Part 2: Whole Population Segmentation Models

The Segmentation Model in Detail

OBH have adapted and extended the ‘Bridges to Health’ segmentation model, to create a data-driven segmentation model, suitable for outcomes measurement and outcomes based commissioning in the UK.

This section describes OBH’s approach to segmentation for outcomes measurement based on the foundations of the ‘Bridges to Health’ model described earlier.

How People Move Between Segments (Inter-Segment Movement)

Though the movement of people from one segment to another does not in itself define the segment, it is helpful in understanding a population’s characteristics, and the outcomes which matter to that segment. It may also be helpful from a number of other perspectives; such as care planning, financial and contractual.

General principles describing how people move between segments:

a) Overview: in order to fulfil several key segmentation principles, every individual should be captured in one or more core population groups (segments 1, 4, 5, 6, 7, 8) at any one point in time. That is, even if they receive care and services in an acute segment- acute and/or maternity groups (segments 2 and 3), they will still be captured in their core group. From an outcomes perspective, movements in the overall numbers of people within each of these segments will continually occur, as people naturally progress through segments. The guidance below documents how people will generally move between segments, but there is no requirement for any of these movements to occur, from an outcomes measurement perspective. With precise inclusion criteria, outcomes values can be accurately monitored while these movements occur on a continual basis, irrespective of what actually happens in practice.

b) Three key ‘super-segments’: each of the base segments are most sensibly grouped into one of three broad super-segments:

i. People who are currently predominantly ‘healthy’ (segment 1) (ie. without known illness as yet), or have a reasonable expectation of returning to being healthy after treatment. The overarching outcomes theme/goal here is staying healthy, promoting healthy lifestyles, and avoidance of illness ie. primary prevention.

ii. People who are currently living with an established long term condition or serious disability (segments 4 & 5). The principal outcome themes/goals here are more focused on secondary prevention; reducing the impact of already established disease, illness or disability, enabling autonomy, and preventing asymptomatic disease from progressing to symptomatic disease.

iii. People who are in one of the main ‘end phases of life’ (segments 6, 7 & 8). The overarching outcome theme/goals for this group is ‘coping with illness at the ‘end of life’, and having a good death, ensuring that care services across multiple care settings meet the needs and priorities of these individuals.

c) Access to acute and maternity services: people in each of the base segments may still require access to the services of one of the more transient care-focused segments, acute (segment 3) or maternity (segment 2), which they may at various times enter and leave, as their health needs dictate. Typically, people move from their base segment into an acute or maternity segment, and back to their original base segment. However, this is not always the case.

d) Transitions over a life course: Individual segments are described in detail below, and while movement within each super-segment may be relatively common, people will generally move to a super-segment with higher acuity over a life course.

e) Transitions between segments: Although it is possible that someone in a higher acuity supersegment moves to a lower acuity, this is the ‘exception rather than the rule’. As ‘Bridges to Health’ states: ‘return from one of the last phase of life trajectories (segments 6,7,8) to other population segments is so unlikely as to justify only an occasional exception to the routine’. Example: someone’s long term condition may completely resolve, meaning that they move from segment 4 to segment 1, but it is far more common that their condition endures and people either remain in that segment for life, or progress to one of the relatively higher acuity segments.

f) Dealing with overlaps: although people are often captured in just one base segment, some overlaps do occur between segments, where people’s characteristics mean they meet the inclusion criteria for more than one segment. This is particularly relevant for those in the three ‘end phases of life’ segments, who may also have a long term condition and/or disability, as it is important to monitor outcomes relevant to different conditions. Another example would be an older person with frailty, who may have a super-imposed, incurable cancer diagnosis, where again outcomes relevant to frailty, such as falls, may be captured, as well as outcomes related to the last period of life and having a good death. The only segment which does not overlap with another is segment 1 (healthy), as by definition, people in this segment are those that do not fit the inclusion criteria for another segment.

g) Dealing with exceptions: there will naturally be exceptions to the above general principles, but segmentation for outcomes based commissioning requires designing for the rule rather than the exception, thereby allowing more resources to manage any exceptions to the rule.

How People Move Within Segments (Intra-Segment Movement)

There are two principal ways in which movement of people within each segment can manifest itself:

a) Intra-segment variation (each segment will contain people with different levels of severity/complexity/acuity of their condition/s)

b) Intra-segment movement (people will move between various levels of severity/complexity/acuity within a segment)

Why is this important for outcomes measurement?
Using a whole population approach, whereby all people need to be accounted for in a segment, and at the same time limiting the number of segments in the approach, means that there will always be some variation in the degree of severity in any given segment.

Within each segment people will vary from ‘mild’ to ‘severe’; mild being that the condition(s), and/or their health circumstances, enable a person to go about their life with little disruption or impact to their life and are able to confidently self-manage.

