Understanding Stroke Risk Profiles in Middle-Aged Adults:
A Genetic Algorithm-Based Feature Selection Aproach
Ligia Ferreira de Carvalho Gonc¸alves
1 a
, Caio Davi Rabelo Fiorini
1 b
, Daniel Rocha Franca
2 c
,
Marta Dias Moreira Noronha
3 d
, Mark Alan Junho Song
3 e
and Luis Enrique Z
´
arate Galvez
3 f
1
Data Science and Artificial Intelligence, Pontif
´
ıcia Universidade Cat
´
olica de Minas Gerais, Rua Claudio Manuel,
Belo Horizonte, Brazil
2
Computer Science, Pontif
´
ıcia Universidade Cat
´
olica de Minas Gerais, Rua Claudio Manuel, Belo Horizonte, Brazil
3
Institute of Exact Sciences and Computer Science, Pontif
´
ıcia Universidade Cat
´
olica de Minas Gerais,
Rua Claudio Manuel, Belo Horizonte, Brazil
Keywords:
Stroke, Machine Learning, Genetic Algorithm, Decision Tree, Middle-Aged, Rules.
Abstract:
Data mining and machine learning techniques have been widely used in the knowledge extraction process of
medical databases, one highlight being their use to improve diagnostic systems. Decision trees are supervised
black box machine learning models that, although simple, are easy to interpret. In this work, we propose the
use of these techniques to describe the profile of middle-aged adults (40-59) diagnosed with stroke, a disease
that in Brazil was one of the main causes of death in previous years. The genetic algorithm was applied to
extract the best characteristics so that the Decision Tree algorithm could then be used in the database provided
by the 2019 National Health Survey to obtain the most comprehensive rules and identify the most relevant
attributes for describing the profile of these individuals. The conclusions indicate that the rules generated for
middle-aged adults are mainly about routine habits, such as work or salt consumption.
1 INTRODUCTION
Cerebrovascular Accident (CVA), also known as a
stroke, according to the Brazilian Stroke Society, can
be characterized by the appearance of a sudden neu-
rological deficit caused by a problem in the ves-
sels/arteries or veins of the brain. In Brazil, accord-
ing to the Civil Registry Transparency Portal, among
cardiovascular diseases, stroke was the main cause of
death in 2023. In 2024 alone, from January to March,
about 20 thousand deaths were recorded (BRASIL,
2024). Although the occurrence of stroke is higher
among individuals in older age groups (Rajati et al.,
2023), there has recently been a worrying increase
in the number of cases among middle-aged adults,
a group that until now, was believed to have a low
predisposition to this condition. The survey con-
a
https://orcid.org/0009-0000-7601-2938
b
https://orcid.org/0009-0005-3606-7623
c
https://orcid.org/0009-0008-9457-1221
d
https://orcid.org/0000-0002-2992-8422
e
https://orcid.org/0000-0001-7315-3874
f
https://orcid.org/0000-0001-7063-1658
ducted by the American Heart Association discusses
the growth of cases among adults under 49 years of
age, a fact that is relevant to understanding the situa-
tion of the disease among the middle-aged population.
Machine learning (ML) is an area that explores
the study and development of computational models
that learn through datasets, and which has received
great attention in the field of medicine (Paix
˜
ao et al.,
2022) being used to improve clinical diagnostic sys-
tems, and specifically in cardiology, it has been re-
sponsible for a large part of the scientific productions
about aiding in the diagnosis of cardiovascular dis-
eases.
Given the data and discussions presented, it is es-
sential to identify and understand the factors that char-
acterize the profile of middle-aged adults, aged 40 to
59, diagnosed with stroke. One way to understand
how the population behaves is by applying computer
learning models, which allow us to discover whether
the same conditions (extracted rules) characterize the
occurrence of stroke in the population.
This work is divided into 5 sections: Introduction,
Related Work, Materials and Methods, Results and
Discussions, and Conclusions. As a data source, the
Gonçalves, L. F. C., Fiorini, C. D. R., Franca, D. R., Noronha, M. D. M., Song, M. A. J. and Galvez, L. E. Z.
