Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Predicting Stress, Anxiety and Depression Among the

University Students of India Post-Covid
Shikha Verma Aman Goyal Santushti Gandhi
Department of Computer Science Department of Computer Science Department of Computer Science
Ram Lal Anand College, Ram Lal Anand College, Ram Lal Anand College,
University of Delhi University of Delhi University of Delhi
New Delhi, India New Delhi, India New Delhi, India

Abstract:- Objective: The study aimed to assess the significant implications for both the student body and the
reliability and validity of the Psychological factors, administration of the university. By identifying the variables
namely Depression, Anxiety, and Stress Scale-21 (DASS- causing stress, anxiety, and depression among students can
21), among University students in Delhi. give us crucial insights into building tailored interventions
Methods: The DASS-21 questionnaire was administered and support systems. These initiatives could include
by conducting a survey where around 100 samples were enhancing counselling services, assisting with efforts to
randomly selected. A comparison and training model raise mental health awareness, and developing a friendlier
was formed using the benchmark dataset along with the and stress-resistant school climate.
original data collected in the study. Three supervised
machine learning models were trained on the same. The This study concludes by issuing an urgent appeal for
best model was selected and tested on the originally reform at the University of Delhi and beyond. We can
collected data. cultivate a more compassionate and empowered generation
Results: A training accuracy of 100%, 98%, and 98% of people who are better able to negotiate the hurdles of both
were achieved for the Stress, Anxiety and Depression academics and life by understanding the struggles students
scale, respectively, using the Random Forest algorithm. have with their mental health and taking proactive measures
The models were cross-validated using 10-fold cross to offer the necessary support.
validation. A cross-validation score of 99%, 98% and
99% were achieved for the scales. I. INTRODUCTION
Conclusion: The factors selected using machine learning
techniques affect an individual's severity. To further Depression, stress, and anxiety are common emotions
verify these factors, practitioners were engaged to that arise from life's challenges. Depression is characterized
identify the specific features that influence these by low mood and loss of interest, while stress results from
psychological parameters. These results helped to feeling overwhelmed due to pressure. Excessive stress can
understand the importance and use of machine learning lead to distress and exhaustion. Anxiety is the feeling that
techniques for analyzing the severity of stress, anxiety something terrible is about to happen and can be general or
and depression scales amongst individuals. The testing specific to a situation, place, or thing. Understanding these
accuracy achieved was similar to the training accuracy distinctions and seeking help when necessary is essential to
indicating the model did not have any anomaly and ensure timely and effective treatment. Early detection is
could be used for predicting the severity of stress, critical, and prevention is better than cure. [1].
anxiety and depression among university students in
India. The Depression, Anxiety, Stress Scales, commonly
known as DASS, is a scale which consists of self-report
Keywords:- Stress, Anxiety, Depression, Correlation, items. DASS-21 is a questionnaire which consists of 21
Supervised Machine Learning, Testing Accuracy. questions that the user must answer. The questionnaire is
divided into three scales Depression, Anxiety and Stress,
PUBLIC SIGNIFICANCE STATEMENT and it consists of equal numbers of questions for each scale.
The score of each scale is added, which is then used to
Mental Health is an essential aspect whose importance determine the severity of Depression, Anxiety and Stress of
in the context of higher education cannot be understated. an individual [2].
This study on the students of the University of Delhi throws
light on a problem that is becoming common each passing Machine Learning is a subfield of Artificial
day. As per the National Crime Records Bureau’s (NCRB) Intelligence defined as a machine's capability to imitate
Accidental Deaths & Suicides in India (ADSI) report, over human behaviour and predict unforeseen events. It plays a
13,000 students died in 2021 in India at the rate of more pivotal role in the field of Medicine, Sales and Operations,
than 35 every day, a rise of 4.5 percent from the 12,526 Customer Service and many more. Machine Learning is of
deaths in 2020 with 864 out of 10,732 suicides being due to great help in the field of psychology. It allows psychologists
“failure in examination.” The findings of this study have to organize data, describe data, and make inferences based

