MIT Team's Cough Detector Identifies 97% of COVID-19 Cases Even In Asymptomatic People 43
Scientists from MIT have developed a new AI model that can detect COVID-19 from a simple forced cough. ScienceAlert reports: Evidence shows that the AI can spot differences in coughing that can't be heard with the human ear, and if the detection system can be incorporated into a device like a smartphone, the research team thinks it could become a useful early screening tool. The work builds on research that was already happening into Alzheimer's detection through coughing and talking. Once the pandemic started to spread, the team turned its attention to COVID-19 instead, tapping into what had already been learned about how disease can cause very small changes to speech and the other noises we make.
The Alzheimer's research repurposed for COVID-19 involved a neural network known as ResNet50. It was trained on a thousand hours of human speech, then on a dataset of words spoken in different emotional states, and then on a database of coughs to spot changes in lung and respiratory performance. When the three models were combined, a layer of noise was used to filter out stronger coughs from weaker ones. Across around 2,500 captured cough recordings of people confirmed to have COVID-19, the AI correctly identified 97.1 percent of them -- and 100 percent of the asymptomatic cases.
That's an impressive result, but there's more work to do yet. The researchers emphasize that its main value lies in spotting the difference between healthy coughs and unhealthy coughs in asymptomatic people -- not in actually diagnosing COVID-19, which a proper test would be required for. In other words, it's an early warning system. The researchers now want to test the engine on a more diverse set of data, and see if there are other factors involved in reaching such an impressively high detection rate. If it does make it to the phone app stage, there are obviously going to be privacy implications too, as few of us will want our devices constantly listening out for signs of ill health. The research has been published in the IEEE Open Journal of Engineering in Medicine and Biology.
The Alzheimer's research repurposed for COVID-19 involved a neural network known as ResNet50. It was trained on a thousand hours of human speech, then on a dataset of words spoken in different emotional states, and then on a database of coughs to spot changes in lung and respiratory performance. When the three models were combined, a layer of noise was used to filter out stronger coughs from weaker ones. Across around 2,500 captured cough recordings of people confirmed to have COVID-19, the AI correctly identified 97.1 percent of them -- and 100 percent of the asymptomatic cases.
That's an impressive result, but there's more work to do yet. The researchers emphasize that its main value lies in spotting the difference between healthy coughs and unhealthy coughs in asymptomatic people -- not in actually diagnosing COVID-19, which a proper test would be required for. In other words, it's an early warning system. The researchers now want to test the engine on a more diverse set of data, and see if there are other factors involved in reaching such an impressively high detection rate. If it does make it to the phone app stage, there are obviously going to be privacy implications too, as few of us will want our devices constantly listening out for signs of ill health. The research has been published in the IEEE Open Journal of Engineering in Medicine and Biology.
Every time (Score:4, Insightful)
What's the sensitivity, what's the specificity? Without both values, "97%" is meaningless.
Re:Every time (Score:5, Informative)
What's the sensitivity, what's the specificity? Without both values, "97%" is meaningless.
Yeah, I can make a cough detector that detects 100% of covid cases, even before someone has been infected with covid.
Lets all pile on without reading the fine article because [ieee.org]
Sounds so exactly like "just declare them all COVID".
Re: (Score:3)
What's the sensitivity, what's the specificity? Without both values, "97%" is meaningless.
Yeah, I can make a cough detector that detects 100% of covid cases, even before someone has been infected with covid.
Lets all pile on without reading the fine article because [ieee.org]
Sounds so exactly like "just declare them all COVID".
Arrgh. I missed that link at the bottom of the sciencealert piece.
OK, well that helps.
Re: (Score:2)
"Lets all pile on without reading the fine article because"
Pass. After all, we're not newbies with a 7 figure uid.
Re: (Score:2)
Sorry to have to break this to you, but you're a newbie. I have a lower UID than you, but even my UID is well into newbie land.
Re: (Score:2)
Re: (Score:2)
Eh, sonny?
Re: (Score:2)
At least a 2 figure uid should wake up, we'll be off your lawn immediately after, Sir.
Re: (Score:2)
Sounds so exactly like "just declare them all COVID".
You are being sarcastic, right? For those that are not statistically literate, the numbers mean that 6-13% of uninfected people will get a false positive result and will need to get a lab test to be sure. That is hardly the same as a test that simply says everyone is infected.
