AI Think That Cough Sounds Bad
AI can rapidly detect COVID-19 infections with your smartphone for essentially no cost.
One of the biggest challenges in controlling COVID-19 is in rapid detection of infected individuals. By detecting cases early, infected individuals can self-quarantine and help to reduce spread of the disease. However, viral and serologic testing is expensive on a large scale — it would cost roughly $8.6 billion to test the entire population of the United States. With that price tag, frequent large-scale testing is not reasonable.
A collaboration between MIT and Harvard University took a different approach to testing that has the potential to drive the cost to near zero and also nearly eliminate the time and effort one needs to invest to get a test. Focusing specifically on asymptomatic carriers of COVID-19, the researchers hypothesized that asymptomatic carriers and healthy individuals may differ in the way that they cough. To test this hypothesis, they gathered audio recordings of coughs from individuals that tested both positive and negative for COVID-19.
The resulting dataset, composed of recordings from over 5,300 individuals, was used to train an artificial neural network to distinguish between COVID-19-positive and healthy individuals. Testing showed the model to accurately classify 98.5% of COVID-19 positives overall, and very impressively classified 100% of asymptomatic coughs (asymptomatic carriers, by definition, do not have a cough, so they were asked to force a cough). These results are considerably more accurate than traditional testing methods.
The research team sees this model being built into a smartphone app for widespread daily use. The cost of each test is essentially zero, and the burden on an individual to complete the test (i.e. force a cough into their phone) is minimal. Lowering the barriers substantially, this tool makes testing large populations on a daily basis realistic. Such rapid, widespread testing has the potential to prevent outbreaks before they happen.
At present, the researchers are working to gather a larger dataset to further fine-tune their model. They note that there are differences in the way people cough that differs by, for example, age and culture. Additional validation will be required to see how well this model will generalize to larger and more diverse populations.