IJISRT23AUG2134 www.ijisrt.com 2244

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
on the data [3]. All these steps will enable a psychologist to questions from the DASS-42 questionnaire, and Srinath et
understand the patients better and help in developing better al. [11] used an online questionnaire to collect data from
tests and practices to help the patients. In the following 30,776 participants between 2017 and 2019. Lastly, Chiong
sections, we will look at the work done in psychology using et al. [12] aimed to detect the severity of depression using
machine learning. pre-processed social media text data from Twitter, and Sun
et al. [13] evaluated depression levels using DASS-42.
II. RELATED WORK
III. PROPOSED WORK
Shayan et al. [4] validated the DASS-42 questionnaire
in the Dari region of Afghanistan, while J.A. Ademuyiwa This study focused on determining the severity level of
and A. A. Adetunji [5] surveyed 200 staff members at the stress, anxiety and depression scale using the Depression,
Federal Polytechnic in Ile-Oluji, Nigeria. Priya et al. [6] Anxiety and Stress Scale questionnaire (DASS-21). The
used various machine learning algorithms to determine the training data was collected from online sources, and the
best one for predicting psychological problems, and Al- testing data was collected from 100 participants via Google
Wesabi et al. [7] used the DASS-21 questionnaire to Forms. Three machine learning algorithms – namely
evaluate stress, anxiety, and depression severity during Gaussian Naïve Bayes, Logistic Regression and Random
COVID-19. Mary and Jabasheela [8] used the DASS-21 Forest were used for the classification.
questionnaire to evaluate different Machine Learning
techniques for classification in a study involving 600 A. DASS-21 Survey
students from Puducherry, India. Marouf et al. [9] examined The data for this study was collected through the
the relationship between the Big Five Personality traits and DASS-21 questionnaire. The questionnaire comprises 21
perceived stress in Bangladeshi computer science questions. Each scale of Stress, Anxiety and Depression is
undergraduate students using machine learning techniques. allocated seven questions. The options available for each
Singh [10] analyzed a dataset of 42,000 instances acquired question in the DASS questionnaire [14] are shown in
from an online source containing 170 features with 42 Figure 1 below:

Fig 1. Values corresponding to the DASS-21 questionnaire.

The questions asked in the questionnaire [14] are described in Figure 2.

Fig 2. The questions for the DASS-21.

After the data collection, the participant's responses to each Stress, Anxiety and Depression scale were grouped separately.
They were then calculated by simply adding up the values associated with each question using the below formula [15]:

𝑠𝑐𝑜𝑟𝑒 = 𝑆𝑢𝑚 𝑜𝑓 𝑟𝑎𝑡𝑖𝑛𝑔 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑒𝑎𝑐ℎ 𝑐𝑙𝑎𝑠𝑠 ∗ 2 (1)

IJISRT23AUG2134 www.ijisrt.com 2245

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
B. Score Parameters
The condensed form of the original questionnaire is called the DASS-21. The dataset used was a benchmark dataset. The
original dataset consists of DASS-21 questions and has data about an individual's personal information such as age, gender,
ethnicity, etc. The severity of stress, anxiety and depression is categorized into five categories and is presented in the figure below
[14].

Fig 3. The score range for the DASS-21 subscales

C. Feature Selection
The original dataset consisted of various features pivotal to the study, such as gender, age etc. However, some attributes, such
as the screen size of the user's device and which hand is dominant, were deemed non-essential. So, to find this, a correlation
between the attributes and the target column was found. The questions about stress, anxiety and depression were stored in
different dictionaries. The severity of Stress, Anxiety and Depression amongst the individuals was calculated using formula 1. We
created three columns to hold the stress, anxiety and depression scores achieved by an individual based on the Taylor Manifest
Anxiety Scale. We categorized them based on the ranges mentioned in Figure 3. Except for the target column, the data was
standardized using standardscaler().

