Mental Health Prediction in Students using Data Mining Techniques

2.1. Review on Factors Influencing Student’s Mental Health Issue

Seera et al. have proposed [5] multiple logistic regressions to determine the associations between selected factors and help-seeking from mental health professionals (MHPs). All analyses have been performed using IBM SPSS, version 21. The dataset involved 593 students, where 311 have been Thai students and 282 Indian students. In Thailand, compared to India, fewer pupils have been reported to seek expert assistance. In both nations, not being willing to seek assistance from MHPs was connected with having a high GHQ score, which indicates a likelihood of having mental health problems.

Lipson et al. studied over 350,000 students’ data from 373 institutions [6]. The data included the dataset between 2013 and 2021. The minimum age requirement to take the survey was 18 years old or older. Through athletics and academic advising, universities play a significant role in detecting risk factors and mental health issues of students. Previous research has shown a higher rate of mental health problems in students of colour. The data date back to the beginning of COVID-19.

In a paper reported previously [7], the authors have collected 300 students’ data from several universities of China. Obsessive-compulsive disorder, interpersonal sensitivity, and depression have been found to be the top three problems at the university level. Despite the fact that numerous studies have looked at mental health difficulties among Chinese college students, little has been said about the connections between mental health and factors, such as personal and familial situations, academic performance, and social support. In order to develop policy or psychological interventions and lessen mental health issues among university students in China, comparative studies are required to better understand the underlying mechanisms of mental health problems in college students and their long-term effects on the students' career development.

Johnson et al. collected datasets from 21 studies (15 randomized controlled trials and 6 non-randomized studies) [8]. The research was done from April 2006 to February 2020. The treatment gap for mental diseases may be bridged in part by digital technologies. Further, the authors have analyzed the research on technology-based interventions to enhance help-seeking for mental health and hypothesized that such interventions could be significant in favorably affecting the same.

Scheel et al. gathered information from 275 students (18–55 years old) [9]. 66% of respondents reported they had never considered suicide; 28.6% said that it had crossed their minds occasionally, 3.7% said it had occurred to them seriously, and 1.5% indicated that they had tried to kill themselves one or more times. The rates of suicidal ideation and attempts have been found to be considerably lower than those observed in studies of American Indian adolescents in general, and the proportion of students reporting current suicidal ideation has been very similar to the 13% discovered. However, 57% of participants reporting suicidal ideation have indicated that they would not seek the assistance of mental health professionals. 14% of individuals considering suicide said they were likely to use none of the available resources for assistance.

Mishra et al. collected datasets on 200 patients (125 males and 75 females) in AIIMS, New Delhi, from July-August 2009, comprising many catchment areas and also distant states [10]. A group of patients with mental health issues most frequently sought out psychiatrists as their first point of contact, followed by non-psychiatric medical professionals, such as family practitioners. The cost of treatment with non-psychiatric doctors was found to be higher than others. The authors suggested that to avoid treatment delays, there is a need to spread awareness about mental health.

Alsahali collected data from 330 students, where more than one-third (36%) of the participants agreed that people with mental illnesses would go to mental health clinics for professional assistance, and 32.4% of them said they would go to a doctor if they experienced emotional problems [11]. 60% of participants said they would not feel at ease discussing their personal troubles with healthcare professionals, and 48.8% said that they would feel ashamed if their friends knew they were obtaining professional assistance for emotional problems. Approximately 59% of the students thought that those with mental illnesses would be more likely to commit crimes.

Nguyen et al. gathered data from 268 students (67 domestic and 201 international) [1]. Intercultural interactions can lead to acculturative stress when international students settle into their new surroundings; thus, they must adjust to resolve any disputes that may arise. The fact that international students experience depression at a higher rate than domestic students may result from a variety of factors. First, acculturative stress is more intense for international students attending an international university, which increases their likelihood of developing depression. Second, compared to domestic students, international students have less access to mental health services. Third, compared to domestic students, international students have fewer options for finding sources of assistance. The main hypothesis that was supported by the study was that social connectedness and acculturative stress are both significantly negatively correlated with depression in both domestic and international students.

After studying the above literature, it has been found that there are several factors that can lead to mental health issues among students, of which depression and anxiety disorders are more prevalent. The symptoms of depression include severe sadness, anger, loss of interest, sleep disorder, appetite disturbances, slow thinking, suicidal thoughts, and guilt. Anxiety disorder is basically termed as an inability to regulate fear or worry. In our survey, the data have been collected based on depression, anxiety disorder, social connectedness, and class performance.

2.2. Machine Learning Approaches Adapted to Predict Depression and Anxiety Disorder

In a previous paper [12], the authors considered 4041 patients’ datasets in STAR*D and 665 patients’ datasets in COMED. A repeated k-fold cross-validation technique has been applied to create multiple test-train folds. 25 top predictors have been extracted after applying a data-driven feature selection approach. The gradient-boosting machine learning algorithm has been trained with the top 25 features to predict the treatment outcomes. In this study, the gradient boosting model showed an accuracy of 64.6%.

