Predictive Clustering Learning Algorithms

for Stroke Patients Discharge Planning

Luigi Lella

1,*

, Luana Gentile

2,†

, Christian Pristipino

3,‡

and Danilo Toni

4,§

ASUR Marche, via Oberdan n.2, Ancona, Italy

Dept. of Human Neurosciences, Sapienza University, Rome, Italy

San Filippo Neri Hospital, ASL1 Roma, Rome, Italy

Dept. of Human Neurosciences, Sapienza University, Rome, Italy

Keywords: Pattern Recognition and Machine Learning, Big Data in Healthcare, Data Mining and Data Analysis,

Decision Support Systems.

Abstract: Stroke patients discharge planning is a complex task that could be carried out by the use of a suitable

decision support system. Such a platform should be based on unsupervised machine learning algorithms to

reach the best results. More specifically, in this kind of prediction task clustering learning algorithms seem

to perform better than the other unsupervised models. These algorithms are able to independently subdivide

the treated clinical cases into groups, and they can serve to discover interesting correlations among the

clinical variables taken into account and to improve the prediction accuracy of the treatment outcome. This

work aims to compare the prediction accuracy of a particular clustering learning algorithm, the Growing

Neural Gas, with the prediction accuracy of other supervised and unsupervised algorithms used in stroke

patients discharge planning. This machine learning model is also able to accurately identify the input space

topology. In other words it is characterized by the ability to independently select a subset of attributes to be

taken into consideration in order to correctly perform any predictive task.

* www.linkedin.com/in/luigi-lella

† https://www.neuroscienze.uniroma1.it/

‡ https://www.aslroma1.it/presidi-ospedalieri/presidio-ospedaliero-san-filippo-neri

§ https://www.neuroscienze.uniroma1.it/

1 INTRODUCTION

According to the Italian Ministry of Health website,

approximately 196,000 strokes occur every year in

Italy, of which 20% are relapses. As defined by the

World Health Organization, stroke is a "neurological

deficit of cerebrovascular cause that persists beyond

24 hours or is interrupted by death within 24 hours"

(Italian Ministry of Health website, 2020).

Stroke is caused by an interruption of

oxygenated blood supply due to an occlusion or a

rupture of the arteries supplying the brain. As a

result, brain functions controlled from that area

(limb movement, language, vision, hearing or other)

are partially or totally impaired or lost (Donnan et

al., 2008). About 10-20% of people with stroke die

within a month and another 10% within the first year

after the event. Only 25% of stroke survivors

recover completely, 75% survive with some form of

disability, and half of these suffer from a deficit so

severe that they lose self-sufficiency.

Major risk factors include age, high blood

pressure, tobacco smoking, obesity, high blood

cholesterol, diabetes mellitus, a previous TIA or

stroke, and atrial fibrillation. Diagnosis is performed

by a physical examination and it is supported by

neuroimages (CT and/or MR).

During hospitalization, different pharmacological

and/or interventional treatments are put in place to

preserve vital functions and minimize brain damage.

The National Institutes of Health Stroke Scale

(NIHSS) is used to assess stroke severity (Putra

Pratama et al., 2019; Lyden et al., 2009).

In the discharge phase, it is important to make a

proper plan, not only to enhance individual recovery,

but also to reduce the high social burden and the use

of health system resources (Mess et al., 2016)

296

Lella, L., Gentile, L., Pristipino, C. and Toni, D.

Predictive Clustering Learning Algorithms for Stroke Patients Discharge Planning.

DOI: 10.5220/0010187502960303

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 296-303

ISBN: 978-989-758-490-9

(Pereira et al.,2014). To help in planning, functional

outcome after stroke is evaluated using the modified

Rankin Scale (mRS) which measures the degree of

disability or dependence in activities of daily living.

The lower the score the higher the likelihood of

being able to live at home with a degree of

independence after being discharged from the

hospital or a long-term care ward (Saver et

al.,2010)(Wilson et al.,2002).

Aim of this paper is to present an innovative and

effective prediction model for discharge planning,

based on a machine learning algorithm derived from

data gathered during hospitalisation in the acute

phase.

