Research on Student Mental Health Problems Based on Machine

Learning Method

Jiawen Luo

College of General Aviation and Flight, Nanjing University of Aeronautics and Astronautics, Nanjing, China

Keywords: Student Mental Health Problems, Machine Learning, Prediction.

Abstract: Students suffer from mental health problems have less adaptability and efficiency, often show some unusual

behaviors. It is important to find these problems early and eliminate them in time. Machine learning

techniques can be fully used to predict mental health problems. Such applications could identify those students

who have potential mental health problems from the results of student questionnaires. Three methods are

proposed to handle student mental health problems. Their performances are compared to identify the best

model and hyper-parameters. The data set includes the academic situation and mental health conditions of

101 university students. There are 11 attributes in the data set. Two new attributes “Total Mental Health Issue”

and “CGPA Midpoint” are created in the research to better analysis the relation between academic

performances and mental health conditions. Random Forest has the best performance with n_estimators is

100 and max_depth is 6. The research shows the feasibility of predicting students’ mental health problems

with machine learning techniques. This technology has a broad prospect and is beneficial to the whole society.

1 INTRODUCTION

Students live in campus, and study is their main task.

They are also in an age of rapid social development

and full of changes and competition. Many students

struggle with the difficult challenges of individuating

from their birth families and managing the pressure

of their studies, while others may have to take care of

a lot of employment and family obligations (Pedrelli

et al 2015). For many students, going to college may

be a stressful time. These pressures from different

aspects lead to students' mental health problems.

Students suffer from mental health problems have

less adaptability and efficiency, often show a lot of

unadaptable, uncoordinated, irrational and even

wrong behaviors. Thus, it is important to find these

problems in the early days and eliminate them in time

to avoid causing serious adverse consequences.

The algorithms of machine learning can be

entirely utilized to forecast mental health issues.

When applied, such applications would benefit the

society as a tool to monitor individual aberrant

behavior (Srividya et al 2018). The results of student

questionnaires can be used to determine recent

changes in their mental state and generate

corresponding labels. Some classifiers are built based

on these labels to predict mental health conditions of

each student. Such applications could help us identify

those students who have potential mental health

problems at early stages. Thus, psychological

counseling can be implemented in a timely manner.

Three methods are proposed to handle student mental

health problems.

The literature review section reviews and

summarizes some previous articles. In the

methodology section, the data set and algorithms are

introduced. The result section shows the result of this

research. The conclusion section is the summary of

the whole research.

2 RELATED WORK

On the basis of machine learning, characteristic

extraction, and data resources, Rohizah Abd Rahman,

et al proposed a crucial assessment analysis in Online

Social Networks on mental health detection (Rahman

et al 2020). The data analysis method, comparison,

challenges, and restrictions of mental health detection

were analyzed to determine its propriety. The research

selected 22 articles from 2770 articles between 2007

and 2018. Big data in OSNs helps detect mental

Luo, J.

Research on Student Mental Health Problems Based on Machine Learning Method.

DOI: 10.5220/0012801400003885

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Data Analysis and Machine Learning (DAML 2023), pages 285-290

ISBN: 978-989-758-705-4

285

health problems, which shows high potential in early

detection as a data source. A. B. R. Shatte, et al.

adopted a scoping analysis method to quickly define

the realm of machine learning in mental health

(Shatte et al 2019). The extraction of data included

information on mental health applications, machine

learning techniques, data types, and research results.

Support vector machine, decision tree, and neural

network were used. The application of machine

learning has shown a series of benefits to mental

health, but most of researches concentrate on the

identification and treatment of mental health

disorders. Machine learning applications still have

plenty of scope to grow in other fields.

Jetli Chung and Jason Teo used the PRISMA

methodology when collecting relevant research

articles and studies by searching reliable databases

(Chung and Teo 2022). Researchers' challenges and

limitations were reflected in the research. In addition,

specific suggestions on potential future research and

development were also provided. Currently, there is

no model that can predict a person's likelihood of

having mental health issues. Machine learning

techniques could improve logistic regression of the

standard prediction modelling technique. Ashley E.

Tate, et al aimed to evaluate whether machine

learning techniques are superior to logistic regression

and create a model to forecast mental health issues in

mid-adolescence (Tate et al 2020). The research used

nearly 500 predictors from register data and parental

report. Finally, the best preforming model is not fit for

clinical use. It is not necessary to seek more complex

methods and forgo logistic regression for similar

studies.

