91av

IF YOU dial the emergency services in Denmark, soon you won’t just get a human operator –an artificially intelligent assistant will be listening in too.

Developed by start-up Corti, the system kicks into action when someone dials 112 in Copenhagen, then it starts listening for signs of a possible cardiac arrest. To do this, it first uses speech-recognition software to transcribe what is being said before analysing the text. Once it is confident of a diagnosis, it flashes an alert on the screen for the operator to see.

Identifying cardiac arrest over the phone is one of the trickiest tasks for an operator. The person reporting an incident is often distressed and lacking medical training, so reading between the lines of what they say is happening is key. Recognising the symptoms early can be the difference between life or death, but because only 1 per cent of emergency calls are for cardiac arrest, it is difficult to pick up on all of the signals all the time.

“If someone falls from a ladder because of a cardiac arrest, it would just look like they had a nasty fall. Recognising the right signs can be very hard and very stressful,” says Lars Maaløe at Corti.

Underpinning the technology are algorithms called neural networks. These consist of many different computational layers, and are inspired by the way neurons connect in the brain. By exposing the network to a large amount of data, it eventually organises in such a way that it can reliably detect patterns in new data. In Corti’s case, it learns to diagnose certain conditions from years of emergency call logs.

Listening in

“There is no particular word that we are searching for,” says Maaløe. The system isn’t given any specific words or phrases to look for, instead it learns everything from the call logs alone.

The team says that during initial experiments, the system was able to spot all of the cases of cardiac arrest that human operators did, as well as some they didn’t. They have also completed a lengthier study with the Danish emergency services, the results of which are currently under peer review.

Although the first version of the system focuses on cardiac arrest, the team is also preparing to add stroke diagnosis to Corti’s list of tricks, as well as an English-language version for the US.

At the moment, Corti relies only on the conversation dialogue, but there are reasons to believe that the system can be improved by analysing the audio further. For example, a common side effect of cardiac arrest is an abnormal breathing pattern called agonal respiration, which to a layperson could be perceived as regular breathing, rather than a serious issue.

“Once we have this up and running, the phone operator could ask the caller to hold the phone near someone to check if they really are breathing properly,” says Tycho Tax at Corti.

Corti has been getting “very exciting results”, says Marine Riou at Curtin University, Australia. She says that breathing assessment is one of the biggest challenges during emergency calls for cardiac arrest.

However, once cardiac arrest is diagnosed that’s not the end of the story.

“The next big issue is guiding callers through resuscitation,” says Riou. A whole series of small improvements in recognising cardiac arrest and then acting on it would make a big difference, she says.

This article appeared in print under the headline “AI can hear a cardiac arrest”