For example, someone with a LTC in segment 4, with a diagnosis such as diabetes or asthma, may have very little disruption to their life in the early stages of the disease, with very little medical intervention and management required. They may even consider themselves to be ‘healthy’, and pursue otherwise healthy lifestyles. However, unless their condition resolves, even ‘healthy’ people in this segment will need some level of condition monitoring, health/care interventions, and surveillance for complications or progression for life ie. secondary prevention. At the other end of the spectrum, another person with poorly controlled diabetes may have multiple complications already arising, such as poor sight or cardiovascular complications, that disrupt their life to a far greater extent, and which likely require a different treatment/management approach by health and care. The diagram below conceptually visualises this intra-segment variation.

Naturally, people with a mild form of the condition will tend to have better outcomes (better health, less illness) than those at the more severe end, at a specific point in time. As well as slowing, or halting, any progression to a higher acuity segment, maintaining people at the lower end of the severity spectrum in any segment will generally lead to better outcomes. From a population-based outcomes measurement perspective, year-on-year improvements would be evidenced by a shift towards more people in each segment living with a ‘milder’ form of a condition, with less impact on their lives – ie. good health, and avoiding illness, as well as fewer people moving to a higher acuity segment.

A Whole Population Segmentation Approach for Outcomes Measurement

Given that ‘Bridges to Health’ is a whole population segmentation model, only a small proportion of outcomes will map perfectly for all people in a given segment. There are three different ways outcomes can be applied to one or more given segments:

• Outcomes that apply to everyone in the segment (ie. one whole segment), or
• Outcomes that apply to a smaller cohort within the segment (ie. a subsegment within a segment), and
• Outcomes that apply to more than one segments (ie. multiple segments).

There may also be a small number of outcomes which could apply to all segments (ie. whole population outcomes), but even these are unlikely to be equally relevant to everyone, and may not adequately address health inequalities.

In addition, even if the segment criteria are helpful in determining what is important to people within a segment, and overall numbers of people within each segment, the segment criteria themselves will often not form the ‘denominator value’ of the outcomes for a given segment.

Segmentation for Outcome Measurement

With defined technical descriptions, parameters and criteria for a segment, the outcome values for any segmented population can be robustly, accurately, and reproducibly measured over time.

The purpose of these outcome values is to:

• understand the baseline and change of the given outcome (compared to previous years),
• establish targets and improvement trajectories (for the future), and
• be able to commission contingent on improvements in that value.

This enables health systems to create a ‘population-level’ view of whether the care being provided is making a difference to specific groups of people, using outcomes that matter to them.

Considerations on segment definitions for outcomes measurement:

• Segment definitions can be locally configured for local priorities and needs, including age, health inequalities, and clinical conditions.
• If no segmentation is applied to CSOMs, all of the outcomes will be measured for all people, but the results would be hard to interpret. Segmentation allows specific outcomes which matter to be measured for specific groups of people.
• What this approach does not necessitate is the ability to identify individual people clinically for direct patient care, or by particular service, care setting, department, or specialty. Segment definitions, in this context, are for population-level outcomes based commissioning. So there is no requirement to ‘allocate’ people to specific segments, as they will routinely be captured in existing clinical data systems. Clinicians and service providers principally just need to be clear what type of outcomes are being measured, and for broadly what types of people.
• Ultimately, one key purpose of segmentation for outcomes based commissioning is to ensure payment structures support established best clinical practice, rather than conflicting with, or hindering best clinical practice in any way.
• There is a distinction between use of segmentation for population-based outcomes measurement, and segmenting a population in a given electronic health record, or for individual patient care, where stratification is more commonly used rather than just segmentation.
• Not all outcomes apply equally within each segment, so while broad segment definitions are helpful, particularly from a finance perspective, each outcome will have a slightly different focus and technical composition, even within a segment.

Segmentation for different types of outcome measures

• Clinical and Social Outcome Measures (CSOMs) use existing data from clinical and administrative systems. With defined technical descriptions, which include coding inclusions, exclusions, and calculations for different datasets across all of health and social care, an accurate list of people included (or excluded) from each segment can be derived. The outcome definition and calculation is further applied to those people who meet the population segment definition.
• Person-Centred Outcome Measures (PCOMs) use data collected through surveys. The segment definitions supports understanding ‘who’ to survey, and how to distribute the survey to the population segment, using data required to identify whether an individual is included or excluded from a segment.

Practical Implementation

OBH have developed a data-driven segmentation model, based on the foundations of ‘Bridges to Health’.

