Using Machine Learning to Analyze the Impact of Lifestyle and

Socioeconomic Factors on the Incidence of Depression Among Young

Brazilians

Thayris G. F. Rodrigues

, Ariane C. B. da Silva

and Cristiane N. Nobre

Institute of Exact Sciences and Informatics, Pontiﬁcal Catholic University of Minas Gerais,

Dom Jos

e Gaspar, Belo Horizonte, Brazil

Keywords:

Depression, Adolescence, Children, Machine Learning, Instance Selection.

Abstract:

Depression is a growing mental health problem among young people in Brazil, with factors such as socioe-

conomic and lifestyle conditions inﬂuencing its prevalence. This study investigates how variables such as

education, family situation, and access to services impact the incidence of depression, using data from the

National Health Survey (PNS) of the Brazilian Institute of Geography and Statistics (IBGE). Using machine

learning algorithms such as Random Forest, XGBoost, SVM, and MLP, the analysis identiﬁed patterns among

the factors, highlighting sleep problems and depressive feelings as the main determinants, with Recall above

70%. These results support the creation of more inclusive mental health policies.

1 INTRODUCTION

Depression is one of the most prevalent mental dis-

orders worldwide, representing a signiﬁcant concern

in the Brazilian context, especially among young peo-

ple, which saw an increase of 11.1.

In Brazil, depression is a public health concern.

The World Health Organization (WHO) estimates that

350 million people live with the disease, with Brazil

being the second country with the highest prevalence

in the Americas (Abelha, 2014), affecting a signif-

icant portion of the population. However, accord-

ing to estimates from the United Nations Children’s

Fund (UNICEF), the understanding of the speciﬁc de-

terminants of depression among Brazilian youth is

still limited, and the number of these young people

is very high, with almost 16 million between the ages

of 10 and 19 having some mental disorder (UNICEF,

2022).

The central problem motivating this research is

the high prevalence of depression among Brazilian

youth and the lack of comprehensive understanding

of the factors contributing to this condition (Fonseca

et al., 2008). Depression not only affects the mental

and emotional health of young people but can also

https://orcid.org/0009-0008-2380-9634

https://orcid.org/0000-0003-2477-4433

https://orcid.org/0000-0001-8517-9852

have signiﬁcant impacts on their personal relation-

ships, academic and professional performance. Addi-

tionally, the stigma surrounding depression often hin-

ders access to proper treatment and the necessary sup-

port for those suffering from this condition. These

factors can be observed in the works of (Santos and

Kassouf, 2007) and (Brito, 2011).

This study aims to investigate the relationship be-

tween depression and socioeconomic and lifestyle

factors among Brazilian youth aged 15 to 29, a demo-

graphic deﬁned as youth by the Youth Statute (da Ju-

ventude, 2015). Using data from the National Health

Survey (PNS) by the Brazilian Institute of Geogra-

phy and Statistics (IBGE), the research seeks to un-

derstand how variables such as residence, education,

family situation, ethnicity, and access to public health

services inﬂuence the prevalence of depression. By

focusing speciﬁcally on the Brazilian context, this

study aims to contribute to mental health policies and

preventive interventions targeted at this population.

This work is organized as follows: Section 2

presents the theoretical framework, addressing the

main concepts and studies on depression in Brazil and

worldwide, and on depression among young people.

Section 3 discusses related works, highlighting re-

search focusing on depression among Brazilian youth.

Section 4 provides a description of the methodology

used, including the data source and the criteria for

variable selection. Section 5 presents the results of

Rodrigues, T. G. F., B. da Silva, A. C. and Nobre, C. N.

Using Machine Learning to Analyze the Impact of Lifestyle and Socioeconomic Factors on the Incidence of Depression Among Young Brazilians.

DOI: 10.5220/0013185600003911

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 639-646

ISBN: 978-989-758-731-3; ISSN: 2184-4305

639

the analysis, followed by the discussion of these re-

sults. Finally, in Section 6, the study concludes, sug-

gesting directions for future research and implications

for public policies.

