I think every general practice in the first world would be able to afford $200 a month for a POC lung ultrasonography Stuie1977.
https://www.lumify.philips.com/web/clinical-solutions/office-practice
Leaving that to one side. I think this announcement today is breaking two little bits of news.
The first news is that RAP’s sensitivity and specificity statistics strictly speaking weren’t.
Sensitivity and specificity estimates are unbiased estimates of accuracy but only when calculated from a perfect reference gold standard.
In the present case the gold standard would be computed tomography (CT) (not chest x-rays). But this exposes children to radiation and so it would be unethical to require children to undergo these.
So RAP have been reporting sensitivity and specificity statistics calculated from a non-reference standard.
The FDA says this is misleading:
View attachment 620847
https://www.fda.gov/downloads/Medic...onandGuidance/GuidanceDocuments/ucm071287.pdf
Thats the first bit of “news”. And from this flows the next bit of news.
Agreement statistics depend on what exactly that non-reference standard is. And in this case there is a little issue that the US clinical trial uses a more rigorous case definition as the comparator.
In contrast the “clinical diagnosis” used in RAP’s previous studies (not trials) is a rag bag of heterogeneity as standard of care always is.
And so when RAP reverse code from the WA cough data to the more robust “clinical diagnosis” that is being used in the US trial they get a nasty surprise - the agreement statistics deteriorate.
With these two bits of news the mystery of the sample size change is unlocked. True sensitivity / specificity statistics calculated from a gold standard are invariant to prevalence. But agreement statistics are very sensitive to prevalence. Prevalence lower than expected means that accuracy will be poorer.
So what has caused all this.
RAP has a IT / Engineering pedigree – not health. The difference between reference and non-reference standards (and the statistical implications) that the FDA think is very important wouldn’t be appreciated by anyone outside health.
In the same way the difference between a “study” and “trial” and how that specification of “clinical diagnosis” changes things wouldn’t be appreciated outside health.
The next potential misunderstanding – or elephant in the room is captured by the Yogi Bear line ... “its tough to make predictions, especially about the future”.
You can machine learn on cough datasets till the cows come home. But can you make a prediction about the future? Much bigger difference than anyone here thinks..