In this article I want to talk a bit about my thesis dissertation that lead me to successfully complete my Masters degree in Information Systems at the National Tsing Hua University in Taiwan. This was a challenging two year journey that helped to pave the way to an exciting career in tech and allowed me to envision the possibilities of this career.

As an active user of Twitter at the time, I could feel the cries of strangers just longing to be heard.

“I never want to wake up”

“I am feeling empty”

“This world is better off without me”

“#mentalhealthawareness”

These are powerful words that bring you to the reality of social media and the possibilities that can be done with this data.

Bipolar disorder is one of several emotional disorders that affect approximately 60 million people worldwide. I wanted to contribute in someway to better understand these behaviours on social media.

My idea was to design a method to auto detect depressive and hypomanic episodes at an individual level in social media with respect to criteria that is based on professional clinical criteria for bipolar disorder (DSM-V).

To quantify behaviours in a social media setting; sentiment, emotion and linguistic features were employed to capture mood disturbances and mentions of a wide range of physical symptoms.

This is a high level overview of the methodology applied:

• Data extraction: Developed a python program to extract data from X (Twitter) REST API based on keywords

• Data exploration: Wrote Python and R scripts to explore and visualize data distributions such as TF/IDF and frequency distributions

• Data Pre-processing: Utilized Python libraries (NLTK, SciPy) to carry out tasks including removing URLs, mentions, replacing emojis and stemming

• Data Preparation: Label user tweets with features to be used in analysis. User tweets were assigned to one of four major classes of emotions based on the Circumplex model of affect. Tweets were also checked for keywords related to symptoms occurring during the mood cycle including religious inspirations, alcohol abuse, medication mention, hyper-active symptoms.

• Experiments: A series of experiments were conducted and compared to a dataset with users who did not have any signs of bipolar behaviour. Results showed that our method was able to detect bipolar behaviours among patients. In addition, results suggest that awareness hashtags and depressive symptoms keywords can serve as additional indicators of depressive behaviour in social media. Emotion features were found to present clear differences in depressive behaviour between bipolar and normal user dataset.

Bipolar behaviours were further explored with respect to time and location and a visualization tool is proposed for exploring bipolar behaviours with respect to spatio-temporal parameters.

This is just the beginning of what is possible with social media. It would be great to see collaborations with local agencies and government departments to further connect with real individuals and take the next steps towards actually helping people.

Github repo: https://github.com/nweat/mental-health-research/tree/master