Further investigation of differences in COVID diagnoses as derived from application of machine learning (MLP) and as recorded in the EPR of an unknown NHS Trust
Yes, the timing is a bit of a giveaway. Interesting that vaccination status sits low in terms of importance and this makes me wonder just how accurate this information is. I've an idea to follow through with a similar approach for predicted vaccination status.
It must be 25 years since I last played with ANNs and MLPs. ISTR that the choice of a “training set” was critical, and overtraining (= effectively memory) was a risk. Forgive me if I’ve missed it but how are these issues dealt with these days and here?
Still the same problem which is why I run a parallel analysis using logistic regression (and discriminant analysis or factor analysis on occasion), but being lazy I don't bother to type everything up! If I'm getting similar results I start to believe what the black box is saying, though I tend to mainly use it as a short cut to variables of interest rather than a means in itself.
If you flip back to 'The PCR Test As A Predictor Of Acute Respiratory Conditions (part 3)' you'll be able to squint at MLP and LR tackling the same problem - it was important to be confident that I wasn't failing about!
We may be looking at Christmas again. Instead of covid tests being done by people whose jobs required them to get tested routinely and frequently, and people with symptoms, you get a large number of them done by people who simply want to know they don't have covid so they can visit their elderly relatives. Lots of these people will end up as false positives. When they then get sick with whatever, they have a recent positive covid test to add to the confusion. Maybe the diagnosing doctors stayed confused and believed the false test results?
I am sure that is so but do remember that the sample I'm dealing with are very sick in-patients, all of whom died, rather than anyone doing any visiting or working!
Are you aware that I'm analysing a dataset of in-hospital deaths only? When you say "died quickly" I presume you mean within a few days of the jab. This is a large area of work that I've scheduled for later since it takes us into the realm of survival analysis.
Yes. I am aware. But what we have to explain is a large number of patients who got the covid death diagnosis when the ML diagnosis would disagree. One hypothesis is that the ML is missing them, but another is that they died of something else. But would telling the people who admitted you to hospital that you tested positive for covid over Christmas be enough to get you the covid death diagnosis, even thouugh it was something else that was killing you and the positive was a false positive anyway? That's why, when you mentioned 'not visiting or working', I was interested in how quickly these people died. If they had been sick enough to be in hospital over Christmas, then I would think that they are unlikely to be getting their false positive test that week. But how many of the people who died in the first 10 weeks weren't in hospital over Christmas? They are potential candidates for 'felt well enough to be visiting and working'. Would getting a false positive over Christmas strongly predispose the doctors to treat you for covid even though it was something else that was killing you and brought you to the hospital? And when you died of whatever that was, call you a covid death because that was what you were being treated for? Pure speculation on my part, of course.
Well, we shall have to wait and see what unfolds over a number of articles! I'm currently checking what MLP thinks against classic logistic regression to rule out any funny business. I'm also updating case detection rate data with something that better approximates disease prevalence within hospitals. Then I'll be looking at that 10 week hump assuming vaccination status is correct. Reports suggest vaccination status is also in error, so I'll be trying to figure a way of predicting any missing cases. Once all this is done I can start looking at survival in terms of delays between vaccination and death.
However, in all this don't forget that I'm not analysing cause of death - that red spike should not be construed as COVID death, these merely being deaths with mention of COVID - the declared cause of death remains unknown. Also unknown is how and when a COVID diagnosis was given. These could be folk testing positive at home and coming in unwell or they could be inpatients that tested positive at some point. Unfortunately I don't have date or reason for admission so we cannot determine the situation.
- Maybe further speciate "disagreement" into its two subcategories (i.e. "EPR says yes, MLP says no" versus "EPR says no, MLP says yes"). Perhaps that might help vaccine Importance pop?
- Curious about the relationship between AcuteResp Dx and disagreement status. The above speciation may help investigate that too.
- In case you missed it, maybe there is a chance of calculating covid prevalence within the trust:
LOL - I did exactly that yesterday before shutting down my workstation and called it COVagree2. Today I'm typing up my logistic regression work to see if we get the same peak discrepancy (MLP is a bit of a black box). Preliminary runs reveal the importance of acute resp.
Jibby-jabbies making folks both more likely to get Covid and increasing mortality?
Possibly, though I'm not so certain that big red COVID spike is real.
More likely jab deaths wrongly attributed to COVID since the symptoms are similar enough and the NPCs have only been programmed in one direction.
Yes, the timing is a bit of a giveaway. Interesting that vaccination status sits low in terms of importance and this makes me wonder just how accurate this information is. I've an idea to follow through with a similar approach for predicted vaccination status.
