Coughing noises from more than 5,300 test persons
Researchers Jordi Laguarta, Ferran Hueto and Brian Subirana, who work at the Massachusetts Institute of Technology (MIT), used machine learning to analyze the coughing noises of more than 5,300 test subjects. They collected over 70,000 recordings on a specially set up website, both of people suffering from COVID-19 and of healthy people. The infected participants also provided precise information about the symptoms they had experienced or about an asymptomatic course. About half of the recordings came from people who had tested positive for SARS-CoV-2.
98.5 percent of symptomatic infections detected
The MIT scientists trained an AI framework with the recordings they had collected, which included mostly deliberately generated coughing noises. This was based on a tried and tested procedure that is used in Alzheimer’s research. After completing the training, according to the researchers, the artificial intelligence was able to detect 98.5 percent of all symptomatic infections. Asymptomatic participants were even 100 percent identified.
Researchers develop a smartphone app
Laguarta, Hueto and Subirana now want to create an app based on the AI model they developed. The application should be available free of charge to owners of iPhones and Android smartphones and thus enable simple regular self-tests. The researchers also assume that their analysis function can also be integrated into smart speakers such as Apple’s HomePod or voice assistants such as Siri or Alexa. In addition, they continue to work on improving their AI process, including in collaboration with a number of US hospitals. The scientists published the results of their work so far in IEEE Open Journal of Engineering in Medicine and Biology.