Understanding Stroke Risk Profiles in Middle-Aged Adults: A Genetic Algorithm-Based Feature Selection Aproach.
DOI: 10.5220/0013184600003911
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 623-630
ISBN: 978-989-758-731-3; ISSN: 2184-4305
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
623

recent study by the Brazilian Institute of Geography
and Statistics (IBGE), National Health Survey (PNS)
2019, a survey carried out through questionnaires in
2019 throughout the national territory
1234
.
2 RELATED WORKS
Given the impact that the chronic disease CVA has
on the age group described in the previous section, a
search was carried out in repositories based on terms
in English and Portuguese, of the keywords: stroke,
risk factors, and machine learning. Several studies
investigate the risk factors that indicate vulnerability
to stroke, while other authors apply machine learning
techniques to identify patterns and predict their occur-
rence.
In the work by Yousufuddin and Young (2019),
the authors discuss the relationship between the ag-
ing process and the occurrence of stroke. They con-
cluded that the presence of factors such as hyperten-
sion and diabetes increases the risk of stroke as the
individual ages, while the relative risk associated with
factors such as cigarette consumption and high blood
pressure decreases over time. In addition, in Noche
et al. (2020), the authors point out the knowledge gap
concerning understanding the risk factors for middle-
aged adults (aged 40-60). Their research discusses
precisely the importance of understanding the phe-
nomenon of stroke recurrence in this age group.
In Dritsas and Trigka (2022), the authors proposed
the use of ML models for stroke prediction using the
following algorithms: Naive Bayes, Random Forest,
Logistic Regression, K-Nearest Neighbors, Stochas-
tic Gradient Descent, Decision Tree, Multilayer Per-
cepton, Majority Voting, and Stacking. The results
obtained showed that the classification by Stacking
had a better performance when compared to the other
methods used, with an AUC of 98.9%, F-Measure,
Precision, and Recall equal to 97.4%, and accuracy of
98%.
1
https://avc.org.br/
2
https://transparencia.registrocivil.org.br/painel-
registral/especial-covid/
3
https://www.heart.org/en/
4
https://www.pns.icict.fiocruz.br/
3 MATERIALS AND METHODS
3.1 Description of the Database
In this work, we use data from the National Health
Survey (PNS) 2019, carried out by the Brazilian In-
stitute of Geography and Statistics (IBGE) in partner-
ship with the Ministry of Health. The survey collected
data on the general health and lifestyle situation of the
population from Brazil. It has 293,726 records and
1,088 attributes. For this study, we made a cut-off for
the chronic disease Cerebrovascular Accident (CVA)
based on the age group, containing 88,861 records re-
ferring to individuals not diagnosed with stroke and
only 1,975 for those with a positive diagnosis. Fi-
nally, we performed a data balance. That is, the final
database obtained has 1108 records, equally divided
into positive and negative diagnoses.
3.2 Understanding the Problem and
Conceptual Selection of Attributes
Understanding the problem domain is an important
step in the process of knowledge discovery, which
aims to build more representative learning models.
The previous understanding allows us to observe the
complexity of each problem and the discovery of use-
ful and non-obvious knowledge about the domain
considered. Also, due to the high dimensionality of
the PNS 2019 database, the conceptual selection of
attributes is a relevant strategy. In Figure 1, it is pos-
sible to see the conceptual model that was developed
for this study. The model was built using the CAPTO
method, recently proposed in Zarate et al. (2023).
The method is an approach that proposes the cap-
ture of knowledge, both explicit and tacit, for the un-
derstanding of a problem domain prior to the applica-
tion of machine learning algorithms to reduce dimen-
sionality, selecting the attributes that best represent
the problem domain. Table 1 shows the selected at-
tributes considering the conceptual model constructed
by the CAPTO method.