2 PREVIOUS WORK

A considerable body of literature exists on the use of

machine learning algorithms based on the

assimilation of the data of previously treated clinical

cases. The accuracy of prediction generally tends to

increase over time, as new data become available

(Bishop, 2006). However, there is no algorithm

capable of providing the best predictive accuracy for

each category of problem (Alpaydin, 2020).

Machine learning algorithms can be subdivided

into supervised and unsupervised. In the supervised

learning human experts select the correct answers of

the machine learning model, while unsupervised

learning does not need the intervention of human

experts (Alpaydin, 2020).

Within the unsupervised learning paradigm, it is

possible to operate a further subdivision between

symbolic models that seek to reach a formal

representation of knowledge (using for example

logical representations, inference rules or decision

trees) and sub-symbolic models where the acquired

knowledge is stored in complex representations such

as artificial neural networks. Clustering learning

algorithms (Van Hulle, 1989; Kohonen, 1988;

Kohonen, 1989; Kohonen, 1990) are subsymbolic

unsupervised models that allow to achieve the best

results in an unsupervised manner. If a class attribute

is chosen, such as the length of the hospital stay or

the outcome expressed using a suitable evaluation

scale, these algorithms are able to divide the clinical

cases that can occur in clusters, corresponding to

hospitalizations that end with the same length of

hospital stay or with the same outcome. Even if

these models are not classifiers, after being trained

with a set of clinical cases used as test sets, they are

able to assign a test set record to a correct class,

achieving a considerable high prediction accuracy

when new treated cases are processed. Among the

learning cluster algorithms, Kohonen's Self

Organizing Maps (Kohonen, 1989) have been

widely used in the health sector, but better results

can be achieved by using a more adaptive algorithm

such as Fritzke's Growing Neural Gas (GNG)

(Fritzke,1994). GNG is an incremental network

model based on a simple Hebb-like learning rule.

Unlike previous approaches like the “neural gas”

method of Martinetz and Shulten (Martinetz and

Shulten, 1991) (Martinetz and Shulten,1994) this

model has fixed parameters and it is able to continue

the learning phase, adding other neural units and

connections, until a performance goal is achieved.

By being able to accurately identify the input space

topology, this model is able to identify the variables

to be considered in order to effectively operate the

prediction of a class attribute.

For example, if the mRS is chosen as class

attribute (that is the study variable to be predicted),

the GNG model could predict that a male individual

aged between 40 and 65, with a particular form of

diagnosed stroke and treated with certain

pharmacological therapies, at the end of the hospital

stay, is able to reach a mRS score of 1. This

prediction could be made without taking into

consideration other variables such as risk factors,

results of instrumental and laboratory tests etc. In

other cases the same model could need a greater

number of variables in order to make an accurate

prediction.

The GNG model has already been successfully

used in similar areas, such as in the prediction of the

length of hospital stay (Lella and Licata, 2017)

Functional outcome after stroke is related to

many variables, from biometric data (gender, age,

weight, pressure levels) to risk profiles, from the

results of laboratory tests to the results of

instrumental tests to therapies.

Given this complex interaction of many

variables, using machine learning appears to be the

best solution choosing an unsupervised model that

can improve its predictive accuracy over time. The

purpose of these algorithms is to identify over the

course of time new possible and clinically

meaningful correlations among the available data.

This approach has proved particularly effective

in the clinical setting, especially to predict the

therapeutic outcome of stroke patients (Zdrodowska,

2019; Zdrodowska et al., 2018; Chen et al., 2017)

(Alaiz-Moreton, 2018). The best results were

obtained through the use of the Random Forest

algorithm (Breiman, 2001; Alaiz-Moreton, 2018)

(Tin Kam Ho, 1998; Tin Kam Ho, 1995), an

Predictive Clustering Learning Algorithms for Stroke Patients Discharge Planning

297

ensemble learning method for classification obtained

through the aggregation of decision trees, and the

PART model, a partial decision tree algorithm that

does not need to perform global optimization like

other rule based learners (Zdrodowska et al., 2018;

Ali and Smith, 2006). Excellent results were also

obtained by using supervised models such as the

Support Vector Machine (SVM) (Zdrodowska et al.,

2018), a model based on a binary non-probabilistic

linear classifier (Ben Hur et al.,2001) (Cortes and

Vapnik, 1995).

