Forecasting the Health Needs of a Changing Population

Background

Decisions around medium and long-term allocation of healthcare resources are fraught with challenges and uncertainties, which explains the use of blunt resource allocations based on across-the-board annual percentage uplifts.

The Bristol, North Somerset, South Gloucestershire Integrated Care Board (BNSSG ICB - we love elaborate acronyms in the National Health Service!), in the south west of England, is part of the local NHS apparatus responsible for planning the current and future health needs of the one million resident population.

Figure 1: A map of the area covered by BNSSG, a space covered by three local authorities, with about 1 million people living inside it.

Population Segmentation

Before tackling the complex problem of forecasting healthcare resources into the future, we first need to understand the current situation regarding the distribution of health needs.

While every individual has a unique set of circumstances, population segmentation is an approach used to help understand overall need by combining individuals into different groups, based on certain criteria.

We use the Cambridge Multimorbidity Score which is a metric designed to summarise the presence of multiple health conditions, known as multimorbidity. Using that score, which applies different weights to different health conditions, we previously found a way of splitting the adult (17+) population into five Core Segments, with Core Segment 1 patients having the lowest score and being the least ill and Core Segment 5 being those with the most multimorbidity.

Applied to the BNSSG adult population (of around 750K individuals), the following interesting properties were found:

1. Halving: Going up one segment results in roughly half the number of people in that segment
2. Doubling: Going up one segment results in roughly twice the NHS monetary spend per person per year

We can see this in Figure 2.

Figure 2: Halving-Doubling Effect of the Core Segments

Creating The Model

To forecast health needs of the population, in terms of how many people will be in which Core Segment in what future year, the Dynamic Population Model (DPM) takes information from two different sources:

1. The Office for National Statistics projections for our area. From this, we get yearly projections for not just the total 17+ population, but also the predicted number of people turning 17 (and so entering our model), deaths, and in- and out-ward migration.

2. NHS patient attribute and activity data, stored in the System Wide Dataset (SWD). This gives us: past and current information on the adult population’s NHS healthcare usage; the Core Segment breakdown of our current and past populations; the proportion of those turning 17, migrating, and dying that are in each Core Segment. From this, we estimate the historical rates of transition within Core Segments, which is essentially the yearly number of people getting sicker or healthier.

By synthesising these pieces of data, we create our DPM forecast. Starting from the most up to date Core Segment population breakdown, the model takes yearly time steps into the future, at each time step using the inputs to estimate how many people are to be in each Core Segment. This modelling approach of having discrete time steps and different movements between states can be set up as a Markov chain, although here we have formulated it as a set of difference equations - through which the outflow of each Core Segment population at each time step is deterministic. The design was led by Zehra and Christos, through a collaboration between the NHS and the Centre for Healthcare Innovation and Improvement (CHI2) at the University of Bath.

The model can be thought of as having the following inputs:

Model Input	Description	Data Source
initial population	The starting number of people in each Core Segment	SWD
inner transition matrix	The yearly proportions of people moving from one Core Segment to another	SWD
births, net migration, deaths - numbers	The yearly number of people moving in and out of the area	ONS
births, net migration, deaths - proportions	The proportion of births/migrations/deaths that come from each Core Segment group	SWD

From these inputs, it deterministically outputs the yearly forecasts for the number of people in each Core Segment. From these yearly Core Segment population figures, we can also forecast use by point of delivery by taking historic SWD information on the activity used by current Core Segment breakdown, under the assumption that stays the same into the future.

We combine these population health segment projections – i.e., how many people will be in which Core Segment in what future year – with recent NHS healthcare usage data to yield forecasted changes for various delivery points, like Emergency Department (ED) visits or maternity service appointments.

Findings

The first output of the model is the population forecast for each Core Segment, as plotted in Figure 3. The visualisation is a type of sankey diagram called an alluvial plot, which shows the proportion of people moving between the Core Segments each year. As it is to be expected, the majority of individuals stay in the same Core Segment year-on-year as the process of acquiring conditions and developing multimorbidity takes places over many years and decades.

The concerning insight shown in Figure 3 is that all Core Segments apart from (the most healthy) Core Segment 1 are due to increase in size, with Core Segment 5 having the largest percentage increase over the next 20 years. While, at first glance, this could be attributed to the effect an ageing population, in which people are staying alive for longer we will see in the next set of results that this itself does not wholly explain the forecasted Core Segment changes.

Figure 3: All Core Segments, except the most healthy (CS1), are forecast to increase in size. BNSSG Population rescaled to have an initial population of 1,000.