3.3 Pre-Processing and Data
Preparation
Due to the high rate of missing data in the selected
attributes, it was necessary to perform combinations
between attributes to minimize the impact caused by
them. For this, 6 new attributes were created they are:
BMI: Body Mass Index, based on the attributes
Weight (P00103) and Height (P00403). The catego-
rization was carried out based on the ranges defined
HEALTHINF 2025 - 18th International Conference on Health Informatics
624

Figure 1: Vertical Conceptual Model.
Table 1: Variables Extracted from the PNS 2019 database from the Concept Map.
Variables Extracted from the PNS 2019 database from the Concept Map
Dimension Aspect PNS Variables
Anthropology
Personal Characteristics C8, C6, C9
Anthropometry P1a, P4a
Consumption
Diet P9a, P15, P18, P20b, P25a, P26a, P26b
Alcohol P27, P28a, P29
Drugs No information available in the database
Smoking P50, P51, P52, P53, P54, P54a, P54b, P54c,
P54d, P54e, P54f, P54g
Health
Access to health services and Health in-
surance
100102
Mental Illnesses Q92, Q110a
Heart Disease Q2a, Q63a
Chronic Diseases Q30a, Q60, Q63a, Q68
Sedentary lifestyle P45a, P45b
Physical Activity
Work P38, P39, P40, P41, P42, P43, P44, E17, E19
Sport P34, P35, P37, P36
by the World Health Organization (WHO) where BMI
= Weight/Height².
• If BMI <18.5: “Underweight”
• If 18.5 <= BMI <24.99: “Normal weight”
• If 35 <= BMI <39.99: “Obesity grade II”
• If 25 <= BMI <29.99: “Overweight”
• If BMI >= 40: “Obesity grade III”
Daily Smoking: To create the attribute contain-
ing information on the daily consumption of products
containing tobacco, the following PNS attributes were
used: P05402 (Industrialized Cigarettes), P05405
(Straw or Hand-Rolled Cigarettes), P05408 (Clove
or Bali Cigarettes), P05411 (Pipes), P05414 (Cigars
or Cigarillos), P05417 (Hookah (sessions)), P05421
(Other). The procedure applied consisted of the
sum of response values between the attributes
(Daily Smoking= P05402 + P05405 + P05408 +
P05411 + P05414 + P05417 + P05421). When an
empty record (NaN) is found, it is considered zero
(Non-smoker), so the individual whose result of the
sum of all the columns associated with it was equal to
zero was considered a non-smoker. The standardiza-
tion proposed by the Government of Canada
5
, entitled
”Tobacco Use Statistics – Terminology”, was used as
a basis for categorization. The procedure adopted is
given below:
• If cigarettes == 0: “Non-smoker”
• If 11 <= cigarettes <= 20: “Moderate Smoker”
• If 0 <cigarettes <= 10: “Light Smoker”
• If cigarettes >20: “Heavy Smoker”
Weekly Alcohol Intake: The strategy to create
this attribute combines two answers: the weekly fre-
quency of consumption (P02801) and the number
5
https://www.canada.ca/en.html/
Understanding Stroke Risk Profiles in Middle-Aged Adults: A Genetic Algorithm-Based Feature Selection Aproach
625

of doses consumed per occasion (P029). When the
weekly frequency is not available, but there is infor-
mation that the individual in question consumes alco-
hol less than once a month, the developed function ad-
justs the amount consumed, dividing it by four, to es-
timate a weekly average of consumption. It was used
to categorize the standards defined by the National In-
stitute on Alcohol Abuse and Alcoholism (NIAAA)
6
.
