A laboratory at EPFL is developing a system, based on artificial intelligence, which will make it possible to diagnose Covid-19 from the sound of a cough. The application, which should be available in early 2021, will allow free, widely available testing.
Given all the different types of cough that are related to infectious diseases, it is difficult for the human ear to distinguish the Covid-19 cough. But there truly is a subtle difference, easier to pick up if one observes it with “Coughvid”, a system developed by an EPFL laboratory, based on machine learning.
“The electronic systems we have today, including mobile phones, are able to perform more sophisticated calculations than we can do ourselves, by listening to the sound of a cough. We can therefore make more detailed analysis than a doctor will do manually,” explained Prof. David Atienza, on RTS’ 19h30.
By breaking down the sound patterns of the cough of a person infected by Covid-19, and those of somebody in good health (or suffering from a cough because of a different infectious disease), the researchers were able to identify two zones of frequency that are unique to SARS-CoV-2.
But the director of the Embedded Systems Laboratory of EPFL wants to make one thing clear: “We can not replace doctors, that is not the idea. They are always with us, to validate our system.”
In order to train the algorithm to detect sound patterns that characterise Covid-19, some 25,000 volunteers, of whom 2,000 were officially tested as positive, have already provided sample sounds of their cough through the Coughvid website. The scientists hope to have 30,000 recordings before the end of the year. “If we manage to do that,” states David Atienza, “We should be able to release the application, early next year.”
Today the system has a success rate of 80% in performing diagnoses, states the EPFL researcher. Unlike other types of test, there is no need to go to the hospital, nor to undergo invasive examinations, such as the PCR test. With Coughvid you just need a smartphone and to cough in the microphone, for 10 to 15 seconds. “The idea is to make pre-diagnostic system, which will be useful for avoiding PCR tests which are expensive, even when used on the grand scale.”
They may be easy to use, fast, cheap and ready for mass adoption, but digital applications will never replace clinical tests.
“We are not attempting to validate the system as a medical device because the process is so long. Our idea is to try to help society as soon as possible,” specifies David Atienza.
The application has a major defect: it can only detect Covid sufferers by their cough. According to the head of Pulmonological services at the CHUV, Christophe Von Garnier, a third to a half of patients with Covid have no cough: “There are asymptomatic patients, others who only have headaches and a sore throat. So this has to be taken into consideration,” he explains.
In spite of this fact, the pulmonologist is curious to see how this application will help the medical teams in their daily routine.
“I can see artificial intelligence playing a role in our clinical activity,” says Christophe Von Garnier. “It is not a tool which will allow us to make a firm diagnosis of a Covid infection. It will be one of a number of indicators, which can be useful, because we can then advise the patient about a possible admission.”
Other labs are also working on similar systems, in particular MIT. Lately, the american institution claimed that their sound analysis system was capable of detecting practically all positive Covid-19 cases, even asymptomatic ones.
This spectacular progress may herald a major step forward for experts in artificial intelligence with respect to Covid diagnosis, and in the long term we can envision – why not? – similar progress against many other types of infectious disease.
Author: Yoan Rithner/fme
Translation: John Maxwell