A machine learning approach to identify the various levels of mental health concerns has been suggested by Ahmed et al. [13]. Five distinct AI techniques, including convolutional neural networks, support vector machines, linear discriminate analysis, k Nearest Neighbour, and linear regression, have been employed in the study to analyse data related to anxiety and depression. The CNN model had shown the highest accuracy, with 96% for anxiety and 96.8% for depression.

Sau et al. proposed a predictive model for automated diagnosis of anxiety and depression among older patients using machine learning algorithms [14]. The authors have evaluated 10 classifiers with a dataset of 510 geriatric patients and tested with 10-fold cross-validation method. Out of 10 classifiers, the random forest has shown the highest prediction accuracy of 89%.

Generalized anxiety disorders were distinguished from healthy people and major depressive disorders by Hilbert et al. using machine learning techniques [15]. They employed a sample of people with generalized anxiety disorders, healthy individuals, and people suffering from serious depression to create the dataset. They used binary support vector machine and discovered that using the clinical questionnaire data made it difficult to predict generalized anxiety disorders. For the classification of normal case and disorder, the input involving the inclusion of cortisol and grey matter volume could achieve an accuracy of 90.10% and 67.46%, respectively.

Mengjun Luo has proposed [16] an improved decision tree, the improved ANN algorithm. In this study, the dataset has been collected from 49 students. The pressure on people with respect to living conditions, opportunities for learning and companionship, and employment conditions have been rising as a result of society's rapid and diverse development, which poses a serious psychological problem.

From the above literature, it can be observed that not a single classifier has given a better result for all the datasets. As classifying mental health data is generally very challenging, the features used in the machine learning algorithms significantly affect the performance of the classification. The performance of the classifier is dependent on the dataset used. Sau and Bhakta [14], Ahmed et al. [13], and Hilbert et al. [15] have applied the random forest and support vector machine in the classification of mental health problems, and they have provided comparatively better results. Most of the literature studies have involved a small sample size, and to the best of our knowledge, logistic regression and linear regression have not been efficiently used for this specific problem.

3.1. Study Site

In this research, three colleges in Odisha have been selected as the study sites. The colleges have been selected because they accommodate students of ages 18-25, both from Odisha and outside of Odisha. These colleges also involve students from various branches and departments.

3.2. Participants

Males contributed to 49.2% of the reported population, while females comprised 59.8%. With 43% of students in their fourth year of college, just 28% in their first year, and 13.1% and 12.1% in their second and third years, respectively, there has been found a significant gap in the years of the students studying there. 78.5% of respondents have reported as not having an intimate partner, and 79.4% have reported to be religious. Only 24.3% of participants have reported to be raised in rural areas, compared to 75.7% who said that they were raised in an urban setting. When asked how often they worked out, only 26.2% of participants reported working out often, compared to 21.5% who said that they worked out seldom (Table 1).

3.3. Instruments of Measurement

The participants were required to fill out questionnaires on socio-demographic information, Patient Health Questionnaire (PHQ-9) [16-18], revised Social Connectedness Scale (SCS) [19, 20], General Anxiety Disorder (GAD-7) [21, 22], and questionnaire on academic performance [23-29].

3.4. Procedure

The research was conducted based on grounded theory, where the data from different university students have been collected in a systematic way using a survey link. The survey was conducted in an online format through a Google Form for two reasons: firstly, it is very common and used by a lot of people, and secondly, it was more cost-effective and easier to manage and maintain. The survey was distributed in various branches and departments of three colleges in Odisha. The survey was made available online and participants could fill it out at their convenience. The total response was 107 (link to the survey: https://forms.gle/usLPUkLFt5ZHbbu3A).

Prior to filling out the form, the participants were asked for their consent. Each participant had to read a consent form outlining the purpose, research methods, research team, their selection as a research subject, and potential benefits, risks, and discomforts of participating in the research. The participants had the option to choose between yes or no to participate in the research. By selecting “no,” the participants had the option to leave the survey at any point. In the case of agreeing to take the survey, a list of questions was shown to the participants.

3.5. Methodology

Microsoft Excel was used to edit the unprocessed data retrieved from the Google form. The various stages of data analysis we implemented are as follows: tables and graphs were first used to compare the data visually. The dataset was transformed into a .csv file, after which data cleaning, normalisation, one hot encoding, label encoding, and feature selection were carried out on the raw data. The dataset was then subjected to multiple logistic regression and multiple linear regression techniques. The pre-processed dataset was additionally subjected to the Chi-squared test and Pearson correlation in order to meet our additional goals.

3.6. Model Diagram

The model diagram depicted in Fig. (1) describes the steps performed to predict the prevalence of depression.

The Google form's raw data were gathered and converted to.csv format (Responses.csv). The dataset was then imported into Jupyter Notebook, where data pre-processing was carried out. One hot encoding and label encoding were used during pre-processing, along with the removal of duplicates and null values from the dataset. Further, feature selection was carried out, where each feature's relevance was determined, and any features that had a value of 0 were deleted (reduced dataset). The data were then split into training and testing data, and both linear and logistic regression models were fitted to the data. Both models' accuracy was examined and compared. For greater accuracy, the logistic regression model's c parameter was hyper-tuned.