D. The Framework of the Proposed Work

The primary aim was to determine the severity of an individual's stress, anxiety and depression. The initial step was creating
different models for stress, anxiety and depression. Firstly, the null records were dropped, and the count of each severity for the
scales was plotted. This helped us understand the balance among the various severity levels, as shown below in Figure 4.

Fig 4. Count plot of the severity of the Stress, Depression and Anxiety scale

From the above graphs, we can infer that the depression and anxiety scale data was imbalanced. The data included cases
where most samples had highly severe depression and anxiety. The data was split into training and testing data in a ratio of 75:25.
In the next stage, Gaussian Naïve Bayes, Logistic Regression and Random Forest were applied to the training data to help us
determine the severity. The performance of the algorithms was tested based on different metrics explained in the following
section. The framework of the model is shown below in Figure 5.

IJISRT23AUG2134 www.ijisrt.com 2246

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

.
Fig 5. The framework of the model

E. Hyperparameter Tuning for the Model

Hyperparameter Tuning is a concept where the existing machine learning models are improved using a set of optimal
hyperparameter values. Combining hyperparameters helps maximize the model's performance and minimize the loss to produce
better results with fewer errors [16]. The technique used in this study was GridSearchCV which helped to find the optimal
hyperparameters for all the algorithms shown in Table 1 below.

Table 1. The value of hyperparameters for the different models

Hyperparameter Models Value
var_smoothening Gaussian Naïve Bayes 1e-06
multi_class Logistic Regression multinomial
n_jobs Logistic Regression -1
criterion Random Forest entropy
max_depth Random Forest 10
n_estimators Random Forest 100
random_state Random Forest 100

IV. RESULTS AND DISCUSSION

The quality of a machine learning model and its ability to generalize to new data is assessed with the help of performance
metrics. The study evaluated the three supervised machine learning algorithms using four metrics, Accuracy, Cross-Validation,
Precision, and Matthews Correlation Coefficient (MCC). The cross-validation score for each algorithm was determined for each
scale separately, and the best model was selected for each scale based on these metrics. The results for different metrics for all the
subscales are shown below in Table 2.

IJISRT23AUG2134 www.ijisrt.com 2247

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Table 2. Result for different metrics for DASS-21 subscales
Algorithm Scale Accuracy Cross-Validation Precision MCC
Stress 0.9 0.89 0.98 0.87
Gaussian Naïve
Anxiety 0.98 0.97 0.97 0.97
Bayes
Depression 0.82 0.81 0.96 0.77
Stress 0.93 0.91 0.98 0.91
Logistic Regression Anxiety 0.91 0.92 0.9 0.87
Depression 0.93 0.91 0.96 0.9
Stress 1.0 0.99 1.0 0.97
Random Forest Anxiety 0.98 0.98 0.96 0.98
Depression 0.98 0.99 0.96 0.97

As inferred from the above observations, Random Forest was the best-performing algorithm for all the scales based on all the
metrics.

Table 3. Results for DASS-21 subscales for validation data.

Algorithm Anxiety Stress Depression
Gaussian Naïve Bayes 0.989 0.967 0.989
Logistic Regression 0.733 0.8 0.733
Random Forest 0.989 1.0 0.967

Random Forest had the best testing accuracy for the scales, and Gaussian Naive Bayes had the same accuracy for Anxiety.
Similar testing and training accuracy indicate no over- or under-fitting, making models perform similarly on new and trained data.
Random Forest outperformed others as it is an ensemble of Decision Trees, using a subset of the data to make different
predictions. Majority voting is used to make final predictions based on the output of each tree.

V. CASE STUDY

We analyzed our machine learning model's performance by engaging practitioners in studying the patterns observed. The
practitioners were given access to all the features present in the dataset and were asked to identify which features were linked to
an individual's levels of stress, anxiety, and depression. The factors selected by the practitioners for the depression scale are shown
below in Table 4.

