‘Romeo + Juliet’

It analyzes not only what you say, but how you say it.

A team of nine scientists successfully used machine learning to determine whether someone is likely to attempt suicide. Sound crazy? Perhaps that’s because historically, it’s been almost impossible to figure out who’s at risk and who’s just fantasizing.

To understand how the team went about cracking the linguistic code of suicidal behavior, it’s important to understand two different types of analyses.

In trait analyses, scientists study characteristics of the patient’s biology. For example, they measure the size of different parts of the brain like the amygdala, cerebellum or hippocampus.

During state analyses, scientists study characteristics like verbal and nonverbal communication.

In the new study, which was published in November, researchers at the American Association of Suicidology used a machine to focus on state analyses?—?specifically, analysis of linguistic patterns (word/phrase recurrence) and acoustic patterns (the way the words are spoken).

Other studies have used written language to predict suicidal behavior with some success. So what sets this new study apart? Well, according to its authors:

“The study described herein is novel because it uses a multisite, multicultural setting to show that machine learning algorithms can be trained to automatically identify the suicidal subjects in a group of suicidal, mentally ill, and control subjects. Moreover, the inclusion of acoustic characteristics is most helpful when classifying between suicidal and mentally ill subjects.”

In other words, this one can differentiate between truly suicidal and “merely” (a poor term) mentally ill patients.

Over 18 months, researchers studied 371 participants in a three-site clinical trial. The group comprised three sub-groups: those who were suicidal, those who were mentally ill but not suicidal and a control group that was neither mentally ill nor suicidal.

To gather data, researchers asked each subject to complete several standard questionnaires (like the Columbia Suicide Severity Rating Scale) as well as the “Ubiquitous Questionnaire,” an interview with the following five questions:

  • Do you have any hope?
  • Do you have any fear?
  • Do you have any secrets?
  • Are you angry?
  • Does it hurt emotionally?

The scientists chose these questions for their open-endedness. By encouraging elaboration, the questions gathered a wider sample of language to feed the machine. All subjects’ answers were recorded and transcribed.

The receiver operating characteristic (ROC) is a fancy name for a number that the scientists calculate to determine the machine’s diagnostic accuracy. In this case, the machine has to achieve a “score” of .80 or higher to be considered an accurate diagnostic tool. It struggled a little more with identifying the adolescent subjects’ suicidal tendencies than it did the adult subjects’. It also showed trouble with pure acoustic analysis (as opposed to linguistic and acoustic analysis combined).

But overall? The experiment was a success. And the study’s authors point out that its use isn’t limited to hospitals: “[This method] could be translated to schools, shelters, youth clubs, juvenile justice centers and community centers, where earlier identification may help to reduce suicide attempts and deaths.”

Now we’ve got a machine that can identify the truly suicidal among a group of suicidal, mentally ill, and control. A suicide sieve, if you will. A machine that can accurately distinguish non-lethal mental illness from true suicidal ideation isn’t just a neat gizmo?—?hopefully, it will save lives.