MIT lab to build AI software to diagnose COVID-19

A team at MIT developed an AI software that should be able to detect at a 98.5% accuracy if someone has been infected with the COVID-19 virus from listening to their coughing.

In order to develop it, they are using three ResNet50 models, a convolutional neural network, analyzing audio. The scientists carried out research on around 5000 people in April and May, listening to their coughing, asking them if they had caught the virus, and what symptoms they had. After trimming the dataset, they used 4/5 of the samples to train the model, while the rest were employed to test the model. However, bear in mind that these data sets are based on human honesty.

With all of this, they were able to create an AI pre-screening test that can differentiate between actual COVID positives from forced coughs, including asymptomatic people. The results seemed rather promising and the team is now working on developing their model into a fully-fledged diagnostic tool.

The scientists believe that creating a smartphone app for their model will provide a fast, easy, and free method to test people to know if they caught the virus or not. In the long run, this could decrease the spread of the pandemic.

Once the app gets approved, it would work as a screening tool. You would only have to cough into the mic and the app will tell you if it thinks you have the virus or not and it’ll advise you to get a proper medical test. The team believes that the virus affects the way we sound and thus, by coughing, the AI would be able to recognize a change in the voice.