Recently, I've been working on a project with Addenbrooke's Hospital (where I'm based as a clinical student) to be able to forecast how many people will be in the Emergency Department (ED, or A&E) as well as the number of admissions and how many times the four-hour waiting limit will be breached.
If you've been to a hospital ED, chances are that you've seen staff stressed about the number of patients they need to see, and perhaps you've had to wait longer than you might have expected to see a doctor. If you haven't, you may have heard of the bed crisis currently affecting the NHS, where there aren't enough free beds on the wards for people being admitted into the hospital, leading to a backlog of people all the way through to ED.
This affects patient care in a huge way. Studies have shown that the worse this crowding is in ED, the worse patient outcomes are for everyone - things get missed, long waits turn simple cases into complex ones, and more.
These problems are set to only increase with time. However, if it was possible to accurately predict exactly how busy ED will be at what times, then rotas could be allocated to match this. If it was possible to predict how many admissions will happen in advance, then we could aim to have a certain number of beds free in advance.
It wouldn't solve the chronic underfunding of the NHS, but it could really help us to make the best of a bad situation.
Traditionally, these forecasts have never been very accurate. For one, they are quite crude, given as a daily summary and often collated manually, and so there are not very many data points to build models from. The statistical tools (e.g. ARIMA) the predictions are built on are laden with assumptions and biases that may or may not be true all the time.
However, I have been working on tackling some of these issues. Addenbrooke's has had an electronic health record system since 2014 (EHR, from EPIC Systems) and this routinely collects a lot of interesting metadata passively. I was able to extract out the lists of patients passing through ED, and the times at which they arrived and left for about the last two years. From this, I could simulate every minute of ED for the last two years. That's a lot of data.
Then I used a type of artificial neural network known as a sequence-to-sequence (or seq2seq) model to help build these predictions using a completely assumption-free approach.
This quickly proved to be a very powerful approach. It also threw up an interesting question - how on earth is this so predictable? Is there a deeper truth here? That will stay a question to be answered another day.