An Introduction to the New Hospital Programme Demand Model

HACA 2023

Tom Jemmett

Jul 11, 2023

The team

A hospital is a place where you can find people…

• having the best day of their life,
• the worst day of their life,
• the first day of their life,
• and the last day of their life.

Planning is hard

• built with enough capacity to replace the existing school
• failed to take into account a new housing estate
• likely needs double the number of spaces within the next decade

BBC article

Review of existing models

Steven Wyatt - NHS-R 2022

Review of existing models

• lots of models
• lots of external consultancies
• lots of similarities

• lots of repetition/duplication
• sufficiently different that comparing results is difficult
• methodological progress slow
• no base to build from

• consultancies don’t tend to offer products, but services
• difficult to compare different models to understand if differences are methodological or due to assumptions
• same issues seen 20/30 years ago
• learning and expertise gathered tends to be trapped within trusts, or kept secret by consultancies

Common issues

• handling uncertainty
• unnecessary/early aggregation
• poor coverage of some changes
• lack of ownership & auditability of assumptions
• conflating demand forecasting with affordability

• most models handle changes like demographic changes and the impact of changes in occupancy rates
• but few try to handle addressing inequities, health status adjustment

“How much, and what types of activity might a hospital need to accommodate in the future?”

• difficult problem
• a question that is routinely asked outside of new hospital planning
• a planning problem, not a forecasting problem, health systems want to influence the types and quantities of activity they deliver in future
• difficult to get a handle in confidence, the results should recognise the inherent uncertainty

Our model

• open source (not quite yet…)
• uses standard, well-known datasets (e.g. HES, ONS population projections)
• currently handles Inpatient admissions, Outpatient attendances, and A&E arrivals
• extensible and adaptable
• covering all of the change factors
• stochastic Monte-Carlo model to handle uncertainty

Project Structure

• Data Extraction (R + {targets} & Sql)
• Inputs App (R + {shiny})
• Outputs App (R + {shiny})
• Model Engine (Python & Docker)
• Azure Infrastructure (VM/ACR/ACI/Storage Accounts)
• All of the code is stored on GitHub (currently, private repos 😔)

flowchart TB
  classDef orange fill:#f9bf07,stroke:#2c2825,color:#2c2825;
  classDef lightslate fill:#b2b7b9,stroke:#2c2825,color:#2c2825;

  A[Data Extraction]
  B[Inputs App]
  C[Model]
  D[Outputs App]


  SB[(input app data)]
  SC[(model data)]
  SD[(results data)]

  A ---> SB
  A ---> SC
  
  SB ---> B
  SC ---> C

  B ---> C

  C ---> SD
  SD ---> D

  B -.-> D

  class A,B,C,D orange
  class SB,SC,SD lightslate

Model Overview

• the baseline data is a year worth of a provider’s HES data
• each row in the baseline data is run through a series of steps
• each step creates a factor that says how many times (on average) to sample that row
• the factors are multiplied together and used to create a random Poisson value
• we resample the rows using this random values
• efficiencies are then applied, e.g. LoS reductions, type conversions

• IP/OP/A&E data
• complex, but not complicated

Model Diagram

flowchart TB
    classDef blue fill:#5881c1,stroke:#2c2825,color:#2c2825;
    classDef orange fill:#f9bf07,stroke:#2c2825,color:#2c2825;
    classDef red fill:#ec6555,stroke:#2c2825,color:#2c2825;
    classDef lightslate fill:#b2b7b9,stroke:#2c2825,color:#2c2825;
    classDef slate fill:#e0e2e3,stroke:#2c2825,color:#2c2825;

    S[Baseline Activity]
    T[Future Activity]

    class S,T red

    subgraph rr[Row Resampling]
        direction LR

        subgraph pop[Population Changes]
            direction TB
            pop_p[Population Growth]
            pop_a[Age/Sex Structure]
            pop_h[Population Specific Health Status]

            class pop_p,pop_a,pop_h orange

            pop_p --- pop_a --- pop_h
        end

        subgraph dsi[Demand Supply Imbalances]
            direction TB
            dsi_w[Waiting List Adjustment]
            dsi_r[Repatriation/Expatriation]
            dsi_p[Private Healthcare Dynamics]

            class dsi_w,dsi_r,dsi_p orange

            dsi_w --- dsi_r --- dsi_p
        end

        subgraph nsi[Need Supply Imbalances]
            direction TB
            nsi_g[Gaps in Care]
            nsi_i[Inequalities]
            nsi_t[Threshold Imbalances]

            class nsi_g,nsi_i,nsi_t orange

            nsi_g --- nsi_i --- nsi_t
        end

        subgraph nda [Non-Demographic Adjustment]
            direction TB
            nda_m[Medical Interventions]
            nda_c[Changes to National Standards]
            nda_p[Patient Expectations]

