(Bloomberg Businessweek) -- A woman in Copenhagen hears a loud crash in the next room and rushes in to discover her father sprawled on the floor, unresponsive. She quickly calls Denmark’s health-emergency hotline, where a person answers the phone—but a computer is eavesdropping. As the operator runs through a series of questions—the patient’s age, physical condition, what he was doing when he fell—the computer quickly determines the man’s heart has stopped and issues an alert. “Those human dispatchers have an amazingly hard job,” says Andreas Cleve Lohmann, co-founder of Corti SA, a Danish artificial intelligence software house that created the program. “This software can help them save lives.”
Corti’s AI employs machine learning to analyze the words a caller uses to describe an incident, the tone of voice, and background noises on the line. The software correctly detected cardiac arrests in 93 percent of cases, vs. 73 percent for human dispatchers, according to a study by the University of Copenhagen, the Danish National Institute of Public Health, and the Copenhagen EMS. What’s more, the software made its determination in an average of 48 seconds, more than a half-minute faster than the humans did. False positives—mistakenly concluding that a person is having a heart attack—were the same for both, 2 percent.
With a rapid diagnosis, dispatchers can quickly give the caller instructions on how to perform CPR or where to locate a defibrillator to shock the heart back into action. That can make the difference between life and death: Danish studies have found that a patient’s 30-day survival rate triples when a dispatcher recognizes cardiac arrest during an emergency call. “Seconds matter,” says Freddy Lippert, head of Copenhagen’s EMS, which provided more than 150,000 recorded calls to test the algorithm. “When I first saw their results, I thought this is too good to be true.” He was so skeptical he asked Corti to do it again with a new batch of calls, which the software also aced. This spring, Copenhagen began a large-scale randomized trial with live calls, and if it’s successful the city plans to use Corti on all its emergency hotlines.
Lohmann launched Corti with money from the 2013 sale of his first startup, which used a text-based chatbot to help medical staff schedule shifts. For his second act he wanted to create a voice-based interface to help doctors make diagnoses. But in machine learning, the data available often dictates the product you can build—and one of the best audio archives in medicine belongs to the Copenhagen EMS. Unlike many other localities, the city logs and analyzes all its emergency calls and tracks patient outcomes, and it had already studied how well its human dispatchers recognized cardiac arrest, giving Corti a good benchmark for judging its success.
This fall, Corti will begin pilot studies in five other European Union countries in conjunction with the European Emergency Number Association. And its software will soon be deployed in emergency call centers in Seattle, Singapore, and Taiwan. The installations will help Corti’s software work with new languages, dialects, and accents—essential because the algorithm involves speech recognition and analysis of the words a caller uses. “Our plan is to plant flags with the best emergency medical departments in the world,” says Lohmann, who’s raised $3 million from Nordic venture capital firms and has 22 employees in Copenhagen, Seattle, and Paris.
Corti’s software uses deep neural networks, a kind of machine learning loosely based on the human brain. These systems can ferret out complex correlations in data, but they’re largely opaque even to their creators: In any individual emergency call it’s difficult to determine the combination of factors that lead to the conclusion that a victim’s heart has stopped. An early version of the software ran via the cloud, but Lohmann soon realized his program needed to run locally in the call center. Otherwise, the system would fall behind on recognizing and analyzing the audio, causing it to miss key moments when it needed to prompt a dispatcher to ask a question or send an ambulance. Corti is now training the system to identify other critical conditions such as strokes and milder heart attacks.
Since many emergency operators already follow a standard script and decision-tree protocol to assess calls, it might appear that Corti aims to replace humans with computers. Lohmann acknowledges such a system could make sense in places such as large cities in Asia and Africa where there’s currently no emergency call handling at all. But he says the company’s research indicates the best outcomes occur when the machine works together with dispatchers. “This is a human augmentation tool,” he says, “not a human replacement tool.”
©2018 Bloomberg L.P.