Analysis of School Performance of Children and Adolescents

with Attention-Deﬁcit/Hyperactivity Disorder:

A Dimensionality Reduction Approach

Caroline Jandre

, Bruno Santos

, Marcelo Balbino

, D

ebora de Miranda

, Luis Z

arate

and Cristiane Nobre

Department of Computer Science, Pontiﬁcal Catholic University of Minas Gerais, Minas Gerais, Brazil

Department of Pediatrics, Federal University of Minas Gerais, Minas Gerais, Brazil

Keywords:

Dimensionality Reduction, Features Selection, Machine Learning, ADHD, School Performance.

Abstract:

Attention-Deﬁcit/Hyperactivity Disorder (ADHD) is deﬁned by harmful inattention, disorganization, and/or

hyperactivity and impulsivity. ADHD can negatively affect an individual’s life, but it is not a deﬁnitive factor

for poor school performance. This work aims to identify classiﬁcation rules that best describe the school

performance in arithmetic, writing, and reading of students with ADHD. For this, information obtained from

the Genetic Algorithm, Random Forest and specialists in ADHD were used so that later the VTJ48 and JRip

algorithms could be applied. It is usual in the health area to collect various information about the individual,

resulting in the frequent need to reduce the base’s dimensionality. The results found were promising, reaching

up to 92% of F-Measure. The discovered rules point to environmental and emotional factors as drivers of

school performance prognosis and reinforce that ADHD is not synonymous with academic failure.

1 INTRODUCTION

The large-scale collection and storage of information

have hampered the analysis and visualization of data

and the discovery of useful knowledge for decision

making (Borges and Nievola, 2012). The high di-

mensionality of data, represented here by the num-

ber of features, can impair the predictive capacity of

Machine Learning (ML) algorithms and increase the

cost of computational processing at the time of the

analysis (Santos et al., 2007). In the health area, it is

common to collect various information about the in-

dividual, such as socioeconomic status, demographic

characteristics, lifestyle, and health conditions, so that

they can be used to assess the prognosis of the pathol-

ogy (dos Santos et al., 2019). Therefore, the reduction

of dimensionality in a health-oriented database is of-

ten necessary.

Present in the Diagnostic and Statistical

Manual of Mental Disorders (DSM), Attention-

Deﬁcit/Hyperactivity Disorder (ADHD) is deﬁned

by harmful levels of inattention, disorganization,

and/or hyperactivity and impulsivity, symptoms that

are excessive when compared with other people of

the same age and degree of development. Population

surveys suggest that ADHD occurs in most cultures

in about 5% of children and 2.5% of adults, being

more frequent in males (APA et al., 2014).

ADHD can have negative repercussions on the in-

dividual’s social, educational, and family life (Mat-

tos, 2015). Reduced academic performance and suc-

cess, social rejection, and relationship difﬁculties are

usually due to the disorder, which leads to consider-

able educational and social losses (Rangel J

unior and

Loos, 2011).

Since ADHD is not a deﬁnitive factor for poor

school performance (Frazier et al., 2007), it is es-

sential to identify which other characteristics can en-

hance or minimize losses in the academic environ-

ment. Thus, this work aims to ﬁnd classiﬁcation rules

that best describe school performance in arithmetic,

writing, and reading (subjects that are part of the ﬁrst

phase of primary education, foreseen in the Standard-

ized International Classiﬁcation of Education (UN-

ESCO, 2012) of students with ADHD. A database

with 266 children and adolescents and 225 features

was used. Two classiﬁcation algorithms were also

used: JRip, which generates rules directly from the

data set, and VTJ48, a decision tree algorithm.

In order to improve the representativeness of the

Jandre, C., Santos, B., Balbino, M., de Miranda, D., Zárate, L. and Nobre, C.

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction Approach.

DOI: 10.5220/0010240401550165

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 155-165

ISBN: 978-989-758-490-9

155

database, the dimensionality reduction process was

applied. As the base has a high degree of dimen-

sionality and the classiﬁcation algorithms are sensi-

tive to the number of features, the reduction was car-

ried out to seek non-compromise of the performance

and the ﬁnal result of the classiﬁcation. Therefore,

three approaches were used to features selection: 1)

Genetic Algorithm (GA), which work to select the

smallest subset of features, representative and rele-

vant, seeking a higher quality model (Pappa et al.,

2002); 2) Random Forest (RF), an ensemble-type al-

gorithm, combining decision trees, which lists the

features in order of importance; and 3) attribute se-

lection performed by an ADHD specialist and infor-

mation from the literature (Ara

ujo, 2002).

It is hoped that the adopted methodology can

be used to create strategies aimed at students with

ADHD, helping them to reduce their daily difﬁculties.

The article follows the following structure: in Sec-

tion 2, the theoretical foundation is presented, which

brings the main concepts related to work. The works

related to the theme are covered in Section 3. Section

4 presents the methodology used, with a detailed de-

scription of the database and the pre-processing steps.

Section 5 presents the discussions regarding the re-

sults found. Finally, in Section 6, the ﬁnal considera-

tions in this article are exposed.

