Emergency Department Admissions: Analysis of CDS Dataset (part 5)

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

In this article I want us to get a feel for some basic relationships within the CDS dataset before the iron curtain came down in spring 2020. I shall call this normal era modelling, and I’ll use January 2017 – February 2020 for my data sample, this giving me 1,373,963 admission records.

Before we get stuck in we need to consider age at admission in a bit more detail, so let’s get binning and counting:

There you go, that’s what I wanted to illustrate over there on the left - a preponderance of infant and neonatal admissions who will have their own blend of woes and issues, the poor wee mites. In the same vein youngsters have their own special blend of reasons, and won’t start falling off mopeds, scooters and motorcycles under 50cc until the age of 16. I thus decided to restrict the sample a little by limiting it to adults aged 18 years and over, this giving me 1,077,116 admissions over the period January 2017 – February 2020.

Age Effects

Aside from its interaction with sex, we’ve got arrival mode and disposal method to consider as well as 9 diagnostic and 7 treatment categories, making a total of 19 independent variables to think about. This seemed like hard work so I decided to run a neural network procedure (multilayer perceptron) to point a few things out in one almighty black box crunch-e-rama drama. If I ignore oodles of printout and cut to the chase then this modest table tells me what I want to know:

Right up there in the top three spots are arrival mode, disposal route and a cardiac diagnosis, so let’s have a look at these three in a bit more detail:

No surprises here with the elderly tending to arrive by ambulance, being at greater risk of death, and more likely to be suffering from a cardiac condition.

Disposal route is worth a bit more contemplation since I’ll be using this later in the development of a model assessing the impact of vaccination on likelihood of hospitalisation. Here’s an error bar plot:

What this is usefully telling me is that referrals and discharges may be lumped together to form a reference category in the prediction of hospitalisation and death (at least in terms of patient age).

Just missing out on the medals is Total Diagnoses, but this is most certainly worth an error bar plot:

Told you so!

As with the in-hospital death data the total diagnoses made per admission is a valuable indicator of health status and, as we see here, of age at admission. We now see that those missing data entries for diagnosis tend to be associated with the younger and stronger age groups.

Given my experience of Friday and Saturday nights in the A&E of a city teaching hospital I did wonder about intoxication. There is a coding entry for this – Poisoning (including ovedose) – other, including alcohol – but if we’re just talking about being drunk and dishevelled then it’s unlikely that this heavy duty coding will be used. In our A&E when full-to-bursting those with a few minor cuts and bruises were usually shown the exit if overly vocal/amorous/aggressive. No coding for them!

Sex Effects

OK, so we’re getting a handle on age, but what about sex? Here’s that table again:

This time the age*sex interaction comes to the fore with the mean age for males fetching-up at 48.5 years at admission, and with the mean age for females fetching-up at 49.5 years at admission, as may be expected. Again the disposal route serves to distinguish so let’s have a crosstabulation to squint at:

Now this is interesting.

Being a stats-bod my eyes tend to whizz about the standardised residuals, and these tell me that females tend to be over-represented in the hospitalised category with an admission rate of 53.8% compared to 46.2% for males. This bias will be addressed when I come to modelling vaccine efficacy at a later stage. Right now I suspect it’s time to raid the larder…

Kettle On!