For the Female Gender
• If Doses == 0: “Don’t drink”
• If 1 <= Doses <= 7 servings: “Low consump-
tion”
• If 8 <= Doses <= 14 servings: “Moderate Con-
sumption”
• If Doses >14: “High Consumption”
For the Male Gender
• If Doses == 0: “Don’t drink”
• If 1 <= Doses <= 14 servings: “Low Consump-
tion”
• If 15 <= Doses <= 28 servings: “Moderate Con-
sumption”
• If Doses >28: “High Consumption”
Weekly Moderate Physical Activity and
Weekly Intense Physical Activity: Using the
WHO definition of moderate (M.Activity Weekly
= ([P04001*(P04101*60)] + P04102) +
([P042*(P04301*60)] + P04302)) and in-
tense physical activity (I.Activity Weekly
= ([P035*(P03701*60)] + P03702) +
([P03904*(P03905*60)] + P03906)), attributes
regarding the weekly minutes dedicated to these
activities were created. The following attributes
were used: P035 (Days per week spent exercising,
practicing physical exercise or sport), P037 (Duration
in hours and minutes – P035), P039 (Days, hours
and minutes per week spent by the interviewee
doing heavy activities), P04001 (Days per week
spent walking or cycling), P041 (Duration in hours
and minutes – P04001), P042 (Days per week that
involve commuting to carry out usual activities) and
P043 (Duration in hours and minutes – P042). For
categorization, the activity guide published by the
Federal Government in 2021
7
is used as a reference.
For Moderate Physical Activity
• If 0 <= Minutes <= 149: “Level 1”
• If 150 <= Minutes <= 299: “Level 2”
6
https://www.niaaa.nih.gov/
7
https://www.gov.br/saude/pt-
br/composicao/saps/ecv/publicacoes/guia-de- atividade-
fisica-paraBrazilian population/view/
• If 300 <= Minutes <= 449: “Level 3”
• If 450 <= Minutes <= 599: “Level 4”
• If Minutes >= 600: “Level 5”
For Intense Physical Activity
• If 0 <= Minutes <= 74: “Level 1”
• If 75 <= Minutes <= 149: “Level 2”
• If 150 <= Minutes <= 224: “Level 3”
• If 225 <= Minutes <= 299: “Level 4”
• If Minutes >= 300: “Level 5”
Working journey: It refers to the number of
hours worked weekly. It resulted from the sum of
the values between the attributes E017 (Total hours
worked per week in the main job) and E019 (Total
hours worked per week in other jobs). If both in-
stances were empty (NaN), this value was considered
zero, so the final attribute was not affected and there
was no change in the data used. In other words, if all
the attributes used were null, the final attribute would
have a balance equal to zero, meaning that this person
is not employed. For its categorization, World Health
Organization (WHO) recommendations were used as
basis.
• If Hours == 0: “Not employed”
• If 41 <= Hours <= 54: “Excessive Working
Hours”
• If Hours <= 40: “Normal Working Hours”
• If Hours >= 55: “High-Risk Excessive Working
Hours”
Diet Classification (Score): Weights were as-
signed to each food item of the respondent and were
based on its importance for constructing a diet that
prevents the occurrence of cardiovascular diseases.
Thus, regular consumption of fruits (P018) and veg-
etables (P00901) was associated with reduced risk of
stroke due to their high content of antioxidants, fiber,
and micronutrients in Dauchet et al. (2006). Regu-
lar consumption of fish (P015) is also responsible for
cardiovascular health due to the anti-inflammatory ef-
fects as it is a food rich in fatty acids according to
Mozaffarian and Rimm (2006). On the other hand,
excessive consumption of sweets (P02501), fast food
(P02602), and soft drinks (P02002) is related to an in-
creased risk of obesity, hypertension, and cardiovas-
cular diseases Malik et al. (2006). To conclude the
treatment of these attributes, the score attribute was
created, which corresponds to the sum of the products
between the frequency of consumption of each food
and its respective weight. To categorize this attribute,
cut-off values were defined in the score attribute to
separate the categories related to the type of food.
HEALTHINF 2025 - 18th International Conference on Health Informatics
626

• If Score <= 0: “Unhealthy Diet”
• If Score >5: “Healthy Eating”
• If 0 <Score <= 5: “Unhealthy Diet”
In all the attributes created, the OrdinalEncoder
encoding method was applied, to encode categorical
attributes in which their elements have a hierarchy
among themselves. The rest of the attributes were al-
ready coded by the way the answers were arranged
in the PNS 2019 questionnaire, maintaining the orig-
inal coding. The only change made was in the coding
of the attribute ”Salt Consumption’”. An inversion in
the order was made, where the value ”5”, which origi-
nally corresponded to the lowest level, now represents
the highest consumption, while the value ”1”, which
previously corresponded to the highest level, now rep-
resents the lowest consumption.