In applying the typical NHS healthcare usage per Core Segment to the projections of Figure 3, we derive the expected future healthcare usage for various healthcare settings (Figure 4). In overlaying to these the equivalent projections due solely to demographic factors (both for total population size and capturing the effect of Age and Sex), we see that the DPM projections for increased resource use are not solely attributable to an ageing and growing population, but also to a population becoming gradually less healthy over time.

Specifically, from Figure 4 we can glean the following insights:

1. In all areas except Maternity, the DPM forecasts an increased use beyond just the growing, aging population. The reason that Maternity can be explained as the exception is due to it closely following the demographic changes forecast, specifically for numbers of women of child bearing age.

2. For Community contacts, with the highest proportion of use from Core Segment 5 patients, the DPM forecasts the highest increase into the future. This is because, relative to current size, the number of Core Segment 5 patients is set to increase the largest and so that has the largest impact on Community contacts, which include home visits to patients to support rehabilitation and services to manage long-term mobility issues such as physiotherapy.

3. Whilst Secondary Elective and Non-Elective activity is forecast to grow at similar rates, the Carbon and Cost values are forecast to grow more for Secondary Non-Elective due to the average Carbon and Cost usage per person in Core Segment 5 being higher. In this context ‘Secondary’ is a hospital stay, with ‘Elective’ being planned and ‘Non-Elective’ being unplanned. For example, a hip replacement is elective whereas an admission following a road traffic accident is non-elective.

Figure 4: Forecasts by activity, carbon, and cost for four different points of delivery.

Limitations

It’s difficult to make predictions, especially about the future.

– Danish Proverb

As with any modelling / forecasting method, there are limitations to be mindful of.

1. The cost and activity usage estimates are made under the assumption that we will continue to deliver services as they are currently being delivered. We know this isn’t going to be true, as healthcare-seeking behaviour evolves over time, with younger people accessing healthcare in different ways to previous generations. On top of that, healthcare advances can result in significant changes in healthcare provision, in ways unaccounted for within this model.

2. The model is tied to ONS forecasts for population change, and robust forecasting is hard. It is difficult to estimate what the population will look like in 20 years’ time, and the influence of uncertain and unknown future local development and housing plans. Having said this, population forecasts tend to be robust, one way to consider this is that everyone who will be an adult by the end of the forecast in 20 years’ time has already been born.

3. The DPM does not explicitly account for the interaction of demand and capacity: it simply predicts future healthcare resource requirement assuming that health needs of a given Core Segment patient are met in the same way they are met now. This is an essential assumption to help ensure legitimate use of the empirically derived Core Segment transition rates. However, it inevitably limits practical use, as flexing demand and capacity assumptions is of importance to planners and service managers.

4. It is not possible to validate the model on historic data, firstly because of point 3. above but also because we only have good quality SWD information for the past two years, so cannot reliably look further back into the past and create a forecast that we can check against what actually happened.

5. Whilst it is possible to use the model in other healthcare systems and geographic areas, the underlying data required to generate the Core Segments is non-trivial, so significant data pipelining may be required to get to create local model inputs, as explained above in Section 3.

What Next

We have already generated local use cases for the DPM in forecasting different geographical areas or specific hospital trusts. We envisage the DPM becoming a standard tool in most forward planning initiatives and will continue to refine the model as more information becomes available both for calibration and validation.

Outside of BNSSG, we are keen to disseminate our modelling approach to others who may be interested, as well as expanding our collaboration. There are also other innovative approaches in this space, such as the Health in 2040 report by the Health Foundation which looks at England-level and uses the same ONS forecasts, but using a different ‘micro simulation’ modelling approach.

If long-term forecasting in the NHS is of interest to you and your work, we’d love to chat! Please get in touch at bnssg.analytics@nhs.net

Summary

Reliably forecasting longer-term population health needs and healthcare resource requirements is essential if the NHS is to effectively plan for tomorrow’s problems today.

While this is undoubtedly a difficult problem – both conceptually and statistically – our modelling, undertaken through an academic-NHS collaboration, demonstrates that there are alternatives beyond the commonly-used but simplistic approaches based only on demographic factors.

Find more case studies

About the authors

Luke Shaw is a Data Scientist working in the NHS.

Rich Wood is Head of Modelling Analytics at BNSSG ICB and Senior Visiting Research Follow at University of Bath School of Management.

Christos Vasilakis is Director of the Centre for Healthcare Innovation and Improvement (CHI2), and Professor at the University of Bath School of Management.

Zehra Onen Dumlu is a Research Associate at CHI2 and Lecturer at the University of Bath.

Copyright and licence

© 2024 Luke Shaw

This article is licensed under a Creative Commons Attribution 4.0 (CC BY 4.0) International licence.

How to cite

Shaw, Luke et al 2024. “Forecasting the Health Needs of a Changing Population” Real World Data Science, May 08, 2024. URL