Plenty of people in 2021 and 2022 had a cough, wondered whether it was COVID-19, and did not get tested, either because test kits were scarce or because a crowded testing centre seemed like a good place to catch the virus. A 2021 paper in Computers in Biology and Medicine by Madhurananda Pahar, Marisa Klopper, Robin Warren, and Thomas Niesler asked whether a phone recording of that cough could stand in for the first screening step.
The problem
Dry cough is one of the most common COVID-19 symptoms, along with fever and fatigue. It is also a symptom of more than a hundred other conditions, so the question is not whether a person is coughing but whether the cough itself carries a signature that separates COVID-19 from everything else. The researchers set out to build a machine learning classifier that distinguishes COVID-19 positive coughs from COVID-19 negative and healthy coughs in smartphone audio.
The data
Two datasets were used. Coswara, a public dataset of crowdsourced recordings, contributed 92 COVID-19 positive subjects and 1,079 healthy ones. Sarcos (SARS COVID-19 South Africa), collected by the researchers themselves, contributed 18 positives and 26 healthy subjects. Both are heavily imbalanced toward healthy coughs.
The approach
The imbalance was addressed first, using the synthetic minority oversampling technique (SMOTE) to generate additional positive examples. In pre-processing the researchers noticed that COVID-19 positive coughs were 15 to 20 percent shorter than the others, so they designed the feature extraction to preserve the time-domain pattern across a whole cough rather than averaging it away. Each pre-processed cough was split into segments and features such as MFCCs, log energies, zero crossing rate, and kurtosis were extracted from each.
Seven classifiers were trained and tuned with leave-p-out cross validation: logistic regression, k-nearest neighbours, support vector machine, multilayer perceptron, a convolutional neural network, a long short-term memory network (LSTM), and a residual network (Resnet50).
Results
On Coswara, all seven classifiers could separate COVID-19 coughs to some degree. Resnet50 did best, with an area under the ROC curve of 0.976, accuracy of 95.3 percent, sensitivity of 93 percent, and specificity of 98 percent. The harder test was transfer: training on Coswara and evaluating on the separately collected Sarcos data. There the LSTM performed best, reaching an AUC of 0.94 using the 13 best features chosen by sequential forward selection.
The authors are careful to say that larger datasets are needed before any of this is relied on. What the results establish is that COVID-19 screening from automatically classified cough sounds is possible.
Why it matters
A classifier like this would let someone screen themselves with nothing but a phone, at home, with no contact. Beyond convenience, it could reduce the load on testing centres and the transmission that happens in the queues outside them.