A machine learning approach for predicting the depression status among undergraduates: A case study from Sri Lanka

doi:10.21203/rs.3.rs-3555106/v1

Download PDF

Research Article

A machine learning approach for predicting the depression status among undergraduates: A case study from Sri Lanka

https://doi.org/10.21203/rs.3.rs-3555106/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Depression is a common mental health disorder that affects a significant portion of the global population. In Sri Lanka, depression is a growing concern, especially among undergraduate students. This study aims to develop a machine learning model to predict the depression status of science undergraduates from a state university in Sri Lanka

Methods

Data was collected from a sample of 363 undergraduates via a Google form which included the Patient Health Questionnaire-9 (PHQ-9) and several questions pertaining to their background. The data collected (through a convenience sampling scheme) was then used to build and evaluate 13 different machine learning models. The accuracy, precision and recall were used to evaluate the performance of each model.

Results

30% of the sample consisted of those who screened positive for depression. The results showed that the Gradient Boosting Classifier with an accuracy of 79% on the test set had the best accuracy as well as the best precision in predicting the depression status. Feature importance analysis identified the overall life satisfaction and level of stress associated with academic activities as the most important predictors of depression. Furthermore, the extent to which the respondents enjoy their university life, their BMI and whether or not they had nutritionally balanced diets were found to be the next three most important factors of depression status.

Conclusions

This study is possibly the first attempt to model and predict the depression status of Sri Lankan undergraduates using machine learning approaches. The developed model can be used as a screening tool to identify students who may be at risk of developing depression and can aid in providing targeted interventions to improve the mental health and well-being of undergraduates.

Depression

PHQ-9

Gradient Boosting Classifier

Machine Learning

Depression is a mental illness that can negatively affect the way you feel and act. It can have ill effects on one’s life such as changes in sleeping patterns (1), lack of energy and a decrease in physical endurance (2). It is also associated with various physical symptoms such as weight gain (3), lower back pain (4) and headache (5). Furthermore, it is found to make a person have suicidal thoughts (6). Major Depressive Disorder (MDD) is a form of depression that can potentially exist over a long period of time, and for some, even be a lifelong condition. However, there are treatments that can help an individual deal with this mental disorder. These include but are not limited to exercise (7), counselling (8), antidepressant medication (9) and psychotherapeutic treatment (10). It is crucial to treat depression at the onset of the illness (11). However, many of those with this condition are unwilling to seek professional help (12). Early identification and proper guidance given to those at risk of developing the illness are of paramount importance as depression ranks as the most common mental disorder not only in Sri Lanka but also in many other countries in the world (13).

University is an exciting place for the youth of Sri Lanka. However, it is common for many undergraduates to experience stress because of the changes in their lives such as moving away from home, studying, learning English due to their mother tongue being Sinhala or Tamil and for some, even becoming financially independent. Having to cope with these changes may have an association with certain undergraduates being at risk of developing depression. Studies show that more students are dealing with depression now than in previous generations (14).

A student’s academic performance may suffer due to depression (15). Studies have shown that students having depression can have their cognitive functions affected (15). Depression in university students can also lead to unhealthy habits. Some turn to alcohol or substance abuse to deal with their stress and other negative feelings (16). This can increase around the end of a semester when deadlines and course work become more demanding. Most alarmingly, depression is strongly related with suicide attempts (17). More than half the deaths due to suicide in Sri Lanka occur among those of the age group 15–29 (18). While there is no cure for depression, treatment is available to control it (11). These treatments can improve the symptoms and minimize their impact on an individual’s daily life.

In Sri Lanka, depression was estimated at 6.6% of the total population in 2010 (19). However, since then, Sri Lanka has had to face various crises including the Easter Sunday bombings, Covid-19 pandemic, and a crippling economic crisis, and hence, the prevalence rates of depression may now be different. This research may be the first case study conducted on depression among the undergraduates of a state University after the Covid-19 pandemic.

In Sri Lanka, university education is offered free of charge at state funded Universities to those who perform outstandingly well at the General Certificate Examination – Advanced Level. Due to this the university community consists of people from different parts of the country, speaking different languages, following different religions, and coming from various economic backgrounds. Furthermore, some of these students may get ragged and harassed by their seniors when they enter. These background factors which differ from student to student could place some at a risk of developing depressive symptoms. Hence, a model to identify the most influential socio-demographic, academic and economic factors associated with depression and eventually to screen students who may be at risk of developing symptoms will be of great benefit to the Sri Lankan youth at the University.

Data Description

A survey was conducted to collect data from the undergraduates studying at this state university. A set of questions was designed to gather the socio-demographic information of the participants, and another set of questions was included from the PHQ-9 Questionnaire. PHQ-9 is a rating scale that can be used to screen the participants for depression. This rating scale has been validated for use both internationally (20) and for the general population in Sri Lanka (21). This study categorized each participant as depressed or not depressed, according to the PHQ-9 score. The process of collecting the data and ensuring that a representative sample was obtained is given below.

