Artificial Intelligence Could Help Prevent Subway Suicide Attempts

Suicide is a serious public health problem in Canada, and it's a sad reality that sometimes, people jump onto a subway track in an attempt to end their lives.

According to researchers, artificial intelligence could help prevent suicides underground by helping to identify those who are at risk.

Suicide crises are often temporary, according to the Canadian Association for Suicide Prevention, which provides a list of resources to those who need help. And subway attempts doesn't always go as planned, said Brian Mishara, director of the Center for Research and Intervention on Suicide and Euthanasia at the Université du Québec à Montréal (UQAM). Oftentimes, he told me over the phone, people who attempt suicide by subway are left maimed and disfigured, and wind up in hospital, where they may only die after much protracted suffering.

Read More: NASA's New Self-Learning AI Could Save First Responders

This encouraged him and a group of students to watch CCTV video of 60 suicide attempts on the Montréal metro and identify behaviours that could indicate when an attempt is about to happen.

"It's really painful to watch someone display a lot of ambivalence before the attempt and then try to change their mind by squashing themselves down to the ground as the train's approaching," said Mishara. "That was the most difficult."

While some cities are approaching suicide prevention by building doors that only open when the train arrives—Toronto has also considered this—the cost can be prohibitive for older metro systems. Instead, Mishara thinks artificial intelligence can help by analyzing live CCTV footage and looking for the telltale signs of an attempt underway.

"A human being can't sit and watch 100 or more screens in real-time to try and see if there's anyone at risk," Mishara said. "The ideal situation would be to have automated, computerized monitoring where if certain patterns of behaviour are observed, the program could then immediately signal to someone in the control room or the driver to look at the video. The person can then make the judgement call."

Mishara considers his study, published on Wednesday in the journal BMC Public Health, a first step towards such a system. After analyzing videos of suicide attempts provided by the Société de transport de Montréal(STM), the city's public transit agency, Mishara and his colleagues identified a set of behaviours that, together, can indicate an attempt is about to happen. These include pacing around the yellow line, placing an object on the platform, and appearing slouched and depressed.

The idea would be to train software to identify these possible warning signs using pixel information from CCTV video feeds, a domain known as computer vision. Great strides have been made in computer vision thanks to a technique known as deep learning, although there's a ton of work left to do in order to make it truly reliable in the real world. Still, universities and even Google have all proposed methods for computers to automatically, and accurately, evaluate human body poses.

"A decision has not yet been made: we must evaluate before deciding on the recommendations," STM spokesperson Amélie Regis wrote in an email, regarding whether or not the STM plans on implementing live CCTV monitoring with artificial intelligence.

As Mishara said, this is just a first step.

Get six of our favorite Motherboard stories every day by signing up for our newsletter.