What Was COVID-19, Exactly? (part 1)

In this miniseries I try to fathom what COVID was supposed to be using a sample of 21,810 adult in-hospital deaths over the period 2020/w1 – 2021/w36 for an undisclosed NHS Trust

In my recent miniseries Needle To Door Time I have once again come a cropper in that very little makes sense. We’ve got asymptomatic COVID (non-respiratory conditions with a COVID designation) fetching-up more serious than symptomatic COVID (respiratory conditions with a COVID designation) in terms of risk of early (28-day) death and we’ve got symptomatic COVID cases after their second dose at less risk of early death than non-COVID cases after the same. Talk about topsy-turvy!

This is not the first time the data have not matched the established narrative, nor the second, nor the third. At one point I abandoned diagnostic coding as presented in the EPR in favour of something I called probable COVID, these being predicted values arising from a neural network model. Your best bet for going back over this work is to start with this article and work both forward and backward in time.

So here I am again but this time I’m not simply questioning accuracy in clinical coding, I’m questioning the whole concept of COVID-19 as a pathogen-driven respiratory disease. I’m sure this will sound wild to some readers but please do remember that the data are driving me down this rabbit hole… or should I say into an Emmental of epic proportion?

Re-Tracing My Steps

Fortified by a rather splendid sausage sandwich and steaming mug of rooibos I set about producing weekly time series for the diagnostic variables I have been using along with a time series for case detection rate (CDR). I first introduced readers to the concept of CDR in this article way back in 2022 but it has popped up regularly ever since. CDR (cases detected per 100 viral tests) is a crude way of accounting for the reliance of declared COVID cases on the number of tests undertaken yet it is a rather useful variable when it comes to soaking up unwanted variance in multivariate modelling.

Once upon a time we thought that the PCR test was telling us something useful about the spread of a novel virus within the population but this narrative has fallen apart at the seams to the point where I’m not quite sure what the PCR test was detecting. If we suppose it was detecting something of value in amongst all those false positives then CDR will give us a reasonable estimate of when that something was increasing and decreasing over time (in theory). This objective measure of when something was happening thus becomes a handy yardstick with which to assess the rise and fall of clinical symptoms prior to death that presumably were associated with that something.

The easiest way to go about this is to count the incidence of various diagnoses each week over the period 2020/w1 – 2021/w36 and correlate these counts with the mean case detection rate for those weeks. When we do this for my raft of key conditions plus a few extras we end with this matrix summary:

PROD = Prior risk of death (see this article).

I’ve ranked these primary diagnostic groups (plus extras) into descending order of Pearson correlation coefficient so we can quickly get to grips with what is hot and what is not. I’m relieved to see acute respiratory conditions come out tops because a great deal of my work has rested on the premise of COVID-19 being a respiratory issue. The immunocompromised and diabetics also feature ‘up top’, this being an established clinical reality. It’s also good to see the COVID diagnosis offering something of value.

However, the first thing that struck me with this table was the sheer number of diagnoses that correlate significantly with CDR. In fact there’s only one diagnostic grouping that failed to elicit a correlation and that is other cardiac conditions (chiefly chronic ischaemic heart disease/atherosclerotic heart disease).

There’s a negative correlation with vaccination status that we need to ignore owing to sample bias (please see my archived work on this), and a negative correlation with prior risk of death that might raise an eyebrow but is merely a pointer to decline in complex cases. So let’s have a look at the strongest positive correlation but as a dual time series:

That’s not a bad correspondence at all, if I may say so! One feature that caught my eye is how closely the initial two peaks during spring 2020 align compared to the two later peaks during late 2020/early 2021. That first spring 2020 peak offered instant death (within a week), being a most peculiar affair, whereas the 2020/21 winter peak is characterised by a four week delay that sounds about right for sorry folk in decline that are being cared for.

We have a very similar situation with the immunocompromised/diabetic cohort:

And with acute myocardial infarction:

And with hypertension:

It is most curious that these four very different disease groups should all be responding in the same way; namely, a peculiarly instant spring death followed by the usual drawn out winter peak. How is it that a virus that is now known to be about as deadly as the flu dished out instant death across a whole raft of medical conditions in spring 2020 yet acted like a normal flu virus in the 2020/21 winter season? Do I smell something fishy? Are we seeing something nasty in the woodshed?

What I need to do now is gather my wits and set about establishing a brand new indicator variable that I may call integrated COVID using all of these correlated diagnoses and a shiny spanner. This may require more than one packet of biscuits!

Kettle On!

Now for a spot of tin rattling…

Being self-employed these days means the time I can dedicate to data preparation, analysis and article writing is directly proportional to the income I derive from subscriptions. My substack channel has to pay its way, and remains viable only if there are sufficient subscribers. My grateful thanks to those who have supported this project thus far.