Sumathi M.R. and B. Poorna identified eight

algorithms and evaluated the efficacy in diagnosing

five basic mental health issues with various measures

(Sumathi and Poorna 2016). In order to train and

detect the accuracy of the algorithms, a data set of

sixty cases was collected in the research. Ayako Baba

and Kyosuke Bunji obtained data from 63%

responses of about 6000 undergraduate students from

a Japanese national university (Baba and Bunji 2023).

The research compared the results of different

machine learning models, including elastic net,

logistic regression, XGBoost, random forest, and

LightGBM. According to the comparison, the

LightGBM model performed the best. Both

conditions and analyses reached adequate

performance in this model.

Konda Vaishnavi, et al identified five machine

learning techniques, including KNN, Random Forest,

Decision Tree, etc. (Vaishnavi et al 2021). The

research used several accuracy criteria to assess the

accuracy in identifying mental health problems.

Finally, they acquired the most accurate model based

on the Stacking technique with the prediction

accuracy 81.75%. Jetli Chung and Jason Teo

evaluated some popular machine learning algorithms

(Chung and Teo 2023). Responses to a survey taken

by Open Sourcing Mental Illness were included in the

data set. Machine learning techniques included

Gradient Boosting, Logistic Regression, KNN,

Neural Networks, and Support Vector Machine, as

well as an ensemble approach based on these

algorithms. Gradient Boosting reached the highest

accuracy, which was 88.80%, providing a highly

promising approach toward automated clinical

diagnosis for mental health professionals.

3 METHODOLOGY

Random Forest, Support Vector Machine, and

Logistic Regression are three machine learning

methods that the study suggests using. Their

performances on this data set are compared in order

to identify the best model.

The classification algorithm Support Vector

Machine (SVM) uses interval maximization that

separates data points of different classes by finding an

optimal hyperplane. Data points are mapped into a

high-dimensional space as the fundamental concept

of SVM, which makes it easier to separate data points.

SVM is a commonly used machine learning algorithm

with high accuracy and generalization ability. Finding

an ideal hyperplane that optimizes the separation

between different categories of data points is the aim

of SVM. This distance is known as the margin, and

support vectors are the most closely linked data points

to the hyperplane. The following stages can be used

to explain the fundamentals of SVM, mapping data

points to a high-dimensional space, finding an

optimal hyperplane in a high-dimensional space to

increase the separation of data points from the

hyperplane in several categories., classifying data

points into different categories according to the

hyperplane, categorizing new data points. In SVM,

the mapping of data points can be implemented using

different kernel functions. Gaussian, polynomial, and

linear kernel functions are examples of kernel

functions that are frequently used. These kernel

functions can map data points into higher-

dimensional Spaces, making it easier to separate data

points in higher-dimensional Spaces.

Random forest belongs to the category of

ensemble learning, which creates a strongly

supervised model by mixing weakly supervised

DAML 2023 - International Conference on Data Analysis and Machine Learning

286

models. If one of the weak models produces a wrong

prediction, the others can correct the error. Bagging

adopts a random sampling with return, that is, a fixed

number of samples are randomly collected from the

training set, but after each sample is collected, the

samples are returned. T weak learners are trained by

T samples, and then strong learners are generated by

combination strategy. Random forest is implemented

on the basis of Bagging, and the weak learner is

decision tree. The fundamental goal of a decision tree,

a supervised classification model, is to choose a

feature value that will result in the greatest

information gain for tree segmentation. The node

segmentation process of the decision tree is similar to

a recursive process, which finds the most critical

feature according to the information gain, splits it

according to this feature, and splits the data nodes of

the subtree in a similar way until the feature is

exhausted or all the data on this node is the same label.

The random forest chooses a few features on the node,

whereas the ordinary decision tree chooses an ideal

feature from all n sample features to segment the

decision tree. To further improve the model's

generalizability, an optimal feature is chosen for tree

segmentation from among the characteristics that

were randomly chosen.