2 BACKGROUND

2.1 Attention-Deﬁcit/Hyperactivity

Disorder

The names and forms of treatment referring to ADHD

have changed over the years. Research shows that

the disorder is related to a deﬁciency of neurotrans-

mitters, known as dopamine and noradrenaline, and

a change in the right region of the frontal cerebral

lobe, affecting the control of inhibitory behavior and

executive functions (ABDA, Associac¸

ao Brasileira

do D

eﬁcit de Atenc¸

ao, 2016). However, it is dis-

cussed about its cause being multifactorial, covering

from biological to environmental aspects (Andrade

and Lohr J

unior, 2017).

The individual with ADHD may have an inatten-

tive, hyperactive/impulsive, or combined proﬁle. The

combined proﬁle consists of the union of characteris-

tics of the other two proﬁles and is present in about

62% of people with ADHD (Cardoso et al., 2019).

The disorder is usually noticed at school age because

concentration becomes more necessary for tasks’ per-

formance. However, its symptomatological signs can

extend to adulthood (Santos and Vasconcelos, 2010).

The diagnosis of the disorder is based on clinical-

behavioral criteria, which hinders the accuracy of the

result. It depends on the patient’s clinical history, the

use of consolidated questionnaires in the classiﬁca-

tion of disorders, and the information provided by

parents and the school, often omitted because they

are not considered essential or for other personal rea-

sons (Andrade and Lohr J

unior, 2017). However, de-

spite the difﬁculties of diagnosis, there is an increase

in cases of ADHD. Social awareness on the subject

brought greater collective knowledge, attracting par-

ticular attention from parents, educators, and health

professionals to the topic (de Azevedo Santos, 2017;

Jou et al., 2010).

The school environment is considered to be of

great help in the cognitive and socio-emotional de-

velopment of human beings, can assist in reducing

to a decrease in these losses resulting from ADHD.

However, due to these students’ peculiar functioning,

educational institutions have found it challenging to

deal with them (Rangel J

unior and Loos, 2011). In

addition to the very characteristics of ADHD, that

can leave them more scattered, students face emo-

tional and psychological factors that directly affect

their academic performance, leading to high numbers

of failures, expulsions, and dropping out of school

(de Lima and Coelho, 2018; DuPaul et al., 2017; Mat-

tos, 2015). Thus, the student with ADHD must be

helped to overcome their difﬁculties in the academic

environment, creating strategies that contribute to the

inclusion process and making these students not be at

a disadvantage about other people who do not have

the disorder (Cortez and Pinheiro, 2018; de Lima and

Coelho, 2018).

2.2 Features Selection with Genetic

Algorithm

When it comes to reducing dimensionality in terms

of the number of features, two methods stand out:

1) compression of features, which encodes or trans-

forms the data to obtain a compact representation

of the originals, as is the case of Principal Com-

ponent Analysis; and 2) features selection, that de-

tects and discards irrelevant, little relevant or redun-

dant features. There are at least three strategies for

features selection: exponential (represented by ex-

haustive search), sequential (represented by direct se-

quential selection), and random (represented by GA)

(Pappa et al., 2002).

GA are mathematical algorithms inspired by the

mechanisms of evolution of populations of living be-

ings. The technique introduced by (Holland, 1975)

HEALTHINF 2021 - 14th International Conference on Health Informatics

156

and popularized by (Goldberg, 1989), provides an

adaptive search engine, which follows the principle of

natural selection and survival of the ﬁttest. This con-

ception is based on the Darwinian maxim that ”The

better an individual adapts to his environment, the

greater his chance of surviving and generating de-

scendants.” (de Lacerda and de Carvalho, 1999). The

Algorithm 1 presents a simpliﬁed representation of

the GA evolution process:

Algorithm 1: GA Pseudocode.

t ← 0;

InitializePopulation(P(t));

f itness(P(t));

while not Stopping Criteria do

parents ← Selection(P(t));

children ← CrossoverMutation(parents);

P(t) ← P(t) + children;

f itness(P(t));

P(t + 1) ← NewPopulation(P(t));

t ← t + 1;

end

In GA, the individual is portrayed by the chromo-

some, a data structure representative of the possible

solutions to the problem. The process begins with the

random generation of a set of chromosomes, form-

ing the so-called population. This population is as-

sessed through ﬁtness, which determines how well

the individual has adapted to the problem in ques-

tion. As long as the stopping criterion is not met,

chromosomes are subjected to an evolutionary pro-

cess involving: selection, which consists of ﬁnding

the most suitable individuals and letting them pass on

their genes to the next generation; and genetic oper-

ators (crossover and mutation), which are applied to

each pair of selected parents and will generate new

individuals, called children. After this process, the

current population receives the new children gener-

ated. The ﬁtness metric again evaluates this popula-

tion. Finally, the next individuals who will be part of

the new population are selected. After several cycles

of evolution, the population should contain the ﬁttest

individuals (Pacheco et al., 1999).

3 RELATED WORKS

ML methods have been widely used in research

related to ADHD and have generated important

advances in searching for knowledge associated with

this disorder.

