Using Artificial Intelligence to Forecast Suicide

Members of the armed forces and young people in the United States have something in common: They think about suicide, or try it, more often than the general public.

Suicide is the second most common cause of death for those ages 15-34, after accidental injury, and third among those ages 10-14, according to the Centers for Disease Control and Prevention. There were just over 44,000 suicide deaths in 2015, making it the 10th leading cause of death overall in the U.S. And for every suicide death, there may be up to 100-200 attempts.

Among students in grades 9-12 in the U.S. during 2013: 17% considered attempting suicide; 13.6% made a plan about how they would attempt suicide; 8% attempted suicide. Among adults aged ≥18 years in the United States during 2013: 3.9% had suicidal thoughts; 1.1% made a plan about how they would attempt suicide; 0.6% attempted suicide. Source: Centers for Disease Control and Prevention

Suicide is a major concern for the military, which has spent millions of dollars for training and research to combat what has become an ongoing problem.

But what if through a combination of social science and artificial intelligence, we could predict when someone is thinking about suicide? That is what a new group at the University of Southern California is trying to do.

The Center for Artificial Intelligence In Society, or CAIS, is in the beginning stages of a project involving members of the military, homeless youth and college students, said Eric Rice, a professor at the USC Suzanne Dworak-Peck School of Social Work and one of the two USC professors who founded the center.

Rice and his co-director, Milind Tambe, an engineering and computer science professor, founded CAIS in 2016 to share ideas about how artificial intelligence can be used “to help solve the most difficult social problems facing our world.”

Rice said that the academic world has a unique responsibility to combat certain social ills. While corporations use data mining to personalize ads and experiences, they are not focused on analyzing data for societal good. 

But ending social isolation is one of the “grand challenges” of social work, he said. Suicide “is one of the most negative aspects of social isolation,” and it’s one of the big negative outcomes that Rice and others in the social work field are trying to prevent.

“We’re saying there are a lot of A.I. tools that could be used to make a difference in the world right now.”

To do that, they are trying to bring together experts from a variety of disciplines and leverage the power of computers.

Rice calls it “social scientists and computer scientists really coming together and trying to do what neither group could do by themselves.”

The social scientists may identify a problem, and their computer scientist counterparts look at the problem and say, “We can solve that with some cool math,” he said. “It’s a wonderful collaboration … [and] brand new.”

Rice said people most often think of artificial intelligence being used for self-driving cars or computer-guided drones. They also worry about “killer robots,” he said half-jokingly. “People’s ideas about A.I. are really science-fiction ideas.”

But most people already encounter artificial intelligence in their everyday lives—during a Google search, for instance, when A.I. predicts what they’re looking for, or when they talk to “digital assistants” like Siri, Alexa or Cortana.

Rice and his peers use artificial intelligence to work on systems that are difficult for humans to quantify. Math, he said, “helps us solve problems we can’t solve in our heads, and A.I. helps us solve math problems we can’t solve on paper.”

Using A.I., researchers can take billions or trillions of possibilities and drill down to a small number of the most likely outcomes. Using “supervised learning,” they can teach a computer to recognize patterns in one set of data, then tell it to apply what it has learned to a new set of data, Rice said.

With “unsupervised learning,” researchers simply say, “Here’s a bunch of data. We don’t know what’s going on in here. See if you can find a pattern.” If only 5 percent of a group of people are consistently doing something, for example, statistics would likely miss it, he said. But computers can look for recurring patterns in small populations.

The center has launched the early stages of the suicide prevention project, but only recently received funding. Although focusing on three different populations—college students, homeless youths and military service members—the project will use two different approaches.

On college campuses, researchers will look at friendship networks and try to figure out how to select students who can be trained to look for depression and suicidal thinking in peers. The rate of suicide at universities is higher than one might expect, Rice said. First-year students are at the greatest risk, perhaps because of the drastic change in their social contacts.

Schools typically train resident advisers and dorm assistants to watch for depression or suicide risk, but they are generally seen by students as authority figures, so they are less effective than students’ friends. Instead, algorithms will do some sophisticated social analysis, looking for the students who have the most influence on their peers. That will help researchers maximize the program’s reach.

“What’s the minimum number of people we need so we can kind of have eyes on everybody so nobody falls through the cracks?” Rice said.

In a similar study of HIV and other sexually transmitted diseases in homeless youth, Rice and Tambe said, A.I.-identified youth were much more effective in reaching their peers and convincing them to get tested for HIV. 

With the other two groups CAIS is focusing on—homeless teens and members of the military—researchers will try to understand the early warning signs of a person at risk for suicidal behavior. One of the classic signs, Rice said, is when a person starts to withdraw from social contacts.

Members of the military are often at highest risk during major transitions: when they first join, when they first return from deployment and when they leave the military. Rice said that may have to do with their connections to their support network.

Others have speculated that war deployments drive the numbers.

Historically, the suicide rate in the U.S. Army was below that of the general population who were of the same gender and age. But that rate increased 80 percent from 2004 to 2008, and in 2008, the suicide rate in the Army surpassed the civilian average for the first time in history. Last year, there were 601 suicides of service members, including those on active duty, with the National Guard or in the reserves, according to Department of Defense data.

The CAIS researchers plan to interview a military unit before deployment and then follow them after they return to see if they can identify a pattern of emerging suicidal thoughts. With the perception, or stereotype, that military members have to be “tough,” Rice said, researchers may have to figure out “what’s the right intervention space with them.” They also may try an outreach plan that is similar to what they will try for college students—maximizing outreach by targeting the most influential peers, perhaps an officer or chaplain.

“There’s a whole system of formal and informal support systems” in the military, Rice said.

A study of homeless teens in New York using CDC data found that those teens were more likely to struggle with depression and to think about suicide. They were also three times more likely to attempt suicide compared to their peers who were not homeless. CAIS researchers may focus on the caseworkers as the influential peer group.

Rice said his team will have to reverse-engineer the social networks of the homeless and the military to see what would work the best and where to “insert the assessment tools” to give them the help that they need.