A random sampling technique was used for this study. The sample size was calculated according to Slovin’s formula (22). Slovin’s formula is given as

\(n=\frac{N}{(1+N\times {e}^{2})}\) 1

where \(n\) is the sample size, \(N\)is the population size and \(e\) is the margin of error.

At the time of this study, the total number of active students of the faculty under consideration was 2251. Taking the margin of error to be 5%, the required sample size was calculated to be 340 students.

A study done to compare the response rates of different modes of data collection for community based surveys found that internet surveys had a response rate of just 2.2% (23).

Wayne’s formula given below can be used to adjust the sample size to account for non-response bias (24).

\(n=\frac{N}{W}\) 2

where \(n\) is the sample size, \(N\) is the number of respondents and \(W\) is the response rate.

Assuming a response rate of 15% and the number of respondents to be 340 (as calculated above), the total number of students the questionnaire needed to be sent out to is 2251.

In general, as stated above, the response rate for internet-based surveys is around 2.2%. However, it is not possible to assume a response rate this small for the current study, because if it was assumed to be 2.2%, then the number of people the survey needed to be sent out to would be larger than the total population size when calculated according to Wayne ‘s formula (Eq. 2). The smallest response rate which could be assumed without the required sample size exceeding the population size is 15.104%. This value was calculated by equating the number of people the survey needed to be sent out to, to the population size, then taking 340 as the required sample size and finally calculating what the response rate would be.

Hence, the questionnaire was sent to all the undergraduates of the Faculty under consideration.

The Google form which was sent out had 360 usable responses, satisfying the originally required sample size of 340. The dataset was separated into two in the ratio 80:20 and used as the training set and testing set respectively. This yielded 288 observations and 72 observations respectively for the two tests. The entire analysis carried out from this point forward was done on the training set.

There are 3 variables which are important to consider when stratifying, to ensure that the random sample obtained is representative of the population. These are gender, stream of study and year of study. Using the Chi-Squared Goodness of fit test it was found that there is no significant difference observed between the distribution of gender over the sample and the distribution of gender over the population. However, when considering the stream and the year, there were significant differences observed. Thus, post stratification weights were calculated using both the year of study and the stream. In Part 1 (the descriptive analysis) of the Results, the graphs were generated by incorporating these weights so as to reflect the population characteristics. The additional data processing techniques given below were only used when the data was fed into machine learning models and is not relevant to e Part 1 of the Results section.

A novel approach for fitting machine learning models utilizing post stratification weights

There is no established and clear-cut way in the literature of utilizing the post stratification weights in fitting classification models.

When fitting a classification model, the goal is often to generalize the model to the population from which the sample was drawn. However, if the sample is not representative of the population in some way, the resulting model may be biased and hence may not accurately reflect the population. In this case, post-stratification weights can be used to adjust the sample data to better reflect the population distribution.

The importance of using post-stratification weights in fitting classification models is to ensure that the resulting model is unbiased and more representative of the population, which can lead to more accurate predictions and classifications. Hence, a novel approach of using complex survey weights is proposed in this study which could be utilized to incorporate the post stratification weights in fitting machine learning models.

What is proposed is to draw a sample without replacement (essentially creating a new dataset) from the original dataset such that the new sample will be representative of the population. This dataset should be selected using probabilities which are proportional to the post stratification weights assigned to each observation. By this, the data points which are under-represented in the sample (which have a larger post stratification weight) will have a higher probability of being selected to the resampled dataset in comparison to the data points which are over-represented in the sample. The observations which are over-represented will have a smaller post stratification weight and thus will have a lower probability of being selected to the sample, which will ultimately produce a sample which is representative of the population.

Using post stratification weights, a dataset which consists of 80% of the observations was drawn (from the full sample of 360 observations). To verify whether the newly drawn sample is representative of the target population, the chi squared goodness of fit test was performed for each of the variables “Year_in_university” and “Stream_in_university” to compare the proportions in the target population with the sample. The test revealed that there was no significant difference in the observed proportions for both the “Year_in_university” (p value = 0.15) and the “Stream_in_university” (p value = 0.055), thus indicating that the novel approach proposed yielded a sample which is representative of the population.

The reason that 80% was selected as the cut-off point was because this was the largest dataset which could be selected such that the sample was representative of the population. When a cut off greater than 80% was selected the sample did not match the population with respect to the stream and year of study.

Once this sample (which is a representative sample) was obtained, train and test sets were obtained from this dataset maintaining an 80:20 ratio.

The method proposed in this study can be important to the research community in the future as well, as machine learning models can be trained on representative samples using this method.

Encoding the variables

One-Hot encoding is appropriate for converting nominal categorical predictor variables into numerical data. However, ordinal encoding is appropriate for ordinal categorical variables as this preserves the order of the variable (25). Hence these two encoding schemes were applied for the relevant variables in the data.

Furthermore, it is useful to standardize ordinally encoded variables as this would assist in placing them on the same scale and prevent any one variable from dominating the model (25).

Thus, the continuous and discrete numerical variables together with the ordinally encoded variables were standardized and combined with the One-Hot encoded variables. Then, the machine learning models were fitted.

