MIT researchers have developed a model which can distinguish people with asymptomatic Covid-19 from healthy people, using the sound of their coughs. These differences are impossible to detect with the human ear.
The MIT team, based at the Institute’s Auto-ID Laboratory, had been developing models to detect signs of Alzheimer’s using these recordings; as the coronavirus pandemic emerged, the researchers decided to investigate whether a similar tool could be applicable to Covid-19.
In April, they collected as many recordings of coughs as possible, setting up a website where anyone could submit recordings of forced coughs, along with information about their symptoms and test results. The researchers collected more than 200,000 forced-cough samples including 2,500 from people confirmed to have Covid-19.
These Covid-19-positive recordings and 2,500 recordings from healthy participants formed a dataset. The scientists used 4,000 to train the model and the remaining 1,000 to test it.
With minimal tweaking to the AI framework intended for detection of Alzheimer’s, the researchers found that they could pick up patterns in four biomarkers (vocal cord strength, sentiment, lung and respiratory performance, and muscular degradation) specific to Covid-19.
When fed new recordings of coughs, the model was able to identify people diagnosed with Covid-19 with 98.5 per cent accuracy, and 100 per cent accuracy for people with asymptomatic Covid-19.
“We think this shows that the way you produce sound changes when you have Covid, even if you’re asymptomatic,” said Professor Brian Subirana, director of the Auto-ID Laboratory.
The team are partnering with several hospitals to collect a larger and more diverse set of cough recordings to boost the accuracy of the model. They are also working with a company to incorporate the AI model into a user-friendly pre-screening app, which could provide a free, quick and non-invasive tool for identifying people with asymptomatic Covid-19. People whose coughs suggest they may have Covid-19 could then confirm with a test.
Eventually, the model could be incorporated into smart speakers and other listening devices.
“The effective implementation of this group diagnostic tool could diminish the spread of the pandemic is everyone uses it before going to a classroom, a factory, or a restaurant,” said Subirana.