Primary care doctors aren’t particularly good at spotting patients with depression. But a new study suggests that an algorithm analyzing your Instagram photos might be. In combination with other tools, like technology that could detect depression from the sound of your voice or how much time you spend in particular locations or what you say in text messages, the algorithm could be used in the future to help patients realize they need to go to the doctor, and give a doctor more data for a diagnosis.
The new machine learning tool was correct 70% of the time, at least in the small sample studied. On the other hand, a meta-analysis of studies found doctors correctly identified patients as depressed–without using a questionnaire or other screening tools–only 42% of the time.
“I think at some point soon we’re going to be able to assess risk factors in people’s behavior without asking them any questions, just based on what’s going on on their phone, what they’re saying and posting and writing on the internet, and then having your physician be aware of what an algorithm thinks your health state is, compared with other people who have been in that same position in the past,” says Christopher Danforth, a professor at the University of Vermont and one of the co-authors of the study. “I think that the future of this technology is not going to be so much in replacing the diagnosis made by professionals, but just getting people in front of them sooner.”
In the study, the researchers created an algorithm that looked for a variety of features in an Instagram post, such as the number of people in a photo. They trained the algorithm by showing it posts from people in the study who had been diagnosed with depression, and asking it to compare those people with those who weren’t depressed; then they tested it by asking it to predict who was depressed out of a new group of Instagram users.
People who were depressed were more likely to post photos that had a shade that was bluer and a darker, grayer tone. Healthier people were more likely to have a larger number of people in their photos. Depressed people were slightly less likely to use Instagram filters, and got slightly fewer likes (though more comments) for their posts. As it looked retroactively through a user’s entire posting history, the technology was able to detect depression even before someone had been diagnosed.
The study only looked at 166 Instagram users, and it doesn’t necessarily mean that it would work for all users; it’s possible, for example, that people who were willing to participate in the study and give access to their accounts don’t well represent everyone who uses the site. But it’s an indication that the technology could work. Danforth envisions that someone could eventually use it to build an app that patients could download to track their mental health, potentially sending them an alert when it was time to go see a doctor.
“I think that in terms that machine learning can help us make people’s lives better, this seems like a no-brainer to me, as long as people are opting into it, and not being forced by the operating system on their phone, for example, to share this kind of data,” he says.
Danforth now hopes to study the potential for a similar tool to predict suicide risk. “That’s one of the hardest prediction problems that there is,” he says. “There’s a lot of risk factors associated with suicide, but trained psychiatrists whose job it is to assess risk still really struggle with it. Maybe there’s something in our social media that could be helpful.”