Such a procedure is logistic regression, which

involves establishing a cost function, iteratively

solving the ideal model parameters using

optimization techniques, testing the results to ensure

the accuracy of our solution. Although the term

"regression" appears in the name of the technique,

logistic regression is essentially a classification

method that is primarily applied to two classification

issues. The generalized linear model is a family of

regressions that includes multiple linear regression

and logistic regression, both of which have many

similarities. The dependent variables in this family of

models are varied, but the model forms are essentially

the same. It is multiple linear regression if the data is

continuous. Logistic regression is used if the

distribution has a binomial shape. Although binary

variables are more widely used and simpler to

understand, binary and multi-categorical dependent

variables are both acceptable in logistic regression.

Binary logistic regression is the most common

method in practice.

Firstly the results of these three models are

generated, which include the accuracy, precision,

recall, and F1 score. Then the best model is identified.

Next some hyper-parameters of the model are set as

different values. Finally the best hyper-parameters

and the best performance of the model are recorded.

4 RESULTS AND DISCUSSION

The data set for this research was collected by Google

forms in a survey in 2020. The survey was aimed to

find the relationship between the academic situation

and mental health of university students. Both basic

information and the mental health conditions of the

total 101 students were recorded in the data set. Table

1 shows the attributes of the data set.

Table 1: The attributes of the data set.

No.

Attribute

Values

Date&Time

8/7/2020 15:14,

13/7/2020 10:34,

etc

Gender

Male, Female

Age

18, 19, 20,

etc

Course

Engineering, Law, IT, etc

Year

Year 1, Year 2,

etc

CGPA

3.00-3.50, 3.50-4.00, etc

Marital Status

Yes, No

Depression

Yes, No

Anxiety

Yes, No

Panic Attack

Yes, No

Treatment

Yes, No

In the preliminary data processing, the

distribution of students of different courses and

genders who have mental health problems is firstly

visualized. Fig. 1 shows the distribution of depression.

The difficulties of engineering, IT, and BCS are

higher that other courses, and students of these three

courses suffer from the most depression. For all

courses, female students are more likely to have

depression than male student. The distributions of

anxiety and panic attack also have similar situations.

It can be concluded that female students and students

of difficult courses such as engineering, IT, and BCS

are more likely to suffer from mental health problems.

Research on Student Mental Health Problems Based on Machine Learning Method

287

Figure 1: The distribution of course, gender, and depression (Picture credit: Original).

The distributions of students of different ages and

years who have mental health problems are showed

in Fig. 2, Fig. 3, and Fig. 4. Different figures

correspond to different mental health problems.

Students in year 1 aged between 18 and 20 suffer from

all these three mental health problems. Students in

year 4 do not experience anxiety, depression, or panic

attack except for those who are aged 24. In year 3,

elder students are more likely to have anxiety, while

younger students are more likely to have depression.

Students of all ages in year 2 experience depression

and panic attack, but only those who are aged between

18 and 20 experience anxiety.

Figure 2: The distribution of year, age, and anxiety (Picture

credit: Original).

Figure 3: The distribution of year, age, and depression

(Picture credit: Original).

Figure 4: The distribution of year, age, and panic attack

(Picture credit: Original).

DAML 2023 - International Conference on Data Analysis and Machine Learning

288

Figure 5: The distribution of year, age, and CGPA (Picture credit: Original).

Table 2: The results of three models.

SVM

Random Forest

Logistic Regression

Accuracy

0.333333

0.904762

0.761905

Precision

0.333333

0.904762

0.761905

Recall

0.333333

0.904762

0.761905

F1 score

0.333333

0.904762

0.761905

The relation among students’ CGPA, years, and

ages is showed in Fig. 5. Combining the distributions

of CGPA and mental health problems, it can be found

that students in year 4 have the best academic

performance, and they suffer from the least mental

health problems. Top students in year 3 also perform

well, but there are still many students have their

CGPA under 2.0. That is why students in year 3

experience much more mental health problems than

students in year 4. Students in year 1 and 2 have their

CGPA all above 2.5, so they experience less mental

health problems than students in year 3. It can be

concluded that those students who have better

academic performances suffer from less mental health

problems.

Through the data processing, it can be found that

the students’ mental health conditions have a close

relation with their academic performances. Thus, two

new attributes are created to better analysis the

relation between academic performances and mental

health conditions. The “Total Mental Health Issue”

attribute is combined with three attributes,

“Depression”, “Anxiety”, and “Panic Attack”. It can

present the total mental health issue of each student.

Another one is the “CGPA Midpoint” attribute from

“CGPA”. It transforms an interval value into a

specific value which can increase efficiency when

dealing with the data.