2 THEORETICAL FRAMEWORK

2.1 Depression in Brazil and the World

Depression is a mental illness characterized by per-

sistent sadness, loss of interest in activities once en-

joyed, and a range of emotional and physical symp-

toms. According to the World Health Organization

(WHO), Brazil stands out with one of the highest de-

pression prevalence rates in the world, being the ﬁfth

country with the most cases of the disease. It is esti-

mated that about 5.8% of the Brazilian population suf-

fers from depression, which represents approximately

11.7 million people, showing that Brazil is above the

global average of around 4.4% (World Health Orga-

nization, 2017).

According to this same study, in the context of the

Americas, Brazil holds an alarming position, being

behind only the United States, where 5.9% of the pop-

ulation is affected by depression. These data highlight

the severity of the situation and the urgent need for ef-

fective public policies to tackle depression.

Moreover, in this same WHO study, it is empha-

sized that depression is one of the leading causes of

disability worldwide and has profound impacts on

productivity and economic costs due to lost workdays

and associated disabilities. The lack of proper treat-

ment and its consequences are also recurring themes

in this research (World Health Organization, 2017).

Furthermore, the article by Rakel (1999) discusses

the prevalence and challenges in treating depression

in primary care. It emphasizes that depression is often

underdiagnosed and undertreated. Patients with de-

pression frequently present physical symptoms, such

as fatigue, insomnia, and unexplained pain, which can

hinder the correct diagnosis. The author also high-

lights that the longer a depressive episode lasts, the

greater the likelihood of recurrence, underscoring the

importance of appropriate and early treatment.

2.2 Depression Among Young People

Depression is one of the most prevalent mental disor-

ders among young people, with a signiﬁcant increase

in incidence rates during adolescence and early adult-

hood. This stage of life is marked by various transi-

tions and challenges that can make young individu-

als particularly vulnerable to mental health problems

(Thapar et al., 2022).

Several risk factors contribute to the development

of depression in this age group, with psychosocial

stress playing a central role. Experiences such as difﬁ-

culties in social relationships are particularly impact-

ful (Smith and Blackwood, 2004). These risk factors

not only increase the likelihood of developing depres-

sion but can also exacerbate the condition in already

vulnerable individuals.

The impacts of depression on young people are

extensive and profound, affecting various aspects of

their lives. Depression can compromise academic

performance, leading to educational difﬁculties that,

in turn, increase the risk of reliance on social assis-

tance and unemployment in adulthood (Thapar et al.,

2022). Beyond the direct effects on mental health,

these challenges can hinder individuals’ social and

economic development, creating a cycle of adversity

that is difﬁcult to break.

A particularly concerning aspect of depression in

young people is its tendency to recur. While many

young individuals may recover from a depressive

episode within a year, depression often becomes a

chronic condition with high relapse rates. Factors

such as comorbidities and social adversity are predic-

tors of a worse prognosis, complicating full recovery

and increasing the need for continuous interventions

(Carr, 2008).

Moreover, the growing prevalence of depression,

both in its milder forms and in Major Depressive Dis-

order (MDD), during adolescence and early adult-

hood underscores the need for more effective preven-

tion and treatment strategies. Studies indicate that

the global prevalence of MDD among children and

adolescents is approximately 1.3%, though this ﬁgure

tends to be higher in the older age groups within this

population (Thapar et al., 2022).

3 RELATED WORK

Depression is an extremely relevant topic and, accord-

ing to Barbosa et al. (2011), still faces resistance in

debate and knowledge production. Healthcare pro-

fessionals often lack adequate information for the de-

tection and management of depression, especially in

severe cases.

The increasing rates of depression among young

adults, as noted by Barbosa et al. (2011), are alarm-

ing. Studies such as Carneiro Pinto (2015) indi-

cate that factors like gender, age, and education level

may not result in signiﬁcant differences in emo-

tional symptoms related to depression, highlighting

HEALTHINF 2025 - 18th International Conference on Health Informatics

640

the importance of comprehensive analyses. In Brazil,

Gonc¸alves et al. (2018) found that low education lev-

els and physical inactivity are associated with depres-

sion in women, while living with a partner and engag-

ing in physical exercise act as protective factors. Ad-

ditionally, Emerson and Llewellyn (2023) point out

that about 20

Furthermore, Maia et al. (2023) demonstrate that

Machine Learning techniques can identify risk fac-

tors for depression among Brazilian youth, provid-

ing support for public policies. However, they high-

light ethical challenges in using such technologies.