Segmentation and outcomes measurement for specific groups of people, across whole care pathways, different care settings and different providers, requires a person-centred dataset. This means that data related to health and care activities for the entire population, that is currently largely held separately by each care setting, needs to be joined together. However, not all data collected and stored by clinical record systems needs to be joined. Only the minimum data required to establish the segmentation model and outcomes selected should be processed.

Data Use and Governance for Segmentation and Outcome Measurement

For population-level segmentation and outcomes measurement, only anonymous data is required. However, a pseudonymised NHS number is required to join records for the same individual, both within a single dataset, and between multiple datasets. This facilitates a person-centred and complete view of an individual and their care.

To enable processing of any data for segmentation and outcomes measurement, the appropriate information governance requirements and data regulations need to first be satisfied. This is a complex and constantly evolving area. The following are some key considerations and principles that need to be applied in this context:

• In an NHS context, outcomes measurement is a ‘secondary use’ of data (for indirect patient care).
• Pseudonymised data should be treated as ‘personal data’ under the EU General Data Protection Regulation.
• The sharing of confidential personal information for the purpose of outcomes measurement must be lawful.
• Any organisations processing data for population-level segmentation and outcome measurement requires a Data Sharing/Processing Agreement with the Data Controller, or to be named as a Data Processor in existing Data Sharing Agreements.

Using the patient-level pseudonymised health and care data

Segments are defined using technical descriptions, and rules based around criteria and parameters, whilst still allowing for local configuration, according to local needs.⁴ These can be set through a number of variables, for example: age parameters, clinical scores, diagnoses, procedure codes, deprivation scores, and so-on. In the ideal scenario, pseudonymised patient-level datasets from each care setting are required for these rules to be applied to.

For the purpose of outcome measurement, the following rules are applied to the linked dataset:

1. Cohort Identification: A set of rules is applied to identify all people who meet the criteria of a segment, at any given time.

Rules or ontologies are applied to this data, so that ‘cohort’ registers are up-to-date. Data to identify a cohort may sit in different care setting specific datasets, and coded using different clinical coding systems. For example, to identify people with physical disability, data from social care and primary care is required; to identify people with mental health conditions, data from primary care and mental health trusts is required. In addition, the definitions of these cohorts if based on clinical codes may vary depending on the dataset.

2. Outcome Measurement: A set of rules defining the outcome of interest is applied to the cohort identified above.

Once the cohort/segment has been identified, a set of rules can be used to identify whether the outcome has or has not occurred for those people. Many adverse outcomes are acute or chronic health complications that require treatment or an intervention from health and care providers. When this is serious, people will most often require treatment in secondary care. Therefore most outcome data is captured in hospital data. Even though, often the health and care providers who can impact that outcome may also include primary, community and other care providers. Typically SUS (Secondary Uses Service) data or HES (Hospital Episode Statistic) data can be used for this secondary care activity. Importantly, this means that any activity happening out-of-area can be captured for people registered at a local GP practice. For example, where an adverse outcome that relates to an acute condition (e.g. a heart attack, or a stroke) presents with a sudden onset and requires emergency acute care services, somewhere away from an individual’s local area.

Data maturity stages for outcome measurement

Not all health and care systems are yet set-up to allow for data to be processed under Data Sharing Agreements, that allow the use of pseudonymised data for ‘secondary uses’ or indirect patient care. As a result, each local area will be at a different stage of data maturity for outcome measurement. However, at all stages of maturity, outcome measurement is possible. For the early stages of maturity, there will be more constraints and limitations to the outcomes available to be measured, and the accuracy of the outcome value, compared with those operating at more mature stages. Although this is not the ideal scenario, this is frequently the initial baseline position in current care systems. The resulting outcome measures will give the most accurate and robust outcome values currently available, and will generally be valid for commissioning purposes, whilst a data maturity process is implemented.

A summary of the maturity stages of health and care systems for outcome measurement is set out below.

Maturity stage 1: Single, publicly available, aggregate datasets
For example, QOF (Quality Outcomes Framework), ONS (Office for National Statistics), and other national outcomes frameworks.

• This enables measurement of only those outcomes that are currently processed and available through an alternate data source, and in the public domain.
• There are challenges and issues around poor matching with local commissioning requirements, and an inability to locally configure measures, long time lags to access data, with only annual figures becoming available. Although these often cause significant challenges when agreeing on and monitoring an outcomes-based contract, it is often the starting point for many organisations.
• See section below on ‘Use of Publicly Available Aggregated Data’ for further details of this

Maturity stage 2: Pseudonymised, patient-level SUS/HES data for all people registered to a group of GP practices. Used alongside other single, publicly available aggregate datasets (e.g. QOF).

