Predicting Depression Severity from Social

Media Posts: A Machine Learning Approach

line 1: 1st Given Name Surname
line 1: 2nd Given Name Surname line 1: 3rd Given Name Surname
line 2: dept. name of organization
line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation)
(of Affiliation) (of Affiliation)
line 3: name of organization
line 3: name of organization line 3: name of organization
(of Affiliation)
(of Affiliation) (of Affiliation)
line 4: City, Country
line 4: City, Country line 4: City, Country
line 5: email address or ORCID
line 5: email address or ORCID line 5: email address or ORCID
line 1: 4th Given Name Surname
line 1: 5th Given Name Surname line 1: 6th Given Name Surname
line 2: dept. name of organization
line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation)
(of Affiliation) (of Affiliation)
line 3: name of organization
line 3: name of organization line 3: name of organization
(of Affiliation)
(of Affiliation) (of Affiliation)
line 4: City, Country
line 4: City, Country line 4: City, Country
line 5: email address or ORCID
line 5: email address or ORCID line 5: email address or ORCID

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Abstract— Mental health problems are on the rise globally, ones, financial insecurity, and social isolation are just a few
affecting a large number of people of all ages. Traditional of the more obvious ones for those who live alone. COVID-
clinical assessments rely on periodic self-reports, which are 19 is a pandemic affecting the United States and many other
time-consuming and may miss the optimal timing for countries, and I strive to stay up to current on the latest
intervention. The ubiquity of smartphones provides an
developments. This opinion was sparked by two headlines
opportunity to continuously monitor mental health states and
enable timely interventions. Machine learning (ML)
and a slew of phone calls and e-mails from coworkers and
techniques, particularly deep learning models such as friends, among other sources. "What We Must Do To
convolutional neural networks (CNNs) and recurrent neural Prevent a Global COVID-19 Depression" and "COVID-19
networks (RNNs), have shown promise in detecting mental Could Lead to an Epidemic of Clinical Depression and the
health states accurately using sensory data collected from Health Care System Isn't Ready for That, Either" are the two
smartphones. However, the black-box nature of these models headlines that recently drew my attention. There is little
hinders their explainability and trustworthiness in healthcare doubt that this pandemic is wreaking havoc on the
applications. Moreover, the multi-modality of sensor data economies of many individuals, as well as putting a strain on
poses challenges in fusion and interpretation. To address these
many people's ability to cope.
issues, we propose an attention-based model for mental health
detection that generates explanations by visualizing the
attention mechanism. CNNs are used to extract features from A deep learning-based technique for assessing the degree of
different modalities, which are then fused using an attention- depression is proposed in this study, based on information
based layer considering their diverse importance. An attention- that the user posts on social media. This is the first study of
based bidirectional LSTM then extracts and fuses important its sort to attempt to gauge the severity of depression, to the
features from multiple time windows. The model's reliability best of our knowledge. We begin by renaming the dataset
and explainability are promoted by generating explanations provided by Shen et al. Self-supervised classification based
based on the visualization of attention weights, and contextual on textual polarity and latent semantic analysis into distinct
information is fused with the multi-modal sensory features.
Experiments on a real-life mobile sensing dataset demonstrate
intensity categories (LSA). Based on various descriptions of
the proposed method's superior performance compared to depression in the literature, we devised a set of 527 features,
other approaches. The proposed model aims to provide a comprising emotional, event-triggered, behavioural, and
scientifically sound and explainable approach for automatic user-level characteristics as well as depression-related ones.
mental health detection using smartphone sensory data, We use the retrieved information to train an LSTM network
potentially enabling early warnings and timely interventions. that can accurately predict depression severity. When
evaluating our intensity estimation, we compare our
Keywords— Social Media, Machine Learning, Depression experimental findings to those of other models. Furthermore,
detection, Medical. our method is compared against and outperforms other
binary classification methods.
II. LITERATURE SURVEY
I. INTRODUCTION
F. Hao, G. Pang et al, The proposed approach is based on
The most frequent cause of disability in the world is the idea that social support can help to prevent depression in
depression. 350 million people worldwide are believed to people who are particularly anxious. In this study,
suffer from depression at some point in their lives. researchers found that AI and DI have a Gaussian
Depression symptoms manifest themselves in unique ways in distribution when compared. As a result, a GMM was used to
depressed individuals. Clinical diagnoses are frequently cluster AD points. This has led to the discovery of a number
made through face-to-face interviews with psychologists of useful clusters, each of which has its own semantics and
who use criteria from the DSM-IV-TR. When clinicians are observations. Analysis of correlations between anxiety,
called upon, they must be able to discern between depression, and SS was conducted. Our findings, which are
demoralisation and depression. According to the criteria, based on real-world data, suggest that students who are under
there are nine main types of depressive symptoms. In spite of a great deal of stress may benefit from high-level SS.
