Is my cough Covid-19 related or not?

The research team from Stellenbosch University is currently building a smartphone app which can in future possibly be used as a Covid-19 classifier.

Coughing can be a symptom of a range of respiratory diseases, not just of Covid-19. However, people don’t necessarily know how to distinguish between Covid-19 positive and Covid-19 negative coughs.

That is why researchers at Stellenbosch University (SU) developed a machine learning tool that can make it easier for people to distinguish the one from the other.

“In our study, we presented a machine learning-based Covid-19 cough classifier that is able to distinguish Covid-19 positive coughs from Covid-19 negative coughs and healthy coughs that have been recorded on a smartphone,” says Professor Thomas Niesler from the Digital Signal Processing Lab in the Department of Electrical and Electronic Engineering at SU.

He conducted the research with colleagues Marisa Klopper, Madhurananda Pahar and Robin Warren. Klopper and Warren are affiliated with the SAMRC (South African Medical Research Council), Centre for Tuberculosis Research, DST-NRF (Department of Science and Technology-National Research Foundation) and the Centre of Excellence for Biomedical Tuberculosis Research in SU’s Division of Molecular Biology and Human Genetics, while Pahar is a postdoctoral research fellow at the Digital Signal Processing Lab.

The findings of their study were published recently in the Computers in Biology and Medicine. As part of their research, the team used data from a global (Coswara) and a national (Sarcos – SARS Covid-19 South Africa) dataset.

The Coswara dataset contains data of Covid-19 positive people as well as physically healthy people, while the Sarcos dataset contains data of Covid-19 positive and Covid-19 negative people. In the Sarcos dataset, only people who have undergone a COVID test were asked to participate, whereas anyone could participate in the Coswara dataset.

These two datasets consist of coughing sounds, gathered from all six continents, recorded either during or after the acute phase of Covid-19. For both datasets, participants were asked to provide cough recordings via a web-based data collection platform using their smartphones.

“Our analysis of the recordings shows that Covid-19 positive coughs are 15% – 20% shorter than non-COVID coughs. Since this type of cough audio classification is cost-effective and easy to utilise, it is potentially a useful and viable means of non-contact Covid-19 screening,” says Dr Pahar.

The researchers point out that the technology being developed should not be seen as an official form of testing nor will it replace testing being done at accredited testing sites.

Classification of the coughs

The researchers cut out the silences between coughs in the recordings, allowing resource-efficient analysis of just the cough sounds, which resulted in highly accurate classification of the coughs.

“The data has been captured on smartphones, and our classifier can also be implemented on these devices. Furthermore, it could be applied remotely, thus avoiding contact with medical personnel,” says Dr Pahar. He also believes that this type of non-contact screening can reduce the workload in testing centres as well as limit transmission by recommending early self-isolation to those who have a cough suggestive of Covid-19.

The researchers explain that even though the systems they describe in their study require more stringent validation on a larger dataset, the results they have presented are very promising and indicate that Covid-19 screening based on automatic classification of coughing sounds is viable.

They also point out that several attempts have been made to identify early symptoms of Covid-19 through the use of artificial intelligence applied to images. “Respiratory data such as breathing, sneezing and coughing can be processed by machine learning algorithms to diagnose respiratory illnesses.”

The research team are continuing to enlarge their dataset and to apply transfer learning in order to take advantage of the other larger datasets.