• This enables measurement of outcomes where the event being measured is one which presents to secondary care, and is diagnosed and treated in secondary care.
• There will generally be significant undercoding of underlying conditions or co-morbidities in the SUS/HES dataset, which are required for cohort identification. At this stage/level, sharing and processing of primary care data from all local GP practices has not been established for secondary purposes, not all people in specific segment can be identified, depending on the outcome and segment of interest. It is common in this scenario to use population figures from ONS, or segment denominators from QOF.
• The outcomes available to be measured will be constrained to those where the acute event (often captured by the numerator) is coded in secondary care data.

Maturity stage 3: Pseudonymised, patient-level datasets for all health and social care, that can be joined using a unique pseudonym.

• This will enable accurate and valid measurement of outcomes.
• Undercoding of underlying conditions and co-morbidities in the SUS/HES dataset will be significantly reduced through the use of data from primary care, social care, community care and mental health trusts.
• All outcomes are available to be measured to a high degree of accuracy.

Use of publicly available, aggregated data

Some examples of these are described below:

Quality Outcomes Framework (QOF): Aggregate information submitted by nearly all GP practices to NHS Digital. Data from QOF extracted from NHS Digital can be used to identify outcome denominator population numbers, as it includes total number of people on disease registers (prevalence), whole population sizes, amongst other indicators. Annual data is provided, released approximately 7 months after the end of the financial year.

Office of National Statistics (ONS): Includes data on resident population numbers, GP registered population estimates and mortality data. Population numbers can be used for outcome denominators.

National Audits: Wide range of clinical audits, several condition-specific with data helpful for benchmarking certain outcomes. Time lag to access data is often prolonged.

National Outcomes Frameworks: NHS Outcomes Framework (NHSOF), CCG Indicator Outcome Specification (CCGIOS), Adult Social Care Outcomes Framework (ASCOF), Public Health Outcomes Framework (PHOF): These frameworks include a number of indicators often using ONS and HES datasets as their sources, where appropriate, amongst others, such as national survey data. Where any of the outcomes framework indicator definitions use a national dataset, such as HES, those indicators can be locally configured and then translated to provide a more accurate, and timely monthly baseline value using locally sourced data.

Limitations of publicly available datasets

Although helpful, there are several important limitations to using these aggregated publicly available datasets, for the purpose of outcomes measurement or outcomes based commissioning:

a) Meeting local commissioning requirements: the data has already been aggregated based on agreed rules, which may not match with local commissioning requirements. For example, GP practices nationally submit data on how many people have a diagnosis of heart failure in people who are 18 years and over, as part of QOF. For local commissioning purposes, if the number required is for people who have heart failure who are 65 years and over, it is not possible to extract a number for a different age band (without access to the underlying source primary care data).

b) Timeliness of data: these data sources often report data with approximately a year (and sometimes considerably longer) time lag. For example, QOF data is collected between April of one year until March the following year. Aggregated annual figures are released in October with a 7-month time lag. Data published by the National Diabetes Audit often has a variable time lag, which can extend to several years, due to data acquisition/collection, processing and validation. Typically, publicly available datasets are subject to such significant delays, that their usefulness for OBC purposes is limited to benchmarking and data validation, rather than as ‘contractable’ outcomes measures.

c) Incompleteness of data: for example, the National Diabetes Audit relies on primary care data and HES data; however, it only received data from 57% of all GP practices nationally in 2014/15, which has more recently increased to 95%. Other examples include indicators in frameworks such as ASCOF, where surveys are used that are distributed to only a sample of the population (though could be statistically significant), or where the data extraction only occurs for 3 months of any year, to produce an annual result.

d) Accuracy of outcome values: When limited to single, unlinked and aggregate publicly available data for outcome measurement, each individual outcome measure could be composed of a numerator or denominator definition that uses a different data source, even for outcomes that exist in the same segment.

There are a number of terms used in this paper – definitions of these are set out below for reference:

*Tier structures refer to Michael Porter’s Outcomes Hierarchy.⁶

1. Lynn J, Straube BM, Bell KM, et al. Using population segmentation to provide better health care for all: the “Bridges to Health” model. The Milbank Quarterly 2007; 85(2): 185-208.
2. Vuik SI, Mayer EK, Darzi A. Patient segmentation analysis offers significant benefits for integrated care and support. Health Affairs 2016; 35(5): 769-775.
3. Outcomes Based Healthcare. Segmentation for outcomes. Approaches to outcomes based commissioning. Part 1 – whole population segmentation models. 2016.
4. Clegg A, Bates C, Young J, et al. Development and validation of an electric frailty index using routine primary care electronic health record data. Age Ageing 2016; 45(3): 353-360.
5. Outcomes Based Healthcare. Segmentation Data Definitions. 2016.
6. Porter ME. What is value in health care? New England Journal of Medicine 2010; 363(26):2481-3