Researchers at EPFL, the University of Cambridge and MIT are developing algorithms using AI to diagnose Covid-19 from the sound of a cough recorded by a smartphone. Find out how this innovation could revolutionize Covid-19 screening, with impressive results but challenges to overcome.
Researchers at EPFL, the University of Cambridge and MIT are exploring an innovative method for diagnosing Covid-19 based on the sound of a cough, captured by a simple smartphone. The aim is ambitious: to enable everyone to detect whether they are a carrier of the virus without having to resort to a traditional test.
In April, scientists at EPFL launched Coughvid, a project aimed at collecting cough recordings to train algorithms. The aim? To develop an application capable of detecting Covid-19 from the sound of a cough, with an accuracy rate of 70% once sufficient data had been collected.
Five months later, the researchers had collected over 20,000 recordings, including around 1,500 from Covid-19-positive individuals. However, the results showed that pulmonologists, although consulted for interpretation of these recordings, were unable to agree on a precise diagnosis, except for a few obvious cases.
The researchers hope that artificial intelligence will surpass human capabilities. They used recordings cleaned of surrounding noise and demographic data (age, sex, etc.) to train their machine learning models. Currently, their model detects 40% of infected cases, with only 3% false positives. A promising result, but insufficient for a reliable diagnostic application.
MIT has published even more impressive results: their diagnostic model achieves 98.5% accuracy, with less than 6% false positives. The model is even capable of detecting asymptomatic cases with a false positive rate of less than 20%. The MIT researchers are already considering a large-scale application, offering free, non-invasive and instant screening, which could complement current methods of combating the spread of the virus.
Despite these remarkable results, Tomás Teijeiro, the leader of the Coughvid project at EPFL, remains cautious. He is concerned about the difficulty of reproducing these results, a common problem in data science. The adjustments required to refine the models make their reproducibility complex.
To overcome this challenge, some researchers are making their data available to the scientific community, as those at EPFL and Cambridge have done. Others are launching challenges on platforms such as Kaggle, to collaboratively test and improve models.
Although impressive advances have been made, the road ahead remains full of pitfalls. However, using artificial intelligence to diagnose Covid-19 from the sound of a cough could well become a revolutionary alternative to traditional screening methods. What remains is to resolve the problems of reproducibility to make these solutions reliable and accessible to all.
Source : ICTjournal
