Emergency Department Admissions: Analysis of CDS Dataset (part 10)

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.

• A neural network model (multilayer perceptron) was developed revealing the relative importance of 42 clinical variables, enabling a patient profile to be formulated for these two unique periods. Total diagnoses made, age at death, COVID-19 diagnosis and diabetes featured as the top four discriminators.

Appetiser…

In part 9 of this series I identified five peculiar periods when the ratio of in-hospital to emergency department death went through the roof, these occurring during two weeks during lockdown (2020/w15 - w16 covering the period 3 April - 17 April 2020) and three weeks immediately after vaccine rollout (2021/w2 - w4 covering the period 9 January – 29 January). The next logical step in my recipe book is to try and figure if these stand out in some way and for that I need to flip back to analysing the in-hospital death database containing 57,557 anonymised electronic patient records (EPR).

A Saucy Beginning

To avoid unnecessary confounding factors I limited the sample to those 21,928 in-hospital deaths occurring between 1 January 2020 and 10 September 2021 (the point where the sample fizzles out). A 3-level indicator variable (RatioFlag) was derived to mark out the two lockdown weeks, the three post-vaccine weeks and everything else with the grand title of Peculiar Death Ratio Indicator.

Whilst my training tells me to slowly and carefully build a picture I couldn’t help shovelling everything I had into one almighty neural network model (multilayer perceptron) in a somewhat saucy fashion to see what would come out in the wash in the prediction of the freshly minted indicator variable. Herewith the listing of a great jumble of independent variables sorted by normalised importance in making that prediction according to the neural network that was built:

Right at the top we see a golden oldie being the total number of diagnoses made per patient. This may take values between 1 and 10, there being up to ten fields into which a clinical coder may enter an ICD10 code. I’m guessing that the relationship is positive in that mean diagnostic count was higher for those deaths occurring during the five peculiar periods, but will check on this in a little while.

Aside from this we have age popping up in second place which, of course, is going to be linked to the total number of diagnoses. I’m going to suggest we are going to see a hike in the mean age during those five peculiar periods, but we’ll have to wait and see!

Now then, we have a COVID diagnosis popping up in third place. Only the naïve and uninformed these days will assume this to be a genuine marker of disease status that is backed by a thoroughly reliable PCR test. I suggest we forget this being an indicator of illness - or even of infection - and think of it as a robust indicator of patient management policy. As we should all know by now some very unpleasant, unethical, immoral and downright idiotic practices have been going on hidden from view, as courageous whistleblowers will testify, whilst the rest still remain silent to this day… and complicit.

That test result singled you out for death regardless of whether you were infected with SARS-COV-2, and regardless of whether you were suffering from COVID-19’s bizarre raft of kitchen sink symptoms from snivel to seizure. If you don’t believe me take a look at where acute respiratory conditions sits in this table; you’ll find it near the bottom. The only saving grace is symptomatic COVID (a positive test result supported by a respiratory diagnosis) though this isn’t saying much because that respiratory diagnosis might point to a chronic condition such as asthma or even incidence of the common cold.

We expect diabetics to be sitting near the top of this table, and their risk status was acknowledged fairly quickly on. What I don’t expect to be sitting near the top is major injury and trauma and I’m wondering if these are arising from falls in the elderly – we shall see!

I shall now stop guessing, order a pot of rooibos and crayon some bakes…

Total Diagnoses

Strike! Yep, there’s a result right there, with a Kruskal-Wallis independent samples non-parametric test offering p<0.001 and confirming what our eyeballs already know. We may conclude that the five peculiar death ratio weeks were periods when particularly complex cases were abundant. We may also note the similarity between mean diagnoses for the two weeks of lockdown (mean = 3.48, Std. Deviation = 1.68) and mean diagnoses for the three weeks immediately following vaccine rollout (mean = 3.51, Std. Deviation = 1.75).

Age

Strike! Yep, there’s a result right there, with a Kruskal-Wallis independent samples non-parametric test offering p<0.001 and confirming what our eyeballs again already know. There’s a fascinating twist in the plot here as we observe a preponderance of elderly in-hospital death during the two lockdown weeks compared to any other period. Did the mysterious virus only strike hard for those exact two weeks? I think not. I think the elevation we see for the post-vaccine three week period is driven by seasonal respiratory illness, and I am going to suggest the difference we see between this level and that for the two weeks of lockdown was driven by labelling and mishandling of the elderly. More than a few nurses have confided this in private communication and I would suggest now is the time for those who’ve seen to also be heard.

COVID-19 Diagnosis

Strike! Yes indeedy, another result. I’ve made this table extra complicated to satisfy your inner geek but you can simply look at the percentages to see what was going down. During ‘normal’ periods rates for positive test results were down at 12.7% of all in-hospital deaths, with these inflating to 39.6% and 40.6% for the peculiar lockdown and post-vaccine periods respectively. A chi-square test of association yielded p<0.001, so this is far from being a fluke. Whilst positive-testing folk are not necessarily going to be dying from COVID-19 or even carry a genuine SARS-COV-2 infection, what we can be sure of is that they are going to be treated differently. The question is whether this difference in treatment enhanced risk of death or even caused it, or whether these unfortunate souls were always going to die.

Diabetes

OK, so not a big result here but it does pass statistical muster with a chi-squared test of association yielding p=0.014. There is some evidence that diabetics were slightly over-represented in the two peculiar periods in question but only by 2 – 3%.

Symptomatic COVID

Strike! We have a chi-square test of association pushing out p<0.001 and our eyeballs note a rise in symptomatic COVID cases from a ‘normal’ 7.7% to 23.7% and 23.9% for the lockdown and post-vaccine periods respectively. The first thing we should note is that this is not necessarily cause of death we are talking about here – cause of death could be anything – all that this case indicator represents is an ICD10 code in the EPR to mark a positive test result and another ICD10 code to represent any kind of respiratory condition whether or not arising from anything infectious.

Injury & Trauma

This is the sort of result where we have to ask what multilayer perceptron was thinking when it declared major injury and trauma to be a predictor of peculiar death ratio indicator. I don’t see a result and neither does a chi-square test of association (p=0.292). Harrumph!

Such unworldly weirdness prompts me to abandon the neural network approach and go for something classical but, alas, we are at the end of another article! Until next time…

Kettle On!