3 METHODOLOGY

On the basis of the above literature review, it was

decided to use GNG networks to predict the status

of stroke patients one day and seven days after

entering the hospital, as well as at hospital discharge

and after three months, comparing the results with

the ones achieved by the best models used in this

kind of study, i.e. the Random Forest model, the

PART model and the SVM model.

The ZeroR model (Witten et al., 2011), the

OneR (Holte, 1993), the Naive Bayes (John and

Langley, 1995) and the J48 (Witten et al., 2011)

were also taken into consideration.

The ZeroR model is used as a benchmark to

verify that all the other tested algorithms have been

configured and used correctly. ZeroR always

predicts the most frequent class variable in the

presence of any combination of input variables.

Given its simplicity, it has generally a much lower

level of prediction accuracy than the other

algorithms. If this does not occur, the found result

may be due to a bad data selection and coding, or to

a bad configuration of the models.

The OneR, which stands for "one Rule", is a

one-level decision tree. In various areas and

predictive tasks, this model has proved to be much

more efficient than other more complex models, and

it is always advisable to check whether the problem

in question can be effectively treated using this

model which requires a reduced amount of

resources.

The Naive Bayes is a simple probabilistic

classifier based on applying Bayes' theorem with

strong (naive) independence assumptions between

the features.

The J48 is a decision tree based on the "divide

and conquer" strategy used recursively. At each

training step, the node characterized by the highest

amount of information is selected and split into a

series of nodes corresponding to some possible

values that the original node can assume. The

process ends when all the considered instances refer

to the same class attribute value.

The study involved a subset of 20,000 samples

taken from the Italian subset of the SITS registry

(Safe Implementation of Treatments in Stroke

website, 2020), a non-profit, research- driven,

independent and international monitoring initiative

for stroke patients.

Of these, just the data that hold the fields of the

outcome 24 hours after the patient's access to the

hospital (Global Outcome 24), the mRS at 7 days

(Rankin at 7 days), the mRS at hospital discharge

(Rankin at hospital discharge) and the mRS at 3

months (Rankin at 3 months) were taken into

consideration. These three variables were chosen as

class attributes for the tests.

The first class attribute is a qualitative clinical

variable that can take 7 possible values

("muchBetter", "better", "unchanged", "worse",

"muchWorse", "dead"). The other three class

variables can only take the scores 0,1,2,3,4,5 and 6.

The first subgroup of records, with the global

outcome at 24 hours specified, consisted of 13008

records characterized by 69 non class attributes; the

second subgroup, with the mRS at 7 days specified,

was made by 10460 records characterized by 99 non

class attributes; the third subgroup, with the mRS at

hospital discharge specified, was constituted by

4989 records characterized by 129 non class

attributes; the fourth subgroup, with the mRS at 3

months specified, was made by 10777 records

characterized by 152 non class attributes. The

reason why the number of non-class attributes is

different in the four subsets is that an higher length

of hospitalization also increases the number of

available data deriving from further tests performed.

Therefore, the number of non-class variables that

can be considered increases.

The characteristics of the four subsets of records

are presented in table 1.

Table 1: Features of the considered data sets.

Data set Total no.(male,female); avg age(min,max);

hemorrhagic stroke; ischemic stroke

1–Global

Outcome 24

13008(6974,6034); 71(14,102); 1902;

11106

2–Rankin at

7 days

10460(5491,4969); 71(14,101); 1376;

9084

3–Rankin at

Hospital

Discharge

4989(2780,2209); 68(14,102); 557; 4432

4- Rankin at

3 months

10777(5685,5092); 71(14,104); 755;

10022

HEALTHINF 2021 - 14th International Conference on Health Informatics

298

Part of the patient data was incomplete, but it

was not necessary to perform any type of data

cleaning due to the data entry performed by a

codified online form.

Data were discretized and normalized before

being processed. All predictive machine learning

models taken into consideration were trained with

60% of the samples and tested with the remaining

40% of the samples. The Weka 3.8.4 platform was

used to test ZeroR, OneR, J48, Naive Bayes, SVM

and Random Forest models, while a Java

implementation was used to test the GNG model.

The sequential minimal optimization algorithm

(Platt, 1998) was used to train the SVM model.

The GNG model was tested with the following

parameters:

λ=100, ε

=0.2, ε

=0.006, α=0.5, α

max

=50, δ=0.995.