its efficacy, many people are embarrassed or unaware of
their sadness. In 2000, depression cost the United States an S. Pappa, V. Ntella et al, "The prevalence of depression,
annual economic burden of 83 billion dollars, the majority of anxiety, and insomnia among healthcare workers during the
which was attributed to poorer productivity and higher COVID-19 pandemic: a systematic review and meta-
medical costs.. About 15 to 20 percent of those with major analysis. When it comes to depression, anxiety, and insomnia
depressive illness commit suicide, according to a study by among healthcare professionals, a systematic review and
Goodwin and Jamison. Depression can be prevented in many meta-analysis are timely and comprehensive. With these
cases with appropriate treatment and early detection of findings, it'll be possible to calculate how much help
depression is the first step in applying these treatments. The employees need, as well as develop tailored treatments to
vast bulk of research into the early detection of depression help people be more resilient and less vulnerable in
relies on patient surveys and self-reports for diagnosis. Only pandemic situations.
30% of countries providing basic health care services have
these systems in place in 2009, making them prohibitively T. Anwar et al, Geo-social-temporal pattern mining is used
expensive. to infer location types. Health care professionals are found to
suffer from depression, anxiety, and insomnia at an alarming
Many factors have contributed to the outbreak of the rate in this study's systematic review and meta-analysis. With
COVID-19 pandemic. Unemployment, bereavement of loved
these findings, it'll be possible to calculate how much help N. Asghar et al, proposed based on the generation of
employees need, as well as develop tailored treatments to emotional neural responses In this paper, three emotional
help people be more resilient and less vulnerable in tactics are used to enhance the development of affectively
pandemic situations. conscious neural encoder decoder conversation systems.. An
affective space is created using a cognitively constructed
G. Shen et al, Multimodal dictionary learning for depression dictionary, and a variety of affect-based heuristic objective
detection via harvesting social media For this research, the functions and decoding algorithms are proposed. Information
ultimate goal is real-time social media diagnosis of retrieval activities such as question-answering and dialogue
depressive disorders. There are two types of datasets used in systems can be retained by interacting with users in a more
this study: one for depression and the other for non- compassionate and human way using these strategies
depression, along with well-defined depression-oriented
feature groups that can be used to discriminate between the R. I. Shader et al , proposed on the basis of depression and
two types of datasets. As part of an extensive study involving the COVID-19. Among other things, the COVID-19
a large number of depression candidates, researchers pandemic encompasses unemployment, death, and isolation.
examined the role played by various feature modalities in the When called upon, clinicians must be able to tell the
identification of depressed individuals and uncovered some difference between demoralisation and depression. Both
of the subtle differences in online behaviour between depression and demoralisation can be treated with this
depressed and non-depressed individuals. statement's remedies, as well as with a warning about the use
of chloroquine or hydroxychloroquine in patients with
T. Shen et al, Based on cross-domain depression detection COVID-19 syndrome. It is imperative that the risks and
via social media harvesting. In this study, the researchers benefits of both treatments are thoroughly evaluated before
looked at how to better identify depression in social media promoting either one over the other.
users by combining data from multiple sources. At the
Twenty-Seventh International Joint Conference on Artificial III. EXISTING SYSTEM
Intelligence, researchers describe the use of Feature Adaptive They had previously shown that the COVID-19 pandemic
Transformation & Combination (DNN-FATC) to transmit had the potential to harm the mental health of healthcare
meaningful information across different domains. workers in earlier studies. Further research is needed on
HCW mental health effects of the COVID-19 pandemic, as
F. Sadeque et al, according to research into how long it all of the studies included in our meta-analysis were cross-
takes to detect depression on social media. Latency and sectional. In any problem where a system must quickly
flatness were used in this study to establish early detection analyse a series of elements connected to an object and make
task evaluation measures, demonstrating that their theoretical a prediction about the object's class, flatency is an all-
behaviour is preferable to the current state of the art for early encompassing measure. Flatency is only considered in the
risk detection error (ERDE). We recreated common models context of identifying early signs of depression on social
and features from the eRisk 2017 shared challenge on the media. Wilcoxon's Signed Rank significance test was used as
early detection of depression in social media and empirically before to compare each model to the baseline (vanilla BS).
demonstrated that our measures capture major variations Affective word embeddings as input, affective loss functions,
between models better than ERDE. and affectively diverse decoding were compared in this
study, which also demonstrated how they could be
M. Trotzek et al, neural networks and linguistic metadata combined. Clinical depression and online social media
for early detection of depression indications have been activities are covered by MDL. Determining whether or not a
proposed in this study ERDE5, in particular, was found to be user is depressed is a binary classification problem for their
a useless metric for the stated shared task by the researcher in multimodal depressive dictionary learning model. MDL is
this paper, which investigated the currently used ERDEo used in conjunction with our labelling method to measure
metric in greater depth. By combining more than two depression intensity.
