Click to Skip Ad
Closing in...

Machine learning can trump humans in depression diagnosis, study says

Could a computer be better at identifying depression than a primary care physician?

That’s the suggestion of a new study that focused on using machine learning to analyze Instagram photos. The study, conducted by a researcher from the department of psychology at Harvard University and another from the University of Vermont, analyzed nearly 44,000 photographs posted to Instagram, exploring factors like what filter was used and how makes “likes” a photo received.

The study included photographs from 166 people, some of whom were depressed, and some of whom were not.

Instagram offers a variety of filters to change how a photo appears, and the researchers discovered that healthy participants were more likely to use a filter than depressed people. But if a depressed person did use a filter, the most common choice was Inkwell, which is one of the filters that turns a photo black and white.

The most common filter choice for healthy users was Valencia, which makes photos brighter. Other popular filters included X-Pro II and Crema, for healthy and depressed photographers, respectively.

“[P]hotos posted by depressed individuals tended to be bluer, darker, and grayer” the study stated. What’s more, photos from depressed users garnered more comments, but fewer likes.

Ultimately, the study concluded that not only was depression discernible in Instagram photos, but their method was actually more accurate than some professionals’ diagnosis success rate. “Our model showed considerable improvement over the ability of unassisted general practitioners to correctly diagnose depression,” the researchers reported.

By “general practitioners,” the study was essentially referring to a primary care physician, and not a psychotherapist or psychiatrist, and by “unassisted,” the researchers meant a doctor who was evaluating a patient simply by talking with him or her, and not using a test, like a questionnaire.

But the researchers caution that the comparision between people and machines in their study was “strictly informal” and included to contextualize their findings, and wasn’t the main point of their research.

“We’d like the emphasis to be more on the fact that this [is] a new way of thinking about improving early detection of mental illness, than on the possibility that this early groundwork is already outperforming trained medical professionals,” Andrew Reece, a doctoral candidate at Harvard University and the study’s first author, told in an email.

The full study, which hasn’t yet been published in a peer-reviewed journal, is available here.

Follow Rob Verger on Twitter: @robverger