Table 4. Stress, Anxiety and Depression features selected by the practitioner

Stress Anxiety Depression
Finding hard to wind down Being aware of dryness of mouth Using a lot of negative energy
Worried about situations where I Experiencing breathing difficulty Not experience any positive feelings at all
might panic and make a fool of myself
Getting agitated Over-reacting to situations Feeling that there is nothing to look forward
to
Feeling close to panic Feel like trembling Feeling downhearted and blue
Intolerant of things that kept me from Aware of the action of my heart in the Difficult to work up the initiative to do
getting on with what I was doing absence of physical exertion things
Scared without any good Difficulty in relaxing No enthusiasm about anything
reason
Feeling that I was rather touchy Feeling that I am not worth much as a
person
Feeling that life wasn't worthwhile.

IJISRT23AUG2134 www.ijisrt.com 2248

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The above table shows a pattern in the factors selected REFERENCES
by the practitioners for DASS-21 subscales. The
practitioners' selection of factors for each scale showed the [1]. "Anxiety, Depression, Stress: Why the Differences
following patterns : Matter | Premier Health."
● Physical characteristics are more related to anxiety due https://www.premierhealth.com/your-
to the release of cortisol during the body's "fight or health/articles/women-wisdom-wellness-
flight" response [17]. /anxietydepression-stress-why-the-differences-matter
● Self-esteem and self-confidence were more related to (accessed Sep. 20, 2022).
depression, while behaviour-based parameters were [2]. "Overview of the DASS and its uses."
more related to stress. http://www2.psy.unsw.edu.au/dass/over.htm (accessed
● These factors can also be interconnected, such as fear Sep. 20, 2022).
causing restlessness and hindering activity and focus. [3]. K. Cherry, "Why Are Statistics Necessary in
● These findings suggest a complex relationship exists Psychology?," p. 2.
between physical, psychological, and behavioural factors [4]. N. Shayan, A. Niazi, A. WASEQ, and H. Ozcebe,
in the development of anxiety, depression, and stress. “Depression, Anxiety, and Stress Scales 42 (DASS-
42) in Dari-Language: Validity and Reliability Study in
Further research can explore the interplay between Adults, Herat, Afghanistan,” Bezmialem Science, vol.
these factors and their impact on mental health outcomes. 9, pp. 356–362, Jun. 2021, doi:
10.14235/bas.galenos.2020.4250.
VI. CONCLUSION & FUTURE WORK [5]. A. A. Adetunji and J. A. Ademuyiwa, "Assessing
DASS-42 Models among Polytechnic Staff," Open
The study used machine learning algorithms to identify Access Library Journal, vol. 6, no. 10, Art. no. 10, Oct.
severity levels of anxiety, stress, and depression using the 2019, doi: 10.4236/oalib.1105334.
DASS-21 questionnaire. Three supervised ML techniques [6]. A. Priya, S. Garg, and N. Tigga, "Predicting Anxiety,
were employed, with Random Forest showing the best Depression and Stress in Modern Life using Machine
performance for DASS-21 subscales' training data. Testing Learning Algorithms," Procedia Computer Science,
accuracy of 0.75, 1.0, and 0.995 was achieved for stress, vol. 167, pp. 1258–1267, Jan. 2020, doi:
anxiety, and depression, respectively. Random Forest and 10.1016/j.procs.2020.03.442.
Gaussian Naïve Bayes performed best for the testing data, [7]. F. N. Al-Wesabi, H. Alsolai, A. M. Hilal, M. A. Hamza,
while Logistic Regression had a significant drop in M. Al Duhayyim, and N. Negm, "Machine Learning
accuracy. Metrics such as precision, Matthew correlation Based Depression, Anxiety, and Stress Predictive
coefficient, and cross-validation were used to evaluate Model During COVID-19 Crisis," CMC-Comput. Mat.
model performance due to the data's imbalanced nature. Contin., pp. 5803–5820, 2022.
[8]. S. T. A. Mary and L. Jabasheela, "An Evaluation of
The primary objective of our paper was to investigate Classification Techniques for Depression, Anxiety and
the levels of stress, anxiety, and depression among students. Stress Assessment," presented at the International
Upon analyzing the survey results, we noted that the Conference for Phoenixes on Emerging Current Trends
participants fell within the age range of 18-22. Expanding in Engineering and Management (PECTEAM 2018),
our study to include various age groups would allow for a Atlantis Press, Feb. 2018, pp. 64–69. doi:
comparative analysis of the prevalence and severity of 10.2991/pecteam-18.2018.13.
stress, anxiety, and depression across different populations. [9]. A. Marouf, A. Ashrafi, T. Ahmed, and T. Emon, "A
This would provide valuable insights into the role of age Machine Learning based Approach for Mapping
concerning stress, anxiety, and depression, as well as the Personality Traits and Perceived Stress Scale of
variations in the frequency of these conditions among Undergraduate Students," International Journal of
different age groups based on their responsibilities. Modern Education and Computer Science, vol. 11, pp.
42–47, Aug. 2019, doi: 10.5815/ijmecs.2019.08.05.
To verify these factors affecting the psychological [10]. T. Singh, "Prediction of Depression, Stress and
parameters, we sought the input of practitioners who Anxiety," International Journal of Research in
identified the features that impacted these parameters. The Engineering, Science and Management, vol. 5, no. 4,
DASS-21 questionnaire reported in this study is a valid and Art. no. 4, May 2022.
reliable tool for assessing the severity of stress, anxiety and [11]. K. S. Srinath, K. Kiran, S. Pranavi, M. Amrutha, P. D.
depression amongst university students in Delhi. This Shenoy, and K. R. Venugopal, "Prediction of
questionnaire can serve as a baseline for a more Depression, Anxiety and Stress Levels Using Dass-
comprehensive validation analysis of the magnitude of 42," in 2022 IEEE 7th International conference for
stress, anxiety and depression amongst university students in Convergence in Technology (I2CT), Apr. 2022, pp. 1–
India and help analyze the reason behind the development of 6. doi: 10.1109/I2CT54291.2022.9824087.
stress, anxiety and depression. [12]. R. Chiong, G. S. Budhi, S. Dhakal, and F. Chiong, "A
textual-based featuring approach for depression
detection using machine learning classifiers and social
media texts," Computers in Biology and Medicine, vol.