It must be 25 years since I last played with ANNs and MLPs. ISTR that the choice of a “training set” was critical, and overtraining (= effectively memory) was a risk. Forgive me if I’ve missed it but how are these issues dealt with these days and here?
Still the same problem which is why I run a parallel analysis using logistic regression (and discriminant analysis or factor analysis on occasion), but being lazy I don't bother to type everything up! If I'm getting similar results I start to believe what the black box is saying, though I tend to mainly use it as a short cut to variables of interest rather than a means in itself.
If you flip back to 'The PCR Test As A Predictor Of Acute Respiratory Conditions (part 3)' you'll be able to squint at MLP and LR tackling the same problem - it was important to be confident that I wasn't failing about!
Oh dear, that 2021 spike looks rather , er , spikey!
Sure does, but it may be a big pile of false positives!
I don't trust anyone absolutely.
Your cosy tea and biscuits cover story may be a scam.
For all we really know, you were fed all that false data by a Russian agent, attempting to discredit our proud nation.
I mean, obviously that is only one conspiracy theory. LOL
Always best to check the colour of a person's socks.
We may be looking at Christmas again. Instead of covid tests being done by people whose jobs required them to get tested routinely and frequently, and people with symptoms, you get a large number of them done by people who simply want to know they don't have covid so they can visit their elderly relatives. Lots of these people will end up as false positives. When they then get sick with whatever, they have a recent positive covid test to add to the confusion. Maybe the diagnosing doctors stayed confused and believed the false test results?
p.s. this is really interesting work, John.
I am sure that is so but do remember that the sample I'm dealing with are very sick in-patients, all of whom died, rather than anyone doing any visiting or working!
That's something I couldn't tell. How many of these 'died in the first 10 weeks' died quickly?
Are you aware that I'm analysing a dataset of in-hospital deaths only? When you say "died quickly" I presume you mean within a few days of the jab. This is a large area of work that I've scheduled for later since it takes us into the realm of survival analysis.
Yes. I am aware. But what we have to explain is a large number of patients who got the covid death diagnosis when the ML diagnosis would disagree. One hypothesis is that the ML is missing them, but another is that they died of something else. But would telling the people who admitted you to hospital that you tested positive for covid over Christmas be enough to get you the covid death diagnosis, even thouugh it was something else that was killing you and the positive was a false positive anyway? That's why, when you mentioned 'not visiting or working', I was interested in how quickly these people died. If they had been sick enough to be in hospital over Christmas, then I would think that they are unlikely to be getting their false positive test that week. But how many of the people who died in the first 10 weeks weren't in hospital over Christmas? They are potential candidates for 'felt well enough to be visiting and working'. Would getting a false positive over Christmas strongly predispose the doctors to treat you for covid even though it was something else that was killing you and brought you to the hospital? And when you died of whatever that was, call you a covid death because that was what you were being treated for? Pure speculation on my part, of course.
Well, we shall have to wait and see what unfolds over a number of articles! I'm currently checking what MLP thinks against classic logistic regression to rule out any funny business. I'm also updating case detection rate data with something that better approximates disease prevalence within hospitals. Then I'll be looking at that 10 week hump assuming vaccination status is correct. Reports suggest vaccination status is also in error, so I'll be trying to figure a way of predicting any missing cases. Once all this is done I can start looking at survival in terms of delays between vaccination and death.
However, in all this don't forget that I'm not analysing cause of death - that red spike should not be construed as COVID death, these merely being deaths with mention of COVID - the declared cause of death remains unknown. Also unknown is how and when a COVID diagnosis was given. These could be folk testing positive at home and coming in unwell or they could be inpatients that tested positive at some point. Unfortunately I don't have date or reason for admission so we cannot determine the situation.
- Maybe further speciate "disagreement" into its two subcategories (i.e. "EPR says yes, MLP says no" versus "EPR says no, MLP says yes"). Perhaps that might help vaccine Importance pop?
- Curious about the relationship between AcuteResp Dx and disagreement status. The above speciation may help investigate that too.
- In case you missed it, maybe there is a chance of calculating covid prevalence within the trust:
https://jdee.substack.com/p/do-covid-vaccines-work-part-13/comment/22375919
It is a longshot, but if the two suggested methods matched closely, then it would be validated.
LOL - I did exactly that yesterday before shutting down my workstation and called it COVagree2. Today I'm typing up my logistic regression work to see if we get the same peak discrepancy (MLP is a bit of a black box). Preliminary runs reveal the importance of acute resp.