            class nda_m,nda_c,nda_p orange

            nda_m --- nda_c --- nda_p
        end

        subgraph mit[Activity Mitigators]
            direction TB
            mit_a[Activity Avoidance]
            mit_t[Type Conversion]
            mit_e[Efficiencies]
            
            class mit_a,mit_t,mit_e orange

            mit_a --- mit_t --- mit_e
        end

        pop --- dsi --- nsi --- nda --- mit

        class dsi,nsi,pop,nda,mit lightslate
    end

    class rr slate
    
    S --> rr --> T

• uses either patient-level data, or minimal aggregation
• row resampling grouped into 5 broad groups
  • population changes address the changes to the structure of the population and health status over the medium term
  • demand supply imbalances: hospitals are currently struggling to keep pace with demand, so we correct for this to not carry forwards these into the future
  • need supply imbalance: addressing gaps in care that currently exist
  • non-demographic: such as the development of new medical technologies
  • activity mitigators: strategies trusts adopt for reducing activity, or delivering activity more efficiently
• some assumptions set nationally, such as population growth via ONS population projections
• other assumptions set locally, with support from a Shiny app

Model Diagram

flowchart TB
    classDef blue fill:#5881c1,stroke:#2c2825,color:#2c2825;
    classDef orange fill:#f9bf07,stroke:#2c2825,color:#2c2825;
    classDef red fill:#ec6555,stroke:#2c2825,color:#2c2825;
    classDef lightslate fill:#b2b7b9,stroke:#2c2825,color:#2c2825;
    classDef slate fill:#e0e2e3,stroke:#2c2825,color:#2c2825;

    S[Baseline Activity]
    T[Future Activity]

    ORANGE[Implemented]
    BLUE[Not yet implemented]

    class ORANGE orange
    class BLUE blue

    class S,T red

    subgraph rr[Row Resampling]
        direction LR

        subgraph pop[Population Changes]
            direction TB
            pop_p[Population Growth]
            pop_a[Age/Sex Structure]
            pop_h[Population Specific Health Status]

            class pop_p,pop_a,pop_h orange

            pop_p --- pop_a --- pop_h
        end

        subgraph dsi[Demand Supply Imbalances]
            direction TB
            dsi_w[Waiting List Adjustment]
            dsi_r[Repatriation/Expatriation]
            dsi_p[Private Healthcare Dynamics]

            class dsi_w,dsi_r orange
            class dsi_p blue

            dsi_w --- dsi_r --- dsi_p
        end

        subgraph nsi[Need Supply Imbalances]
            direction TB
            nsi_g[Gaps in Care]
            nsi_i[Inequalities]
            nsi_t[Threshold Imbalances]

            class nsi_g,nsi_i,nsi_t blue

            nsi_g --- nsi_i --- nsi_t
        end

        subgraph nda [Non-Demographic Adjustment]
            direction TB
            nda_m[Medical Interventions]
            nda_c[Changes to National Standards]
            nda_p[Patient Expectations]

            class nda_m,nda_c,nda_p blue

            nda_m --- nda_c --- nda_p
        end

        subgraph mit[Activity Mitigators]
            direction TB
            mit_a[Activity Avoidance]
            mit_t[Type Conversion]
            mit_e[Efficiencies]
            
            class mit_a,mit_t,mit_e orange

            mit_a --- mit_t --- mit_e
        end

        pop --- dsi --- nsi --- nda --- mit

        class dsi,nsi,pop,nda,mit lightslate
    end

    class rr slate
    
    S --> rr --> T

Monte Carlo Simulation

• We run the model N times, varying the input parameters each time slightly to handle the uncertainty.
• The results of the model are aggregated at the end of each model run
• The aggregated results are combined at the end into a single file

flowchart LR
  classDef orange fill:#f9bf07,stroke:#2c2825,color:#2c2825;
  classDef red fill:#ec6555,stroke:#2c2825,color:#2c2825;
  
  A[Baseline Activity]
  Ba[Model Run 0]
  Bb[Model Run 1]
  Bc[Model Run 2]
  Bd[Model Run 3]
  Bn[Model Run n]
  C[Results]

  A ---> Ba ---> C
  A ---> Bb ---> C
  A ---> Bc ---> C
  A ---> Bd ---> C
  A ---> Bn ---> C
  
  class A,C red
  class Ba,Bb,Bc,Bd,Bn orange
  

Inspired by

• MapReduce (Google, 2004)
• Split, Apply, Combine (H. Wickham, 2011)

Model Parameters

• We ask users to provide parameters in the form of 90% confidence intervals
• We can then convert these confidence intervals into distributions
• During the model we sample values from these distributions for each model parameter
• All of the parameters represent the average rate to sample a row of data from the baseline

Model Parameters

“We expect in the future to see between a 25% reduction and a 25% increase in this activity”

