DESIGN OF THE ARTIFICIAL NEURAL NETWORK MODEL

FOR THE PREDICTION OF OUTCOME AFTER STROKE

Jiri Polivka Jr.

, Petr Kratochvil

, Vladimir Rohan

, Jiri Polivka

and Jana Kleckova

Department of Computer Science and Eng., University of West Bohemia, Univerzitni 8, Plzen, Czech Republic

Department of Neurology, Charles University Medical School and University Hospital in Plzen

Alej svobody 80, Plzen, Czech Republic

Keywords: Neural network, Stroke, Predictive medicine, Preventive medicine.

Abstract: In our contemporary research we are trying to develop the artificial neural network (ANN) model for the

prediction of outcome after the occurrence of stroke. This paper mentions some important facts about stroke

as well as the urgent need for Computer Assisted Decision Support (CAMS) systems in the relation to

clinical practice. The short review of related studies of ANN in medicine is included. The model input and

output parameters were selected and are also described. The basic ANN design for the predictive model is

mentioned together with the future directions of our research.

1 OVERVIEW

Computer Assisted Decision Support (CADS) in

medicine should enhance the consistency of medical

care in the future. Today there is an expanding range

of medical information stored in electronic form for

each patient, which could be effectively used in

computer-assisted diagnoses systems or preventive

and predictive models. CADS systems are also

excellent tools to cover rare conditions, since no

clinical expert can be expected to possess

encyclopedic knowledge of all of the exceptional

manifestations of diseases.

In our proposed work we tried to set up a new

predictive biomedical model which could be able to

make a prediction of stroke outcomes from the

analysis of various medical input parameters

acquired after patient´s hospitalization. Our

biomedical model uses the Artificial Neural

Networks (ANN) as a new technology for CADS

systems. Neural networks are very universal

instrument of approaching problems. The results

could be used for performing prediction if the output

of the network is continuous or classification if the

outputs are discrete values.

2 STROKE

Stroke is the third leading cause of morbidity and

mortality in the Western world, following ischemic

heart disease and cancer. There are more than 50

million stroke and transient ischemic attack (TIA)

survivors all over the world. More than 1 in 5

survivors may have a subsequent stroke in the next 5

years. The worldwide economic cost of stroke

including direct as well as indirect costs could be

approximately $68.9 billion. Permanent disability

remains a big problem, between 15% and 30% of

stroke survivors suffer permanent disability, 20% of

victims require institutional care within 3 months

after the stroke event (Lloyd-Jones et al., 2009).

One third of stroke patients are under the age of

65 that means a variety of populations are at the risk

and the disease should no longer be considered

confined to the elderly. Women are at a greater risk

for stroke than men. In 2005, women accounted for

60.6% of stroke deaths in the US. The increased

lifespan is the main factor for the increase in stroke

occurrence. However there are many others medical

risk factors including myocardial infarction,

coagulopathies, peripheral vascular disease,

hypertension, atrial fibrillation, or diabetes mellitus.

2.1 Classification of Stroke

The main stroke pathophysiological entites include

thrombosis, embolism, and hemorrhage. Stroke can

be classified as ischemic or hemorrhagic types, with

ischemic stroke accounting for approximately 85%

of the total number. Ischemic stroke occurs due to

467

Polivka Jr. J., Kratochvil P., Rohan V., Polivka J. and Kleckova J..

DESIGN OF THE ARTIFICIAL NEURAL NETWORK MODEL FOR THE PREDICTION OF OUTCOME AFTER STROKE.

DOI: 10.5220/0003875304670470

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 467-470

ISBN: 978-989-8425-88-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

either intracranial thrombosis or extracranial

embolism. Intracranial thrombosis is joined to

atherosclerosis, whereas extracranial embolisms

arise from the extracranial arteries or from the

myocardium, often because of concurrent

myocardial infarction, mitral stenosis, endocarditis,

atrial fibrillation or congestive heart failure. The

classification of hemorrhagic stroke can be done as

either intracerebral hemorrhage (ICH) or

subarachnoid hemorrhage (SAH). The common

causes for both ICH and SAH contain hypertension,

trauma, drug use, or vascular malformations (Adams

et al., 1993, Lloyd-Jones et al., 2009). In our case

the TOAST (Trial of Org 10172 in Acute Stroke

Treatment) classification of subtypes of acute

ischemic stroke is used. The acute ischemic stroke

subtypes than include Large-artery atherosclerosis

(embolus/thrombosis), Small-vessel occlusion

(lacune), Stroke of other determined etiology and

Stroke of undetermined etiology. The hemorrhagic

stroke type is also included in the model.