Therefore, policies that consider the social determi-

nants of depression and train healthcare professionals

for proper interventions are essential (Santos and Kas-

souf, 2007). Early intervention programs and contin-

uous support are crucial, especially for young people

(Brito, 2011).

After reviewing related studies, it is essential to

acknowledge that each offers valuable insights into

youth depression. However, it is important to high-

light that these studies have certain limitations that

our work can address more comprehensively. It is

necessary to recognize limitations such as small sam-

ple sizes, speciﬁc focuses, or the absence of longitudi-

nal data. Table 1 presents comparisons and limitations

of each study.

This study’s proposal stands out by speciﬁcally

addressing depression among Brazilian youth and

identifying speciﬁc factors that may cause this con-

dition. It aims to contribute to the ﬁeld by broadening

existing perspectives, overcoming the limitations of

related works, and exploring new aspects of youth de-

pression, such as violence, socioeconomic conditions,

and access to mental health services.

4 MATERIALS AND METHODS

4.1 Database Description

The research was based on the National Health Sur-

vey (PNS) Database, a nationwide household sur-

vey conducted by the Ministry of Health (MS) and

the Brazilian Institute of Geography and Statistics

(IBGE). For this study, the most recent version of the

PNS, from the year 2019, was utilized. This version

offers comprehensive information on various sociode-

mographic, behavioral, and health characteristics, in-

cluding data related to depression. The original 2019

PNS database comprises 1,087 attributes and 293,726

instances.

The primary objective of this study is to analyze

the occurrence of depression among Brazilian youth.

For this purpose, the central attribute used for ﬁltering

the instances was Q092, which indicates whether a

physician or mental health professional has ever diag-

nosed the respondent with depression. All instances

where this attribute was absent were excluded from

the analysis. Additionally, the attribute C008, which

refers to the age of the household member on the ref-

erence date, was used as a ﬁltering criterion to restrict

the analysis to youth aged 15 to 29 years, as deﬁned

by the Statute of Youth (da Juventude, 2015).

From the original 1,087 attributes, the most rel-

evant ones were selected based on the risk factors

associated with depression, as identiﬁed in previous

studies such as Sim

oes (2021). After ﬁltering the

instances and selecting the attributes of interest, the

resulting dataset contained 63,260 instances (62,334

without a depression diagnosis and 926 with a depres-

sion diagnosis) and 32 attributes. Table 2 provides a

detailed breakdown of these attributes and their re-

spective descriptions.

4.2 Methodology

During Step 0, we performed several preprocessing

steps to reduce noise in the database, including han-

dling duplicates, removing outliers, and managing

missing values. We also analyzed the correlation of

attributes with the class to eliminate redundancies.

These actions were carried out carefully to ensure that

the database was clean and organized before applying

machine learning algorithms. This ensured that the

models were trained on more representative and ro-

bust data, minimizing distortions.

In Step 1, we further analyzed the dataset and

made additional adjustments by removing attributes

with incomplete information or few responses. We

also merged some attributes and their responses to ob-

tain more consistent variables, improving the quality

of the data. By the end of this step, the dataset con-

tained 27,701 instances, of which 26,775 had no de-

pression diagnosis, 926 had a positive diagnosis, and

24 attributes.

After preprocessing, we split the dataset into 80%

for training and 20% for testing, with stratiﬁcation to

maintain the correct proportion between cases with

and without a depression diagnosis. We applied strat-

iﬁed cross-validation with 10 iterations (StratiﬁedK-

Fold), using the average of the results to represent

model performance. This process was essential for

a more reliable evaluation of the models during hy-

perparameter tuning, improving the robustness of the

results.

Additionally, we used the Ant Colony-based in-

stance and attribute selection technique (RantIFS). In-

Using Machine Learning to Analyze the Impact of Lifestyle and Socioeconomic Factors on the Incidence of Depression Among Young

Brazilians

641

Table 1: Comparison of related works.

Reference Main Contributions Limitations

Santos and

Kassouf (2007)

Need for comprehensive mental health policies that include social

determinants

Broad focus, with little em-

phasis on speciﬁc strategies for

young people

Fonseca et al.