• grey highlighted section: 90% confidence intervals
• black line: confidence intervals into distributions
• yellow points: sampled parameter for a model run

Model Parameters

“We expect in the future to see between a 20% reduction and a 90% reduction in this activity”

• grey highlighted section: 90% confidence intervals
• black line: confidence intervals into distributions
• yellow points: sampled parameter for a model run

Model Parameters

“We expect in the future to see between a 2% reduction and an 18% reduction in this activity”

• grey highlighted section: 90% confidence intervals
• black line: confidence intervals into distributions
• yellow points: sampled parameter for a model run

Model Run Example (1)

id	age	sex	specialty	los	f
1	50	m	100	4	1.00
2	50	m	110	3	1.00
3	51	m	120	5	1.00
4	50	f	100	1	1.00
5	50	f	110	2	1.00
6	52	f	120	0	1.00

Start with baseline data - we are going to sample each row exactly once (column f).

Model Run Example (2)

id	age	sex	specialty	los	f
1	50	m	100	4	1.00
2	50	m	110	3	1.00
3	51	m	120	5	1.00
4	50	f	100	1	1.00
5	50	f	110	2	1.00
6	52	f	120	0	1.00

age	sex	f
50	m	0.90
51	m	1.10
52	m	1.20
50	f	0.80
51	f	0.70
52	f	1.30

f
1.00 × 0.90 = 0.90
1.00 × 0.90 = 0.90
1.00 × 1.10 = 1.10
1.00 × 0.80 = 0.80
1.00 × 0.80 = 0.80
1.00 × 1.30 = 1.30

We perform a step where we join based on age and sex, then update the f column.

Model Run Example (3)

id	age	sex	specialty	los	f
1	50	m	100	4	0.90
2	50	m	110	3	0.90
3	51	m	120	5	1.10
4	50	f	100	1	0.80
5	50	f	110	2	0.80
6	52	f	120	0	1.30

specialty	f
100	0.90
110	1.10

f
0.90 × 0.90 = 0.81
0.90 × 1.10 = 0.99
1.10 × 1.00 = 1.10
0.80 × 0.90 = 0.72
0.80 × 1.10 = 0.88
1.30 × 1.00 = 1.30

The next step joins on the specialty column, again updating f. Note, if there is no value to join on, then we multiply by 1.

Model Run Example (4)

id	age	sex	specialty	los	f	n
1	50	m	100	4	0.90	1
2	50	m	110	3	0.90	0
3	51	m	120	5	1.10	2
4	50	f	100	1	0.80	1
5	50	f	110	2	0.80	0
6	52	f	120	0	1.30	3

id	age	sex	specialty	los
1	50	m	100	4
3	51	m	120	5
3	51	m	120	5
4	50	f	100	1
6	52	f	120	0
6	52	f	120	0
6	52	f	120	0

Once all of the steps are performed, sample a random value n from a Poisson distribution with λ=f, then we select each row n times.

Model Run Example (5)

id	age	sex	specialty	los	g
1	50	m	100	4	0.75
3	51	m	120	5	0.50
3	51	m	120	5	1.00
4	50	f	100	1	0.90
6	52	f	120	0	0.80
6	52	f	120	0	0.80
6	52	f	120	0	0.80

id	age	sex	specialty	los
1	50	m	100	2
3	51	m	120	1
3	51	m	120	5
4	50	f	100	0
6	52	f	120	0
6	52	f	120	0
6	52	f	120	0

After resampling, we apply efficiency steps. E.g., similar joins are used to create column g, which is then used to sample a new LOS from a binomial distribution.

How the model is built

• The model is built in Python and can be run on any machine you can install Python on
• Uses various packages, such as numpy and pandas
• Reads data in .parquet format for efficiency
• Returns aggregated results as a .json file
• Could also output full row level results if needed

How the model is built

• Code is built in a modular approach
• Each activity type (Inpatients/Outpatients/A&E) has its own model code
• Code is reused where possible (e.g. all three models share the code for demographic adjustment)

How the model is deployed

• Deployed as a Docker Container
• Runs in Azure Container Instances
• Each model run creates a new container, and the container is destroyed when the model run completes

Data Extraction

• Uses principles of RAP, using R + {targets} and Sql
• All of the data required to run the model
• Data is extracted from various sources
  • Sql Datawarehouse (HES data)
  • ONS population projections + life expectancy tables
  • Central returns, e.g. KH03
  • ODS data (organisation names, successors)
• Extracted data is uploaded to Azure storage containers

Inputs App

A {shiny} app that allows the user to set parameters, and submit as a job to run the model with those values.

Inputs App

Outputs App

A {shiny} app that allows the user to view the results of model runs.

Outputs App

Questions?

Contact The Strategy Unit

• strategy.unit@nhs.net
• The-Strategy-Unit

Contact Me

• thomas.jemmett@nhs.net
• tomjemmett