Cough as efficient and rapid diagnostic tool for COVID-19
We learnt that, to prevent the uncontrolled spread of COVID-19, it is necessary to identify new cases in good time. At the moment the tests are performed by carrying out nasal swabs, from which the presence of the virus is measured in the laboratory. In addition, it is possible to check for the presence of antibodies in the blood, which would confirm a previous infection.
Are there other ways to diagnose COVID-19? The artificial intelligence is offering new insights. Artificial intelligence is defined as computerized devices capable of being trained to perform tasks belonging exclusively to human intelligence. In our case, we will see an example of how a machine can be trained to accurately diagnose COVID-19.
At the Massachussetts Institute of Technology (MIT) in the USA, an algorithm has been developed that can recognize the typical cough of a patient suffering from COVID-19. The difference in sound between a “normal cough” and a “COVID cough” is imperceptible to the human ear, but the algorithm developed by Brian Subirana and colleagues is able to correctly diagnose the latter in 98.5% of patients positive for the virus (giving a false negative result in only 1.5% of cases). Furthermore, the specificity, which is the probability that a healthy individual will test negative, is 94.2% (so there is a 5.8% probability of a false positive result).
How was the algorithm trained? In April 2020, researchers set up a site to collect (forced or spontaneous) cough samples from volunteers (https://opensigma.mit.edu/, still accessible and available in English or Spanish). In few months, they collected about 70,000 samples, of which 2,660 from patients diagnosed with COVID-19. After sampling, participants were asked for general and other information about any symptoms, including: fever, fatigue, sore throat, difficult breathing, persistent pain or pressure in the chest, diarrhea and cough. For the study, all the samples of people affected by COVID-19 and a representation of healthy people equal to the number of sick people (2660 + 2660) were selected. Of the 5320 selected samples, 80% (4256) were used for training, while the remaining 1064 were tested as “final exam”.
If this test is approved, it could be translated into a smartphone app to be used for daily screening of students and workers (given the immediacy of the response), or for random sampling in order to quickly detect new outbreaks.