(2008)

Gender differences in the perception of depression and the inﬂuence

of the socio-affective context

Focus on reactive depression

and limited sample; suggests

broader scope in future studies

Barbosa et al.

(2011)

Identiﬁcation of resistance in the debate on depression and the

alarming increase in depression rates among young adults

General approach, without a

speciﬁc focus on Brazilian

youth or cultural/regional

factors

Carneiro Pinto

(2015)

Study in Portugal showing that gender, age, and education do not

signiﬁcantly inﬂuence emotional symptoms

Study conducted in another

country; results may not be di-

rectly applicable to the Brazil-

ian context

Gonc¸alves et al.

(2018)

Identiﬁcation of speciﬁc risk factors for depression in Brazilian

women

Limited focus on women aged

20 to 59 years and speciﬁc geo-

graphic area

Emerson and

Llewellyn

(2023)

Detailed analysis of the relationship between depression, disability,

and lifestyle factors in young people from low and middle-income

countries.

The study presents a global

approach, without an analy-

sis focused on the cultural

and socioeconomic speciﬁcs of

Brazilian youth.

Maia et al.

(2023)

Use of Machine Learning to identify proﬁles and key factors of de-

pression in Brazil.

Ethical challenges and limi-

tations in data interpretation;

no speciﬁc focus on Brazilian

youth

stances from the majority class were selected with

a 70% probability to ensure representative samples,

and attribute selection was performed after combining

the selected instances. This approach helped reduce

the impact of class imbalance. The use of the Ant

Colony was crucial for optimizing the most relevant

instances and features. After this step, the database

contained 16,723 instances (15,797 without depres-

sion diagnosis and 926 with a positive diagnosis) and

15 attributes. Table 3 shows the number of instances

per class for training/validation and testing.

In Step 2, we trained various machine learning

algorithms, including Decision Tree Classiﬁer, Ran-

dom Forest Classiﬁer, Gradient Boosting Classiﬁer,

XGBoost Classiﬁer, MLPClassiﬁer, and SVM (Sup-

port Vector Machine). To optimize performance,

we applied advanced hyperparameter tuning tech-

niques such as GridSearchCV combined with strat-

iﬁed cross-validation (StratiﬁedKFold). These tech-

niques ensured an efﬁcient search for the best hy-

perparameters, signiﬁcantly improving model perfor-

mance.

Additionally, we addressed class imbalance us-

ing a combination of Undersampling and Oversam-

pling. First, we applied Random Undersampling

to reduce the majority class samples without exces-

sive loss of relevant information. Then, we used

the G SM1 method to generate synthetic samples for

minority classes, considering nearby neighbors and

adding controlled noise for diversity.

After balancing, we normalized the data with

StandardScaler to standardize the variable scales,

promoting better convergence of the models during

training.

Step 3 consisted of evaluating the trained classi-

ﬁers using metrics such as Precision

, Recall

and F1-

Score

to assess model performance. Precision is the

proportion of instances correctly classiﬁed as positive

out of all predicted as positive, assessing the model’s

precision in predictions. Recall measures the percent-

age of positive instances correctly identiﬁed by the

model, indicating its ability to detect relevant exam-

ples. F1-Score is the harmonic mean between Pre-

cision and Recall, balancing these metrics, especially

when there is a trade-off between them. These metrics

provided a detailed analysis of the algorithms’ effec-

tiveness in correctly predicting the classes.

Finally, in Step 4, we focused on interpreting

the generated models, analyzing the results obtained.

This phase was crucial for evaluating the performance

of the algorithms after applying data preprocessing

and balancing techniques. We examined the impact

of these techniques on performance metrics such as

precision, recall, and F1-score. The analysis allowed

us to identify the best model and understand the in-

ﬂuence of each technique, enabling adjustments for

future iterations.

Precision =

T P

T P+FP

Recall =

T P

T P+FN

F1-Score = 2 ×

Precision×Recall

Precision+Recall

HEALTHINF 2025 - 18th International Conference on Health Informatics

642

Table 2: Description of Attributes.

Attribute Description Reference

Individual Characteristics

Age Refers to the chronological age of individuals. C008

Race Refers to the racial identiﬁcation of individuals. C009

Gender Refers to the gender of the individual, which can be male or female. C006

Marital Status Indicates the current marital status of the individual (single, married, divorced,

widowed, etc.).