The training was stopped when the mean square

error, i.e. the main of the local square error related to

each unit (expected distortion error), dropped below

the threshold of E=0.7.

4 RESULTS

The results in terms of prediction accuracy (i.e. the

number of correct predictions over the total number

of predictions) are shown in table 2.

All the tested models reached an higher

prediction accuracy than the ZeroR model, and the

OneR model resulted to be the second worst

algorithm, confirming the complexity of the

prediction task. Naive Bayes, J48, SVM and PART

performed better with a low number of training

records characterized by more non class attributes.

The best results were obtained by the GNG

model and by the Random Forest model, followed

by PART and SVM. The result confirmed the

correctness of the choice of the unsupervised models

over the supervised ones and the chosen clustering

learning model proved to be more performing than

the Random Forest ensemble model.

Once trained, it is also possible to use the GNG

model to identify which non-class attributes are

linked to particular values of the selected class

attribute. For example, it is possible to identify

which clinical variables are linked to the worsening

of patients during the first 24 hours of

hospitalization.

Using the Girvan-Newman algorithm (Girvan

and Newman, 2002) it is possible to identify

communities of nodes starting from the portion of

the trained model of self-organizing neural network

associated with a deterioration of state. In the case of

Table 2: Prediction accuracy of the tested models.

Tested

Models

Prediction Accuracy on data set 1; data

set 2; data set 3; data set 4

ZeroR

30.67; 19.55; 24.05;

27.73

OneR 38.72; 29.96;

37.36;

43.73

Naive

Bayes

29.44; 46.21;

80.36;

56.76

J48

68.9; 71.48;

82.08; 44.88

SVM 41.48; 78.46; 99.19; 82.57

PART 72.26; 76.54; 83.04; 46.89

Random

Forest

97.99; 97.54;

97.25;

84.89

GNG

99.17; 99.64; 99.05; 89.88

the first GlobalOutcome24 subset the selected nodes

are those in which the values of the class attribute

code corresponding to "worse", "muchWorse" and

"dead" exceed a threshold value that has been set

equal to 0.7. Girvan-Newman's algorithm identifies

the communities of nodes by eliminating those

connections characterized by the greatest number of

shorter paths that link each pair of nodes. After

training the first model with the records of the first

subset of input records (GlobalOutcome24) and

selecting the network portion corresponding to the

values of the class attribute "muchWorse" and "dead",

the Girvan-Newman algorithm was used to remove

the first 100 connections characterized by the greatest

number of shorter paths between pairs of nodes.

Table 3: Clinical variables related to death in the first 24

hours of hospitalization.

Cluster

no.

Label (weight)

Hypertension(0.99); HighTemperature(0.97)...

NIHSS1A=3(1.00); GenderMale(1.00)...

NIHSS5A=4(0.99); NIHSS6A=4(0.99)...

LowAPTTvalues(1.00); GenderMale(1.00)...

NIHSS11=9(1.00); GenderMale(1.00)...

GenderMale(1.00);Hypertension(0.96)

Hypertension(0.99); NIHSS5B=4(0.98)...

Hypertension(0.99); Age>=80(0.96)...

Hypertension(0.99); Diabetes(0.95)...

Hypertension(0.99); HighTemperature(0.96)...

CurrentInfarct(1.00); GenderMale(1.00)...

NIHSS5A=4(0.98); Age65-80(0.83)...

NIHSS1C=2(0.95); NIHSS6A=4(0.90)...

Hyperlipidaemia(1.00);GenderMale(1.00)...

NIHSS4=2(1.00); GenderMale(1.00)...

PreviousStroke<3months(1.00)...

NIHSS8=2(1.00); GenderMale(1.00)...

CerebralOedema(1.00); GenderMale(1.00)...

NIHSSB=2(0.93); NIHSS1C=2(0.88)...

NIHSS7=4(1.00); GenderMale(1.00)...

Predictive Clustering Learning Algorithms for Stroke Patients Discharge Planning

299

Subsequently, the non-class attributes associated

with the nodes of the individual communities were

extracted, that is, those characterized by

corresponding code values higher than the threshold

value of 0.7.

For each node, using a weight function tf-idf

(Baeza, 1999), the attribute most frequent in its own

community and less frequent in the overall set of

extracted communities was chosen. In this way, all

the non-class attributes associated with the

worsening of the clinical picture were identified.

