This artificial intelligence is able to detect Covid-19 by analyzing the...

Building on previous work analyzing an individual’s voice to identify signs of Alzheimer’s disease, MIT researchers have designed an artificial intelligence capable of detecting Covid-19 from a simple cough.

An innovative diagnostic method

If the difference between the cough of a person with Covid and that of a healthy person is indistinguishable to the human ear, it can be identified by a machine learning algorithm. As part of the work presented in the IEEE Open Journal of Engineering in Medicine and Biology, a team of researchers from WITH took thousands of cough and voice samples to train his artificial intelligence, which is now able to detect people with Covid-19 with an overall accuracy of 98.5% (with 100% of asymptomatic cases identified).

Before the pandemic, researchers used this type of technology to detect signs of the disease.Alzheimer. Best known for its detrimental effects on human memory, it also tends to weaken vocal cords and often involves heightened frustration and reduced emotional expressiveness. Designed to function like a human brain, the algorithm ResNet50 was trained to distinguish sounds based on their intensity, and two other neural networks to detect emotions in the voice, as well as variations in lung and respiratory performance due to coughing.

The combination of these three systems, coupled with an algorithm capable of determining the level of muscle breakdown, gave the researchers an artificial intelligence model capable of detecting signs of the disease.Alzheimer, and a second specifically adapted to the diagnosis of Covid-19.

« The sounds of speech and cough are both influenced by the vocal cords and surrounding organs. This means that when you speak, part of your conversation is like coughing, and vice versa. », note Brian Subirana, co-author of the study. ” There is actually a feeling built into the way you cough. So we thought, why not see if Alzheimer’s disease biomarkers are relevant for Covid. »

200,000 cough samples collected

In total, the researchers collected more than 70,000 recordings of spoken words and 200,000 cough sound samples, of which 2,500 were from subjects who previously tested positive for Covid. These, along with 1,500 others, were used to train the model ofHE, and a thousand more samples used to test its accuracy. This enabled researchers to identify four biomarkers (vocal cord strength, emotion, lung and respiratory performance, muscle degradation) specific to Covid-19, whether symptomatic or asymptomatic.

The team indicated that they are actively working on the development of a smartphone application based on this technology, which must first be approved by the Food and Drug Administration, agency authorizing the marketing of food and drugs to United States, before it can be offered for download. If this is the case, then the user will just have to cough in front of their phone in order to instantly obtain information about their potential infection.

« The effective implementation of this group diagnostic tool could help stem the spread of the pandemic if everyone uses it before going to a classroom, work or restaurant. », Concludes Subirana.