Multidisciplinary analytical pipelines: Productionising a business critical simulation model

Chris Beeley

4 December 2025

Intro

  • This presentation is about building a model to predict the future
  • Doing it well required high quality collaboration from a range of staff in different teams with very different skillsets
  • I think the stuff “around” the modelling is non-obvious and this presentation is intended to help others treading a similar path

Why?

  • New Hospital Programme came to the Strategy Unit c.2020
  • Predict demand for the future of the hospitals c.2041
  • We built on existing work and knowledge in the SU as well as the literature
  • I was not here!

The principles

So what does it do?

'Gif of Homer Simpson pressing a button and saying 'do something cool'

So what does it do?

  • Demographic change
  • Non-demographic change
  • Types of potentially mitigable activity
  • More to come

The model

  • Sample the parameters (assume normal)
  • Calculate demand at IP, OP, A&E level
  • Do this 256 times and plot the distribution
  • The results are conceptually at row level, but not in practice

Structure of the project

  • Not going to show a big confusing diagram
  • There is one if you want one
  • Three places that stuff happens, partly because of IG
  • We store and process data in Azure (SQL -> databricks; Docker)
  • We host the reports and dashboards separately, with no data on this server, only results
  • And we have a plethora of scripts, notebooks, and other gubbins on laptops

Versioning, RAP, and audit

  • Versioning is really challenging
  • We learned a lot!
  • Are we Apple or Microsoft?
  • RAP is important for audit too
  • We continue to wrestle with versioning of data to support old model runs

RAP is awesome… right?

Challenges

  • All models are wrong
  • Consistency versus accuracy
  • “It is difficult to make predictions, especially about the future”
    • “This model is simple”

The national elicitation exercise

  • We asked experts to predict likely levels of mitigation in the future, in a structured way
  • We also made some whizzy data science tools to do it with- which ended up being really important and useful
  • It’s not my area so I won’t say any more- this was Prof Mohammed and team
  • We show these values to trusts to help them make better guesses about the future

Challenges- users

  • Complexity correctness versus accessibility correctness
  • Who is a user? What do they want? (I’ll talk more about this in a talk later on)
  • Hospital vs ICB vs NHP vs director vs data scientist vs… perspective

Challenges- decision making

  • Different analytical teams
    • Theoretical ideas to…
    • Pragmatic interpretations
    • All delivered with data science
  • NHP
  • Strategic versus operational
  • Prioritising

Challenges- openness and transparency

  • The work it takes to actually make this happen
  • Running the code
  • IG and sharing

The future

  • National and regional model runs
  • Bring your own data (FDP?)
  • Understanding more about categories of potentially mitigable activity, who thinks what’s possible, and why it matters
  • Increasing understanding of the shift from hospital to the community