The results of this processing are shown in table 3

and table 4.

Table 4: Clinical variables related to the worsening in the

first 24 hours of hospitalization.

Clus

ter

no.

Label (weight)

NIHSS1B=2(1.00); Hypertension(1,00)...

Hypertension(1.00); NIHSS5A=4(0.88)...

Hypertension(0.97); NIHSS5B=4(0.87)...

Hypertension(0.92); NIHSS4=2(0.88)...

NIHSS1B=2(1.00); Hypertension(0.94)...

Hypertension(0.97); GenderFemale(0.96)...

Age65-80(0.98); NIHSS3=2(0.98)...

NIHSS5B=4(0.99); NIHSS4=2(0.99)...

Hypertension(1,00); LowAPTTvalues(0.93)...

NIHSS4=2(0.99); NIHSS11=2(0.88)...

Hypertension(0.98); CurrentInfarct(0.92)...

NIHSS10=1(0.99); Age18-65(0.87)

Hypertension(0.86); GenderMale(0.81); ...

For each cluster of nodes extracted by the use of

Girvan-Newman's algorithm, the non-class attributes

associated to the corresponding nodes are listed. Due

to the GNG model training method, it is reasonable

to assume that each identified cluster is related to a

subset of analysed clinical cases.

The clinical variables associated with the various

clusters which are related to cases of serious

worsening and death patient are displayed in the

results tables 3 and 4. In the first lines the most

important clusters are represented, i.e. those

associated with a greater number of neural units and

therefore with a greater number of clinical cases.

For each cluster the non-class attributes, i.e. the

clinical variables responsible for the worsening or

the death of the patient, are sorted in descending

order by weight associated with the attribute. The

weight is a coefficient between 0 and 1 which

indicates how much the neuronal units associated

with the selected cluster are activated when the

attribute is present in the considered clinical case.

Values close to 0 mean that the attribute is not very

relevant for the worsening or death of the patient,

those close to 1 are instead considered the main

responsible factors.

The obtained results show that the presence of

hypertension, assessed on the basis of the systolic

and diastolic blood pressure values, is considered an

important potential factor for the worsening the

patient's state which can also lead to death. The

presence of hypertension alone, however, is not

sufficient to infer the risk of worsening or death.

Considering for example the first cluster of table 3

related to the death cases, hypertension is an

important factor having a weight equal to 0.99, but it

must also be accompanied by other factors such as

hightemperature, hyperlipidaemia, genderfemale,

NIHSS4 = 2 and age65-80.

Using the tf-idf algorithm, the most

representative clinical variables of the considered

cluster were selected. These are represented in bold

and they allow to highlight which are the attributes

most related to the worsening or death of the patient.

For example, the attributes of table 3 most

related to death cases with the relative weights are

PreviousStroke<3months (1.00), Hyperlipidaemia

(1.00), LowAPTTvalues(1.00), CerebralOedema

(1.00), GenderMale (1.00), NIHSS8 = 2 (1.00),

NIHSS4 = 2 (1.00), NIHSS1A = 3 (1.00), NIHSS7 =

4 (1.00), Hypertension (0.99), NIHSS5A = 4 (0.99),

NIHSS5B = 4 (0.98), HighTemperature (0.96),

Age> = 80 (0.96), Diabetes (0.95), NIHSS6A = 4

(0.90), NIHSS1C = 2 (0.88), Age65-80 (0.83). This

means that the presence of a previous stroke that

occurred no more than three months before,

accompanied by hyperlipidaemia or a cerebral

oedema or a low level of APTT, especially if the

patient is male can be considered a quite worrying

clinical picture. An age greater than 80 is to be

considered a more important risk factor than an age

between 65 and 80.

The clinical variables of table 4 related to the

severe worsening of the patient with the relative

weights are NIHSS1B = 2 (1.00), NIHSS5B = 4

(0.99), NIHSS10 = 1 (0.99), Age65-80 (0.98),

GenderFemale (0.96), Hypertension (0.94),

LowAPTTvalues(0.93), NIHSS11 = 2 (0.88),

NIHSS4 = 2 (0.88), NIHSS5A = 4 (0.88), NIHSS5B

= 4 (0.87), GenderMale (0.81). This means that the

presence of a low sense of orientation, of plegia or

dysarthria especially if the patient is aged between

65 and 80 and hypertensive is to be considered a

potential worsening factor. Female patients are

considered more at risk of serious worsening than

male patients.