Anuradha et al. (2010) conducted a study for di-

agnosing ADHD in 100 children aged 6 to 11 years

using the Support Vector Machines (SVM) algorithm.

The database was collected through the students’ re-

sponses to a questionnaire and medical diagnoses in-

dicating whether the children had ADHD. The work

used GA to features selection in order to identify the

essential features and increase the accuracy of the

model. The study achieved an 88.7% success in diag-

nosing ADHD, which was considered a satisfactory

result from the authors.

Rahadian et al. (2017) used GA to improve

a Learning Vector Quantization 2 Neural Network

(LVQ2NN) method to classify data about the type of

ADHD in patients. The GA was used to optimize the

weight vector in the training process. The tests per-

formed without the GA reached 80% of correctness,

while the model with the GA improved the perfor-

mance to 89.5%.

Another research line that associates ML algo-

rithms with ADHD is related to solutions that seek

to diagnose the disorder through Magnetic Reso-

nance Images (MRI). Sachnev (2015) presents a

classiﬁcation approach combining Meta - Cognitive

Neuro-Fuzzy Interface System (McFIS) with an fea-

tures selection mechanism based on Binary Coded

Genetic Algorithm and Extreme Learning Machine

(BCGA-ELM). The experiments show the possibil-

ity of achieving good results in the classiﬁcation of

ADHD based on the hippocampus images. Already

Aradhya et al. (2020), describe an approach based on

Metaheuristic Spatial Transformation (MST) and hy-

brid GA. The purpose of the work was to use the MST

to perform the extraction of MRI features from the

ADHD-200 base that served as input for classiﬁcation

using the Projection Based Learning - Meta-cognitive

Radial Basis Function Network (PBL-McRBFN).

The literature also presents research that relates

to data mining and the school environment. This

combination has become more and more frequent,

giving rise to the ﬁeld of Educational Data Mining

(EDM) (Angeli et al., 2017). EDM includes works

that use ML algorithms to acquire knowledge on ed-

ucational topics, such as prediction of student perfor-

mance (Ahmed and Elaraby, 2014), learning through

remote teaching (He, 2013), among other issues re-

lated to the school environment.

However, in the surveys carried out, no studies re-

lated to the use of ML, ADHD, and school perfor-

mance, object of study of this work. The research

results associating ML and ADHD presented studies

that focus on the diagnosis of the disorder, rather than

its impact on the school environment.

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction

Approach

157

4 MATERIALS AND METHODS

4.1 Description of the Database

The database was made available by the Depart-

ment of Pediatrics of the Federal University of Minas

Gerais - Brazil. The sample consists of 266 students

(instances), aged between 6 and 18 years. Among

them, 196 are diagnosed with ADHD and 70 have a

negative diagnosis.

The database, based on 225 characteristics (fea-

tures), it was built from the individual and family

responses present in the questionnaires. Also, inter-

views were conducted with those responsible and pa-

tients who are followed up at the hospital linked to

the college. The base contains information on health,

ﬁnancial conditions, parental care, education, among

others, in addition to the notes for the arithmetic, writ-

ing, and reading tests of each present in the database.

4.2 Pre-processing

With the aim to improve the quality of the data,

the pre-processing of the database was performed

through the steps described below.

1. Exclusion of instances that did not present the

grade of the Test of School Performance (TSP)

(Stein, 1994) in arithmetic, writing, or reading.

2. Transformation of the grades obtained in the TSP,

comparing them with the average grade of a

Brazilian state, in High and Low performance.

For this, the Standards Tables present in the TSP

manual were used. The tables in the manual for

the classiﬁcation of notes range from the 1st to the

6th grade, which corresponds from the 2nd to the

7th year in the current academic category. There-

fore, the years before the 2nd year were classi-

ﬁed according to the classiﬁcation criteria of the

2nd year, and those after the 7th year followed the

classiﬁcation parameters of the 7th year.

3. Exclusion of irrelevant features (e.g., name and

telephone number), and those who presented the

same information or were complementary (e.g.,

age in months and age in years), concatenating

them and maintaining only one feature.

4. Transformation of features belonging to the same

category into a single feature (e.g., Conduct Dis-

order and Opposition Disorder are part of Behav-

ior Disorders, so a single feature called ”Behav-

ior Disorders” has been created, which indicates

whether the individual has ”Conduct Disorder” or

”Opposition Disorder”).

5. Filling in the missing data by the average, in nu-

merical data, or by mode, in categorical features.

6. Binarization of non-ordinal nominal features, that

is, they were coded as the presence or absence of

the characteristic.

7. Random manual separation of 15% of each class’s

instances to carry out the testing stage.

8. Balancing of 85% of the remaining data using

the SpreadSubsample algorithm present in the

Waikato Environment for Knowledge Analysis

- WEKA

. The SpreadSubsample algorithm fol-

lows the undersampling approach, randomly re-

ducing instances of the majority class. Uniform

distribution was applied. After balancing, the

Randomize ﬁlter was executed in order to shufﬂe

the instances.