Part I

Prevalence of depression

Prevalence of depression is found to vary across regions with 2% of the population screening positive for MDD in some Asian countries and up to 21% in some European nations (26).

Undergraduates are known to have greater depressive symptoms with the highest rates being observed in some African regions with prevalence rates as high as 40.1% and lower-middle income countries with rates as high as 42.5%, considering studies conducted both before and after the corona virus pandemic (27). In Sri Lanka however, MDD was observed in 9.3% of the undergraduates in some state universities in 2015 (28). There is only one study conducted during the pandemic, where depression rates were found to be as high as 87.85% for the students in the radiology unit of the University of Peradeniya (30). This may be due to the Covid 19 pandemic.

This is the only study of this nature conducted among undergraduates in Colombo University after the pandemic. As shown in Fig. 1, it was observed that 29.7% of the respondents screened positive for depression.

Gender and depression

Figure 2 presents the distribution of depression for the two different genders of the respondents. The study has found that depression seems to be more prevalent in females than in males, with 34.2% of the females screening positive for MDD and only 23.9% of the males screening positive for MDD. However, the phi coefficient indicated that there was no significant relationship between the two variables (p value = 0.062).

Even from an international perspective, females are found to be more likely to develop depression in comparison to males. A study which reviewed the prevalence of depression across 29 countries found that in all countries, females were more likely to develop depressive symptoms, indicating that this is a gender issue rather than a cultural problem. (31)

Alcohol intake and depression

Figure 3 depicts how the level of depression in an undergraduate varies with the frequency of consuming alcohol. According to the figure, there does not seem to be any association between the 2 variables.

Since certain international studies which focused on the relationship between alcohol consumption of women and depression found a significant association between these two variables (32), in this study too this was considered. The percentage component bar chart corresponding to the alcohol intake of the female respondents of the study, with the level of depression, is given below in Fig. 4.

When considering the females respondents, it can be clearly seen that as the intake of alcohol increases, so does the percentage of those who have depression. As observed, only 31% of those who never consume alcohol screened positive for MDD while 85% of those who occasionally consume alcohol screened positive for MDD. Hence the pattern indicated by the female respondents seems to follow the pattern of undergraduate women in international universities, whereas such a relationship was not observed when both genders were considered together.

The rank-biserial correlation coefficient was used to check for any significant associations between the two variables mentioned above. When both the genders were considered, there was no significant association to be found (p value = 0.06). However, when only the female respondents were considered, the two variables were significantly associated (p value = 0.005). Thus, it seems that undergraduate females who deal with MDD have a higher tendency of consuming alcohol more frequently.

Year in university versus depression

Considering the year of study, as seen in Fig. 5, 2nd second year students seem to be screening positive for MDD more than students in the other years, indicating an association between the two variables, year of study and depression. Furthermore, the Chi Squared test of independence indicates that the relationship between the two variables is significant. (p value = 0.04).

The difference in depression levels observed for different academic years could be due to the level of stress experienced by the undergraduates. Studies have shown that exposure to stressful life events is a factor that is significantly associated with depression (33, 34).

In this study, second year students tend to have more stress, most likely due to the pressure of getting selected to an honours degree programme of their choice. This could be the reason for the above obtained result. To verify this, the relationship between the academic year and the level of stress associated with academic activities was investigated, and the results are shown below.

Academic year versus the stress associated with academic activities

Figure 6, confirms the reasoning given above, with 2nd year students having the highest level of stress associated with academic activities. However, the 4th year students also seem to have a high level of stress, possibly due to their final year research projects and striving to achieve a good class. As observed, more than 60% of both the 2nd year and 4th year students have stated that they experience high levels of stress when dealing with academic activities while less than 50% of the 1st and 3rd year students have indicated the same.

Furthermore, the Chi-squared test of independence indicates a significant relationship between the year of study of an undergraduate and the level of stress related to their academic work (p value = 0.045).

Satisfaction with academic achievements versus depression

As observed in Fig. 7, the percentage of depressed individuals seems to increase with dissatisfaction with their academic achievements, with only 17.6% of those who are satisfied with their academic achievements screening positive for MDD whereas 46.4% of those who are not satisfied with their academic achievements screened positive for MDD. Furthermore, the Goodman and Kruskal’s lambda indicated a significant relationship between the two variables (p value = 0.001)

While the results obtained in this study are not directly related to the academic performance of individuals, in general, those who are less satisfied with their academic achievements are more likely to have obtained poorer grades than those who are more satisfied with their academic achievements, indicating that students with depression tend to do more poorly in terms of academics when compared with those without depression.

The causes for this could be those discovered in international studies that students who are dealing with depression tend to have poor memory (35), difficulty concentrating (36) and cognitive impairment (37), which could all contribute to poor academic performance.

Income and depression

As observed in Fig. 8, those in the higher income brackets (family income of more than 250,000 rupees per month) show lower depressive symptoms in comparison to those whose family income is less than 250,000 rupees. This could be because the economic crisis is more likely to have more severely affected the poorer households.