2.2 Clinical Diagnosis of Stroke

Stroke is a medical emergency. The successful

treatment relies especially on its right and well-

timed clinical diagnosis. Great effort in acute stroke

management is focused on correct and rapid

diagnosis and maximal shortening of “onset to

needle time”. It is critical for determining eligibility

for thrombolytic therapy, as the window of

opportunity for therapeutic effectiveness of stroke is

only a few hours (Morgenstern et al., 2004).

There are many new imaging techniques

available which leads to the potential for earlier

opportunities for therapeutic intervention in stroke

patients. The neurological imaging can be used for

the differentiation between hemorrhagic and

ischemic stroke. Important features gained from

brain imaging include detecting early infarction and

determining the location and degrees of infarct and

vascular distribution of the lesions. Computed

tomography (CT) is routinely used in the initial

acute assessment of stroke patient. In acute stroke

case, MRI diffusion-weighted imaging (DWI)

techniques have the ability to differentiate between

various stroke subgroups.

2.3 Therapeutic Intervention in Stroke

The main goal in stroke therapeutic intervention is to

salvage as much cerebral tissue as possible.

Therefore effective thrombolytic therapy must be

initiated rapidly. In 1996, US Food and Drug

Administration (FDA) approved revolutionary

therapeutic intervention with intravenous

recombinant tissue plasminogen activator (rtPA). It

has been used consistently for thrombolysis in acute

stroke. The window of opportunity is less than 4.5 h

from the onset of symptoms (Hacke et al., 2008).

3 ARTIFICIAL NEURAL

NETWORKS IN MEDICINE

The neural networks application in the diagnosis of

cardiovascular disease, primarily in the detection

and classification of at-risk people from their ECG

waveforms was done (Nazeran and Behbehani,

2001). Anoher study uses neural networks to classify

normal and abnormal ECG waveforms and the

abnormal ECG and is described in (Celler and

Chazal, 1998). It made classification of the

waveforms with 70.9% accuracy.

In the next study the ANN which uses non-linear

statistics for pattern recognition was used in

predicting one-year liver disease-related mortality

with the initial clinical evaluation information.The

application of ARTMAP in medicine include

classification of cardiac arrhythmias was described

in (Ham and Han, 1996). The selection of treatment

for schizophrenic and unipolar depressed in-patients

was also made (Modai et al., 1996). Another study

described using of ANN to predict patients with

colorectal cancer more accurately than

clinicopathological methods.

Anothe work based on ANNs is able to detect

ischaemic episodes in long duration ECG recordings

(Papaloukas et al., 2002). The use of the ANN

model as a data mining tool was made to model

complex behaviour of different molecular markers

of dialysis treatment (Elmer et al., 2005). The ANN

model was also used for prediction of

tromboembolic stroke (Shanthi et al., 2010).

4 THE PROPOSED MODEL

The main goal of our proposed system should be the

correct prediction of the stroke outcomes in patients

who were admitted to the hospital with the stroke

diagnosis. The outcome prediction will be made

from the various input medical data processed in the

model. The output describes the overall medical

condition of the stroke patient which is represented

as a grade on some international summarizing scale.

The model outputs (stroke outcomes prediction) are

HEALTHINF 2012 - International Conference on Health Informatics

468

computed for the time points 7 and 90 days after the

stroke occurrence (patient´s admittance to the

hospital).

In the proposed model design we use two

different international scale systems for the stroke.

First is the 42-point National Institutes of Health

Stroke Survey (NIHSS) scale. It was developed to

assist with diagnostic consistency among physicians

and was designed to be completed within 5 to 8 min

(Goldstein and Samsa, 1997). The NIHSS quantifies

neurological deficits in stroke patients. The second

one is The Modified Rankin Scale (mRS). It is a

commonly used scale for measuring the degree of

disability or dependence in the daily activities of

people who had a stroke. This scale is widely used in

clinical outcome measures for stroke clinical trials.