Re: (Score:2)
Sounds so exactly like "just declare them all COVID".
You are being sarcastic, right?
Yes of course. But just let me put that to rest knowing how it's become impossible to be sure any more.
Re:Every time (Score:5, Informative)
When validated with subjects diagnosed using an official test, the model achieves COVID-19 sensitivity of 98.5% with a specificity of 94.2% (AUC: 0.97). For asymptomatic subjects it achieves sensitivity of 100% with a specificity of 83.2%.
.
Re:Every time (Score:5, Interesting)
Those numbers look pretty good.
Ideally we would test everyone every day. That's impractical but testing 20% of people in high risk areas like offices is possible.
Re: (Score:3)
Re: (Score:2)
Never trust a Percent.
Percentages rarely ever show us anything really interesting, they are other numbers that we really need for a breakdown
If their cough sounds different... (Score:4, Insightful)
Re: (Score:2)
A new AI model that can detect COVID-19 from a simple forced cough.
Re: (Score:2)
Re: (Score:2)
Ok, I thought you meant that if you cough naturally, you are not asymptomatic (I asked myself the same question).
But in this case, people don't notice the difference, only the AI does.
Evidence shows that the AI can spot differences in coughing that can't be heard with the human ear
It is up the the definition of a symptom.
My understanding is that a symptom is apparent to a patient as opposed to something that is measured by a machine. For example, low SpO2 is a sign of COVID-19, but not a symptom. It is something that is measured but that is not directly apparent to the patient. Related shortness of brea
Re:If their cough sounds different... (Score:4, Informative)
Right. Their forced cough sounds different. How is that not a symptom?
Technically that'd be a sign, not a symptom - a subtle and fairly unimportant difference, but since you asked.
Generally symptoms are those things that patients notice and complain about to their doctor (e.g. I have a cough). Signs are the things the doctors look/test for (e.g. ask them to cough and pick up on characteristic sounds).
Re: (Score:2)
"A new AI model that can detect COVID-19 from a simple forced cough."
My boss is able detect such fake sickness coughs over the phone for years now.
And he ain't no artificial intelligence, much less a natural one.
Re: If their cough sounds different... (Score:2)
I wondered if there isn't some noise in the data from the fact that the validated patients know whether they have COVID. Perhaps one's forced coughs sounds different when one thinks one's sick. It would be ideal to record the coughs just before people are tested.
Re: If their cough sounds different... (Score:2)
I thought that as well. However it does say that it is a forced cough. Just like some do when the doctor tells you to cough during certain exams.
But but... (Score:1)
Re: But but... (Score:1)
Re: (Score:2)
Re: (Score:2)
FFS (Score:2, Funny)
Re: (Score:2)
Why do people think this will be some passive "always listening" type of thing? There's absolutely no reason for it to be doing that. Not only is it unnecessary to have 24/7 monitoring, it's going to be overwhelmed by ambient noise if you're not coughing directly at it.
If this becomes an app, it's going to be something where you turn it on, cough at it, get the results, and turn it off. They'll suggest you check it once or twice a day, and if you get a positive to go in for a real test.
Great (Score:2)
When can I get an app that can listen to ambient coughs and do the same thing? Every time I go shopping there's tons of people coughing and sneezing.
Small sample size (Score:2)
Very valuable if can be used with voluntary speech (Score:2)
Of most value would be to be able to test noises a person voluntarily makes such as voluntary cough or speech, so you can have people use a phone app or such to test themselves regularly. If reliable, it could be a game changer for reducing the R0 of the virus and making the numbers go down.
This is such bullshit-"science" (Score:5, Insightful)
Really, this paper is an emberrasment of bullshit-science. But hey, "it's got an app and AI!".
Re: This is such bullshit-"science" (Score:2)
Also need to add confounders, such as other respiratory illnesses (RSV, influenza, etc). Iâ(TM)m willing to bet the specificity reported is more respiratory infection vs healthy rather than SARS-CoV-2 vs not SARS-COV-2. It may still be useful in some scenarios, but this definitely isnâ(TM)t the rapid definitive test we want it to be.
It's not the cough that carries you off... (Score:2)
It's the coffin they carry you off in.
"In Asymptomatic People" (Score:2)