However, a major difference in the prevalence of depression cannot be observed for the different categories of income. Furthermore, the association between the 2 variables was found to not be statistically significant using the rank biserial correlation coefficient (p value = 0.657)

While low levels of income are associated with certain life stresses which may lead to the development of depressive symptoms, high income earners also have their own problems to deal with such as long work hours (38), meeting deadlines and achieving goals which increase stress levels (39). Studies have found that undergraduates who come from such households have a lot of pressure put on them to overachieve. Furthermore, they tend to feel a sense of isolation from their parents (40). These factors may lead those from higher income families to report similar levels of depression as those from low-income families, and this may be the reason for not observing a significant relationship between income and the level of depression.

Effect of being harassed in university on depression

It is clear when observing Fig. 9 that there is an association between developing depression and being harassed in university, with 67.4% of those who have been harassed screening positive for MDD while only 27.9% of those who have not been harassed screened positive for MDD. The relationship between the 2 variables was found to be statistically significant, using the phi coefficient (p value = 0.002)

Literature reveals that those who have been ragged develop psychological problems, physiological problems and behavioural problems due to the stress associated with ragging. Apart from these negative effects, harassment of any form is found to be associated with an increased risk of developing depression (41). This seems to be the case in the current study as well as there is a large difference between the percentage of students who screened positive for MDD when comparing those who have faced some form of harassment in university with those who have not. Hence, being harassed in university either by a student or a staff member seems to have a relationship with the development of depressive symptoms. However, while these two variables were significantly associated, it was also found that only 4.6% of the respondents reported being ragged or harassed in university.

Effect of a breakup son depression

The breaking up of a relationship has an association with MDD as observed in Fig. 10, with 77.1% of those who have recently broken up screening positive for MDD while only 28.1% of those who have not recently broken up screened positive for MDD. Furthermore, the association between the level of depression and whether or not a respondent recently broke up is found to be statistically significant using the phi coefficient (p value = 0.001).

The reason why such a large difference was observed in the prevalence of depression when comparing those who had recently broken up with those who had not, may be due to the fact that many relationships which are formed in the university are among students who are at the age of marriage (The average age of marriage in Sri Lanka is 24.1 years for females and 27.3 years for males (42)) and hence, it is reasonable to assume that the relationships were formed in the hopes of getting married. These types of relationships lead to higher levels of depressive symptoms when they end (43).

The effect of breakups on developing depressive symptoms is more for females than for males, as seen below (See Fig. 11 (a) and 11 (b)). While 91.1% of the women who have recently broken up screened positive for MDD only 64.3% of the men who have gone through a breakup had MDD. However, among those who had a recent break up, gender and depression are not found to be significantly associated (p value = 0.151 for phi coefficient).

How being in a satisfying or unsatisfying relationship impacts depression

As observed in Figs. 12 (a) and 12 (b), 37.6% of those who are in a relationship which is not satisfying screened positive for MDD whereas only 21.0% of those who are in a relationship which is satisfying screened positive for MDD.

Negative relationships are seen to increase the prevalence of depression (44). This can be seen to be true in the current study as well by observing the prevalence rates of depression in those who are in satisfactory and unsatisfactory relationships. Furthermore, there is a significant association between the level of depression and whether or not someone is in a satisfying relationship (p value = 0.045 according to the phi coefficient). However, such a significance is not observed when checking whether or not a respondent is in an unsatisfactory relationship with the level of depression (p value = 0.617 for the phi coefficient).

Interestingly, 57.5% of female respondents who are not satisfied with their relationship screened positive for MDD whereas this was true for only 8.8% of the male respondents (Figs. 13 (a) and 13 (b)).

However, with the use of the phi coefficient, the association between gender and depression for those in an unsatisfying relationship was found to be not statistically significant (p value = 0.058)

Satisfaction with physical appearance and depression

Figure 14 shows that the percentage of those who screen positive for MDD tends to decrease as satisfaction with their physical appearance increases. This is found to be further validated through the rank biserial correlation coefficient which verified that the two variables were significantly associated (p value = 0)

Part II

Results of model fitting

Once the representative sample was derived from the full sample of 360 observations using the novel approach mentioned above, it left 288 observations. This, when separated into a training and testing set, left 230 training observations and 58 testing observations. The selected models were fitted on the training set using 5-fold cross validation, and the results of the model fitting process is given in Table 1.

Table 1

Machine learning models fitted on an imbalanced representative sample (Results) – Fitted on the training set
Machine Learning Model	Accuracy	Precision	Recall
K Nearest Neighbor Classifier	68%	51%	26%
Linear Discriminant Analysis	59%	34%	30%
Quadrating Discriminant Analysis	67%	20%	4%
Logistic Regression	60%	35%	25%
Decision Tree Classifier	60%	38%	37%
Support Vector Classifier	60%	34%	27%
Random Forest	68%	57%	19%
Gradient Boosting	61%	39%	32%
Light Gradient Boosting	67%	48%	30%
Extreme Gradient Boosting	66%	45%	23%
Gaussian Naive Bayes	41%	34%	86%
CatBoost	68%	50%	16%
Artificial Neural Networks	60%	35%	28%

Fixing the problem of class imbalance

It was observed that there is a class imbalance problem in the dataset, with 32% of undergraduates screening positive for MDD and 68% screening negative. Addressing this issue may yield better scores for the recall. Hence the SMOTE technique was used to fix the class imbalance problem.