After the steps described above, the resulting
database was divided into a new database referring
to the age group of 40 to 59 years considered in this
study. A correlation and entropy analysis are applied
to analyze respectively the existence of attributes with
little variance and information. In other words, iden-
tify attributes that present little information and at-
tributes that are highly related to the target attribute, to
avoid direct classification of a dominant attribute. Af-
ter this procedure, it was not possible to eliminate at-
tributes based on these analyses, because the entropy
values were very close, making it difficult to establish
a cut-off value and the correlation analysis. It means
that no attribute had a high correlation with the class
attribute.
Finally, it was checked if there were inconsisten-
cies after coding, in order to identify instances that
have the values of all the identical attributes, but be-
longing to different classes. Records that met this
condition were eliminated from both databases. Ta-
ble 2 describes the attributes present in the databases
after pre-processing and data preparation.
3.4 Multi-Objective Genetic Algorithm
Through all the steps applied as conceptual selec-
tion, pre-processing and data preparation, the origi-
nal dataset that contained 1088 attributes gave rise to
a dataset with 17 attributes. As a final step to ob-
tain the learning model, the new dataset was submit-
ted to an attribute selection process based on a multi-
objective genetic algorithm (GA), NSGA-II (Non-
dominated Sorting Genetic Algorithm II). The objec-
tives were to select the least number of attributes,
with the lowest error rate in the decision tree-based
classification model. The algorithm implementation
was carried out in Python, with the DEAP library
(https://github.com/DEAP ).
3.4.1 Coding of Individuals
Candidate individuals (containing selected attributes)
were treated in a binary manner, assembling the chro-
mosomes with values 0 and 1. The value One repre-
sents the presence of the attribute and the value Zero
represents the absence of it. The chromosome is rep-
resented by 17 positions.
3.4.2 Objective Function
For this study, two aspects of the prediction problem
for the data set were considered by the objective func-
tion Lowest Percent Error Rate of Decision Tree Clas-
sifier Prediction: the accuracy of the model, and the
smallest possible number of characteristics.
Lowest Percent Error Rate = 1 – Accuracy = 1 –
TP) / TP) + FP), in which TP is True Positive and FP
is False Positive.
As a tiebreaker between two individuals, the one
with the highest fitness was chosen, which conse-
quently means the one with the lowest error rate of
Precision, followed by the one with the smallest num-
ber of characteristics. To increase the robustness and
reliability of the result, a cross-validation was per-
formed for each evaluated individual, with the num-
ber of folds equal to 10.
3.4.3 Defining the Size of the Population
Preliminary tests were carried out to define which pa-
rameter value ranges would be the most appropriate to
carry out the experiments and to establish a stopping
criterion based on the number of generations. This
was done to have a comprehensive understanding of
the experiments in relation to the computational effort
required.
During the preliminary experiments, significant
changes in the individuals of the population between
generations were analyzed, seeking to observe the
number of generations from which the convergence
of the results is observed. If there were no signif-
icant changes among individuals in the population,
this could indicate that the results were in the direc-
tion of convergence. The values considered adequate
for the population size were 200, 400, and 600. For
the number of generations, the values of 200 and 400
were chosen.
Other parameters of the genetic algorithm, such
as the crossover probability (Pc), ranged between 0.7
and 0.9. The mutation probability (Pm) was set at
0.1, since the goal of the preliminary tests was to find
a range of values for the population size and the num-
ber of generations. It is worth noting that these val-
ues used as experimental parameters were based on
Understanding Stroke Risk Profiles in Middle-Aged Adults: A Genetic Algorithm-Based Feature Selection Aproach
627

Table 2: Attributes in the Database After the Pre-Processing steps.
Attribute Attribute description
Gender Refers to the gender of the interviewee.