HEALTHINF 2021 - 14th International Conference on Health Informatics

300

A similar study was carried out to identify the

clinical variables most correlated with a serious level

of disability after 3 months (corresponding at a mRS

of 3, 4 or 5). The results are shown in table 5.

As it can be seen, high blood pressure levels are

always related to the appearance of a disabling

stroke. The mRS at three months is also strongly

affected by the patient's state 7 days after

hospitalization. Generally there are only slight

differences in outcomes between male and female

patients, although clusters 1, 4 and 6 show a greater

correlation with the male sex while clusters 3 and 5

are more correlated with the female sex.

The first cluster is represented by patients aged

between 65 and 80 characterized by the presence of

occlusions that lead to the appearance of an

ischaemic penumbra (perfusion infarct mismatch).

The second cluster is represented by underweight

patients over the age of 65 characterized by diabetes

and hyperlipidaemia and occlusions and a

hyperdensity of the arteries. The third cluster relates

to underweight patients characterized by a mRS of 4

both after the first 7 days and upon discharge from

the hospital. Such patients are characterized by low

APTT levels. The fourth cluster is made up of

patients aged between 18 and 65, with high blood

cholesterol levels, diabetes and occlusions. The fifth

cluster is represented by patients over eighty

characterized by hyperlipidaemia and low levels of

APTT. The sixth cluster is associated with pre-

diabetic patients aged between 65 and 80 years.

Table 5: Clinical variables related to the worsening in the

first 24 hours of hospitalization.

Cluster

no.

Label (weight)

mRS_4_dis(0,90); mRS_3_7d(0,99); ...

LowApttValues(0,96); Hypertension(1,00); ...

Age65-80(0,85); Hypertension(0,97); ...

GenderMale(0,91); Hypertension(0,99); ...

Age>=80(0,92); GenderFemale(0,97); ...

Prediabetes(0,83); Hypertension(0,97); ...

The age range most at risk of incurring

permanent disabilities at 3 months after treatment is

consistent with the fact that in this cohort 78% of

patients with a 3-month mRS equal to 3, 4 or 5 are

over 65-year-olds.

Analysing the subset of patients with a 3-month

mRS of 3, 4 or 5 aged between 65 and 80, in 72% of

these cases a significant ischaemic penumbra is

detected (perfusion infarct mismatch) and in 44% of

cases also an hyperdense artery sign. This subset of

patients is clearly identifiable with cluster 1.

Considering instead the subset of patients with a

3-month mRS of 3, 4 or 5 aged between 65 and 80

characterized by low APTT levels, in 85% of cases

there are also high serum glucose levels. This subset

of cases can be associated with clusters 2 and 3.

Analysing the subset of patients with a 3-month

mRS equal to 3, 4 or 5 aged between 18 and 65

years it is found that in this case 67% of the patients

are male, 81% have low APTT levels and 80% had

high serum glucose levels. This subset of patients is

clearly identifiable with cluster 4.

The 87% of the subset of patients with a 3-month

mRS equal to 3, 4 or 5 over the age of 80 with

hyperlipidaemia are patients with low APTT levels.

This group can be associated with cluster 5.

5 CONCLUSIONS

The findings of this study suggest that the use of

clustering learning algorithms allows to identify in

an unsupervised way a set of clinical variables

which can be taken into consideration in order to

carry out a good prediction of the clinical outcome.

The Growing Neural Gas model has proved

particularly effective in predicting the patient

outcome compared to other algorithms used in the

same application area. The best result in terms of

predictive accuracy achieved by this model is due to

its ability to exactly identify the input space

topology, which also makes it particularly robust to

noise and lack of data. By analyzing the final

configuration of the trained GNG network, it is also

possible to obtain useful information on the

attributes which are most correlated with certain

outcomes. Statistical analyses carried out on the data

used as training set and test set seem to confirm the

consistency of the extracted knowledge. The

developed model is ready to be tested in prospective

studies in the real world.