However, applying SMOTE creates synthetic data, and hence, applying it to the training set and then using 5-fold cross validation on the same set would likely yield inflated results (Since the validation set would also then contain synthetic data). Thus, the training set was separated into 2, where 80% of the observations were considered when applying SMOTE, and the model was validated on the remaining 20% of observations. Table 2 represents the results of the model fitting process which is fitted on the data after having applied the SMOTE Technique.

Table 2

Machine learning models fitted on a balanced representative sample (Results)
Machine Learning Model	Accuracy	Precision	Recall
K Nearest Neighbor Classifier	48%	36%	63%
Linear Discriminant Analysis	57%	33%	25%
Quadrating Discriminant Analysis	66%	nan	0%
Logistic Regression	60%	42%	31%
Decision Tree Classifier	58%	29%	13%
Support Vector Classifier	61%	42%	31%
Random Forest	66%	50%	13%
Gradient Boosting	70%	60%	38%
Light Gradient Boosting	63%	44%	25%
Extreme Gradient Boosting	61%	38%	19%
Gaussian Naive Bayes	52%	38%	56%
CatBoost	63%	44%	25%
Artificial Neural Networks	63%	46%	38%

Final model

As observed, the highest accuracy as well as precision was obtained for the Gradient Boosting Classifier after fixing the problem of the class imbalance. Hence, this is the model that was used for further analysis.

Feature Selection

In total there were 113 independent variables after encoding. Variables were dropped one by one in the model fitting process, starting from the least important variable to the most important variable, as identified by the variable importance plot. At each stage, the model was fitted and evaluated on the test set to identify how many variables should be kept to obtain the best model. It was found that the best accuracy, sensitivity and F1 score are obtained after removing the 67 least important variables, leaving the model to work with 46 variables. The variable importance plot of the most important variables in fitting the gradient boosting classifier is shown below in Fig. 15.

Hyperparameter Tuning

The hyperparameters were tuned to maximize the recall as this is the most important score to consider in the model being built. This is because the recall would give the percentage of those with depression who were identified correctly, and it is of utmost importance that an individual who has a potential of developing depression is identified and directed to properly diagnose themselves. Identifying an individual without depression as having a risk of developing symptoms is not a major problem, as once they are directed to the relevant healthcare professional, they can get themselves evaluated to know whether they are at risk or not. However, if someone who is at risk is identified as not being at risk, then he/she will not seek professional help. Thus, the model to be developed will focus on maximizing the recall scores. Table 3 depicts the best hyperparameters to be used in fitting the gradient boosting classifier so as to obtain the maximum possible recall.

Table 3

Best hyperparameters of the gradient boosting classifier
Hyperparameter	Values that were tried	Best value for the hyperparameter
learning_rate	0.1, 0.05, 0.01	0.05
n_estimators	50, 100, 250	50
max_depth	2, 3, 4	2
min_samples_split	2, 4, 8	2
min_samples_leaf	1, 2, 4	1
subsample	0.5, 0.8, 1.0	0.8
max_features	None, sqrt, log2	None

Once the optimal model was selected, after having tuned the hyperparameters, the model was evaluated on both the train and the test set. The scores obtained are shown in Table 4.

Table 4

Performance metrics of both train and test sets in the tuned model
Score	Training Set Values	Test Set Values
Accuracy	87%	79%
Precision	87%	78%
Recall	88%	72%
Specificity	87%	85%
F1 score	87%	75%

The model seems to perform quite well on both the train and the test set.

Comparison with international studies

The model scores obtained in this research are comparable with international studies which have tried to predict depression using socio-demographic factors.

A study conducted to predict depression in U.S. adults with hypertension found a model with 77% accuracy. (45)
A study which focused on children developed a model that can predict depression with an accuracy of 82% (46)
A study conducted on citizens of Bangladesh from different age groups developed a model which had an impressive accuracy of 93% in making predictions. (47)

As observed, the model developed in this research is comparable with the models spoken of in international studies.

This research is subject to the following two limitations:

The PHQ-9 questionnaire categorizes the subjects into 5 different classes (None, Mild, Moderate, Moderately Severe and Severe depression). However only two classes have been validated. (None and Mild as not having Major Depressive Disorder and Moderate, Moderately Severe and Severe Depression as having Major Depressive Disorder). This validation has been done internationally as well as for Sri Lanka. As a result, there are only two classes which the model can predict (whether a student is at risk of developing depressive symptoms or not).

Importance and Relevance of the study

This study is a novel contribution to the field of machine learning and mental health, specifically with regard to predicting depression among undergraduates in Sri Lanka. The use of post-stratification weights in the fitting of classification models is a unique approach that has not been explored in previous studies. Additionally, this study is the first to examine the feasibility of using machine learning models to predict depression in the University of Colombo, filling a significant gap in the current literature.