Color/Race Refers to the interviewee’s color/race.
Salt Consumption Interviewee’s perception of their salt consumption.
Heavy Domestic Work If the interviewee does heavy cleaning in their domestic activities.
Hypertension Whether the respondent has been diagnosed with hypertension.
Diabetes Whether the respondent has been diagnosed with diabetes.
High Cholesterol Whether the respondent has been diagnosed with high cholesterol.
Heart Disease Whether the respondent has been diagnosed with any heart disease.
Depression Whether the respondent has been diagnosed with depression.
Mental Illness Whether the interviewee has been diagnosed with a mental disorder.
BMI category Respondent’s Body Mass Index category.
Working Hours Level of hours in the interviewee’s working day.
Smoker category Interviewee’s level of tobacco consumption.
Weekly alcohol category Interviewee’s weekly alcohol consumption.
Diet Classification Type of diet of the interviewee.
Category moderate cent Moderate physical activity level practiced by the interviewee.
Category vigorous cent Interviewee’s level of intense physical activity.
STROKE Target column: Whether the respondent has been diagnosed with stroke.
the articles by Bento and Kagan (2008), Santos et al.
(2018), Oh et al. (2004), and Fern
´
andez et al. (2019)
After carrying out these experiments, the popula-
tion size was defined as ranging between 200 and 400.
The Table 3 shows the parameters and their adjusted
values for the global experiments.
4 RESULTS AND DISCUSSION
A total of 240 experiments were conducted, repre-
senting the different possible combinations between
the 20 seeds and the elements listed in Table 3. For
each of the seeds, 12 experiments were carried out,
ensuring that all the parameterized possibilities for
that seed were tested. Different seeds were used to
expand the search space where the algorithm began to
track its analysis through the genetic space of individ-
uals. For each experiment, a set of 10 non-dominant
candidates (hall of fame) was found. The most fre-
quent attribute is gender, and the least frequent but
not less important, is ethnicity. It is worth noting that
frequency is merely a heuristic for the creation of a
synthetic individual and does not influence whether
that attribute is considered important for the analyzed
cause or not.
To evaluate the contribution of each attribute more
frequently, chromosomes were created and used to
build learning models based on Decision Trees. For
the training process, cross-validation with k = 10 was
used. The best results were achieved with 14 to 17
attributes, from the attribute Gender to Hypertension.
After testing the synthetic chromosomes to determine
the best results, we concluded that the chromosome
containing 14 attributes achieved the highest F1 score.
These attributes are: Gender, Weekly Alcohol Cat-
egory, Diabetes, Physical ActivityI Category, Diet
Classification, Smoking Category, Mental Illnesses,
Heavy Housework, Physical ActivityM Category, Salt
Intake, High Cholesterol, Depression, Heart Disease,
and Hypertension.
After discovering the best attributes, a decision
tree was created based on these attributes. Conse-
quently, the following values were obtained for the
evaluation metrics: average precision, average ac-
curacy, average sensitivity, and average F1 score:
0.6955, 0.6931, 0.6955, and 0.6919, respectively.
The model generated earlier with all the attributes
achieved an average accuracy of 0.65. For compar-
ative purposes, only this metric will be used. There-
fore, we can conclude that the genetic algorithm pro-
duced better results with fewer attributes, demonstrat-
ing its effectiveness.
In Table 4 , it is possible to observe that the ab-
sence of heart disease as well as the presence of con-
HEALTHINF 2025 - 18th International Conference on Health Informatics
628

Table 3: Parameters of the Genetic Algorithm for the Experiments.
Parameters Values
Population Initialization Random
Representation Binary
Crossover Operator Two Points
Crossover Probability (PC) 70%, 90%
Mutation Operator One Point
Mutation Probability (PM) 1%
Population Size [200, 400, 600]
Number of Generations [200, 400]
Crossover Selection Method Tournament = 2
Composition of the New Generation Not dominated individuals
Stop Criterion Number of generations
Table 4: Main Rules and their Respective Coverage.