REFERENCES

Alaiz-Moreton, H., Fernandez-Roblez L., Alfonso

Cendon, J., Castejon-Limas, M., Sanchez-Gonzalez L.,

Perez H., 2018. Data mining techniques for the

estimation of variables in health-related noisy data.

Advances in intelligent systems and computing, 649,

482-491.

Ali S., Smith K.A., 2006. On learning algorithm selection

for classification. Applied Soft Computing, 6, 119-

138.

Alpaydin E., 2020. Introduction to Machine Learning. 4

Predictive Clustering Learning Algorithms for Stroke Patients Discharge Planning

301

Edition. MIT Press.

Baeza Y., 1999. Modern Information Retrieval. Addison-

Wesley Longman Publishing Co., ISBN 0-201-39829-

Ben-Hur, A., Horn, D., Siegelmann, H.T., Vapnik, V.N.,

2001. Support vector clustering. Journal of Machine

Learning Research. 2, 125–137.

Bishop, C. M., 2006. Pattern Recognition and Machine

Learning, Springer.

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

45(1), 5–32. doi:10.1023/A:1010933404324.

Chen Y.C., Suzuki T., Suzuki M., Takao H., Murayama

Y., Ohwada H., 2017. Building a classifier of onset

stroke prediction using random tree algorithm.

International Journal of Machine Learning and

Computing, 7(4), 61-66.

Cortes, C., Vapnik, V.N., 1995. Support-vector networks.

Machine Learning. 20(3), 273–297.

CiteSeerX10.1.1.15.9362. doi:10.1007/BF00994018.

Donnan, G.A., Fisher M., Macleod M., Davis S.M., 2008.

Stroke. The Lancet. 371(9624), 1612–23.

Fritzke B., 1994. A Growing Neural Gas Network Learns

Topologies. Part of: Advances in Neural Information

Processing Systems 7, NIPS.

Girvan M., Newman M.E.J., 2002. Community structure

in social and biological networks, Proc. Natl. Acad.

Sci. USA 99, 7821–7826.

Holte,R.C., 1993. Very simple classification rules perform

well on most commonly used datasets. Machine

Learning.

Italian Ministry of Health website, 2020. http://www.

salute.gov.it/portale/salute/p1_5.jsp? lingua=italiano&

id=28&area=Malattie_cardiovascolari, last accessed

2020/04/24.

John, G.H.; Langley, P., 1995. Estimating Continuous

Distributions in Bayesian Classifiers. Proc. Eleventh

Conf. on Uncertainty in Artificial Intelligence. Morgan

Kaufmann. 338–345. arXiv:1302.4964

Kohonen T., 1988. An introduction to neural computing.

Neural Networks, 1, 3-16.

Kohonen T., 1989. Self-Organization and Associative

Memory, Berlin: Springer-Verlag.

Kohonen T., 1990. The Self Organizing Map. Proc of the

IEEE, 78(9).

Lella L., Licata I., 2017. Prediction of Length of Hospital

Stay using a Growing Neural Gas Model. In

Proceedings of the 8th International Multi-Conference

on Complexity, Informatics and Cybernetics (IMCIC

2017), 175-178.

Lyden P., Raman R., Liu L., Emr M., Warren M., Marler

J., 2009. National Institutes of Health Stroke scale

certification is reliable across multiple venues. Stroke,

40(7), 2507-2511. doi:10.1161/STROKEAHA.116.

015434.

Martinetz, T. M., Schulten, K J., 1991. A "neural-gas"

network learns topologies. In Kohonen, T., Makisara,

K, Simula, 0., and Kangas, J., editors, Artificial Neural

Networks, North-Holland, Amsterdam, 397-402.

Martinetz, T. M., Schulten, K J., 1994. Topology

representing networks. Neural Networks, 7(3), 507-

522.

Mess M., Klein J., Yperzeele L., Vanacker P., Cras P.,

2016. Predicting discharge destination after stroke: A

systematic review. Clin Neurol Neurosurg. 142(15-

21). doi:10.1016/j.clineuro.2016.01.004.

Pereira S., Foley N., Salter K., McClure J.A., Meyer M.,

Brown J., Speechley M., Teasell R., 2014. Discharge

destination of individuals with severe stroke

undergoing rehabilitation: a predictive model. Disabil

Rehabil. 36(9), 727-731. doi:10.3109/09638288.

