additionally, T. Stamkopoulos et al. proposed an

approach using nonlinear Principle Component

Analysis (PCA) and neural networks in the

identification of ischemic beats. The accuracy of this

approach was 80% for healthy and 90% for ischemic

beats when applied to the European ST-T Database

(Stamkopoulos 1998).

Similarly, Victor-Emil Neagoe applied a

Gaussian Neuro-Fuzzy Approach and PCA toward

the classification of myocardial ischemia. The

accuracy shown in the paper was 100% for 50

features. However, Neagoe dealt with only

identifying ischemic and normal patients. Moreover,

the number of training and testing data were 40

patients, half used for training and half for testing

(Victor-Emil Neagoe 2003).

2 DATA SET AND

PRE-PROCESSING

Various ECG and intracardiac datasets are available

for the use of modelling and detecting myocardial

ischemia. The data sets preserve the privacy of the

subjects as there are no direct or indirect identifiers

linking back to them.

2.1 Long Term ST Database

The Long-Term ST Database from PhysioNet

contains 86 Holter ECG recordings from 80

independent patients. Holter recordings are ECG

recordings recorded using portable recording

devices, generally taken over a long period. These

recordings were selected from the Holter libraries at

Beth Israel Deaconess Medical Center in Boston,

Physiolab (Laboratory of Biosignal Processing) of

the Institute of Clinical Physiology in Pisa, Brigham

and Womens Hospital in Boston, and the Zymed

company. The recordings vary in length from 20 to

24 hours. Each record contains either two or three

ECG leads. The records are digitized at 250 Hz with

12 bit resolution (Moody 2001).

Complete annotations have been provided for the

database. These annotations label the significant ST

shifts and episodes, the beginning (3-point) of most

ST segments has been annotated along with R wave

annotations using a 16 second averaging window.

The beats were detected using WQRS function as

part of the WFDB package supplied by the

Physionet (Moody 2001).

To aid in the development of an ischemia

classification algorithm, complete ST level

annotations have also been provided. These

annotations give the ST level, ST reference function,

and the calculated ST deviation. The ST reference is

expertly labelled moving average of the important

ST shifts. The ST deviation is calculated by

subtracting the ST level from the ST reference

function shown in Figure 2 (Jager, Taddei & Moody

2003).

Figure 2: Example of ST deviation calculation.

The data consists of 43 free records from 42

patients and 43 fee records from 38 patients. The

total number of beats used in this work is 20,528 for

HEALTHINF 2008 - International Conference on Health Informatics

both healthy and ischemic. The number of ischemic

beats is 16,794, while that of the healthy beats is

3734.

In order to evaluate the proposed classifier, a ten

fold cross validation is applied to the dataset. The

ten fold cross validation is described as follows:

1. Divide data into 10 set of size n/10

2. Train on 9 sets and test on 1 set

3. Repeat the process 10 times and take the

mean of the accuracy.

2.2 Signal Pre-processing

As mentioned in the previous section, the beats are

obtained automatically from the records using the

‘WQRS’ function provided by the Physionet

Toolkit. Each beat is then anchored such that the iso-

electric line prior to the P wave is set to zero. A

wavelet decomposition approach is used to denoise

the signal from high frequency noise (GD. Clifford

2005).

3 METHOD

The classification approach utilizes a heart cell

group model fitted to the patient’s ECG signal along

with the principle component analysis of the signal.

The method is described in the block diagram shown

in Figure 3.

A heart cell group model is used to generate a

template ECG signal. Then, using a nonlinear

constrained optimization technique, the model

parameters are updated until reaching a certain error

with the patient’s signal beat. The estimated model

fitting the ECG signal are then used with the PCA

components as features in the C4.5 decision tree

classifier.

Figure 3: Block diagram of the Ischemia diagnosing

method.

3.1 Heart Cell Group Model

Electrocardiograms indicate the electric activity of

the heart over the body surface. In general, two types

of model have been developed to characterize the

ECG signal. The first type is a model used for

interpolating experimental data and can be fitted to

ECG signals without having a reference to the

physical system. The second type is a model that can

characterize the ECG signal and can be related back

to the heart activity. The objective of this work is the

latter modelling approach, focusing on development

of a model that can estimate the activation sequences

of the heart cells from real patient ECG signals. This

objective is called the inverse problem. The

difficulty of this problem is that unless it is stated in

a particular manner, the solution will not be uniquely

defined.

