Predicting Depression via Social Media

De Choudhury, M., Gamon, M., Counts, S., & Horvitz, E. (2013). Predicting Depression via Social Media. Proceedings of the International AAAI Conference on Web and Social Media, 7(1), 128-137. https://doi.org/10.1609/icwsm.v7i1.14432

---

Summary

- Introduction
The paper explores the potential to leverage social media to detect and diagnose MDD (major depression disorder) in individuals.

- Methodology
The authors employ crowdsourcing to compile a set of Twitter users who report being diagnosed with clinical depression, based on a standard psychometric instrument. Through their social media postings over a year preceding the onset of depression, the authors measure behavioral attributes.

- Result
The authors leverage these behavioral cues to build a statistical classifier that provides estimates of the risk of depression, before the reported onset.
Useful signals were a decrease in social activity, raised negative affect, highly clustered egonetworks, heightened relational and medical concerns, and greater expression of religious involvement.

- Conclusion
They find that social media contains useful signals for characterizing the onset of depression in individuals. The findings may be useful in developing tools for identifying the onset of MDD, enabling those suffering from depression to be more proactive about their mental health.

---

Comments

Things to study more

• How the threshold for the prediction task is being determined?
  The study used CES-D (Center for Epidemiologic Studies Depression Scale) questionnaires to determine the depression levels of the crowdworkers. Out of a scale from 0 to 60 of depression severity, the study used 22 as a chosen threshold to minimize type 1 and type 2 errors in the prediction task. It referred to one study (Park et al., (2012)) for their choice of threshold.

Possible future research ideas

1. Adaptation to Korea

• Validation of the result in different cultures
  The way users express their thoughts could hugely differ due to cultural differences. (e.g. South Korea belongs to the high-context culture, while the U.S.A. belongs to the low-context culture.) It could also affect important cues for depression prediction.

• "Linguistic style" adaptation to Korean
  The study leverages the linguistic style analysis as one of the most important signals to capture the risk of suffering from depression. (e.g. articles, conjunctions, adverbs, personal pronouns... etc.) How can this be adapted to the Korean Language?

2. Ideas for Factors to Investigate

• Idea of using "diurnal pattern" as a measurement
  Patterns of posting made by a user during the course of a day could be an important factor that differentiates people who are at risk versus people who are not.

• The concept of activation and dominance as a measure of the emotional state
  Not only PA (positive affect) or NA (negative affect), but also activation and dominance can be used as measurements of the emotional state of users.

• Investigating the trend of behaviors
  In the study, momentum seems to be a feature that shows statistical significance. Not only is the absolute degree of behavioral change important, but also the trend of changes over time bears useful markers of distinguishing depressive behavior.

Questions

• Why did they choose these embedding methods?
  To build the "depression lexicon", the study used Pointwise mutual information (PMI) and log-likelihood ratio (LLR) to compute each word's association with the regex "depress*". Would this be the best method?

• What is a Support Vector Machine classifier?
  The classifier the authors chose was a Support Vector Machine classifier with a radial-basis function (RBF) kernel.

---

Things to note

• The vantage point of using social media rather than self-report:
  Social Media provides data about people's social and psychological behavior about mental status in an unobtrusive and fine-grained manner, while self-report is a high-level summary of experiences over long periods of time.

• The way the study built the lexicon of depression symptoms in online settings:
  The authors mined a 10% sample of a snapshot of the "Mental Health" category of Yahoo! Answers. They extracted all questions and the best answer for each question, resulting in 900,000 question/answer pairs.

• The way the study interpreted behavioral tendency in relation to depression symptoms:
  For example, a decrease in user engagement measures such as volume and replies indicates a possible loss of social connectedness. Also, higher expression of NA possibly reflects mental instability and helplessness, and higher presence of first-person pronouns could illustrate higher attention to self.