Thus, Table 1 shows the total number of instances,

of each class, for testing and creating models, after

the pre-processing steps. It should be noted that the

separation of the instances for testing was carried out

before the database was balanced.

Table 1: Number of balanced instances for creating the

model and unbalanced for the testing phase.

Discipline

Training/Validation Test

High Low High Low

Arithmetic 59 59 11 29

Writing 50 50 09 31

Reading 39 39 08 27

As for the features, at the end of the pre-

processing, 130 remained to represent the base. Fig-

ure 1 presents an overview of the database, with

the categories of data and the number of features in

each category, in the form ”Number of initial fea-

tures/Number of features after pre-processing”.

4.3 Dimensionality Reduction with

Genetic Algorithm, Random Forest,

and Specialist

A combination of three methods - GA, RF, and spe-

cialist - was used to select the essential features

in learning among the 130 features present in the

database.

The Non-dominated Sorting Genetic Algorithm II

(NSGA-II) algorithm was chosen to ﬁnd the best sub-

set of features maximizing your ﬁtness, in this case,

WEKA is open source software issued under the

GNU General Public License that contains a collec-

tion of ML algorithms (Garner, 1995). Available at

http://www.cs.waikato.ac.nz/ml/weka

HEALTHINF 2021 - 14th International Conference on Health Informatics

158

Figure 1: Overview of the database, grouped by categories.

the F-Measure. To evaluate the F-Measure, the K-

Nearest Neighbors (KNN) classiﬁer was used. His

choice was motivated by the low computational cost

for adjusting his parameters. In this work, the pa-

rameter k (number of closest neighbors) was adjusted,

varying its value range [1-10]. GA was implemented

in the Python language, using the DEAP library, avail-

able from Universit

e Laval (Fortin et al., 2012).

Aiming at the adequate deﬁnition of the necessary

parameters for the experiments’ execution, intervals

of values were analyzed, having as a stopping crite-

rion the number of GA generations. For each set of

parameters, 10 different random seeds were used. Ta-

ble 2 shows the ranges of values used.

Table 2: Range of parameter values.

Population initialization Random

Representation Binary

Crossover operator Two Points

Crossover probability 70%, 75%, 80% and 85%

Mutation operator One Point

Mutation probability 1%, 5% and 10%

Population size 100, 300 and 500

Number of generations 100, 300 and 500

Crossover selection method Tournament (size = 2)

Composition of the new generation Non-dominated individuals

Stopping criteria Number of generations

To automatically adjust the Random Forest algo-

rithm’s parameters, CVParameterSelection

was used

(Kohavi, 1995b), varying the number of trees and fea-

tures, and the depth of the trees. The RF calculates the

feature’s importance, indicating which are the most

important for creating the model. Only features with

relevance as from 70% were considered.

Regarding specialist, the knowledge of the medi-

cal expert in ADHD, who provided the database and

Available in the WEKA tool.

is a collaborator in this work, was used to ﬁnalize the

features selection, in conjunction with the informa-

tion present in the work of Ara

ujo (2002), focused on

the school performance.

Figure 2 shows the features selected by GA and

RF, separated for each of the three disciplines (Fig-

ures 2 a, b and c). The ﬁgure also shows the features

considered important, for the three disciplines (Figure

2 d), according to the specialist.

4.4 Description of Methods

To assist in discovering the rules that lead the stu-

dent to obtain a high or low performance, the VTJ48

algorithm (tree algorithm) and the JRip (rule algo-

rithm) were used. VTJ48 (Stiglic et al., 2012) has

the same functioning as J48, developed by Quinlan

(1993), which is a Java adaptation of C4.5. That is,

the VTJ48 also builds a decision tree from the in-

stances, but its difference is that it automatically ad-

justs the conﬁdence for pruning and the minimum

number of instances per leaf. JRip is a Java imple-

mentation of the Ripper algorithm, proposed by Co-

hen (1995) as an optimized version of the IREP algo-

rithm. JRip constructs the rules seeking to represent

the model as compactly as possible with as much in-

formation of the data; that is, it seeks to consistently

explain which features are relevant to the pattern(s)

found in the database, with the minimum of rules.

4.5 Model Quality Assessment Metrics

To evaluate the quality of the obtained models, Preci-

sion, Recall, and F-Measure metrics were used.

Precision

is the percentage of instances correctly

Precision =

V P

V P+FP

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction

Approach

159

Figure 2: Description of the features of the three databases, considering the three selection methods.

HEALTHINF 2021 - 14th International Conference on Health Informatics

160

classiﬁed in a class, out of all those that were classi-

ﬁed in the class.

Recall

is the percentage of instances of a class

that have been correctly predicted to belong to the

class.

F-Measure

represents the harmonic mean be-

tween Precision and Recall.

All classiﬁers were built and validated using the

k-fold cross-validation process, with k = 10 (Kohavi,

1995a).

5 RESULTS AND DISCUSSIONS

Figures 3 and 4 present the results in the phase of cre-

ating the model and test, respectively. The values re-

fer to the experiments with the selection of the GA

and the insertion of the features of the RF and special-

ist. Results are presented by discipline and by class.