Extracted Rules Coverage
If the individual does not have hypertension, but has heart disease, high choles-
terol, depression, and does not smoke, then they do not have a stroke.
[102 cases without stroke, 86 cases
with stroke]. Total: 188 cases.
If the individual has hypertension, heart disease, depression, does not smoke, and
their level of intense physical activity is low, then they have a stroke.
[83 cases without stroke, 146 cases
with stroke]. Total: 229 cases.
If the individual has hypertension, heart disease, depression, does not smoke, and
their level of intense physical activity is high, then they have a stroke.
[46 cases without stroke, 148 cases
with stroke]. Total: 194 cases.
If the individual does not have hypertension, heart disease, high cholesterol, de-
pression, and does not smoke, then they do not have a stroke.
[74 cases without stroke, 14 cases
with stroke]. Total: 88 cases.
If the individual does not have hypertension, heart disease, high cholesterol, de-
pression, does not smoke, and their level of intense physical activity is low, then
they do not have a stroke.
[9 cases without stroke, 5 cases with
stroke]. Total: 14 cases.
ditions such as depression, hypertension, smoking
habits, and the practice of intense physical activity
are related to the rules used to describe the profile of
individuals diagnosed with Stroke. As for the rules
associated with the profile of individuals with nega-
tive diagnoses, a variation is noticeable in the pres-
ence and absence of certain attributes, for example,
tobacco consumption and high cholesterol. In addi-
tion, hypertension is indeed a condition that is not as-
sociated with these individuals; however, the presence
of the attributes heart and mental diseases is guaran-
teed.
Thus, for positive diagnoses, the attributes present
in the rules are more consistent in the sense that there
is no variation in their state. In other words, there are
no discrepancies regarding the presence or absence
of these attributes in the classification of instances,
which enhances the predictability and the description
of their profile. On the other hand, the rules asso-
ciated with negative diagnoses exhibit greater varia-
tion in the state of the attributes, meaning that the at-
tributes may be present in some instances and absent
in others. This variability makes it more challenging
to accurately extract patterns and characteristics for
this group of individuals.
5 CONCLUSIONS
The use of Genetic Algorithms (GA) in this study
proved to be advantageous because it allowed the ex-
ploration of a wide variety of combinations of at-
tributes, generating consistent rules for stroke predic-
tion. One of the main advantages of GA is its abil-
ity to find near-optimal solutions in complex search
spaces, which has proven effective in identifying rel-
evant patterns. The rules extracted, such as those
related to hypertension, heart disease, and lifestyle
habits, provided a clearer picture of the factors that
contribute to stroke risk, showing that GA can be a
useful tool to support medical diagnoses.
However, parameterizing the algorithm was a sig-
nificant challenge. The choice of parameters, such
as the number of generations and the mutation rate,
Understanding Stroke Risk Profiles in Middle-Aged Adults: A Genetic Algorithm-Based Feature Selection Aproach
629

directly impacts the effectiveness of GA. Because
NSGA-II scales rapidly depending on the parameters,
it was not possible to exploit the entire search space
efficiently. This limited the number of subjects as-
sessed, which could be improved with methods that
allow better exploration of the start space. This chal-
lenge reflects the need for further investigation into
the best parameter setting to maximize coverage of
individuals of interest.
Another important point was the transformation of
categorical data to numerical data, a necessary action
for the use of the decision tree. Some inaccuracies
brought by the transformation may impact the qual-
ity of the extracted rules. Still, in this work, it is
minimized by using standard measures from reliable
sources to preserve data quality. Thus, the interpre-
tation of the results is not so affected by the catego-
rization performed on the numerical data. Even so,
the use of a classification algorithm capable of deal-
ing directly with categorical data, in Python, without
the need for transformation, would be a solution to re-
duce these inaccuracies and improve the accuracy of
the model. Finally, the rules extracted were consistent
with the theory about the disease, as demonstrated by
the rules that relate hypertension, heart disease, and
the practice of physical activities to the occurrence
of stroke. However, the limitations imposed by the
dataset, which did not include a wider range of 17 at-
tributes, restrict the predictive potential of the model.