models, and then measuring the probabilities that result, this
approach appears to hold promise. Combining word IV. PROPOSED SYSTEM
embedding models in a single neural network has not been Depression intensity analysis does not have a large-scale
tested either. benchmark dataset that is publicly available. A dense
labelling strategy is now being devised to re-label the
T. Cai et al, proposed on the basis of Targetaware holistic original sparsely labelled dataset with depression intensity.
influence maximisation in spatial social networks. For the The technique can be applied to a wide range of users
purposes of this paper, we investigate the new research because we define attributes that correspond to each one of
problem of Holistic Influence Maximization (HIM), which is them. If you want to predict a person's level of depression,
an important add-on to the more traditional IM problem and you'll need to process social data, extract features from it,
has the potential to improve numerous real-world and train an LSTM network in the proposed system. Using
applications. By analysing four datasets, we found that one our feature set and model, we can use binary labelling to
or two orders of magnitude can be gained in efficiency in our determine whether a person is depressed (i.e., whether the
experiment. person is depressed or not). For binary classification, we
extract features from weights that have been trained to
predict depression intensity. In terms of accuracy, our
approach beats theirs by more than 2%. Our extensive levels were all done in a self-supervised manner. Numerous
research into the proposed strategy's efficacy is the result of tests on a standard dataset showed that our method was
these tests. We outperform other comparable models, as well superior to the others in terms of estimating intensity.
as existing binary classification approaches, in terms of
intensity estimates. References

System Architecture
[1] F. Hao, G. Pang, Y. Wu, Z. Pi, L. Xia, and G. Min, “Providing
appropriate social support to prevention of depression for highly
anxious sufferers,” IEEE Trans. Comput. Social Syst., vol. 6,
no. 5, pp. 879–887, Oct. 2019.
[2] S. Pappa, V. Ntella, T. Giannakas, V. G. Giannakoulis, E.
Papoutsi, and P. Katsaounou, “Prevalence of depression,
anxiety, and insomnia among healthcare workers during the
COVID-19 pandemic: A systematic review and meta-analysis,”
Brain, Behav., Immunity, vol. 88, pp. 901–907, Aug. 2020.
[3] T. Anwar, K. Liao, A. Goyal, T. Sellis, A. S. M. Kayes, and H.
Shen, “Inferring location types with geo-social-temporal pattern
mining,” IEEE Access, vol. 8, pp. 154789–154799, 2020
[4] G. Shen et al., “Depression detection via harvesting social
media: A multimodal dictionary learning solution,” in Proc.
27th Int. Joint Conf. Artif. Intell., Aug. 2017, pp. 3838–3844.
[5] T. Shen et al., “Cross-domain depression detection via
Figure 1. System Architecture har16vesting social media,” in Proc. 27th Int. Joint Conf. Artif.
Intell., Jul. 2018, pp. 1611–1617.
[6] F. Sadeque, D. Xu, and S. Bethard, “Measuring the latency of
depression detection in social media,” in Proc. 11th ACM Int.
Conf. Web Search Data Mining, Feb. 2018, pp. 495–503.
[7] M. Trotzek, S. Koitka, and C. M. Friedrich, “Utilizing neural
networks and linguistic metadata for early detection of
Conclusion depression indications in text sequences,” IEEE Trans. Knowl.
Data Eng., vol. 32, no. 3, pp. 588–601, Mar. 2020.
For this research, the ultimate goal is real-time social [8] T. Cai, J. Li, A. S. Mian, R. Li, T. Sellis, and J. X. Yu,
media diagnosis of depressive disorders. During the ongoing “Targetaware holistic influence maximization in spatial social
networks,” IEEE Trans. Knowl. Data Eng., early access, Jun.
COVID-19 pandemic and numerous lockdowns, mental 17, 2020, doi: 10.1109/TKDE.2020.3003047.
health has been a major source of concern. We proposed a [9] T. Anwar, K. Liao, A. Goyal, T. Sellis, A. S. M. Kayes, and H.
deep learning method based on social media data to estimate Shen, “Inferring location types with geo-social-temporal pattern
the severity of depression. It is the goal of this project to use mining,” IEEE Access, vol. 8, pp. 154789–154799, 2020.
[10] R. I. Shader, “COVID-19 and depression,” Clin. Therapeutics,
social media in real time to estimate the severity of vol. 42, no. 6, pp. 962–963, 2020.
depression in order to assist in the selection of the most
appropriate therapy. Relabeling a benchmark depression
dataset, creating a rich collection of discriminative
depression-related variables for users, and proposing an
LSTM network to detect depressed Twitter users of various