No, AI can't tell if you've got COVID-19 by listening to your coughs Machine learning algorithms cannot accurately predict whether someone has COVID-19 by analyzing the sound of their coughs, according to a study led by the UK's Alan Turing Institute. Claims that AI could detect the difference in cough sounds between those with and without COVID-19 with up to 98.5 per cent accuracy were first …
Yes, investigated...
But the original claim already had severe problems passing the sniff test. It was a too good to be true story from the beginning and it seemed to cough up results that were better than the medical profession's collective wisdom.
But hey, AI/ML systems are so good at blinding. They have blinkenlight so the results must be fantastic.
/s(igh)
No, peer review did not spot it - it was published in a peer-reviewed journal in 2020 (and has not been retracted). The particular statistical fail in the article - confounding due to recruitment bias - is a well-known "gotcha" that should have been picked up by a competent reviewer*.
The results were challenged - as the Reg article says - by a much later, independent study in a (presumably as-yet unreviewed) preprint.
* As it happens, I have personal reason to be unimpressed with the quality of peer review at an IEEE journal (which will remain nameless - not the one cited above).
Peer review is what's done at the point of it still being a research paper by peers of the researchers testing the idea and checking it works. It's supposed to be done before you publish claiming that it worked. The idea is that objectively false things don't get published in reputable journals.
If peer review had of spotted the problem, then it wouldn't have been published. If it hadn't have been published then UK PLC wouldn't have blown a hundred grand trying to make something based on the results of the paper.
What it demonstrates is that the IEEE "rigorous peer review" is in fact not particularly rigorous, which makes the "journal" more akin to a magazine.
Indeed. I alluded to this in a comment above. I have been on both sides of the review process, and had a particularly infuriating experience reviewing for an IEEE journal (there are many - it's a huge organisation). I had, on invitation, reviewed a paper for one of their journals, and pointed out many flaws in the submission (largely, as it happened, of a statistical nature). The editor chose to completely ignore my review, and had the effrontery to suggest I collaborate with the original author on a watered-down commentary on the submission. Obviously I refused. The paper was published anyway, unrevised. I tried to escalate the matter up the IEEE hierarchy (they're extremely bureaucratic), but was stonewalled at every stage. There was more to the story which I can't comment on publicly, but it left a nasty taste.
To be fair, this was an unusual experience. Most journals I've published in or reviewed for take the review process seriously.
Covid is just one of them. Listening to some people around here, they sound like they'll be corpses in a matter of days, but they've just been out for (yet another) fag.
That they come back sounding like a plague victim is just their body failing at telling them to stop.
Ironically, the people that I know around here who had Covid tended not to cough. They just ran a fever, felt awful, and looked like they'd pass out any minute. At least before they realised they really ought not stay at work... maybe the coughing comes later?
The most annoying thing is the time the cough lasts because it trains your body into coughing for the slightest tickle in your throat afterwards. I discovered that after having to suppress it during video calls (not for the noise, but once it starts it takes a few seconds) - it got easier after a while.
I guess there's truth after all in the use of a strong laxative as cough medicine* :)
* Because you wouldn't dare coughing then..
"But as we continued to analyse the results, it appeared that the accuracy was likely due to an effect in statistics called confounding – where models learn other variables which correlate with the true signal, as opposed to the true signal itself"
I'm reminded of that chapter in Asimov's Foundation, where the declarations of some diplomat were passed through a logical analysis procedure to conclude that said diplomat had said absolutely nothing of substance or significance.
I still think we need a tool like that.
I believe we're endowed with that tool already. It's honed by listening to Sir Humphrey!
Can you imagine the legislation MPs would wanted pushed through if that tool were proposed for the Commons, or used by broadcasters for politicians' gibberings? Non-partisan cries of "parliamentary privilege"!
"I'm reminded of that chapter in Asimov's Foundation, where the declarations of some diplomat were passed through a logical analysis procedure to conclude that said diplomat had said absolutely nothing of substance or significance."
Oddly enough, that's one of the bits I remember from Asimov too. maybe we both read it at a time in our lives where it was highly appropriate and so it stuck. Semantic analysis was the term IIRC.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