Several techniques have been employed for

generating models to solve the forward and inverse

problem. These techniques overcome the uniqueness

problem by modelling the heart as a small number of

moving dipoles. Some of these techniques apply the

solution of Green’s theorem (Method of Moments)

or Multi-Pole technique to determine the scattering

of the electric waves over the heart. These methods

are considered accurate. However, the main

drawback of these techniques is the computational

complexity (Gulrajani 1998).

McSharry et al. presented a “dynamic ECG

model” that incorporates the ECG features as a

combination of Gaussian functions. Although this

model is easy to build, it cannot be related to the

heart cell activity (GD. Clifford 2005).

3.1.1 Proposed Cell Group Model

A Heart Cell Model (HCM) is proposed in this work

based on the reconstruction of the ECG signal using

a cell group model. This model accounts for the

wave propagation of the SA node, the AV node, the

bundle branches, Purkinje fibers, and left and right

ventricles. We hypothesize that the electric activity

of a heart cell group can be represented by a

difference of two sigmoid functions.

The electric activity of the myocardial cells is

caused by the variation of the positively and

negatively charged ions of the cells. As presented by

researchers (Andrew J. Pullan 2005), the electric

activity of the cell is given in

Figure 4 and it can be

approximated by a difference of two sigmoid

functions as shown in

Figure 5.

A HEART CELL GROUP MODEL FOR THE IDENTIFICATION OF MYOCARDIAL ISCHEMIA

Figure 4: Conduction activity of the heart.

Figure 5: Proposed heart cell activity.

The cell group activity is modelled as the difference

between two sigmoid functions:

()

11 2 2

,,,,

atc

ftacac k

−

=−

⎛⎞

⎜⎟

⎝⎠

, (1)

where k represents the magnitude of the wave, a

and a

control the rising slope, and c

and c

control

the translation in the direction of the time axis.

We hypothesize that the cumulative ECG signal

is generated from the atrial and ventricular

conduction activity. In this work, the P wave is

assumed to be generated from the SA node activity;

the PR interval from the AV node activity, and the

QRS complex and T wave are generated from the

activation of the bundle branches, the Purkinje fibers

and, the right and left ventricles.

3.1.2 ECG Generation

As presented above, the ECG signal can be

generated from the activation sequences of the heart

cell groups. The same steps are used to generate the

ECG from the modelled activation sequences. The

model divides the heart into groups or nodes. Each

node consists of a combination of cells at the SA

node, the AV node, the bundle branches, the

Purkinje fibers and, the right and left ventricles.

Each node activation and deactivation sequence is

represented as the difference between two sigmoid

functions. The variables in the sigmoid functions

consist of the magnitude, inflection (activation)

point and the inclination slope. By summing the

potential difference of the node signals at the

positive and negative terminals of each lead, the

ECG signal is generated:

()

[]

,,,,,

ECG

i SAAVBbPfLVRV

∈

+−

=−

∑

, (2)

where:

• SA and AV represent the activity of the

SA and the AV node respectively.

• Bb and Pf represent the activity of the

bundle branches and Purkinje fibers

respectively

• LV and RV represent the activity of the

Left and right ventricles respectively

•

and

−

are the difference between

two sigmoid functions as presented in (1)

for each of the nodes at the positive and

negative probes respectively

The following sections presents how the ECG

wave features are generated. The features are the P

wave, the PR segment, the Q wave, the R wave, and

the S wave (QRS complex), ST segment, and T

wave.

3.1.3 P Wave Generation

The P wave is generated from the potential

difference between the electric conduction activity

measured at the atrial cells at the positive and

negative probes. In this approach, the atrial

conduction activity at a single probe is estimated by

equation (1). Moreover, it is hypothesized that the P

wave can be generated from the conduction activity

of the SA node:

(

)

wave SA SA

Pff

+−

=−

, (3)

The generation of the P wave using the

difference of sigmoid estimation is shown in

Figure

0 50 100 150 200 250

-50

100

150

time (s)

SA activity

0 50 100 150 200 250

time (s)

P Wave

Figure 6: P wave generation using the differential sigmoid

model.

HEALTHINF 2008 - International Conference on Health Informatics

3.1.4 PR Segment Generation

The PR segment occurs as the impulse travels from

the AV node through the conducting tissue (bundle

branches, and Purkinje fibers) towards the

ventricles. Most of the delay in the PR segment

occurs in the AV node. The PR segment is generally

at the baseline; however, variations might occur due

to certain heart diseases. Thus, by modelling the

electric activity of the AV node as proposed in (1),

and similar to the procedure shown in (3), we are

able to generate the variations in the PR segment as

shown in Figure 7.