C011

Socioeconomic Conditions

Literacy Indicates whether the individual is literate, that is, whether they can read and

write.

D001

Education Level Refers to the highest level of formal education achieved by the individual. VDD004A

Health Insurance Indicates whether the individual has health insurance. I00102

Lifestyle

Frequency of Sluggish-

ness or Agitation

Refers to the frequency with which the individual feels sluggish or agitated in

their daily life.

N015

Sleep Problems Refers to difﬁculties in falling asleep or maintaining continuous sleep. N010, N011

Depressive Feelings Indicates the presence of feelings related to depression, such as persistent sad-

ness.

N016, N017,

N018

Screen Time Refers to the amount of time the individual spends using electronic devices with

screens (cell phone, computer, television, etc.).

P04501, P04502

Alcohol Consumption Refers to the frequency and quantity of alcohol consumed by the individual. P02801

Smoking Indicates whether the individual is a smoker and the frequency of tobacco con-

sumption.

P050

Health Conditions

Disabilities Refers to the presence of any type of physical, sensory, or mental disability. G033, G048,

G066

Mental Health Indicates the general mental health condition of the individual. Q092

General Health Condi-

tion

Refers to the individual’s perception of their own health, both physical and men-

tal.

J001, J00101,

J001

Violence

Insults or Threats Refers to the frequency with which the individual experiences verbal insults or

threats.

V00201,

V00202

Intimidation or Aggres-

sion

Indicates the frequency of intimidation or physical aggression experienced by

the individual.

V00204,

V01401,

V01402,

V01403

Sexual Aggression Refers to the occurrence of any type of sexual violence experienced by the indi-

vidual.

V02801,

V02802

Table 3: Dataset at different processing stages.

Class Original Dataset Filtered by Age Post Preprocessing Unbalanced Training Balanced Training Test Set

Diagnosed 8,332 926 926 741 1,895 185

Not Diagnosed 82,514 62,334 26,775 12,637 1,895 3,160

Total 90,846 63,260 27,701 13,378 3,790 3,345

4.3 ML Algorithms

As for the ML algorithms, after the preprocessing

stage, the training set was used for the training of six

classiﬁcation algorithms: DT (criterion: entropy, max

depth: 4, max features: None, min samples leaf: 14,

min samples split: 14); XGBoost (colsample bytree:

0.54, gamma: 0.1, learning rate: 0.04, max depth: 5,

min child weight: 1, n estimators: 300, reg alpha:

0.01, reg lambda: 1, subsample: 0.631); Gradient-

Boost (learning rate: 0.01, max depth: 12, n estima-

tors: 300, subsample: 0.9, min samples leaf: 5, min

samples split: 20); SVM (C: 0.502, class weight: bal-

anced, degree: 2, gamma: auto, kernel: rbf); MLP

(activation: relu, alpha: 0.1, hidden layer sizes: 512,

learning rate: adaptive, max iter: 500, solver: adam).

All of them were built using the Scikit-learn library

version 1.0.2 (Pedregosa et al., 2011). The experi-

ments were conducted on the Windows 11 operating

system using an Intel(R) Core(TM) i5-1035G1 pro-

cessor, 1.00 GHz, 8 GB of RAM, and the Jupyter tool

v2022.1.3

Using Machine Learning to Analyze the Impact of Lifestyle and Socioeconomic Factors on the Incidence of Depression Among Young

Brazilians

643

5 RESULTS AND DISCUSSIONS

The results presented in Table 4 show that almost all

the Machine Learning algorithms achieved more than

70% in the Recall metric, indicating signiﬁcant efﬁ-

ciency in detecting individuals with a depression di-

agnosis. The high Recall, especially in algorithms

such as Random Forest (90%), SVM (85%), and XG-

Boost (77%), suggests that these models were effec-

tive in correctly identifying depression cases based on

the provided socioeconomic, lifestyle, and health at-

tributes.

On the other hand, when observing precision, the

models displayed behavior that suggests a high rate of

false positives, with low values such as 16% in Ran-

dom Forest and 19% in SVM. This means that while

the models correctly identify many true depression

cases (high recall), a signiﬁcant number of individ-

uals without depression are incorrectly classiﬁed as

positive for the condition.

