Academics claim their AI software can detect, with 98.5 per cent accuracy, whether or not someone has caught the COVID-19 coronavirus, just from the sound of their coughing.
To build this software, the MIT team used three ResNet50 models, a popular convolutional neural network designed by Microsoft. They’re normally used to process images for computer vision, though in this case they’re analyzing audio.
The boffins produced a dataset of 5,320 people, who in April and May submitted audio clips of themselves coughing. Participants also had to fill out a questionnaire that asked if they had caught the coronavirus or not, if they had confirmed this with an official test or not, and what symptoms they had. Thus this experiment relies on the honesty of these human data sources, so bear that in mind.
The dataset was trimmed down to 5,000 recordings, half of them by those who said they tested positive for COVID-19, and the other half negative. Four-fifths of the samples were used to train the model, and the remaining clips were to test the model.
“We have created an AI pre-screening test that discriminates 98.5 per cent COVID-19 positives from a forced-cough recording, including 100 per cent of asymptomatics, at essentially no cost and with an accompanying saliency map for longitudinal explainability,” the researchers wrote in a paper, published in the IEEE Journal of Engineering in Medicine and Biology, detailing the software project.
IBM: Our AI correctly predicts onset of Alzheimer’s 71% of the time, better than standard clinical tests
The results appear promising enough that the team said they are working with a Fortune 100 company to flesh out their model into a fully fledged diagnostic tool.
“We think this shows that the way you produce sound changes when you have COVID, even if you’re asymptomatic,” added Brian Subirana, co-author of the paper and director of the AutoID Lab at MIT, this week. They believe a smartphone app will provide a quick and easy way to test for the coronavirus that is free to run, and requires no invasive nasal swabbing or however which way you want to directly detect the presence of the virus.
“The effective implementation of this group diagnostic tool could diminish the spread of the pandemic if everyone uses it before going to a classroom, a factory, or a restaurant.”
This sort of application, once it gets the necessary regulatory approval, is most likely going to end up being a screening tool: cough into the mic, and if it thinks you have the virus, it’ll advise you to get a proper medical test. The uni’s AutoID Lab earlier developed a speech-analyzing model to screen people for Alzheimer’s testing, though turned their attention to COVID-19 as the pandemic took hold.
“The sounds of talking and coughing are both influenced by the vocal cords and surrounding organs,” Subirana said.
“This means that when you talk, part of your talking is like coughing, and vice versa. It also means that things we easily derive from fluent speech, AI can pick up simply from coughs, including things like the person’s gender, mother tongue, or even emotional state. There’s in fact sentiment embedded in how you cough.
“So we thought, why don’t we try these Alzheimer’s biomarkers [to see if they’re relevant] for COVID.” ®