Could a Machine Identify Suicidal Thoughts?

Despite decades of effort, it has proven frustratingly difficult to predict who is most at risk of dying by suicide. Relying on patients to reveal their intentions doesn’t work. Nearly 80 percent of those who die by suicide hide their suicidal thoughts from doctors and therapists during their last visits. Yet suicide rates are increasing among middle-aged Americans and it is the second-leading cause of death for young people. That’s why researchers have been urgently searching for a reliable biological predictor of suicidal thoughts and behavior.

A report published this week in Nature Human Behaviour suggests an intriguing new possibility. The study combined neural imaging with machine learning to explore whether the brains of suicidal people respond differently to positive and negative words related to life and death. “It turns out they do,” says co-author Matthew Nock, a clinical psychologist at Harvard University. “We can predict with a pretty surprising degree of accuracy who’s had thoughts of suicide and who hasn’t—and even among those with thoughts of suicide, who has made an attempt and who hasn’t.”

Although the study was small, the findings are remarkable, says Barry Horwitz, chief of the Brain Imaging and Modeling Section at the National Institute on Deafness and Other Communication Disorders, who wrote a commentary that accompanies the study. Horwitz was particularly impressed by the technique’s ability to correctly classify the nine out of 17 suicidal subjects who had made attempts at taking their own lives, a group experts say are as difficult to find as needles in a haystack. “It’s hard to imagine any other method or risk factor allowing you to make that kind of distinction,” Horwitz says.

Earlier this year, machine learning was reported to detect suicide risk based on health records with 80 to 90 percent accuracy—considered an encouraging result. But this new study stands out because it reveals a potential biological marker for suicidal thinking. “It’s not just a reported behavior,” says co-author Marcel Just, a cognitive neuroscientist at Carnegie Mellon University, “we get the actual thoughts they have about suicide, and we see how they’re altered.”

The study joined two separate lines of research. Nock had previously used implicit association tests to determine suicide risk. For example, he paired words related to death and life with “like me” and “not like me.” Suicidal people were about three times more likely to respond more quickly than controls when death and me were paired. That result has been replicated repeatedly, and has proven to be a relatively strong predictor of suicidal thoughts and behavior compared to other approaches such as medical assessments.

Meanwhile, Just has been using functional magnetic resonance imaging (fMRI) to look inside the mind. “We can see the pattern that corresponds to thoughts,” he says. His technique, known as neuro-semantics, recognizes not words but concepts. In response to the sentence the old man threw a stone into a lake, for example,brain activation patterns indicate that a person is involved, that movement occurs, and that there’s an outdoor setting in the visualized scene. In separate studies, Just asked subjects (they were method actors) to conjure up emotions such as anger and jealousy, and found recognizable patterns for each. “Emotions have neural signatures,” Just says. “We have an archive of them on our computers.”

When Just read about Nock’s research, he wondered if he could see the thoughts of suicidal people in his scanner. First, the researchers asked a machine learning classifier, which tries to find predictive measures for specific outcomes, if it could distinguish between 17 subjects who had thought about suicide and 17  who had not. Inside the fMRI scanner, participants were asked to think about a series of words related to suicide and specific positive and negative thoughts and feelings. (The most significant proved to be death, trouble, carefree, cruelty, praise, and good.) After measuring the neural patterns of response, the researchers trained the machine on data for 33 out of 34 subjects. They then asked the machine to determine whether the mystery subject was suicidal or not. The classifier accomplished this task with 91 percent accuracy, correctly identifying 15 out of 17 suicidal subjects and 16 out of 17 controls.

In a second phase of the study, the researchers used Just’s archive of emotional signatures to assess how much of each of four emotions—anger, shame, sadness and pride—was associated with each word. “The classifier could tell whether someone was in the suicidal or the control group by how much of each emotion was evoked,” says Just. “Death evoked more sadness and shame among the suicidal ideators.” The classification based on these emotional neural signatures was 87 percent accurate in predicting suicidal thoughts.

In response to death-related words, suicidal people activated more areas of the brain associated with self-referential thinking. That was consistent with previous findings, Nock says. But the emotional response to words like carefree or trouble, was a surprise. “People who think about suicide have patterns that suggest less pride when the word carefree was shown, and more shame when the word death was shown,” says Nock. “There’s this emotional component about how people are thinking about themselves that we hadn’t identified before.”

Other suicidologists say the idea of using neural signatures to predict suicide risk is promising, but stress how preliminary the study is. “These results leave open many questions about the mechanisms at work and the clinical implications given the intensive nature of the task and the expense of fMRI,” says Alexis May, a clinical psychologist at the University of Utah, who was not involved with the research.

The study’s authors agree. They hope first to replicate the study and then to investigate the technique’s clinical viability, perhaps through adopting another technique, electroencephalography or EEG, which monitors brain electrical activity. While it’s clearly of scientific interest to better understand the brains of suicidal people, it remains to be seen if the method is realistically useful. “If you can get good prediction using a three-minute behavioral test, why spend a thousand dollars to put someone in a scanner?” Nock says. “What we’re doing is seeing if these approaches give us a different piece of the puzzle.”