It is noteworthy that in the experiments where the

features were selected only by the GA, the number

of features used in predicting arithmetic, writing, and

reading performance were only 16, 29, and 7, respec-

tively. In the experiments where the features are a

combination of the non-identical features present in

selecting the GA, RF, and specialist, 51 features re-

mained for arithmetic representation, 54 of writing,

and 43 of reading.

Analyzing the model creation phase, it is consid-

ered that the best results in the three disciplines come

from the JRip algorithm. Regarding the disciplines of

’arithmetic’ and ’reading’, the best results were ob-

tained only with the GA’s features. In the case of

the discipline ’writing’, the best results were obtained

with the combination of the features identiﬁed by the

GA, RF, and specialist. It is observed that the per-

formance obtained by the two algorithms remained

close, with F-Measure ranging from 64% to 84%.

To verify the efﬁciency of the generated models,

tests were performed with instances not seen during

the creation of the models. In terms of algorithms, the

disciplines ’arithmetic’ and ’reading’ obtained better

results with the VTJ48 algorithm, while in ’writing’,

the use of JRip was more efﬁcient. Regarding the fea-

tures, in the discipline ’arithmetic’, the best results

were obtained with the GA, RF, and specialist fea-

tures. In ’writing’ and ’reading’, the best results were

obtained with the features found only by the GA. An

F-Measure ranging from 48% to 92% is perceived for

both algorithms.

Recall =

V P

V P+FN

F −Measure =

2×Recall×Precision

Recall+Precision

5.1 Generated Rules

Table 3 presents the rules found, which explain the

reasons for a student to have a high or low perfor-

mance in the subjects, with coverage from 10%. Cov-

erage is the percentage of instances that the rule clas-

siﬁes correctly, out of all class instances.

Regarding arithmetic, analyzing the ﬁve rules, it

is clear that performance in writing vigorously inﬂu-

ences students’ performance in this discipline. It is

observed that the inﬂuence is mainly in rule 3, where

only good writing performance is enough for high

performance in arithmetic. This connection is pos-

sible since the student’s ease or difﬁculty decoding

and/or understanding mathematical symbols can af-

fect calculations. However, detailing each rule, the

relevance of other features is observed. Rules 1 and

2, for example, indicate that if the student does poorly

in writing and is of middle or vulnerable socioeco-

nomic class, according to the Secretariat for Strate-

gic Affairs (Kamakura and Mazzon, 2016), the stu-

dent will do poorly in arithmetic, indicating that the

ﬁnancial aspect is important for student performance.

These two rules alone classiﬁed 76% of the instances

of the lower class. With rule 4, it is possible to no-

tice several aspects of the individual’s life, because

if his performance in writing is average, and he was

never below the ideal weight for his age and was born

by normal childbirth, and his parents give him aver-

age or low autonomy (highest present value in the

base is 25) and there was no health problem during

pregnancy, the performance in arithmetic is high. The

parental style factor referring to autonomy may indi-

cate that if the father has a balance between listen-

ing to his son, but not letting him dictate the house’s

rules, or doing everything he wants, it has a positive

impact on performance. Rule 5 reveals that if the per-

formance in writing is average, and the student was

never below the ideal weight for his age and was born

by cesarean childbirth, the performance in arithmetic

is high, pointing out the relevance of the type of child-

birth.

In writing, rule 1 indicates that if the student has

a poor performance in arithmetic, he also does poorly

in writing, reinforcing the strong correlation between

subjects. This single rule classiﬁes 70% of instances

of the low class. Rule 2 shows that poor performance

in reading and gender = male also leads to poor writ-

ing performance. The fact that males appear in this

rule may indicate a reproduction of social stereotypes

where girls are said to be quieter, and boys are messy

and undisciplined, which leads to a negative assess-

ment of the child’s behavior and inﬂuences the ﬁnal

note (Carvalho, 2007). Rule 3 reveals that if the indi-

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction

Approach

161

Figure 3: Evaluation of model creation.

Figure 4: Evaluation of model quality through tests.

vidual’s life characteristics do not ﬁt into any previous

rules, then his performance will be high in writing.

That is, if the performance in arithmetic and reading

is high and the student is female, then the writing per-

formance will be high, with coverage of 88% of the

instances of this class.

Regarding reading, it is noted that the perfor-

mance in writing has a strong inﬂuence on the stu-

dent’s performance in this discipline, especially in

rule 2, where if the student does well in writing, he

also does well in reading. However, the inﬂuence of

some other features on reading performance is also

observed. Rule 1, for example, with coverage of 80%,

points to the importance of the parental behavior, be-

cause if the student has a low writing performance

and his parents are less indulgent (the highest present

value in the base is 24), that is, generally do not for-

give mistakes or praise successes, their reading per-

formance will also below. This observation indicates

that family support is essential, especially in times of

HEALTHINF 2021 - 14th International Conference on Health Informatics

162

Table 3: Generated rules and coverage.