0 50 100 150 200 250

-10

time (s)

AV activity

0 50 100 150 200 250

-25

-20

-15

-10

-5

time (s)

PR interval

Figure 7: PR interval generation using the differential

sigmoid model.

3.1.5 QRS Complex and T Wave Generation

The QRS complex and the T wave denote the

interval for the beginning and end of the ventricular

activation. When generating the QRS complex, the

activity of the cell groups of the bundle branches,

Purkinje fibers, and left and right ventricles are

modelled during the ventricular cycle. The

representation of the model for the QRS complex

and T wave in an ECG signal is dependent on the

difference between the positive and negative

electrodes at the modelled cell groups.

Figure 8

through Figure 10 show how each wave of the QRS

complex and the T wave are generated using the

differential sigmoid model.

0 50 100 150 200 250

-10

time (s)

AV activity

0 50 100 150 200 250

-25

-20

-15

-10

-5

time (s)

PR interval

Figure 8: R wave and T wave generation.

0 50 100 150 200 250

-10

time (s)

LV activity

0 50 100 150 200 250

-20

-15

-10

-5

time (s)

ST segment

Figure 9: R wave and T wave generation.

0 50 100 150 200 250

-100

100

200

300

time (s)

RV activity

0 50 100 150 200 250

-100

-50

100

time (s)

R wave ; T wave

Figure 10: S wave and T wave generation.

0 50 100 150 200 250

-10

-5

time (s)

LV activity

0 50 100 150 200 250

-20

-15

-10

-5

time (s)

ST segment

Figure 11: ST segment generation.

3.1.6 Parameter Estimation and Signal

Fitting

This section discusses how to determine the

parameters of the activation sequences in order to

generate a real patient ECG signal. In order to

achieve this task, a parameter estimation of the

proposed model (1) and (2) is performed using the

minimization of the least squares with the real ECG

A HEART CELL GROUP MODEL FOR THE IDENTIFICATION OF MYOCARDIAL ISCHEMIA

signal. This process was performed with the help of

the fmincon function, in Matlab, which finds a

constrained minimum of a function for several

variables. The function being minimized is given:

()

ECG

signal

Error ECG f= −

∑

, (4)

The constraints applied to the function are that

the atrial activity occurs prior to that of the

ventricles. Moreover, the activation of the cell

activity is constrained to occur prior to that of the

deactivation. Additionally, the slopes of the

activation are higher than those of the deactivation

curves.

A template initial condition with known

parameters for

is used to set the initial condition

for the optimization process. Additionally, a

dynamic template is generated for each beat. This

choice of the template depends on the sign of the R

peak. This allows more accuracy during the

nonlinear optimization process. The highest cross-

correlation point between the initial template the

patient signal is then chosen.

Figure 12 shows the real and estimated ECG

signal. It can be seen that the ‘fitted’ signal

generated from the model matches the original

patient signal. The model parameters used to

generate the fitted signal are used as features in the

classification process.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

-100

-50

100

150

200

250

300

350

400

time (s)

ECG

Real Signal

Estimated Signal

Figure 12: Estimated signal and original ECG signal.

3.2 Principle Component Analysis

Principle Component Analysis (PCA) is a linear

transform where the basis functions are taken from

the statistics of the signal, and can thus be adaptive.

It is optimal in the sense of energy compaction, i.e it

places as much energy as possible in as few

coefficients as possible. The PCA is typically

implemented using Singular Value Decomposition.

The transform is generally not separable, and thus

the full matrix multiplication must be performed:

Uxx UX==, (5)

where the U is the basis for the transform. U is

estimated from a number of x

where

[

]

0:ik∈ 

[

]

()

UV xx x A

UeigvecAA

…

. (6)

3.3 C4.5 or J48 Decision Tree

Decision trees represent a supervised approach to

classification. A decision tree is a simple structure

where non-terminal nodes represent tests on one or

more attributes and terminal nodes reflect decision

outcomes. Generally, a decision tree algorithm

chooses the attributes that best differentiates the

output attribute values. The Weka classifier package

(Eibe Frank 2007) has its own version of C4.5

known as J48. Weka’s J48 is used in this work to

solve the classification problem.

