Results and future of Covid-19 detection in cough project

Published on 11/11/2021

Detecting Covid-19 by simple voice analysis is a field the Luxembourg Institute of Science and Technology (LIST) has been studying for the past few years with interest peaking last year, including in the media, as the public was invited to participate in an oral survey.

Known as "COVID-19 Detection by Cough and Voice Analysis" or CDCVA, in partnership with the Luxembourg Institute of Health (LIH) and University of Luxembourg, the project’s initial phase has come to an end with the results published, as LIST’s project leader Muhannad Ismael explained. “The project finished on 15 April. We had extensions as it was a project for 6 months, and the results have now been published in Computers in Biology and Medicine journal”.

Indeed, the paper with the results entitled “Detection of COVID-19 from voice, cough and breathing patterns: Dataset and preliminary results” was sent to ScienceDirect in August and published online in October.

So, what were the results? And what was the uptake of the CDCVA survey by the public?

Muhannad laid out three major achievements of the project:

New crowd-sourced dataset of coughs, speech and breathing samples from people can diagnose positive cases of Covid-19 and healthy control individuals
Preliminary results for detection of Covid-19 could be ascertained from cough patterns using standard acoustic features sets, wavelet scattering features, as well audio embeddings extracted from low level feature representation
The CDCVA models achieved an accuracy of 88.52%, with sensitivity of 88.75% and specificity of 90.87%.

The data was collected from five different types of sound: coughing, counting, reading a sentence, breathing in/out, and the vowel sound “ahh”. It came from 1,103 samples of volunteers who participated in the survey.

“We have five different types of sound, but for the moment we have just analysed the cough and not done investigation in other data, so now what would be really interesting is to merge the sounds for more data,” said Muhannad. “What is important to mention is that after removing items with either missing audio files or sound with background noise, a total of 1,103 participants were identified in the study, therefore, we listened to more than 1,103 for coughs and didn’t have time to do for all the data set. Why coughs? Because it is very revealing for respiratory diseases. One of the main symptoms of Covid-19 is a dry cough so this is why we decided to focus on coughs for the moment”.

Once the oral data is collected and before it is classified, the project possesses tools that permit the visualisation of signal features derived from coughs of people diagnosed with Covid-19 and healthy subjects in the form of graphs .

Muhannad explained further. “We have an application where we have participants - not a really high number but up until 25 March there were 1,103 people - taking part with 7.62% of them proving to be positive. After that, different methods were applied. We took the sound signal, and then produced spectrograms or extracted the standard acoustic features sets or wavelet scattering features, after which they were applied in different methods of classification”.

The next step - need more data!

While the project has initial results, it has not been validated, “but we have made a step forward, and our data engine is already complete,” Muhannad said. “We need more data to continue to validate our method. What was slightly negative is we did not have a huge number of people taking part. Currently We have a data set of people who are healthy or have Covid-19, so what we now need is more data from more people with other illnesses like the common flu for example”.

Having more people with normal flu or other repertory illnesses taking part would help to identify or isolate them from the data and therefore allow the study to focus purely on Covid-19.

The next step would then be is to propose a complete prototype, outlined Muhannad. “So we could have an application that takes a cough sample sound at the start, and at the end give a result. The idea would be that we inform people to try our application to determine if they have Covid-19 or not”.

How to take part

You are invited to take part in the voluntary online survey which is still open and totally anonymous. It ensures personal data security and integrity in compliance with the applicable data protection legislation. By taking part, you will be supporting LIST, LIH and Luxembourg University with the CDCVA project by supplying more data to help move it to the next step.

Full details are explained on the CDCVA website: cdcva.list.lu