Emergency Department Admissions: Analysis of CDS Dataset (part 12)

I analyse an anonymised data dump of 1.9 million admissions records to the emergency departments of an undisclosed NHS Trust for the period June 2017 – September 2021

Summary

• Five weekly periods have been identified when the ratio of in-hospital to emergency department death was substantially and remarkably elevated, these occurring during lockdown (2020/w15 - w16 covering the period 3 April - 17 April 2020) and immediately after vaccine rollout (2021/w2 - w4 covering the period 9 January – 29 January).

• The anonymised electronic patient records for a sample of 21,928 adult in-hospital deaths occurring between 1 January 2020 and 10 September 2021 for an undisclosed NHS Trust were subject to statistical analysis in order to identify correlates for the peculiar periods.

• Staged multivariate logistic regression modelling was employed to verify correlates among 18 carefully chosen clinical variables. Acute respiratory conditions (OR = 0.79, p=0.001), other cardiac conditions (OR = 1.26, p=0.038), cancer diagnosis (OR = 0.79, p=0.003) and COVID-19 diagnosis (OR = 4.43, p<0.001) were identified as statistically significant discriminators for the three weeks immediately following vaccine rollout.

• COVID-19 diagnosis was found to be a confounding factor for acute respiratory condition owing to correlation, there being inexplicable differences in such correlation between study periods that point to vaccine harm and/or misuse of the PCR test.

Continued from part 11…

Logistic Regression Results: Post-Vaccine Period

Herewith the resulting model structure:

I’m going to ignore the five ‘balancing’ terms arising from stage 1 and go straight in with the stage 2 medical indicators. Some may ask why two statistically insignificant terms are showing and the answer is that they were statistically significant back at stage 1 but variables incorporated at stage 2 have made them redundant as predictors.

The same readers may go on to point out that this indicates a correlation between certain stage 1 and stage 2 independent variables that could distort the coefficients that were estimated. Technically speaking this is correct but in practice the changes in estimates can be utterly butterly negligible and to prove it I shall run the model as a giant single stage. Have a taste of this…

Told you so! The same four medical diagnoses are present as with the staged model, and with highly similar estimates for their odds ratios as promised. So let us have a look at these four horsemen one by one in the tabulated flesh, noting the commonality between the lockdown peculiarity and post-vaccine peculiarity boils down to acute respiratory conditions and other cardiac conditions – but with the effects reversed!

Acute Respiratory condition

There we go, my beauties, but what the heck happened here? This table is showing an increase in the incidence rate for acute respiratory conditions from 20.8% to 26.8% which is at odds with the odds ratio of 0.786 (p=0.001). WERGH is once again the word.

What has happened here is that the above table cannot take account of what the multivariate logistic regression can take account of and that is all that balancing of the sample stuff arising from incorporation of the pillar 1 lagged case detection rate, total diagnoses made, prior risk of death and their interactions at stage 1. I guess the way to look at this is that the case profile changed between normal and elevated vaccine periods such that adjustments have to be made before we can say anything about the incidence of acute respiratory conditions. Once this is done then incidence of acute respiratory conditions actually declined.

But this opens a can of worms. How can I be sure that my stage 1 modelling is doing the right thing? What if other variables were included or omitted, would I get different results? Would my conclusions be exactly the opposite if I changed things? Well, I can’t be sure and yes, I could easily get different results with different variables. Ouch. Another model on another day with other ingredients may well give entirely different results and this is why we ought to be cautious about any model offered up as a true reflection of reality. No model can ever be that no matter how sophisticated. But it isn’t just models that can deceive us – that simple crosstabulation above looks nice and solid and real but it may also be leading us astray. In fact ALL tables of anything can lead us astray since they always assume that all other things are equal or constant or have negligible impact. This is rarely the case, especially in the medical and biological sciences.

So what can we do? Well, in my experience the first thing to do is get the kettle on and munch on some well-buttered toast. Suitably prepared we can go back and start building the logistic regression one variable at a time to see how and where things go into reverse gear for acute respiratory conditions. My gurgling gut is telling me that acute respiratory conditions must surely be elevated for the post-vaccine period as well as the lockdown period, so what is causing this to reverse and churn out an OR of 0.786?

Found It!

I confess to being excited sufficient for a dry mouth as I rebuilt the exact same model from scratch one term at a time. As independent variables flew in and were thrown out acute respiratory conditions stood its ground with an OR that bobbed about 1.3 – 1.4 with the p-value down at p<0.001 – all rather jolly good and reassuring! And then I entered the very last term under consideration and everything flipped inside out. Suddenly acute respiratory conditions were showing as OR = 0.79 again.

Can anybody hazard a guess as to what, among my collection of 18 clinical variables listed in part 11, caused acute respiratory conditions to do a back flip?

Yes indeed, it was COVID-19 diagnosis!

Here’s the original staged logistic regression model re-run but without the COVID indicator…

That’s rather neat and tidy, and those insignificant interactive terms (age * total diagnoses; age * prior risk of death) that we once saw looking ugly are now looking lean and mean at p<0.001 and p=0.001 respectively. And look at that juicy odds ratio of 1.247 (p=0.001) for acute respiratory conditions – does that make more sense or what? Admittedly other cardiac conditions has disappeared (p=0.569), but cancer diagnosis remains and with a very similar odds ratio of 0.776 (p=0.001).

So what is going on?

Well, whatever the ICD10 entry marking a COVID-19 diagnosis on these 21,928 patient records is supposed to be doing isn’t making any sense. For one thing there was no sign of an interaction with acute/chronic respiratory conditions in the original model as we might expect for a genuine respiratory illness, and to make matters worse we also saw it working against acute respiratory conditions sufficient to render an odds ratio less than unity. In a nutshell a COVID-19 diagnosis is acting as a proxy for acute respiratory conditions but it’s doing so in a rather dodgy manner. Let me furnish a crosstabulation so you can see the problem for what it is:

Let us start by considering the normal period. Some 31.8% of in-hospital deaths with a diagnosis of an acute respiratory condition were also given a COVID diagnosis, with 7.7% of in-hospital deaths not suffering from an acute respiratory condition receiving the same. Now look at how the numbers jump during those three peculiar weeks immediately post-vaccination. We now have 72.9% of in-hospital deaths with a diagnosis of an acute respiratory condition as well as a COVID diagnosis – and this is supposed to be an effective therapy! There’s also a jump in the COVID diagnostic rate to 29.1% for those not suffering from an acute respiratory condition.

If I scrubbed the labels on these two periods any reasonable person would assume the peak pandemic to have occurred during those three weeks at the start of 2021! How crazy is that? Either the vaccines were making things worse or COVID labels were being used to hide what was really going on (or both).

Time to stop and have a think, methinks!

Kettle On!