The results of this study have important implications for both academic research and practical applications. On one hand, the findings will contribute to the existing body of knowledge in the fields of machine learning and mental health by demonstrating the potential of using post-stratification weights in the prediction of depression. On the other hand, the results of this study will be of great use to mental health professionals, educators, and policy-makers in Sri Lanka, providing them with valuable insights into the predictors of depression among undergraduates.

Moreover, the successful implementation of the machine learning models in this study highlights the potential for similar models to be used in other settings and with other populations. The results of this study will serve as a starting point for further research and development of machine learning models for the prediction of depression, both in Sri Lanka and internationally.

Main Conclusions

This research mainly aimed at identifying factors that contribute towards depression in undergraduates of the faculty under consideration and developing a machine learning model (The Gradient Boosting Classifier) to predict the depression status of these students. In line with these objectives the following conclusions can be drawn:

Considering female respondents, a significant association is seen between the level of depression and the consumption of alcohol. However, such an association is not observed when both the genders are considered together. This association would however require further study as the alcohol consumption among females in Sri Lanka depends on the family background and social status.

Considering university life, satisfaction with academic achievements, having faced harassment and/or ragging in university as well as the academic year are all found to be variables which are significantly associated with MDD.

Those who had not experienced a recent break-up, are in a satisfactory relationship and are satisfied with their own physical appearance show significantly lower rates of depression in comparison to their opposing counterparts.

The Gradient Boosting Classifier which had an accuracy of 79% on the test set was found to be the most suitable model in predicting the depression status of an undergraduate.

MDD - Major Depressive Disorder

PHQ-9 – Patient Health Questionnaire 9

Ethics approval and consent to participate

Ethical approval was obtained from the Institute of Biology, Sri Lanka. Registration Number – ERC IOBSL 278 10 2022.

The data of the participants was collected through a google form, and as such, the first question of the google form was on obtaining the consent of the respondent to use their data for the study.

Consent for publication

Not applicable – All the data was compiled and analyzed as a group. Data that would lead to student identity was not obtained.

Availability of data and materials

The datasets generated and/or analyzed during the current study are not publicly available due to the sensitive nature of the data collected but are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests.

Funding

This study received no funding.

Authors' contributions

MDT provided guidance on conducting the survey and collecting a representative sample whereas SD provided guidance on fitting machine learning models and coming up with a novel approach. DS conducted the entire study from collecting and analyzing the data to developing the model and writing the manuscript. All authors read and approved the final manuscript.