The accuracy, precision, recall, and F1 score of

three models are all carried out and showed in Table

2. Considering false positives and false negatives are

both important in this research, the F1 score is the

decisive target. The value of SVM is the lowest, only

0.33. The value of Logistic Regression is much higher

and reaches 0.76. Random Forest obtains the highest

value, which is 0.90. Random Forest might be a good

choice for a precise model.

In order to obtain the highest F1 score of

Random Forest, four different values of n_estimators,

50, 100, 200, 300, and four different values of

max_depth, 2, 4, 6, 8 are used. The values of various

n_estimators and max_depth can be found in Table 3.

For different values of max_depth, when n_estimators

is 100 the F1 score is always higher than others. The

highest point appears when max_depth is 6. The F1

score even reaches 1. Thus the best hyper-parameters

are also identified, n_estimators is 100 and

max_depth is 6.

The top Random Forest model has an extremely

high F1 score. That is probably because the data set is

too small and the model is over-fitting. The data could

not present general conditions in daily life. When

putting other data sets in the best model, the results

may not be so satisfying. To solve this problem, some

larger data sets should be adopted in the research.

Research on Student Mental Health Problems Based on Machine Learning Method

289

Table 3: The values of different n_estimators and max_depth.

N_estimators

Max_depth

100

200

300

0.714286

0.857143

0.809524

0.761905

0.904762

0.952381

0.904762

0.952381

0.904762

0.952381

0.904762

5 CONCLUSION

The research shows the feasibility of predicting

students’ mental health problems with machine

learning techniques. Nowadays, students are under

increasing pressure, more and more students suffer

from mental health problems. Students have potential

mental health problems can be identified at early

stages and treated in time. This technology has a

broad prospect and is beneficial to the whole society.

However, the data set is too small, which could not

present general conditions. The research just selects

three models from machine learning techniques.

There are still many machine learning models and

even deep learning models. In the future, an online

questionnaire will be made in order to collect more

students’ information and obtain a larger data set. The

research will adopt more models, especially deep

learning models.

REFERENCES

P. Pedrelli, M. Nyer, A. Yeung, C. Zulauf, and T. Wilens,

“College Students: Mental Health Problems and

Treatment Considerations,” Academic Psychiatry, vol.

39, pp. 503–511, 2015.

M. Srividya, S. Mohanavalli, and N. Bhalaji, “Behavioral

Modeling for Mental Health using Machine Learning

Algorithms,”Journal of Medical Systems, vol. 42, no.

88, 2018

R. A. Rahman, K. Omar, S. A. M. Noah, M. S. N. M.

Danuri, and M. Al-Garadi, “Application of Machine

Learning Methods in Mental Health Detection: A

Systematic Review.” IEEE, 2020, pp. 183952 -

183964.

A. B. R. Shatte, D. M. Hutchinson, and S. J. Teague,

“Machine learning in mental health: a scoping

review of methods and applications,” Psychological

Medicine, vol. 49, no9, pp. 1426 - 1448, 2019.

J. Chung, and J. Teo, “Mental Health Prediction Using

Machine Learning: Taxonomy, Applications,

and Challenges,” Applied Computational Intelligence

and Soft Computing, pp. 9970363, 2022.

A. E. Tate, R. C. McCabe, H. Larsson, S. Lundström, P.

Lichtenstein, and Ralf Kuja-Halkola, “Predicting

mental health problems in adolescence using machine

learning techniques,” PLoS ONE, vol. 15, pp.

e0230389, 2020.

Sumathi M.R., and B. Poorna, “Prediction of Mental Health

Problems Among Children Using Machine Learning

Techniques,” International Journal of Advanced

Computer Science and Applications, vol. 7, pp. 552-

557, 2016.

A. Baba, and K. Bunji, “Prediction of Mental Health

Problem Using Annual Student Health Survey:

Machine Learning Approach,” JMIR Ment Health, vol.

10, pp. e42420, 2023.

K. Vaishnavi, U. N. Kamath, B. A. Rao, and N. V. S. Reddy,

“Predicting Mental Health Illness using Machine

Learning Algorithms,” AICECS, 2021.

J. Chung, and J. Teo, “Single classifier vs. ensemble

machine learning approaches for mental health

prediction,” Brain Informatics, vol. 10, pp. 1, 2023

DAML 2023 - International Conference on Data Analysis and Machine Learning

290