By using thousands of recordings of coughing sounds, computer programs are able to identify those that are related to Covid with a better rate of success than humans. A Swiss team is soon going to make available an app that will allow self-testing in a matter of seconds.
“Cough, please.” Stethoscope in place, the doctor listens and tries to gauge the condition of the patient’s lungs. This gesture dates back to 1815, with the invention of the stethoscope by René Laennec – but computers are now trying to reproduce it, in their own way. Many teams of scientists are currently putting finishing touches to computer programs capable of identifying those infected with SARS-CoV-2 by simply listening to their cough.
Digital tools using algorithms of Artificial Intelligence (AI) are taking to the field in the fight against the Covid-19 pandemic. Most often the technologies use knowledge-based learning or machine learning, a form of AI in which a program learns to recognise a pattern, for example in sounds or images, from a large number of examples. In the case of this pandemic, tools like these are providing a diagnostic and prognostic system for patients in hospital, using scans of the thorax.
Although useful in such serious cases, AI has not yet been used for early diagnosis of the illness. This could change, if cough-based apps are shown to be effective. Very low-cost, readily available on the grand scale, they would represent a major advance in national strategies of diagnosis. Many similar projects have recently appeared in India, the UK and the USA, but also in Switzerland – at EPFL (Swiss Federal Institute of Technology, Lausanne).
In order to earn their diploma as an expert in coughing, machines have had to learn and pass exams. Swiss researchers have gathered thousands of recordings of coughs through a website called Coughvid. Volunteers, healthy or otherwise, record their coughing and then answer a few questions, anonymously, about their age, their sex and their state of health.
These sounds are stored, and together they form the basis of a knowledge-based system. “We have around 24,000 individual recordings,” from the scientific community and the public at large, says David Atienza, director of the Laboratory of Embedded Systems at EPFL, and member of the Coughvid team.
The machines are then set to work, going through all of the sound waves with a fine-toothed comb. Frequency, intensity, signal length, etc. – these computers analyse the sounds with a level of precision beyond that of humans, who usually describe coughs as dry, wet, or chronic. This high level of precision allows researchers to identify the vocal patterns specific to Covid patients, which are different to the coughs of other illnesses.
In order to prove their effectiveness, the machines were put to the test. They were made to listen to and diagnose new cough recordings, gathered from confirmed Covid patients from the CHUV (Lausanne University Hospital). “We established that the program scored a level of precision between 80 and 85%, greater than that of expert doctors who carried out the same test,” states David Atienza, who encourages everybody to build up the model by recording their coughs at the Coughvid website. “The more data we have, the more precise the model will be.”
In the USA a team from MIT (Boston) has gone a step further. Their program, the MIT Open Voice Model, claims to be able to recognise asymptomatic and low symptomatic patients, with a success rate of 100% in the latter, as stated by the authors in the Open Journal of Engineering in Medicine and Biology. These numbers mystify David Atienza, who is in contact with this team to find out more.
However it all plays out, the potential of these applications is huge, especially in early diagnosis: they make it possible to rule out Covid, if the cough being analysed actually corresponds to bronchitis or the ‘flu. David Atienza hopes to publish a mobile app – free of charge – at the end of this year or in the first quarter of 2021. He may even provide his algorithms to the World Health Organisation – who are keenly awaiting the new app. Discussions are underway.
Author: Fabien Goubet
Translation: John Maxwell
Originally published:
https://www.letemps.ch/sciences/applis-mobiles-depister-covid-toux
Publié mardi 10 novembre 2020 à 20:37
Modifié mercredi 11 novembre 2020 à 01:40