IJISRT23AUG2134 www.ijisrt.com 2249

Volume 8, Issue 8, August – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
135, p. 104499, Aug. 2021, doi:
10.1016/j.compbiomed.2021.104499.
[13]. Y. H. Sun, Q. Liu, N. Y. Lee, X. Li, and K. Lee, "A
novel machine learning approach to shorten depression
risk assessment for convenient uses," Journal of
Affective Disorders, vol. 312, pp. 275–291, Sep. 2022,
doi: 10.1016/j.jad.2022.06.035.
[14]. "Depression Anxiety Stress Scales (DASS)."
https://www.psytoolkit.org/surveylibrary/depression-
anxiety-stress-dass.html (accessed Sep. 20, 2022).
[15]. P. F. Lovibond and S. H. Lovibond, "The structure of
negative emotional states: comparison of the
Depression Anxiety Stress Scales (DASS) with the
Beck Depression and Anxiety Inventories," Behav Res
Ther, vol. 33, no. 3, pp. 335–343, Mar. 1995, doi:
10.1016/00057967(94)00075-u.
[16]. "Anyscale - What is hyperparameter tuning?,"
Anyscale. https://www.anyscale.com/blog/whatis-
hyperparameter-tuning (accessed Sep. 29, 2022).
[17]. "Physical Symptoms of Anxiety: What Does It Feel
Like?" https://www.healthline.com/health/physical-
symptoms-of-anxiety#fa-qs (accessed Jan. 06, 2023).

IJISRT23AUG2134 www.ijisrt.com 2250