Send to

Choose Destination
Transl Behav Med. 2019 Nov 25;9(6):1197-1207. doi: 10.1093/tbm/ibz028.

A computational study of mental health awareness campaigns on social media.

Author information

School of Interactive Computing, College of Computing, Georgia Institute of Technology, Atlanta, USA.
Division of Digital Psychiatry, Department of Psychiatry, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, USA.


As public discourse continues to progress online, it is important for mental health advocates, public health officials, and other curious parties and stakeholders, ranging from researchers, to those affected by the issue, to be aware of the advancing new mediums in which the public can share content ranging from useful resources and self-help tips to personal struggles with respect to both illness and its stigmatization. A better understanding of this new public discourse on mental health, often framed as social media campaigns, can help perpetuate the allocation of sparse mental health resources, the need for educational awareness, and the usefulness of community, with an opportunity to reach those seeking help at the right moment. The objective of this study was to understand the nature of and engagement around mental health content shared on mental health campaigns, specifically #MyTipsForMentalHealth on Twitter around World Mental Health Awareness Day in 2017. We collected 14,217 Twitter posts from 10,805 unique users between September and October 2017 that contained the hashtag #MyTipsForMentalHealth. With the involvement of domain experts, we hand-labeled 700 posts and categorized them as (a) Fact, (b) Stigmatizing, (c) Inspirational, (d) Medical/Clinical Tip, (e) Resource Related, (f) Lifestyle or Social Tip or Personal View, and (g) Off Topic. After creating a "seed" machine learning classifier, we used both unsupervised and semi supervised methods to classify posts into the various expert identified topical categories. We also performed a content analysis to understand how information on different topics spread through social networks. Our support vector machine classification algorithm achieved a mean cross-validation accuracy of 0.81 and accuracy of 0.64 on unseen data. We found that inspirational Twitter posts were the most spread with a mean of 4.17 retweets, and stigmatizing content was second with a mean of 3.66 retweets. Classification of social media-related mental health interactions offers valuable insights on public sentiment as well as a window into the evolving world of online self-help and the varied resources within. Our results suggest an important role for social media-based peer support to not only guide information seekers to useful content and local resources but also illuminate the socially-insular aspects of stigmatization. However, our results also reflect the challenges of quantifying the heterogeneity of mental health content on social media and the need for novel machine learning methods customized to the challenges of the field.


Machine learning; Mental health; Public health; Social media; Twitter

[Available on 2020-03-05]

Supplemental Content

Full text links

Icon for Silverchair Information Systems
Loading ...
Support Center