For future work, the first focus would be the use
of classification algorithms or techniques that, in con-
junction with genetic algorithms, can better explore
the search space in an efficient and computationally
feasible manner. Ideally, this would include the pos-
sibility of covering the entire search space or most of
it, ensuring the discovery of more varied solutions.
Another crucial point would be the use of algo-
rithms that allow working directly with categorical
data, without the need to transform them into numeri-
cal data. This approach avoids the inaccuracies intro-
duced by the transformation and contributes to more
reliable results. Furthermore, a data transformation
methodology that preserves as much information as
possible from the original dataset, especially records
related to the occurrence of stroke, could be devel-
oped. Such a methodology would help to preserve the
richness of the data, providing more robust and reli-
able analyses.
ACKNOWLEDGEMENTS
The authors thank The National Council for Scientific
and Technological Development of Brazil (CNPQ);
The Coordination for the Improvement of Higher Ed-
ucation Personnel - Brazil (CAPES) (Grant PROAP
88887.842889/2023-00 – PUC/MG, Grant PDPG
88887.708960/2022-00 – PUC/MG - INFORMAT-
ICA and Finance Code 001); Minas Gerais State Re-
search Support Foundation (FAPEMIG) under grant
number APQ-01929-22, and the Pontifical Catholic
University of Minas Gerais, Brazil.
REFERENCES
Bento, E. P. and Kagan, N. (2008). Algoritmos gen
´
eticos e
variantes na soluc¸
˜
ao de problemas de configurac¸
˜
ao de
redes de distribuic¸
˜
ao. Revista Controle & Automac¸
˜
ao,
19(3):302–305.
Dauchet, L., Amouyel, P., Hercberg, S., and Dallongeville,
J. (2006). Fruit and vegetable consumption and risk
of coronary heart disease: A meta-analysis of cohort
studies1. The Journal of Nutrition, 136(10):2588–
2593.
Dritsas, E. and Trigka, M. (2022). Stroke risk prediction
with machine learning techniques. Sensors, 22(13).
Fern
´
andez, C., Pantano, N., Godoy, S., Serrano, E., and
Scaglia, G. (2019). Parameters optimization apply-
ing monte carlo methods and evolutionary algorithms.
Revista Iberoamericana de Autom
´
atica e Inform
´
atica
Industrial, 16(2):89–99.
Malik, V. S., Schulze, M. B., and Hu, F. B. (2006). In-
take of sugar-sweetened beverages and weight gain: a
systematic review. The American journal of clinical
nutrition, 84 2:274–88.
Mozaffarian, D. and Rimm, E. (2006). Mozaffarian d, rimm
eb. fish intake, contaminants, and human health: eval-
uating the risks and the benefits. jama 296, 1885-1899.
JAMA: the journal of the American Medical Associa-
tion, 296:1885–99.
Noche, R., Biffi, A., Sansing, L., Shoamanesh, A., Be-
navente, O., Falcone, G., and Sheth, K. (2020). Ab-
stract 156: Recurrent stroke in middle-aged lacunar
stroke survivors: Understanding risk factors and vul-
nerability in an important target population. Stroke,
51.
Oh, I.-S., Lee, J.-S., and Moon, B.-R. (2004). Hybrid ge-
netic algorithms for feature selection. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
26(11):1424–1429.
Santos, B. C., Nobre, C. N., and Z
´
arate, L. E. (2018). Multi-
objective genetic algorithm for feature selection in a
protein function prediction context. In 2018 IEEE
Congress on Evolutionary Computation (CEC), pages
1–8.
Yousufuddin, M. and Young, N. (2019). Aging and is-
chemic stroke. Aging, 11.
Zarate, L., Petrocchi, B., Maia, C. D., Felix, C., and Gomes,
M. P. (2023). Capto - a method for understanding
problem domains for data science projects. Concil-
ium.
HEALTHINF 2025 - 18th International Conference on Health Informatics
630