Table 4: Performance of ML Algorithms (in percentage).

Algorithm Precision Recall F1 Class

DT 98 86 91 No diagnosis

AUC 0.86 22 68 33 With diagnosis

RF 99 72 84 No diagnosis

AUC 0.88 16 90 27 With diagnosis

GradientBoost 98 88 92 No diagnosis

AUC 0.85 23 63 33 With diagnosis

XGBoost 98 83 90 No diagnosis

AUC 0.89 21 77 33 With diagnosis

SVM 99 79 88 No diagnosis

AUC 0.87 19 85 31 With diagnosis

MLP 98 85 91 No diagnosis

AUC 0.85 22 73 34 With diagnosis

In studies on depression in young people, sensi-

tivity (Recall) is one of the most important metrics,

as the primary goal is to minimize the number of in-

dividuals incorrectly diagnosed as not having depres-

sion. Failing to identify a depression case can worsen

the patient’s condition, negatively affecting treatment

success. Therefore, it is essential for the model to

have a high Recall rate, ensuring that most individu-

als with a depression diagnosis are detected.

Additionally, discrepancies between Recall and

Precision may be related to the complexity of the

dataset. Depression analysis involves various factors,

such as socioeconomic and health conditions, which

can be difﬁcult to distinguish clearly. The presence

of overlapping attributes may have made it challeng-

ing for the models to differentiate between individuals

with and without a depression diagnosis, resulting in

false positives.

Figure 1 illustrates the instances from the dataset

after dimensionality reduction to 2 dimensions us-

ing the non-linear reduction technique t-distributed

stochastic neighbor embedding (t-SNE). Even after

data balancing, the plots show that the two classes

(with and without a depression diagnosis) still signif-

icantly overlap. This indicates that the separability

between the classes in the represented dimensions is

not very clear, which contributes to the difﬁculty in

the classiﬁcation task, directly affecting the models’

performance in terms of Recall and Precision.

5.1 SHAP Chart Analysis

In addition to the quantitative metrics, we performed

an analysis using SHAP (SHapley Additive exPlana-

tions) plots on the model with the best performance

(XGBoost) to interpret the individual impact of each

variable on the model’s predictions. This interpreta-

tive approach was important for understanding how

the features inﬂuence the model’s decisions at differ-

ent levels. Figure 2 presents this analysis visually,

showing the relative impact of the variables.

The SHAP graphs showed that the most important

variables for prediction are related to mental health

aspects, especially depression. They played a cen-

tral role in the model’s results, helping identify fac-

tors inﬂuencing the classiﬁcation between diagnosed

and undiagnosed cases. This provided a deeper un-

derstanding of the model’s behavior.

• Sleep Problem Frequency: This variable has a

signiﬁcant impact on the model’s prediction. Indi-

viduals who report a high frequency of sleep prob-

lems (values in red) are strongly associated with a

higher likelihood of belonging to the positive class

(depression diagnosis). This relationship aligns

with the study by M

uller and Guimar

aes (2007),

which indicates that sleep disturbances are impor-

tant markers of mental health conditions.

• Frequency of Feeling Depressed: Individuals

who frequently report feeling depressed showed

a clear correlation with the positive class (diag-

nosed with depression), conﬁrming that this is one

of the most important indicators.

• Frequency of Slowness or Agitation: This vari-

able also stands out as a strong predictor. Be-

havioral rhythm changes, such as extreme slow-

ness or episodes of agitation, are often associated

with depressive disorders. High frequency (in red)

greatly increases the probability of classiﬁcation

as positive for depression.

Other variables also played a relevant role in the

model’s prediction. The variable ”Hours on Devices

for Leisure” showed an ambiguous impact: exces-

sive use may be linked to isolation and emotional de-

cline, while moderate use can promote socialization

HEALTHINF 2025 - 18th International Conference on Health Informatics

644

Figure 1: Spatial representation of training and testing sets. (a) Unbalanced database. (b) Balanced database. c) Test Set.

Figure 2: SHAP explanations for the XGBoost model.

and leisure, highlighting that context matters.

Experiences of Physical Violence” showed high

positive SHAP values, indicating an increased prob-

ability of depression in individuals who experienced

physical violence. Such traumas are recognized as

risk factors for mental health issues, with studies

like Bontempo and Pereira (2012) and Pedrosa de

Medeiros (2018) exploring how violence can impact

mental health, particularly in adolescents and Brazil-

ian women, leading to depression.

