How Twitter Could Help Detect Depression
The Theory and Reality of a Model That Can Identify Early Signs of Depression
Whenever you hear about social media in the news, it's almost always something bad. There's a reason for this of course—studies show social media can interfere with teens’ sleep, contribute to anxiety, and reinforce filter bubbles—but all the doom and gloom can get a bit depressing.
For better or worse, though, social media isn’t going anywhere. So let’s change the perspective: As we work to improve the platforms, why not put their reach to good use?
Researchers at Microsoft had a similar line of thinking in 2013 when they published a study called "Predicting Depression via Social Media." In it, Munmun De Choudhury, Michael Gamon, Scott Counts, and Eric Horvitz explained how they created a model that used Twitter to predict depression.
If you’re thinking, a model that predicts depression through Twitter? How is that possible? Well, so did we. And then we got to unpacking the De Choudhury study to answer this and other questions below.
How Could You Tell if Someone is Depressed on Twitter?
The researchers started off by using crowdsourcing to create a list of 171 Twitter users with severe depression and 305 users who showed little or no signs of the conditions based on a common screening questionnaire. De Choudhury’s team then compared the tweets that the severely depressed respondents posted in the year before their symptoms began to tweets the other respondents posted in the year before they completed the questionnaire.
In total, they analyzed a whopping 2,157,992 tweets. The verdict? During the year leading up to the onset of their depression, the individuals with the illness exhibited less social activity as well as more negative emotion, focus on self, relationship and medical concerns, and interest in religion.
They came to this conclusion by comparing several factors, including the use of depression-related terms and the ways in which participants engaged with Twitter.
Graphic by Victoria Fernández, via Creative Commons license BY-NC-ND 4.0.
As an example, a tweet from a study participant with severe depression read, "Having a job again makes me happy. Less time to be depressed and eat all day while watching sad movies."
Finally, De Choudhury’s team used this data to create a model that could predict the likelihood that participants would develop depression with a 70% accuracy.
Would a Depression-Detecting App Work Today?
While these findings sound promising, we were unsure whether this model would work in our current climate. For one thing, the tendency to only post your best moments online—your highlight reel—has intensified since this research was published. If people are not being honest about how they’re feeling online, would a model like the one created by the Microsoft researchers even work?
After all, the model in question analysed the word choice of tweets, which could potentially be skewed by individuals not tweeting in tune with how they’re feeling. However, the model also considered measures like the number of followers and followees, the amount of outgoing and incoming messages, and the time of day when users posted, which are less affected by the highlight reel phenomenon.
So, it sounds like it could be possible to work around the mirage most people display online. Before it could stand a chance at being trialed in real life, though, the model would have to be tweaked to be even more accurate in its predictions; it would have to still work when people are masking how they’re doing.
Graphic by Victoria Fernández, via Creative Commons license BY-NC-ND 4.0.
Privacy is another concern. A model that uses Twitter to flag depression around the world would be incredible, but would people want something trawling through their tweets? We can imagine some readers thinking, the amount of data that social media tracks is scary enough without this.
De Choudhury’s team saw their findings being put to use in a privacy-preserving app that provides users with information about professional help and the value of support when it predicts that an extreme change in their mood is likely. This could involve having users opt in and ensuring that no personal information is recorded.
A somewhat similar concept is already in place on TikTok, where users are shown a video that encourages them to take a break if they’ve been scrolling on the app for too long. Regardless, deploying a depression-detecting app would require a lot of careful consideration and stakeholders from a variety of backgrounds would need to be involved.
Could Social Media Help People with Depression?
De Choudhury, Gamon, Counts, and Horvitz concluded that “social media activity provides useful signals that can be utilized to classify and predict whether an individual is likely to suffer from depression in the future.”
We have some reservations, namely accuracy and privacy, about a feature on social media that could identify users who may be on the verge of developing a mental illness. Yet, it sounds like an avenue worth exploring further.
According to the Canadian Mental Health Association, 50% of people will have or have had a mental illness by the age of 40. And according to some studies, there’s a link between social media and anxiety and depression, which absolutely needs to be addressed.
While we work to make social networking platforms better, we could use them to meet people where they’re at and equip them with knowledge that could truly effect change. Now that’s something worth tweeting about.
Study Objective & Methods
Predicting Depression via Social Media
Munmun De Choudhury, PhD; Michael Gamon, PhD; Scott Counts, PhD; Eric Horvitz; PhD
Published in Proceedings of the Seventh International AAAI Conference on Weblogs and Social Media, 2013.
In this study, De Choudhury, Gamon, Counts, and Horvitz aimed to examine the potential of using social media as a tool to detect and predict depression. To do this, De Choudhury’s team used crowdsourcing to compile a list of 476 Twitter users who (1) agreed to give the researchers access to their Twitter feeds, (2) indicated that they began experiencing depression symptoms in the last year, (3) reported that they experienced at least two depressive episodes in the last year, and (4) filled out two depression screening tests that yielded similar results. One of the two tests, the CES-D questionnaire, typically calculates three groups of depression severity: low, mild to moderate, and high. The researchers decided to use the high category as the threshold for having depression in their study in order to minimize false negatives and false positives; 171 survey respondents fell into this category. De Choudhury’s team then compared the tweets they posted in the year before their symptoms started to tweets the 305 remaining survey respondents posted in the year before taking the survey. The measures they used to compare the tweets were: user engagement, emotion, egocentric social graph, linguistic style, and depressive language use. The researchers used the results to build a Support Vector Machine classifier that predicted the likelihood of depression ahead of the onset with an accuracy of 70% and a precision of 0.74.
Social Media and Well-Being Training
This research (and all our social media and well-being articles) have laid the foundation for our 3-course program designed for anyone wanting to approach social media and communications in a way that protects well-being and puts people first. Learn more here.