The scale are from 0-6, running from perfect health

without symptoms to death (0 - no symptoms, 1 - no

significant disability, able to carry out all usual

activities, despite some symptoms, 2 - slight

disability, able to look after own affairs without

assistance, but unable to carry out all previous

activities, 3 - moderate disability, requires some

help, but able to walk unassisted, 4 - moderately

severe disability, unable to attend to own bodily

needs without assistance, and unable to walk

unassisted, 5 - severe disability, requires constant

nursing care and attention, bedridden, incontinent, 6

– dead).

4.1 Patient Data and Feature Selection

The various medical inputs for our model were

selected with the help of neurological experts from

the department of neurology, Charles University

Medical School and University Hospital in Plzen.

All stroke patient´s data either for the training set or

the validation set of the model will come from the

University Hospital in Plzen.

The input parameters of the model are listed in

the table 1 (SITS Parameters) and table 2 (non-SITS

parameters). SITS (Safe Implementation of

Treatments in Stroke) is an academic-driven, non-

profit, international collaboration. It is an initiative

by the medical profession to accelerate clinical trials

and to certify excellence in acute and secondary

prevention stroke treatment and to develop

knowledge and leading research. The SITS Network

includes a broad range of hospitals, as well as the

University Hospital in Plzen. The SITS Stroke

Registry is an internet-based interactive stroke

registry developed by SITS. It serves as an

instument for stroke centres to compare own

treatment results with other stroke centres. The basic

parameters which can be found in SITS protocols

were enriched with some other inputs, such as new

laboratory markers of acute stroke or stroke type

classification. This expert´s medical input analysis

and selection of model parameters should ensure

superior predictive accuracy of our biomedical

model.

Table 1: SITS input parameters.

SITS Parameters

In.

No.

Input Name Input range

Modified Rankin Scale before stroke

1 mRS score 0 – 6

Prior treatments

2 Antiplatelet tr. (Dypiridamol,

Clopidogrel)

Yes/No

3 Anticolagulants (Heparin) Yes/No

4 Anti – diabetic (Insulin) tr. Yes/No

5 Antihypertensive tr. Yes/No

Risk factors

6 Hypertension Yes/No

7 Diabetes (Dg. of diabetes) Yes/No

8 Hyperlipidemia (Dg.of

hyperlipidemia)

Yes/No

9 Current Smoker Yes/No

10 Previous Smoker Yes/No

11 Previous Stroke (earlier than 3

months)

Yes/No

12 Previous Stroke (within 3 months) Yes/No

13 Previous TIA / Amaurosis fugax Yes/No

14 Atrial fibrillation Yes/No

15 Congestive heart failure Yes/No

Laboratory Indicators

16 Glucose (mmol/l) Number

17 Cholesterol (mmol/l) Number

NIHS

18 NIH Score 0 – 42

Imaging – CT

19 CT current infarct Yes/No

20 Local haemorrhage Yes/No

Other Parameters

21 Age Number

22 Sex M / F

23 Weight Number

24 Systolic blood pressure (mmHg) Number

25 Diastolic blood pressure (mmHg) Number

5 THE ARCHITECTURE OF ANN

The architecture of the artificial neural network is

the multilayered feed-forward network with 37 input

nodes (in the future 41). The first experimental case

uses 20 hidden nodes. The output is designed as one

node which could be able to calculate the overall

patient’s stroke outcome in the international scale of

NIHSS and RANKIN.

DESIGN OF THE ARTIFICIAL NEURAL NETWORK MODEL FOR THE PREDICTION OF OUTCOME AFTER

STROKE

469

Table 2: Non-SITS input parameters.