Finally, variables like ”Health Plan” and ”Disabil-

ity” reﬂect socioeconomic conditions and personal

challenges that can inﬂuence depressive symptoms,

albeit with a subtler impact on the model. These ﬁnd-

ings underline the importance of considering both in-

dividual and contextual aspects in the interpretation

of results.

6 CONCLUSION

This study investigated the prevalence of depression

among young Brazilians aged 15 to 29, along with

the socioeconomic and lifestyle factors that inﬂuence

this condition. Using data from the 2019 National

Health Survey (PNS) and various machine learning

techniques, it was possible to identify relevant pat-

terns among the analyzed variables.

Although the machine learning models performed

well in correctly classifying both young individuals

diagnosed with depression and those without a di-

agnosis, they still faced challenges in identifying all

cases accurately, as reﬂected in the low Precision val-

ues for undiagnosed cases. However, nearly all mod-

els consistently showed Recall values above 70%,

suggesting that these algorithms were relatively ef-

fective in identifying the majority of diagnosed cases

based on socioeconomic, lifestyle, and health at-

tributes provided as input.

The low Accuracy indicates a higher rate of false

positives, which can be attributed to the complex-

ity of the dataset and the overlap between the ”diag-

nosed” and ”undiagnosed” classes. This overlap com-

promised the algorithms’ ability to clearly distinguish

between individuals with and without a depression di-

agnosis. Furthermore, contextual and subjective fac-

tors inﬂuencing the diagnosis may not have been fully

captured by the quantitative variables used, suggest-

ing that more comprehensive data or reﬁned model-

ing techniques are necessary to improve the balance

between Recall and Precision.

The SHAP plot analysis was essential for inter-

preting the importance of variables in the model,

highlighting that factors such as the frequency of

sleep problems and the frequency of feeling depressed

showed a strong correlation with the depression diag-

nosis. On the other hand, variables such as the time

spent on leisure activities with devices showed a more

ambiguous impact, suggesting that contextual and en-

vironmental factors inﬂuence the diagnosis in a more

complex manner. These results emphasize the im-

portance of interpretable methods and improving data

quality to more accurately capture the psychological,

social, and contextual aspects that affect depression

diagnosis.

The results of this study highlight the importance

of public policies that expand access to education and

mental health services for young people, especially in

vulnerable areas. Preventive programs and interven-

tions that promote mental well-being are essential to

Using Machine Learning to Analyze the Impact of Lifestyle and Socioeconomic Factors on the Incidence of Depression Among Young

Brazilians

645

reduce depression rates and their impacts.

Future studies should investigate factors such as

culture and social support in youth depression, as well

as use longitudinal data to better understand the evo-

lution of mental health over time. It is also impor-

tant for future work to analyze cases that were in-

correctly classiﬁed as positive by the algorithms but

ofﬁcially do not have the diagnosis. This process

could reveal patterns and characteristics present in

the database that make it difﬁcult to correctly sepa-

rate cases, paving the way for adjustments in models

or data preprocessing, improving the balance between

Recall and Precision.

ACKNOWLEDGMENTS

The authors would like to thank the National Coun-

cil for Scientiﬁc and Technological Development of

Brazil (CNPq – Code: 311573/2022-3), the Co-

ordination for the Improvement of Higher Educa-

tion Personnel - Brazil (CAPES - Grant PROAP

88887.842889/2023-00 - PUC/MG, Grant PDPG

88887.708960/2022-00 - PUC/MG - Informatics and

Finance Code 001), the Foundation for Research

Support of Minas Gerais State (FAPEMIG – Codes:

APQ-03076-18 and APQ-05058-23).

REFERENCES

Abelha, L. (2014). Depress

ao, uma quest

ao de sa

ude

ublica. Cadernos Sa

ude Coletiva, 22(3):223–223.

Barbosa, F. d. O., Macedo, P. C. M., and Silveira, R. M.

C. d. (2011). Depress

ao e o suic

ıdio. Revista da

SBPH, 14:233–243.

