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

SMLR modelling of an ECDS extract of 237k admission records to the emergency departments of an undisclosed NHS Trust: the healthy vaccinee effect confirmed using risk of hospitalisation

In part 4 of this series I confirmed the healthy vaccinee effect for adult admissions to the ED of an unknown NHS Trust for the period 2021/w1 – 2021/w38 in terms of risk of treatment within the department. We saw a higher risk of treatment within the department for those coming in unvaccinated compared to those arriving already vaccinated, with the risk for those who got jabbed after admission falling between the two. The effect is not apparent for all age groups and chiefly makes itself known in the 70 – 79 and 80 – 89 year decadal age bands. The effect is subtle, exhibiting a 3.5% increase in risk of treatment at best (70 – 79 year group). Today I’m going to repeat the analysis for risk of hospitalisation, so strap yourselves in…

A Risk Score For Hospitalisation

We start with a near identical list of variables but this time round we have to include Treatment Status [0=Not treated; 1=Treated] as an independent, bring the total to 21 independent variables. The dependent variable of interest – Disposal Route - sits down at the bottom with a mean of 0.219, which tells us that 21.9% of admissions were hospitalised. Please refer back to part 4 for an explanation of the remaining variables:

Modelling Strategy

The sample used was all adult admissions to the ED over the period 2021/w1 – 2021/w38 (n=236,856). In the first stage all 21 independent variables listed above were entered as main effects using a conditional forward selection process. In the second stage an additional 74 two-way interactions were submitted to a conditional forward selection process to further explore age (20 two-way), sex (19 two-way), mode (18 two-way) and treatment (17 two-way) interactions with various diagnostic categories. Beyond this level of parameterisation diminishing returns for model performance were pronounced.

The final model absorbed 70 of the 95 available terms before the run came to a halt at step 49 with the following rather spiffing classification results:

As we may observe the model was particularly good at predicting true negatives, with a performance of 93.6% of discharges correctly identified. It wasn’t quite as good at predicting true positives with performance dropping to 48.5% correct classification. Once more there are two ways of looking at this: either the model is weak in some aspects or it’s telling us something valuable about the mechanics of an emergency department in the middle of vaccine rollout in post-lockdown Britain.

Risk Score Basics

Here’s the ROC for those who enjoy these things but do bear in mind the sample for this was restricted to the 27,320 admissions for May 2021 to stop my graphing module from melting down:

An overall classification performance of 0.833 is rather splendid!

Herewith the distribution of scores for May 2021 revealing two heads and a rather long tail of high risk cases:

The BIG FAT Result

This is a rather tasty slide even if I say so myself:

So there we have it – the healthy vaccinee effect once more, but this time for risk of hospitalisation during the critical period of 2021/w1 – 2021/w38. It really couldn’t be any clearer: those coming into the emergency department already juiced up were healthier than their unvaccinated counterparts (if we go by this measure of ‘health’).  Falling in between these two are those who came in unvaccinated but decided to get jabbed later, with the effect diminishing into nothingness with younger and younger age groups.

A BIG question that follows this BIG FAT result is where did those pale blue intermediaries get jabbed? Was it in the ED during the same admission, was it a while later after a scare and some encouragement? This brings us back round to survival analysis, a topic that I visited back in January of this year. Those with mince pies to spare might like to flip back to the first article entitled Needle To Door Time (part 1).

Meanwhile…

Meanwhile Christmas is coming and the goose is getting fat, so I’m going sit and figure what fresh and fruity line of analysis to pursue in the New Year. With my unique NHS data dump now well past its sell-by date and much of the publicly available COVID stats not to be trusted then perhaps it’s time to move on to new areas of work - do type a comment below and tell me what you’d like to see. Perhaps Santa will drop something rather interesting down the chimney!

Kettle On!