4.1. Dataset Description

The initial dataset is shown in Table 2, where the row represents the samples and the columns represent the attributes for the study. Since the entire dataset could not be used in the test, only 5 samples are displayed in Table 2.

4.2. Feature Selection

The importance of each feature has been measured by the Chi-square test. Based on the test result, Table 3 represents the feature importance values for each feature. With respect to the feature importance value, some of the features have been selected as relevant features, and these are shown in Table 4. The table contains the list of features included in the reduced dataset.

4.3. Descriptive Results

PHQ-9 results showed 61.9% of all respondents to have depression. Males had a higher rate of depression than females did. While mild depression was shown to be more prevalent in females (36.58), minimum depression was more prevalent in males (21.87). Among all students, mild depression was common (33.33%). In total, 7.61% of students reported severe depression, with nearly an equal number of males (7.81%) and females (7.31%). The rate of moderately severe depression was comparatively higher in males (12.5%) than females (4.87) (Table 5).

Table 6 displays the prevalence of depression based on socio-demographic characteristics, as reported by PHQ-9. Second-year college students reported having a higher depression rate than those in any other year. The lowest depression rate was reported by fourth-year students. However, female students in the third grade reported the highest rate of despair, at 87.5%. Compared to students who replied no, students with partners were shown to have higher levels of depression. Comparatively more women than men reported having depression. Students who answered “no” to being religious were found to be melancholy in 81.81% of cases. Males who declined to identify as religious had higher rates of depression than females, both among men and women, the level of depression was higher in male students who were born in rural areas. However, females were recorded as having a higher depression rate if they were born in an urban area. It was observed that students who exercised often showed lower depression rates as compared to students who exercised seldom or occasionally.

Table 7 displays the prevalence of social connectedness based on socio-demographic factors, as reported by PHQ-9. Fourth-year college students reported having a higher social connectedness than those in any other year. However, female students in the first grade reported the highest rate of social connectedness, at 87.5%. Compared to students who replied yes, students with partners were shown to have higher levels of social connectedness. Comparatively more women than men reported having social connectedness. The level of social connectedness was higher in female students who were born in rural areas. It was observed that students who exercised often showed higher social connectedness rates as compared to students who exercised seldom or occasionally.

Table 8 displays the prevalence of anxiety disorder based on socio-demographic factors, as reported by PHQ-9. Second-year college students reported having a highest anxiety disorder rate than those in any other year. However, female students in the 2nd grade reported the highest rate of anxiety disorder, at 77.77%. Compared to students who replied no, students with partners were shown to have higher levels of anxiety disorder. Comparatively more women than men without partners reported higher anxiety disorder. Students who answered “no” to being religious were found to have an anxiety disorder in 77.27% of cases. Females who declined to identify as religious had higher rates of anxiety disorder than males. However, females were recorded to have a higher depression rate if they were born in an urban area. It was observed that female students who exercised seldom showed the highest anxiety disorder rates as compared to students who exercised occasionally.

4.4. Main Analysis

The dataset was split into 75% training and 25% testing sets. The dataset was also split into 50% training and 50% testing for validation. Prior to feature selection, the data were first fitted into a logistic regression model to determine accuracy. After feature selection, the data were once more fitted to test the results, and the accuracy has been found to be significantly improved. To improve accuracy and reduce log loss, the c parameter was hyper-tuned as well. (Tables 9 and 10). It was discovered that while the testing and validation accuracy improved, the training accuracy remained constant. The linear regression model provided an accuracy of 70.03% after feature selection. Table 11 displays the accuracy of the suggested model as well as the performance metrics for several models that have been used to analyze data on mental health issues in the literature. From the table, it is evident that the proposed model has provided better accuracy in comparison to other models, except for CNN. On the other hand, linear regression has not given satisfactory results for our data set.

Table 11 shows that, for our dataset, the hyper-parameter tuned logistic regression model has outperformed all other models in the literature, excluding CNN.

4.5. Factors of Depression

The dataset in this study has involved both the numerical and categorical features. Social connectedness and anxiety disorder as numerical variables have been checked using Pearson correlation. The results have shown anxiety and depression to be correlated positively and social connectedness and depression to be correlated negatively. Chi-squared testing has been used for categorical variables. The findings have indicated exercise and the year of college to be linked to depression. The confusion matrix for the Pearson correlation coefficient is depicted in Table 12. Table 13 displays the results of the Chi-squared test on the category variables.

4.6. Effects of Depression

The association between academic achievement and depression was determined using the Cumulative Grade Point Average (CGPA). “There is no association between academic success and depression,” was taken as the null hypothesis. “There is a relationship between academic success and depression,” was taken as the alternative hypothesis. Through the Chi-squared test, the p-value was observed to be less than the significance level (α), rejecting the null hypothesis and indicating a relation between depression and academic performance (Table 13). The Chi-squared statistic χ2 was also found to be greater than the critical value, again rejecting the null hypothesis (Table 14).