4 RESULTS

The HCM-PCA/C4.5 classifier is applied to the

Long Term ST-Database. The proposed approach is

compared to the technique proposed in

(Stamkopoulos 1998). As mentioned before, the beat

is detected using an automatic tool ‘wqrs’ provided

by Physionet. The high frequency noise in the signal

is removed using wavelet decomposition (Clifford

2006). The model is fitted to the model by

minimizing the sum squared error using a

constrained optimization process. The constraints

are used to maintain the order of the heart’s

activation sequences. That is, the atrial activation

occurs prior to that of the ventricles and the

depolarization event occurs prior to the

repolarization. The model parameters are used in the

classification process, i.e. as features to determine

whether a beat is ischemic or healthy.

A C4.5 decision tree is used in the classification

process. As mentioned above, a 10 fold cross

validation is performed. The classification method is

applied with and without using the PCA components

as features. Using the model parameters without the

PCA features, the accuracy is 87.83% with

sensitivity and specificity of 92.62% and 65.69%,

respectively. Using the PCA features without the

model parameters leads to an accuracy of 87.83%

with sensitivity and specificity of 93.8% and 72.7%.

However, when using the PCA features in addition

to the model parameters, the accuracy increases to

HEALTHINF 2008 - International Conference on Health Informatics

91.62% with sensitivity of 94.89% and sensitivity of

75.66%. Sensitivity and specificity are defined as the

accuracy of detecting the ischemic beat and the

accuracy of detecting the non ischemic beat

respectively. The confusion matrices for the

proposed approaches are given in Table 1, Table 2,

and Table 3 respectively. Confusion matrix is a

visualization tool that presents the instances

classified as ischemic or healthy in its columns and

the actual classification in its rows.

Table 1: Confusion Matrix for HCM /C4.5 approach.

Classified as

Ischemic Healthy

Ischemic 15608 1255

Healthy 1243 2421

Table 2: Confusion Matrix for PCA/C4.5 approach.

Classified as

Ischemic Healthy

Ischemic 15877 986

Healthy 1044 2620

Table 3: Confusion Matrix for HCM-PCA/C4.5 approach.

Classified as

Ischemic Healthy

Ischemic 16035 828

Healthy 892 2772

It can be seen from Table 1 and Table 2 that the

sensitivity of the proposed approach increases by

10% when using the PCA components in addition to

the model parameters as features in the C4.5

decision tree classifier.

As mentioned above, the proposed approach is

compared to the techniques of (Stamkopoulos 1998)

as applied to the LT-ST database.

Table 4: Comparison between the proposed approach and

previous methods.

Approach Accuracy Sensitivity Specificity

HCM-PCA/C4.5 91.62% 94.89% 75.66%

Stamkopoulos 86.76% 91.73% 63.86%

It can be appreciated from Table 4 that the

proposed HCM-PCA/C4.5 approach performs better

than the previous methods by (Stamkopoulos 1998)

for the LT-ST database. However, we have not been

able to replicate the results of (Victor-Emil Neagoe

2003).

The importance in the proposed model, HCM, is

that it can be related back to the heart’s physical and

electrical activity. It can be seen that the parameters

of the HCM can be used in the detection of ischemic

and healthy heart beats. This is due to the fact that

the model parameters captured the information

regarding the ECG waves and segments, such as

slope, interval duration, magnitude and segment’s

variation.

5 CONCLUSIONS

A HCA-PCA/C4.5 approach is presented in this

work to diagnose ischemic and healthy beats. The

proposed approach is applied to the LT-ST database

provided by Physionet. The approach showed

excellent results when diagnosing ischemic and

healthy beats. The proposed modelling approach

provides a method to identify the features of ECG

signals and an estimate to the cellular eclectic

activity useful for ischemia detection. Finally, the

proposed classification approach can be extended to

detect different cardiac diseases.

REFERENCES

AHA 2005, Heart Disease and Stroke Statistics, Dallas,

Texas: American Heart Association.

Andrew J. Pullan, M.L.B., Leo K. Cheng 2005,

Mathematically Modeling the Electrical Activity of the

Heart from Cell to Body Surface and Back Again

World Scientific, New Jersey.

Clifford, G. 2006, 'ECG Bag',

<http://www.mit.edu/~gari/CODE/ECGtools/>.

Eibe Frank, M.H., Geoff Holmes, Mike Mayo, Bernhard

Pfahringer, Tony Smith, Ian Witten 2007,

WEKA, The

University of Waikato.

GD. Clifford, P.M. 2005, 'Method to Filter ECGs and

Evaluate Clinical Parameter Distortion Using Realistic

ECG model Parameter Fitting', Computers in

Cardiology

Gulrajani, R.M. 1998, 'The forward and inverse problems

of electrocardiography', vol. IEEE Engineering in

Medicine and Biology Magazine, pp. 84 - 101.

