Feeling blue: Instagram posts can indicate depression
Social media might have some interesting applications for monitering mental health. University of Vermont statistician Chris Danforth is developing computer algorithms that he says can spot people with depression just by looking at the content and compositions of photographs they’re posting onto the social media platform Instagram. In tests he’s achieving up to a 70% pick-up rate. Georgia Mills spoke to Chris about his work.
Chris - There have been a number of studies in the last few years looking at aspects of our behaviour that we reveal through social media. In particular, studies that we’ve been involved in have been looking at trying to assess someone’s mental health from aspects of images they post to Instagram and messages they write on Twitter.
Georgia - How have you looked into this?
Chris - What we did was recruit individuals who had been diagnosed with depression by a psychiatrist and were active on social media at the time. We asked them for access to their social media feeds, so we compared picture’s they took with pictures taken by a control group who had not been diagnosed with depression.
What we found was that individuals who had been diagnosed with depression, their pictures tend to be bluer, and darker, and greyer. They tended to have less faces in them and these results are consistent with what psychiatrists know about how depression affects people. It literally causes them to see the world with less colour in it, and they spend less time in large groups of people and social interactions.
While those findings weren’t particularly surprising, what’s really exciting about that particular study from our perspective is that when we restricted our computer model to only have access to pictures posted by people prior to the date that they were formally diagnosed, it still picked up on these differences. That’s an indication that we could be able to get individuals who are suffering from depression in front of a doctor sooner.
Georgia - Oh wow! So you could quantify these pointers like having blacker and bluer pictures, and these things appeared before people had actually got a clinical diagnosis?
Chris - That’s right. And the results, they’re consistent with what people are finding looking at other sources of social media as well. We looked in a separate study at depression and PTSD, and identified predictors of those health states from words that people were using on Twitter and their behaviour: frequency of their posts, how likely they were to include links in their posts. And there were indicators, as well, that there were differences prior to the date that they were formally diagnosed.
So I think that the results of our study are promising, but they’re really just a proof of concept. There’s a small group of a few hundred people in each of these studies and it’s not clear that this would translate to the average person using Instagram or Twitter. But the goal is to try and figure out how to leverage all of the information that we give our mobile phone about us: our tone of voice, the words we use, the people we communicate with. This sort of data it’s incredibly private so if we can protect it but yet, at the same time, give algorithms access to information that’s predictive about our health state, we could be connecting with our doctor sooner. For example, the doctors could have access to a lot more information than they would get from a typical yearly checkup interview.
Georgia - Do you see this being used by the social media companies themselves?
Chris - Certainly companies know a lot about you, even if you’re not using it, just from what your friends are doing on the site. I do expect that there are going to be situations where social media companies are selling advertising spots for people who are suffering from mental health problems potentially without even intending to, or realising that they’re doing so. That’s something that I think is really important right now is that we start having a conversation with each other about how to protect this sort of data.
This particular study was done by a few people with limited resources but certainly, if we could do it, then a company like Instagram or Facebook or Twitter, they could use this data to try and make money. We’d like for them to use it to try and help us be happier and healthier.
Georgia - I suppose that is a worry, people will be worried about their privacy; that people could see that they’re depressed before maybe they know themselves?
Chris - I think that we reveal a lot more about ourselves online with our digital footprint than we’re aware of. Part of doing research studies like this in communicating the results of these studies to people is to help educate the population about what it is that they need to be aware of when they’re on the internet.
Georgia - Is there a way the social media companies could use this information for good?
Chris - Certainly Facebook and Instagram are active in trying to provide mental health resources to people who, for example, search for the word depression or use words that have been seen to be predictive of self-harm. They have teams working on trying to get individuals who are suffering from the problems access to people sooner. And that is a combination of artificial intelligence and actual people who are then called in to virtually try and address whatever’s going on and maybe, potentially, dispatch police officers to attempt to rescue someone. So companies are working on that sort of thing.
Stepping back to the picture of public health, there are a number of other studies, some of which we’ve done, showing that things that you might like to quantify about how well a city or a society are functioning, those things can be inferred from the words that are used if you look at particular geographic areas. We’ve shown, for example, that population scale, health rankings, like the percentage of people who are obese, or who suffer from diabetes, those things can be inferred from just the words that people use on Twitter.
Well-being surveys that are done by Gallup can be inferred without asking any questions simply by seeing how often people use happy and sad words in different states, for example, in the US. So there are going to be applications for this sort of public health instrument building that people are going to develop.