Arithmetic

Number Rule Coverage

1 If writing performance = low and social class = average then Low 64%

2 If writing performance = low and social class = vulnerable then Low 12%

3 If writing performance = high then High 42%

4 If writing performance = medium and gained little weight = no and

childbirth = normal and autonomy <= 20 and health problem in preg-

nancy = no then High

22%

5 If writing performance = medium and gained little weight = no and

childbirth = cesarean then High

20%

Writing

Number Rule Coverage

1 If arithmetic performance = low then Low 70%

2 If reading performance = low and gender = male then Low 18%

3 If it doesn’t ﬁt any previous rule then High 88%

Reading

Number Rule Coverage

1 If writing performance = low and indulgence <= 15 then Low 80%

2 If writing performance = high then High 44%

3 If writing performance = average and fear of sleeping alone = no

then High

26%

difﬁculty. Rule 3 suggests that when writing perfor-

mance is average, and the student is not afraid to sleep

alone, he has a high reading performance. One of the

justiﬁcations for fear of sleeping alone may be linked

to phobic anxiety, which causes a lack of conﬁdence,

security, and support, leading to feelings of helpless-

ness, fragility, and dependence (da Rocha Antony,

2009). Therefore, not being afraid to sleep alone can

represent certain self-conﬁdence, which results in a

positive performance.

When analyzing the rules, it is noticed that knowl-

edge in other disciplines directly inﬂuences the stu-

dent’s performance in the discipline being predicted.

Therefore, other experiments were carried out to

verify the model’s behavior without the disciplines

as an entry feature, following the same methodol-

ogy described previously. However, for the ’read-

ing’ and ’writing’ disciplines, the algorithmic perfor-

mance was not promising, showing a loss of up to 47

percentage points in the high class. As for the disci-

pline of ’arithmetic’, there was again in algorithmic

performance of up to 11 percentage points in the low

class. In this discipline, the rules found relate some

features to the high class: 1) the fact that the mother

has completed a higher education course, 2) the stu-

dent is less than or equal to eight years old, and the

parents practice physical coercion in a moderate to

low way.

With the rules found, it is noted that performance

in the disciplines can be affected by family ﬁnancial

and health conditions during pregnancy, factors re-

lated to the student’s weight, parental behavior, sit-

uations surrounding birth, circumstances associated

with fear and insecurity, mother’s education, gender

and age of the student, in addition to her performance

in other subjects. Therefore, a multifocal approach

must be performed so that students with ADHD can

improve their academic performance.

6 FINAL CONSIDERATIONS

With the possible unfavorable prognosis of students

with ADHD, it becomes signiﬁcant to identify which

characteristics inﬂuence the individual’s experience

to alleviate their difﬁculties.

To identify the classiﬁcation rules, ”white box” al-

gorithms were used. The purpose of using these al-

gorithms was to obtain, more directly and objectively,

the knowledge acquired about the school performance

pattern of students with ADHD. The use of this tech-

nique proved to be positive for the database used. Re-

garding the reduction of dimensionality, it is noted

that the Genetic Algorithm performed quite satisfac-

torily since in two (writing and reading) of the three

disciplines evaluated, the best performance of the al-

gorithms was obtained only with the features identi-

ﬁed by this method.

Furthermore, the rules obtained reveal that differ-

ent contexts can inﬂuence school performance, not

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction

Approach

163

being ADHD synonymous with academic failure.

Therefore, it is necessary to understand the envi-

ronment of that child or adolescent with ADHD be-

fore creating strategies that aim to remedy or alleviate

the educational problems faced by them, since envi-

ronmental and emotional factors directly affect their

school performance, contributing to the difﬁculties al-

ready existing in the daily lives of people with ADHD

are expanded.

As future work is intended to evaluate new bal-

ancing distributions with the undersampling approach

since the one used in this work was the uniform distri-

bution. Besides, it aims to investigate what the results

would be like if the problem were treated as a multi-

label. In other words, it is intended to evaluate the

quality of the models in predicting the performance

of students in all subjects, jointly.

ACKNOWLEDGEMENTS

This study was ﬁnanced in part by the Coordination

for the Improvement of Higher Education Personnel

- Brasil (CAPES) - Finance Code 001. The authors

thank the National Council for Scientiﬁc and Techno-

logical Development of Brazil (CNPq - Conselho Na-

cional de Desenvolvimento Cient

ıﬁco e Tecnol

ogico)

and the Foundation for Research Support of the Mi-

nas Gerais State (FAPEMIG). The work was devel-

oped at the Pontiﬁcal Catholic University of Minas

Gerais, PUC Minas in the Applied Computational In-

telligence laboratory – LICAP.

REFERENCES

ABDA, Associac¸

ao Brasileira do D

eﬁcit de Atenc¸

(2016). Motivac¸

e coisa s

eria. https://tdah.org.br/

motivacao-e-coisa-seria/. Online; access October 5,

2020.

Ahmed, A. and Elaraby, I. S. (2014). Data mining: A pre-

diction for student’s performance using classiﬁcation

method. World Journal of Computer Application and

Technology, 2(2):43–47.