Jager, F., Taddei, A. & Moody, G. 2003, 'Long-term ST

database: A reference for the development and

evaluation of automated ischaemia detectors and for

the study of the dynamics of myocardial ischaemia.'

Medical and Biological Engineering and Computing,

vol. 41, pp. 172-171 182.

Long, C. 1980, Prevention and rehabilitation in ischemic

heart disease

, Baltimore: Williams & Wilkins.

Moody, J.M. 2001,

Preventive cardiology : strategies for

the prevention and treatment of coronary artery

disease

, Humana Press, Totowa, N.J.

N. Maglaveras, T.S., C. Pappas, and M. Strintzis 1994,

'Use of neural networks in detection of ischemic

A HEART CELL GROUP MODEL FOR THE IDENTIFICATION OF MYOCARDIAL ISCHEMIA

episodes from ECG leads', Neural Networks for Signal

Processing

, pp. 518-524.

Pardee, H. 1920, 'An electrocardiographic sign of coronary

artery obstruction', Arch Int Med, vol. 26, pp. 244-257.

R. Gianrossi, R.D., D. Mulvihill, K. Lehmann, P. Dubach,

A. Colombo, D. McArthur, and V. Froelicher 1989,

'Exercise- induced ST depression in the diagnosis of

coronary artery disease. A meta-analysis', Circulation,

vol. 80, pp. 87-98.

R.V. Andreao, B.D., J. Boudy, J.C.M. Mota 2004, 'ST-

segment analysis using hidden Markov Model beat

segmentation: application to ischemia detection',

Computers in Cardiology, pp. 381 - 384.

Stamkopoulos, T.D., K. Maglaveras, N. Strintzis, M.

1998, 'ECG analysis using nonlinear PCA neural

networks for ischemia detection',

IEEE Transactions

on Signal Processing

, vol. 46, no. 11.

Victor-Emil Neagoe, I.-F.I., Sorin Grunwald 2003, 'A

Neuro-Fuzzy Approach to Classification of ECG

Signals for Ischemic Heart Disease Diagnosis',

AMIA

Annu Symp Proc

, p. 494–498.

APPENDIX

The cost function for the constrained optimization

function is obtained by replacing (1) into (2):

()

,,,

min , ,

iii

iSAAVBb

Pf LV RV

ECG k g t

−

⎛⎞

⎜⎟

⋅

⎜⎟

⎝⎠

∑

ac , (7)

(

)

()

,1 ,1 ,2 , 2

,3 ,3 ,4 ,4

,, , , , ,

,,,,

ii i i i i

ii ii

tdtacac

dta c a c

−

, (8)

()

(

)( )

11 2 2 11 2 2

,,,, ,, ,,dtacac stac stac=−

, (9)

()

at c

stac

−−

, (10)

Subject to the constraints:

() ()

1,3 2, 4ii

cc< ,

() ( )

1,3 1,3SA AV

cc< ,

() ( )

1,3 1,3SA Rvep

cc<

() ( )

2,4 1,3SA AV

cc< ,

() ( )

2,4 1,3AV Rvep

cc< ,

() ( )

2,4 1,3AV Rven

cc< ,

() ( )

2,4 1,3AV Lvep

cc<

() ( )

2,4 1,3AV Lven

cc< ,

() ()

2,4 1,3Rven Lvep

cc< ,

() ()

2,4 1,3Rven Rvep

cc<

() ()

1,3 1,3Rvep Lvep

cc< ,

() ()

2,4 2,4Lvep Rvep

cc< ,

() ()

1,3 1,3Rvep Lven

cc< ,

() ()

2,4 2,4Lven Rvep

cc< ,

() ()

1,3 1,3Lvep Lven

cc<

() ()

2,4 2,4Lven Lvep

cc< ,

() ()

0.05 0.12

SA SA

cc<−<

() ()

0.05 0.12

SA SA

cc<−<

() ()

0.05 0.10

AV AV

cc<−<,

() ()

0.05 0.08

Rvep Rvep

cc<−<,

() ()

0.05 0.10

Rvep Rvep

cc<−<,

() ()

0.05 0.03

Rven Rven

cc<−<,

() ()

0.05 0.03

Lvep Lvep

cc<−<

() ()

0.05 0.10

Lvep Lvep

cc<−<

() ()

0.05 0.10

Lvep Lvep

cc<−<.

HEALTHINF 2008 - International Conference on Health Informatics