Acknowledgements

Not applicable

Fang H, Tu S, Sheng J, Shao A. Depression in sleep disturbance: A review on a bidirectional relationship, mechanisms and treatment. J Cell Mol Med [Internet]. 2019 Apr 7;23(4):2324–32. Available from: https://onlinelibrary.wiley.com/doi/10.1111/jcmm.14170
Targum SD, Fava M. Fatigue as a residual symptom of depression. Innov Clin Neurosci [Internet]. 2011 Oct;8(10):40–3. Available from: http://www.ncbi.nlm.nih.gov/pubmed/22132370
Sutin AR, Zonderman AB. Depressive symptoms are associated with weight gain among women. Psychol Med [Internet]. 2012 Nov;42(11):2351–60. Available from: http://www.ncbi.nlm.nih.gov/pubmed/22475128
Robertson D, Kumbhare D, Nolet P, Srbely J, Newton G. Associations between low back pain and depression and somatization in a Canadian emerging adult population. J Can Chiropr Assoc [Internet]. 2017 Aug;61(2):96–105. Available from: http://www.ncbi.nlm.nih.gov/pubmed/28928493
Lampl C, Thomas H, Tassorelli C, Katsarava Z, Laínez JM, Lantéri-Minet M, et al. Headache, depression and anxiety: associations in the Eurolight project. J Headache Pain [Internet]. 2016 Dec 1;17(1):59. Available from: https://thejournalofheadacheandpain.biomedcentral.com/articles/10.1186/s10194-016-0649-2
Izadinia N, Amiri M, Jahromi R ghorban, Hamidi S. A study of relationship between suicidal ideas, depression, anxiety, resiliency, daily stresses and mental health among Tehran university students. Procedia - Soc Behav Sci [Internet]. 2010;5:1615–9. Available from: https://linkinghub.elsevier.com/retrieve/pii/S187704281001709X
Belvederi Murri M, Ekkekakis P, Magagnoli M, Zampogna D, Cattedra S, Capobianco L, et al. Physical Exercise in Major Depression: Reducing the Mortality Gap While Improving Clinical Outcomes. Front psychiatry [Internet]. 2018;9:762. Available from: http://www.ncbi.nlm.nih.gov/pubmed/30687141
Carta M, Petretto D, Adamo S, Bhat K, Lecca M, Mura G, et al. Counseling in primary care improves depression and quality of life. Clin Pract Epidemiol Ment Health [Internet]. 2012;8:152–7. Available from: http://www.ncbi.nlm.nih.gov/pubmed/23173011
Avasthi A, Grover S, Aggarwal M. Research on antidepressants in India. Indian J Psychiatry [Internet]. 2010 Jan;52(Suppl 1):S341-54. Available from: http://www.ncbi.nlm.nih.gov/pubmed/21836704
Cuijpers P, Quero S, Dowrick C, Arroll B. Psychological Treatment of Depression in Primary Care: Recent Developments. Curr Psychiatry Rep [Internet]. 2019 Dec 23;21(12):129. Available from: http://link.springer.com/10.1007/s11920-019-1117-x
Duval F, Lebowitz BD, Macher J-P. Treatments in depression. Dialogues Clin Neurosci. 2006 Jun 30;8(2):191–206.
Yoshikawa E, Taniguchi T, Nakamura-Taira N, Ishiguro S, Matsumura H. Factors associated with unwillingness to seek professional help for depression: a web-based survey. BMC Res Notes [Internet]. 2017 Dec 4;10(1):673. Available from: http://www.ncbi.nlm.nih.gov/pubmed/29202791
Müller A. Mental health disorders: prevalent but widely ignored in academia? J Physiol [Internet]. 2020 Apr;598(7):1279–81. Available from: https://onlinelibrary.wiley.com/doi/10.1113/JP279386
Bin MA, Ranizal M, Binti WQ, Zabidi W, Aziera S, Mohd B, et al. Depression Among University Students [Internet]. Available from: www.ejoms.com
Awadalla S, Davies EB, Glazebrook C. A longitudinal cohort study to explore the relationship between depression, anxiety and academic performance among Emirati university students. BMC Psychiatry. 2020 Sep 11;20(1).
Geisner IM, Mallett K, Kilmer JR. An examination of depressive symptoms and drinking patterns in first year college students. Issues Ment Health Nurs. 2012 May;33(5):280–7.
Brådvik L. Suicide risk and mental disorders. Vol. 15, International Journal of Environmental Research and Public Health. MDPI AG; 2018.
Jayasekara JGARS. Community nurses: An urgent need. Nurs Heal Sci. 2001 Jun;3(2):101–4.
Ball HA, Siribaddana SH, Kovas Y, Glozier N, McGuffin P, Sumathipala A, et al. Epidemiology and symptomatology of depression in Sri Lanka: a cross-sectional population-based survey in Colombo District. J Affect Disord [Internet]. 2010 Jun;123(1–3):188–96. Available from: http://www.ncbi.nlm.nih.gov/pubmed/19762085
Kroenke K, Spitzer RL. The PHQ-9: A New Depression Diagnostic and Severity Measure. Psychiatr Ann. 2002 Sep;32(9):509–15.
Hanwella R, Ekanayake S, De Silva VA. The validity and reliability of the sinhala translation of the patient health questionnaire (PHQ-9) and PHQ-2 screener. Depress Res Treat. 2014;2014.
rono L. MICROCREDIT AND ITS RELATIONSHIP TO THE GROWTH OF SMALL AND MEDIUM ENTERPRISES IN KONOIN SUBCOUNTY, KENYA. Int J Adv Res. 2018 Apr 30;6(4):961–8.
Sinclair M, O’Toole J, Malawaraarachchi M, Leder K. Comparison of response rates and cost-effectiveness for a community-based survey: postal, internet and telephone modes with generic or personalised recruitment approaches. BMC Med Res Methodol [Internet]. 2012 Dec 31;12(1):132. Available from: https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-12-132
Wayne I. American Association for Public Opinion Research Nonresponse, Sample Size, and the Allocation of Resources. Vol. 39, Source: The Public Opinion Quarterly. 1975.
Raschka S. Python Machine Learning, 1st Edition. Packt Publishing; 2015. 454 pages.
Gutiérrez-Rojas L, Porras-Segovia A, Dunne H, Andrade-González N, Cervilla JA. Prevalence and correlates of major depressive disorder: a systematic review. Rev Bras Psiquiatr [Internet]. 2020;42(6):657–72. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32756809