Andrade, A. L. M. and Lohr J

unior, A. (2017).

Quest

oes atuais acerca do transtorno de d

eﬁcit

de atenc¸

ao/hiperatividade. Psicologia Argumento,

25(48):73–83.

Angeli, C., Howard, S. K., Ma, J., Yang, J., and Kirschner,

P. A. (2017). Data mining in educational technol-

ogy classroom research: Can it make a contribution?

Computers & Education, 113:226 – 242.

Anuradha, J., Tisha, Ramachandran, V., Arulalan, K. V.,

and Tripathy, B. K. (2010). Diagnosis of adhd using

svm algorithm. In Proceedings of the Third Annual

ACM Bangalore Conference, COMPUTE ’10, New

York, NY, USA. Association for Computing Machin-

ery.

APA, A. P. A. et al. (2014). DSM-5: Manual diagn

ostico e

estat

ıstico de transtornos mentais. Artmed Editora.

Aradhya, A. M. S., Sundaram, S., and Pratama, M. (2020).

Metaheuristic spatial transformation (mst) for accu-

rate detection of attention deﬁcit hyperactivity disor-

der (adhd) using rs-fmri. In 2020 42nd Annual In-

ternational Conference of the IEEE Engineering in

Medicine Biology Society (EMBC), pages 2829–2832.

Ara

ujo, A. P. d. Q. C. (2002). Avaliac¸

ao e manejo de crianc¸a

com diﬁculdade escolar e dist

urbio de atenc¸

ao. Jornal

de Pediatria, 78:S104–S110.

Borges, H. B. and Nievola, J. C. (2012). Comparing

the dimensionality reduction methods in gene expres-

sion databases. Expert Systems with Applications,

39(12):10780–10795.

Cardoso, L., Mollica, A. M. V., Sales, A. M., and Ara

ujo,

L. C. (2019). O l

udico e a aprendizagem de crianc¸as

com transtorno de d

eﬁcit de atenc¸

ao/hiperatividade.

Revista Cient

ıﬁca FAGOC-Multidisciplinar, 3(2).

Carvalho, M. P. d. (2007). Por que tantos meninos v

ao mal

na escola? crit

erios de avaliac¸

ao escolar segundo o

sexo. Cadernos de Pesquisa.

Cohen, W. W. (1995). Fast effective rule induction. In

Twelfth International Conference on Machine Learn-

ing, pages 115–123. Morgan Kaufmann.

Cortez, M. T. and Pinheiro,

A. M. V. (2018). Tdah e escola:

incompatibilidade? Paid

eia, 13(19).

da Rocha Antony, S. M. (2009). Os ajustamentos cria-

tivos da crianc¸a em sofrimento: uma compreens

da gestalt-terapia sobre as principais psicopatologias

da inf

ancia. Estudos e pesquisas em Psicologia,

9(2):356–375.

de Azevedo Santos, J. L. (2017). Tdah-transtorno de d

eﬁcit

de atenc¸

ao e hiperatividade: Intervenc¸

ao psicope-

dag

ogica. Ideias e Inovac¸

ao-Lato Sensu, 4(1):115.

de Lacerda, E. G. and de Carvalho, A. (1999). Introduc¸

aos algoritmos gen

eticos. Sistemas inteligentes:

aplicac¸oes a recursos hıdricos e ci

encias ambientais,

1:99–148.

de Lima, C. B. and Coelho, C. L. M. (2018). Transtorno de

eﬁcit de atenc¸

ao/hiperatividade-um olhar sob a per-

spectiva da educac¸

ao especial inclusiva. Cadernos de

Pesquisa em Educac¸

ao, 20(47):172–192.

dos Santos, H. G., do Nascimento, C. F., Izbicki, R.,

de Oliveira Duarte, Y. A., Chiavegatto Filho, A. D. P.,

et al. (2019). Machine learning para an

alises predi-

tivas em sa

ude: exemplo de aplicac¸

ao para predizer

obito em idosos de s

ao paulo, brasil. Cad. Sa

ude

ublica, 35(7):e00050818.

DuPaul, G. J., Dahlstrom-Hakki, I., Gormley, M. J., Fu, Q.,

Pinho, T. D., and Banerjee, M. (2017). College stu-

dents with adhd and ld: Effects of support services

on academic performance. Learning Disabilities Re-

search & Practice, 32(4):246–256.

Fortin, F.-A., De Rainville, F.-M., Gardner, M.-A. G.,

Parizeau, M., and Gagn

e, C. (2012). Deap: Evolu-

tionary algorithms made easy. J. Mach. Learn. Res.,

13(1):2171–2175.

HEALTHINF 2021 - 14th International Conference on Health Informatics

164

Frazier, T. W., Youngstrom, E. A., Glutting, J. J., and

Watkins, M. W. (2007). Adhd and achievement: Meta-

analysis of the child, adolescent, and adult literatures

and a concomitant study with college students. Jour-

nal of learning disabilities, 40(1):49–65.