Bontempo, K. d. S. and Pereira, A. R. (2012). Sa

ude mental

de crianc¸as e adolescentes v

ıtimas de viol

encia: uma

revis

ao cr

ıtica da literatura. Revista de Terapia Ocupa-

cional da Universidade de S

ao Paulo, 23(2):130–136.

Brito, I. (2011). Ansiedade e depress

ao na adolesc

encia.

Revista Portuguesa de Medicina Geral e Familiar,

27(2):208–14.

Carneiro Pinto, Joana Martins, P. B. P. T. C. O. A. (2015).

Ansiedade, depress

Ao e stresse: Um estudo com

jovens adultos e adultos portugueses. Psicologia,

ude e Doenc¸as.

Carr, A. (2008). Depression in young people: Descrip-

tion, assessment and evidence-based treatment. De-

velopmental Neurorehabilitation, 11(1):3–15. PMID:

17943506.

da Juventude, E. (2015). Perﬁl da juventude de ribeir

preto 2015. https://www.ribeiraopreto.sp.gov.br/port

al/pdf/2015-jovem-04202109.pdf. Acesso em: 23 de

maio de 2024.

Emerson, E. and Llewellyn, G. (2023). Parental report of

signs of anxiety and depression in children and adoles-

cents with and without disability in middle- and low-

income countries: Meta-analysis of 44 nationally rep-

resentative cross-sectional surveys. Child Psychiatry

& Human Development.

Fonseca, A. A. d., Coutinho, M. d. P. d. L., and Azevedo, R.

L. W. d. (2008). Representac¸

oes sociais da depress

em jovens universit

arios com e sem sintomas para de-

senvolver a depress

ao. Psicologia: Reﬂex

ao e Cr

ıtica,

21:492–498.

Gonc¸alves, A. M. C., Teixeira, M. T. B., Gama, J. R.

d. A., Lopes, C. S., Silva, G. A. e., Gamarra, C. J.,

Duque, K. d. C. D., and Machado, M. L. S. M. (2018).

Preval

encia de depress

ao e fatores associados em mul-

heres atendidas pela estrat

egia de sa

ude da fam

ılia.

Jornal Brasileiro de Psiquiatria, 67:101–109.

Maia, C., Nobre, C., Gomes, M., and Z

arate, L. (2023).

Using machine learning to identify proﬁles of indi-

viduals with depression. In Anais do XI Symposium

on Knowledge Discovery, Mining and Learning, pages

105–112, Porto Alegre, RS, Brasil. SBC.

uller, M. R. and Guimar

aes, S. S. (2007). Impacto dos

transtornos do sono sobre o funcionamento di

ario e a

qualidade de vida. Estudos de Psicologia (Campinas),

24(4):519–528.

Pedrosa de Medeiros, Mariana Zanello, V. (2018). Relac¸

entre a viol

encia e a sa

ude mental das mulheres no

brasil: an

alise das pol

ıticas p

ublicas. Estudos e

Pesquisas em Psicologia.

Rakel, R. E. (1999). Depression. Primary Care, 26(2):211–

224.

Santos, M. J. d. and Kassouf, A. L. (2007). Uma

investigac¸

ao dos determinantes socioecon

omicos da

depress

ao mental no brasil com

enfase nos efeitos da

educac¸

ao. Economia Aplicada, 11(1):5–26.

Sim

oes, R. F. (2021). Preval

encia de sintomas depressivos

e fatores associados em usu

arios da atenc¸

ao prim

aria

a sa

ude de belo horizonte. Dissertac¸

ao de mestrado,

Universidade Federal de Minas Gerais, Belo Hori-

zonte, MG, Brasil. Dispon

ıvel no Reposit

orio Insti-

tucional da UFMG, http://hdl.handle.net/1843/44207.

Smith, D. J. and Blackwood, D. H. R. (2004). Depression

in young adults. Advances in Psychiatric Treatment,

10(1):4–12.

Thapar, A., Eyre, O., Patel, V., and Brent, D.

(2022). Depression in young people. The Lancet,

400(10352):617–631.

UNICEF (2022). Sa

ude mental de adolescentes.

World Health Organization (2017). Depression and other

common mental disorders: Global health estimates.

Accessed: 2024-05-17.

HEALTHINF 2025 - 18th International Conference on Health Informatics

646