2014.902510.

Platt,J., 1998. Sequential Minimal Optimization: A Fast

Algorithm for Training Support Vector Machines.

Technical Report MSR-TR-98-14.

Putra Pratama A., Tresno T., Wahyu Purwanza S., 2019.

Development the national institutes of health stroke

scale (NIHSS) for predicting disability and functional

outcome to support discharge planning after ischemic

stroke. Journal Ners, 14(3).

Saver J.L., Filip B., Hamilton S., Yanes A., Craig S., Cho

M., Conwit R., Starkman S., FAST-MAG

Investigators and Coordinators, 2010. Improving the

reliability of stroke disability grading in clinical trials

and clinical practice: the Rankin Focused Assessment

(RFA). Stroke. 41 (5): 992–

doi:10.1161/STROKEAHA.109.571364. PMC

2930146. PMID 20360551

Safe Implementation of Treatments in Stroke website,

2020. https://sitsinternational.org, last accessed

2020/04/24.

Tin Kam Ho, 1998. The Random Subspace Method for

Constructing Decision Forests. In IEEE Transactions

on Pattern Analysis and Machine Intelligence, 20(8),

832–844, DOI:10.1109/34.709601.

Tin Kam Ho, 1995. Random Decision Forests.

Proceedings of the 3rd International Conference on

Document Analysis and Recognition, Montreal, QC,

278–282.

Van Hulle M. M., 1989. Self Organizing Maps. Handbook

of Natural Computing, 585-622.

Wilson J. L., Hareendran A., Grant M., Baird T., Schulz

U.G., Muir K.W., Bone I., 2002. Improving the

Assessment of Outcomes in Stroke: Use of a

Structured Interview to Assign Grades on the Modified

Rankin Scale. Stroke. 33 (9): 2243–2246.

doi:10.1161/01.STR.0000027437.22450.BD. PMID

12215594

Witten,I. H., Frank,E., Hall,M.A., 2011. Data Mining

Practical Machine Learning Tools and Techniques.

Morgan Kaufmann Publishers.

Zdrodowska M., 2019. Attribute selection for stroke

prediction. Sciendo. Doi 10.2478/ama-2019-0026.

Zdrodowska M., Dardzinska M, Chorazy M., Kulakowska

A., 2018. Data Mining Techniques as a tool in

neurological disorders diagnosis. Acta Mechanica et

Automatica, 12(3), 217-220.

HEALTHINF 2021 - 14th International Conference on Health Informatics

302

APPENDIX

For the input data preprocessing the following coding

rules were used:

The age was codified in four main classes: age>=80;

65<=age<80; 18<=age<65 and 0<=age<18.

The clinical variables “Hypertension” and

“Hyperlipidemia” refer to the risk factors which are

specified in the data entry form. They are not

automatically computed by the use of a codifying

formulas.

If the glucose level is lower than 69 the value is codified

as “Hypoglycemia”, if the glucose level is between 70

and 99 the value is codified as “Normoglycemia”, if

the glucose level is between 100 and 124 the value is

codified as “Prediabetes”, if the glucose level is higher

than 125 the value is codified as “Diabetes”.

If the cholesterol level is lower than 199 the value is

codified as “NormalCholesterol”, if the cholesterol

level is higher than 200 the value is codified as

“HighCholesterol”.

If the temperature level is lower than 98.5 Fahrenheit the

value is codified as “NormalTemperature”, if the

temperature level is higher than 98.6 the value is

codified as “HighTemperature”.

If the APTT is lower than 29 the value is codified as

“LowApttValues”, if the APTT level is between 30

and 39 the value is codified as “NormalApttValues”, if

the APTT level is higher than 40 the value is codified

as “HighApttValues”.

If the BMI is lower than 18.4 the value is codified as

“Underweight”, if the BMI is between 18.5 and 24,9

the value is codified as “Normweight”, if the BMI is

between 25 and 29,9 the value is codified as

“IncreasedOverweight”, if the BMI is between 30 and

34,9 the value is codified as “ModerateOverweight”, if

the BMI is between 35 and 39,9 the value is codified

as “SevereOverweight”, if the BMI is higher than 40

the value is codified as “VerySevereOverweight”.

Predictive Clustering Learning Algorithms for Stroke Patients Discharge Planning

303