Can AI Predict Mental Illness?

Artificial intelligence (AI) may one day be able to detect mental illness by reading just a few of your private messages or public posts. In early December, a group of researchers used Facebook data to predict psychiatric illnesses. The data included personal messages sent up to 18 months before an official diagnosis in 223 volunteers.

The research sounds like it can be expanded by introducing more social media platforms and more types of user-generated content, but these findings are not without criticism. Experts not involved with the research say that this medical application of AI may seem promising, but that it does not and should not replace clinicians in diagnosing illness.

The Algorithm’s Skeleton

Using an AI algorithm, the researchers extracted specific details from messages and photos that each volunteer posted. For example, swear words in messages indicate mental illness in general, while words related to negative emotions and words of perception (see, touch, listen) were linked to schizophrenia. And blue tones in photos were indicative of mood disorders. Using these existing traits, the researchers grouped each person into a prediction category: schizophrenia, mood disorder (like depression or bipolar), or no mental health issues.

To test the algorithm’s efficacy, the team used a metric that calculates a trade-off between false negatives and false positives. An algorithm with no false positives and no false negatives (a perfect algorithm) gets a score of 1. On the other hand, an algorithm that guesses randomly gets a score of 0.5. The researchers’ algorithm performed in the range of 0.65 and 0.77, which indicates that the algorithm is on the right track.

When the research team restricted the algorithm from looking at content unless it was posted in the past year, the performance was still better than randomly diagnosing users. H. Andrew Schwartz is a computer science professor at Stony Brook University who did not help with the study. He says that the results of the researchers’ algorithm are similar to those you would get from taking the PHQ-9, which is a 10-question survey used to screen for depression.

The research team was led by Michael Birnbaum, a professor at the Feinstein Institutes for Medical Research located just outside of New York City. Birnbaum says that this type of AI algorithm can have a huge positive impact on the treatment of psychiatric illnesses. According to Birnbaum, psychiatrists have the tendency to work with patients after they’re diagnosed, but the research could help find people who are feeling symptoms much earlier, allowing them to get help before things get worse.

Mining Social Media

Although Birnbaum’s team isn’t the first to use an AI application to detect mental illnesses, they’re the first to work directly with the users as patients. Other studies asked their patients to take the PHQ-9 survey, self-diagnose, or take their word for their diagnosis, but Birnbaum’s team worked with patients who had an official diagnosis from a clinician. Using the date of the professional diagnosis, the team worked backward to find indications of mental illness before the patient had a confirmed diagnosis.

Although Sharath Guntuku, a computer science professor at the University of Pennsylvania, doesn’t think that we’ll be using social media to mine for clues of mental illness as a professional diagnosis any time soon, he says that the algorithm designed by Birnbaum’s team could be a crucial tool in the mental health field. And “what we are increasingly looking at is using these as a complementary data source to flag people at risk and to see if they need additional care or additional contact from the clinician,” he says.

Schwartz agrees with this sentiment, adding that diagnosing mental illness isn’t an exact science, but that more data could definitely improve the accuracy of each diagnosis. “The idea is, you’re triangulating mental health. Assessing mental health is an exercise that can’t just rely on one single tool,” Schwartz says. Adding social media data into the mix can help clinicians bolster the hour per week they spend with their patients, allowing them to get a better (and maybe even clearer) look at their patient’s thoughts and feelings.

And according to Guntuku, some social media platforms are already taking it upon themselves to help catch users who may be feeling symptoms. For example, if someone searches for suicide-related keywords on Google, the search engine will show the phone number for the National Suicide Prevention Lifeline before the search results. Facebook is constantly watching for posts that indicate suicide risk and flagging them for human moderation. If the moderator agrees that the user may be at risk, Facebook can contact law enforcement while sending the user resources for suicide prevention. Either way, Guntuku says, “Any sort of public, large-scale mental health detection, at the level of individuals, is very tricky and very ethically risky.”

Digitalizing Mental Health Awareness

Birnbaum is a clinician, and he sees social media as an impactful data source for therapists to hone in on a diagnosis. After the diagnosis, social media data can be used to monitor and follow patients while they engage in long-term treatment plans. In psychiatry, Birnbaum says, clinicians are lucky to get a snapshot of the patient’s progress once a month, so “incorporating this type of information really allows us to get a more comprehensive, more contextual understanding of somebody’s life.”

Birnbaum has hope that social media will become a part of psychiatry in the next five to ten years. He envisions a day when “digital data and mental health will really combine, and this will be our X-ray into somebody’s mind.”