AI Has Learned to Spot Suicidal Tendencies from Brain Scans

Suicide is the second-leading cause of death among young people between the ages of 15 and 34 in the United States, and clinicians have limited tools to identify those at risk. A new machine-learning technique documented in a paper published today in Nature Human Behaviour (PDF) could help identify those suffering from suicidal thoughts.

Researchers looked at 34 young adults, evenly split between suicidal participants and a control group. Each subject went through a functional magnetic resonance imaging (fMRI) and were presented with three lists of 10 words. All the words were related to suicide (words like “death,” “distressed,” or “fatal”), positive effects (“carefree,” “kindness,” “innocence”), or negative effects (“boredom,” “evil,” “guilty”). The researchers also used previously mapped neural signatures that show the brain patterns of emotions like “shame” and “anger.”

Five brain locations, along with six of the words, were found to be the best markers to distinguish the suicidal patients from the controls. Using just those locations and words, the researchers trained a machine-learning classifier that was able to correctly identify 15 of the 17 suicidal patients and 16 of 17 control subjects.

The researchers then divided the suicidal patients into two groups, those that had attempted suicide (nine people) and those that had not (eight people), and trained a new classifier that was able to correctly identify 16 of the 17 patients.

The results showed that healthy patients and those with suicidal thoughts showed markedly different reactions to words. For example, when the suicidal participants were shown the word “death,” the “shame” area of their brain lit up more than it did in the control group. Likewise, “trouble” also evoked more activity in the “sadness” area.

This is just the latest effort aimed at bringing AI into psychiatry. Researchers are working on machine-learning projects that span from analyzing MRIs to predict major depressive disorder to picking out PTSD from people’s speech patterns. Earlier this year, Wired wrote about researchers who built a system that can sift through health records to flag someone at risk of committing suicide, with between 80 and 90 percent accuracy. Facebook is using text mining to identify users at risk of suicide or self-harm and then pointing them to mental health resources (see “Big Questions Around Facebook’s Suicide-Prevention Tools”).

Artificial intelligence has already made waves in the medical field at large. There are algorithms so good at detecting tumors and other problems in CT scans that Geoffrey Hinton, one of the foremost researchers in deep learning, told the New Yorker that radiologists will eventually be out of a job. Indeed, he said, “they should stop training radiologists now.”

In this case, the research is more likely to inspire new human-driven therapies than put a whole field’s worth of doctors out of a job. The paper pointed out that identifying different patterns and areas could suggest new regions to target for brain stimulation techniques. Identifying particular emotional responses to suicide-related terms could also be useful to psychotherapists treating their patients.