Garner, S. (1995). Weka: The waikato environment for

knowledge analysis. Proceedings of the New Zealand

Computer Science Research Students Conference.

Goldberg, D. E. (1989). Genetic algorithms in search. Op-

timization, and MachineLearning.

He, W. (2013). Examining students’ online interaction in

a live video streaming environment using data min-

ing and text mining. Computers in Human Behavior,

29(1):90 – 102. Including Special Section Youth, In-

ternet, and Wellbeing.

Holland, J. (1975). Adaptation in natural and artiﬁcial sys-

tems: an introductory analysis with application to bi-

ology. Control and artiﬁcial intelligence.

Jou, G. I. d., Amaral, B., Pavan, C. R., Schaefer, L. S., and

Zimmer, M. (2010). Transtorno de d

eﬁcit de atenc¸

ao e

hiperatividade: um olhar no ensino fundamental. Psi-

cologia: reﬂex

ao e cr

ıtica, 23(2):409–409.

Kamakura, W. and Mazzon, J. A. (2016). Crit

erios de

estratiﬁcac¸

ao e comparac¸

ao de classiﬁcadores socioe-

con

omicos no brasil. Revista de administrac¸

ao de em-

presas, 56(1):55–70.

Kohavi, R. (1995a). A study of cross-validation and boot-

strap for accuracy estimation and model selection. In

Proceedings of the 14th International Joint Confer-

ence on Artiﬁcial Intelligence - Volume 2, IJCAI’95,

pages 1137–1143, San Francisco, CA, USA. Morgan

Kaufmann Publishers Inc.

Kohavi, R. (1995b). Wrappers for Performance Enhance-

ment and Oblivious Decision Graphs. PhD thesis,

Stanford University, Department of Computer Sci-

ence, Stanford University.

Mattos, P. (2015). No mundo da lua: perguntas e respostas

sobre transtorno do d

eﬁct de atenc¸

ao com hiperativi-

dade em crianc¸as, adolescentes e adultos. Associac¸

Brasileira do D

eﬁct de Atenc¸

ao.

Pacheco, M. A. C. et al. (1999). Algoritmos gen

eticos:

princ

ıpios e aplicac¸

oes. ICA: Laborat

orio de In-

telig

encia Computacional Aplicada. Departamento

de Engenharia El

etrica. Pontif

ıcia Universidade

Cat

olica do Rio de Janeiro. Fonte desconhecida,

page 28.

Pappa, G. L., Freitas, A. A., and Kaestner, C. A. (2002).

A multiobjective genetic algorithm for attribute selec-

tion. In Proc. 4th Int. Conf. on Recent Advances in

Soft Computing (RASC-2002), pages 116–121, Not-

tingham. Nottingham Trent University.

Quinlan, R. (1993). C4.5: Programs for Machine Learning.

Morgan Kaufmann Publishers, San Mateo, CA.

Rahadian, B. A., Dewi, C., and Rahayudi, B. (2017).

The performance of genetic algorithm learning vec-

tor quantization 2 neural network on identiﬁcation of

the types of attention deﬁcit hyperactivity disorder.

In 2017 International Conference on Sustainable In-

formation Engineering and Technology (SIET), pages

337–341.

Rangel J

unior,

E. d. B. and Loos, H. (2011). Escola e

desenvolvimento psicossocial segundo percepc¸

oes de

jovens com tdah. Paid

eia, 21(50):373–382.

Sachnev, V. (2015). An efﬁcient classiﬁcation scheme for

adhd problem based on binary coded genetic algo-

rithm and mcﬁs. In 2015 International Conference

on Cognitive Computing and Information Process-

ing(CCIP), pages 1–6.

Santos, J. C., Oliveira, J. d. F., Dutra, L. V., Santa’Ana,

S., and Renn

o, C. D. (2007). Selec¸

ao de atribu-

tos usando algoritmos gen

eticos para classiﬁcac¸

ao de

regi

oes. XIII Simp

osio Brasileiro de Sensoriamento

Remoto, Florian

opolis, Brasil. INPE–Instituto Na-

cional de Pesquisas Espaciais, pages 6143–6150.

Santos, L. d. F. and Vasconcelos, L. A. (2010). Transtorno

do d

eﬁcit de atenc¸

ao e hiperatividade em crianc¸as:

uma revis

ao interdisciplinar. Psicologia: Teoria e

Pesquisa, 26:717 – 724.

Stein, L. M. (1994). Tde: teste de desempenho escolar:

manual para aplicac¸

ao e interpretac¸

ao. S

ao Paulo:

Casa do Psic

ologo, pages 1–17.

Stiglic, G., Kocbek, S., Pernek, I., and Kokol, P. (2012).

Comprehensive decision tree models in bioinformat-

ics. PloS one, 7(3):e33812.

UNESCO, I. f. S. (2012). International standard classiﬁ-

cation of education: ISCED 2011. UNESCO Institute

for Statistics Montreal, Quebec.

Analysis of School Performance of Children and Adolescents with Attention-Deﬁcit/Hyperactivity Disorder: A Dimensionality Reduction

Approach

165