Non SITS parameters

In. No. Input Name Input

range

Treatment

26 I.V. Trombolysis Yes/No

27 Stroke care unit Yes/No

Stroke Type

28 Large-artery atherosclerosis Yes/No

29 Cardioembolism Yes/No

30 Small-vessel occlusion Yes/No

31 Stroke of other determined

etiology

Yes/No

32 Haemorrhalgic stroke Yes/No

Laboratory indicators

33 CRP Number

34 Platelets Number

35 Leukocytes Number

36 Fibrinogen Number

37 Vitamin D Number

Other possible laboratory indicators (in the future)

38,39,40,41 IL 6, SAA, NSE, PARK 7 Number

6 CONCLUSIONS

In this paper we have discussed a design of our new

biomedical model for the stroke outcome prediction

which is based on artificial neural network

architecture. First important part of the model

creation was the selection of input and output

parameters. This task was done with the help of

neurological experts from the University Hospital in

Plzen. Than the architecture of an ANN was

designed and also briefly referred in this paper. Now

the model is prepared for training patient´s data set.

The ANN training and optimization of the model are

our main research tasks to the future. The work

presented in this paper is supported by The Czech

Science Foundation project 106/09/0740 dealing

with brain perfusion modelling.

REFERENCES

Adams H. P. Jr., Bendixen B. H., Kappelle L. J., Biller J.,

Love B. B., Gordon D. L., Marsh E. E., 3

, 1993.

Classification of subtype of acute ischemic stroke:

definitions for use in a multicenter clinical trial.

TOAST. Trial of Org 10172 in Acute Stroke

Treatment. Stroke 1993;24:35– 41.

Celler, B. G., Chazal, P., 1998. Low computational cost

classifiers for ECG diagnosis using neural networks.

Proceedings of the International Conference of

Engineering in Medicine & Biology Society (EMBC

1998), pp. 1337–1340.

Elmer Andres Fernández, Rodolfo Valtuille, Jesus

Rodriguez Presedo, and Peter Willshaw (2005)

Comparison of Standard and Artificial Neural Network

Estimators of Hemodialysis Adequacy. International

Center for Artificial Organs and Transplantation.

29(2):159–165, Blackwell Publishing, Inc.

Goldstein L. B., Samsa G. P. Reliability of the National

Institutes of Health Stroke Scale: extensit to non-

neurologists in the context of a clinical trial. Stroke

1997;28:307–10.

Hacke W., Kaste M., Bluhmki E., Brozman M., Da´ valos

A., Guidetti D. et al. Thrombolysis with alteplase 3 to

4.5 hours after acute ischemic stroke. N Engl J Med

2008;359:1317–29.

Ham, P. M. and Han, S., "Classification of cardiac

arrhythmias using fuzzy artmap" IEEE Transactions

on Biomedical Engineering, 43(4): 425–430 (1996).

Lloyd-Jones D., Adams R., Carnethon M., De Simone G.,

Ferguson T. B., Flegal K., et al. Heart disease and

stroke statistics—2009 update: a report from the

American Heart Association Statistics Committee and

Stroke Statistics Subcommittee. Circulation

2009;119:e21–181.

Modai, I., Israel, A., Mendel, S., Hines, E.L. and

Weizman, R., "Neural network based on adaptive

resonance theory as compared to experts in suggesting

treatment for schizophrenic and unipolar depressed in

patients," Journal of Medical Systems, 20(6): 403–412

(1996).

Morgenstern L. B., Lisabeth L. D., Mecozzi A. C., Smith

M. A., Longwell P. J., McFarling D. A., Risser J. M.

A population-based study of acute stroke and TIA

diagnosis. Neurology 2004;62:895–900.

Nazeran, H., Behbehani, K., 2001. Neural networks in

processing and analysis of biomedical signals. In M.

Akay (Ed.), Nonlinear biomedical signal processing:

Fuzzy logic, neural networks and new algorithms, pp.

69–97.

Papaloukas, C.,Fotiadis, D. I.,Likas, A., and Michalis, L.

K., 2002. An ischemia detection method based on

artificial neural networks. Artificial Intelligence in

Medicine, 24, 167– 178.

Shanthi, D., Sahoo, G. and Saravanan, N., 2010. Designing

an Artificial Neural Network Model for the Prediction

of Thrombo-embolic Stroke, International Journals of

Biometric and Bioinformatics (IJBB)

, Volume (3)

SITS International, https://sitsinternational.org

HEALTHINF 2012 - International Conference on Health Informatics

470