Li W, Zhao Z, Chen D, Peng Y, Lu Z. Prevalence and associated factors of depression and anxiety symptoms among college students: a systematic review and meta‐analysis. J Child Psychol Psychiatry [Internet]. 2022 Nov 16;63(11):1222–30. Available from: https://onlinelibrary.wiley.com/doi/10.1111/jcpp.13606
Amarasuriya SD, Jorm AF, Reavley NJ. Prevalence of depression and its correlates among undergraduates in Sri Lanka. Asian J Psychiatr. 2015 Jun 1;15:32–7.
Weerakkodi R, Gamagedara T, Nawarathna L. Prevalence of Depression and Associated Factors among Final Year Undergraduates in Three Health-Related Faculties of University of Peradeniya. 12th Int Res Conf - Gen SIR JOHN KOTELAWALA Def Univ. 2019;236.
Weerakoon BS, Chandrasiri NR. Anxiety, depression, and stress among radiography undergraduates during the COVID-19 pandemic. J Med Imaging Radiat Sci [Internet]. 2022 Dec; Available from: https://linkinghub.elsevier.com/retrieve/pii/S1939865422006737
Hopcroft RL, Bradley DB. The sex difference in depression across 29 countries. Soc Forces. 2007;85(4):1483–507.
Mushquash AR, Stewart SH, Sherry SB, Sherry DL, Mushquash CJ, MacKinnon AL. Depressive symptoms are a vulnerability factor for heavy episodic drinking: A short-term, four-wave longitudinal study of undergraduate women. Addict Behav [Internet]. 2013 May;38(5):2180–6. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0306460312003814
Clay DL, Anderson WP, Dixon WA. Relationship Between Anger Expression and Stress in Predicting Depression. J Couns Dev [Internet]. 1993 Sep 10;72(1):91–4. Available from: https://onlinelibrary.wiley.com/doi/10.1002/j.1556-6676.1993.tb02283.x
Kuiper NA, Olinger LJ, Lyons LM. Global Perceived Stress Level as a Moderator of the Relationship between Negative Life Events and Depression. J Human Stress [Internet]. 1986 Dec;12(4):149–53. Available from: http://www.tandfonline.com/doi/abs/10.1080/0097840X.1986.9936781
Dillon DG, Pizzagalli DA. Mechanisms of Memory Disruption in Depression. Trends Neurosci [Internet]. 2018 Mar;41(3):137–49. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0166223617302503
Andrews PW, Thomson JA. The bright side of being blue: Depression as an adaptation for analyzing complex problems. Psychol Rev [Internet]. 2009 Jul;116(3):620–54. Available from: http://doi.apa.org/getdoi.cfm?doi=10.1037/a0016242
Perini G, Cotta Ramusino M, Sinforiani E, Bernini S, Petrachi R, Costa A. Cognitive impairment in depression: recent advances and novel treatments. Neuropsychiatr Dis Treat [Internet]. 2019 May;Volume 15:1249–58. Available from: https://www.dovepress.com/cognitive-impairment-in-depression-recent-advances-and-novel-treatment-peer-reviewed-article-NDT
Smeets P, Whillans A, Bekkers R, Norton MI. Time Use and Happiness of Millionaires: Evidence From the Netherlands. Soc Psychol Personal Sci. 2020;11(3):295–307.
Damaske S, Zawadzki MJ, Smyth JM. Stress at work: Differential experiences of high versus low SES workers. Soc Sci Med [Internet]. 2016 May;156:125–33. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0277953616301101
Luthar SS, Latendresse SJ. Children of the Affluent. Curr Dir Psychol Sci [Internet]. 2005 Feb 23;14(1):49–53. Available from: http://journals.sagepub.com/doi/10.1111/j.0963-7214.2005.00333.x
Houle JN, Staff J, Mortimer JT, Uggen C, Blackstone A. The Impact of Sexual Harassment on Depressive Symptoms during the Early Occupational Career. Soc Ment Health [Internet]. 2011 Jul 31;1(2):89–105. Available from: http://journals.sagepub.com/doi/10.1177/2156869311416827
Premawardhena AP, De Silva ST, Goonatilleke MDDC, Ediriweera DS, Mettananda S, Rodrigo BKRP, et al. Marriage patterns in Sri Lanka and the prevalence of parental consanguinity in patients with β-thalassaemia: a cross-sectional descriptive analysis. J Biosoc Sci [Internet]. 2020 Jul 25;52(4):573–84. Available from: https://www.cambridge.org/core/product/identifier/S0021932019000658/type/journal_article
Rhoades GK, Kamp Dush CM, Atkins DC, Stanley SM, Markman HJ. Breaking up is hard to do: The impact of unmarried relationship dissolution on mental health and life satisfaction. J Fam Psychol [Internet]. 2011;25(3):366–74. Available from: http://doi.apa.org/getdoi.cfm?doi=10.1037/a0023627
Whisman MA, Bruce ML. Marital dissatisfaction and incidence of major depressive episode in a community sample. J Abnorm Psychol [Internet]. 1999 Nov;108(4):674–8. Available from: http://doi.apa.org/getdoi.cfm?doi=10.1037/0021-843X.108.4.674
Lee C, Kim H. Machine learning-based predictive modeling of depression in hypertensive populations. Cymbalyuk GS, editor. PLoS One [Internet]. 2022 Jul 29;17(7):e0272330. Available from: https://dx.plos.org/10.1371/journal.pone.0272330
Pang G, Pham NTA, Baker E, Bentley R, Hengel A van den. Deep Multi-task Learning for Depression Detection and Prediction in Longitudinal Data. 2020;(Mdd). Available from: http://arxiv.org/abs/2012.02950
Zulfiker MS, Kabir N, Biswas AA, Nazneen T, Uddin MS. An in-depth analysis of machine learning approaches to predict depression. Curr Res Behav Sci [Internet]. 2021 Nov;2:100044. Available from: https://linkinghub.elsevier.com/retrieve/pii/S2666518221000310

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

A machine learning approach for predicting the depression status among undergraduates: A case study from Sri Lanka

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Introduction

Methods

Results

Part I

Part II

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1