When your smartphone knows you have COVID-19

The fact that smartphones know an awful lot about us is nothing new. And many people have long since accepted that their devices — or more precisely the providers behind the apps  that they use — have probably learned more about them than the users know.

Soon, a new app could be launched that will allow users to diagnose an asymptomatic coronavirus infection based on the sound of their coughing or speaking.

Though the AI technology has already proven itself to be able to identify the coughs of infected individuals, it still has to learn more to avoid misdiagnosing noninfected people. If it succeeds in doing so, the app could at some point complement Germany's existing coronavirus tracing app,  which traces contacts with infected persons. 

 

Coughing sounds contain biomarkers

Three computer scientists from the Massachusetts Institute of Technology (MIT)   came up with the idea of using sound analysis to detect infections. Jordi Lugarta, Ferran Hueto and Brian Subriana took sound recordings of 5,320 people in April and May - both participants infected with the coronavirus and participants without. In addition to coughing sounds, they analyzed speech. First, the researchers fed the sound data of 4,256 participants into the computer, which analyzed them with the help of a convolutional neural network.

The researchers were looking for acoustic biomarkers. Those include certain characteristic features in the sounds that they had already found in earlier studies on Alzheimer's patients. They then tested what the machines had learned from samples on the remaining 1,064 participants of the study.

Many hits — but please do not panic in case of false positive results!

The results were quite promising. "When validated with subjects diagnosed using an official test, the model achieves COVID-19 sensitivity of 98.5%," the researchers wrote in their study, which was submitted to the IEEE Open Journal of Engineering in Medicine and Biology  for publication.

The specificity was 94.2% in the group. This would mean that about 1 in 20 people taking such a test would have received a false positive result.

"For asymptomatic subjects, it achieves sensitivity of 100% with a specificity of 83.2%," the authors report. This would mean that every otherwise potentially undetected COVID-19 case was correctly diagnosed. However, almost one in five participants received a false positive warning. 

In other words, the specificity would certainly have to be improved in order to make such an app really workable in practice. If it is used by a large number of people every day, the sheer number of false positive results could otherwise quickly lead to a run on COVID-19 test laboratories and exhaust their capacities.

It is possible, however, that the specificity values could be further improved if the computers were fed with more data and learned more about coughing sounds with deep learning technology.

No replacement but supplement to laboratory tests

In any case, the researchers concluded that "AI techniques can produce a free, non-invasive, real-time, any-time, instantly distributable, large-scale COVID-19 asymptomatic screening tool to augment current approaches in containing the spread of COVID-19."

The MIT scientists suggest that the method could be used to test students, workers and office employees on a daily basis. Should a warning come up, the people concerned could still do a laboratory test to find out if the app was right.