Value-Adding Approach to Lot 5 Service Delivery

Through providing risk stratification and impactability modelling to CCGs and primary care covering a population of more than six million, SCW is experienced at ensuring delivery of effective analytics to support care co-ordination and management. We provide the scale and capability to validate algorithms and apply machine learning (MS analytics/decision tree). This can be evidenced by the large number of different risk models that we have already built for systems in the south of England beyond a simple ‘risk of emergency admission’. Examples include Risk of Persistent High Cost Top 10%, Top 20%, Emergency Bed Days, Total Bed Days, and Urgent Care Contacts. We have also already linked our risk stratification data to socio-economic indicators. We deliver these services through a workforce of 240 business intelligence specialists and 130 digital transformation specialists, supported by a long-standing partnership with Johns Hopkins University, Baltimore and their US originated Adjusted Clinical Groups (ACG) system adapted by SCW for use in UK health and social care. 

We also work with the academic sector (including AHSNs) for interpretation and advanced analytics. Our commercial partners include Cerner and their HealtheIntent programme for delivery of real-time support tools in clinical settings, and Beautiful Information, the first NHS/private partnership of its kind providing unique real-time information to NHS trusts to help them plan and resource clinical services to meet hourly fluctuations in patient demand. Systems therefore gain value from our ability to provide an in-depth risk and impactability modelling approach, supported by key partners, to deliver:

  • An extensive impactability toolkit that includes a range of validated algorithms, integrated data sets covering health and social care partners, and a flexible approach to analysis focussed on local needs and context
  • Measurement, evaluation and learning to provide constant feedback on what works for patients
  • Identification of the most effective case management programmes tailored to individuals’ needs, including social needs, and providing analytics and evaluation support that helps ‘service improvers’ deliver programmes at the level of individuals
  • A review of risk stratification and population segmentation outputs to identify those cohorts with the greatest scope for improving care (looking at risk cohorts and opportunities together)
  • Using the right model for each intervention – for example, high persistent risk models for preventive interventions with younger patients, mortality risk models for end-of-life patients
  • Profiles for different cohorts in terms of clinical issues to identify the most effective interventions
  • Support for care providers to build risk stratification into workflows and incorporate reviews by multi-disciplinary teams.

The case study that follows this response describes our partnership work with Johns Hopkins University and Slough CCG, part of the Frimley STP, where we worked collectively to understand the delivery of risk and impactability modelling and where, as a result of system changes arising from this analysis, the CCG was able to reduce emergency admissions by 19%.

Extended Case Study: Risk Stratification, Case Management and Population Profiling in Slough

Slough CCG in Berkshire is part of the Frimley STP. It is responsible for commissioning and managing most hospital and health services for a population of 150,000 and providing primary care services through 16 GP practices. 

As the population is aging and the prevalence of long-term conditions increases, Slough CCG recognised the increasing importance of understanding the population’s multi-morbidity profile and identifying the best ways to deliver interventions and care to the highest-cost and highest-risk patients. As well as understanding general health trends across the population, the CCG specifically wished to identify how they could improve outcomes in terms of reducing unplanned hospital admissions.

 

In December 2015, SCW started collaborative work with the health system and our strategic partner Johns Hopkins University (JHU) to undertake a population profiling and risk stratification exercise in the health economy. We applied our standard methodology using the Johns Hopkins Adjusted Clinical Groups (ACG) System to gain an insight into morbidity patterns and factors. We particularly explored the impact of multi-morbidity on resource use and the risk of adverse outcomes, to identify cohorts of patients for intervention and to support the development of a primary care based complex case management service.

To achieve this, we used the ACG System as a morbidity-based patient classification system to analyse costs and overlap for 150,000 patients at risk of different outcomes. Specific applications of the ACG System in this case included risk stratification and cost distribution, case mix variation across GP practices (applying Resource Utilization Bands, see below), and segmentation of the population for concurrent and predictive morbidity and cost risk. To facilitate the analysis and improve accuracy, we divided the population into 22 segments based on age and multi-morbidity. The multi-morbidity groups were based on the chronic condition count within the ACG System.

Using the ACG System’s predictive models, our analysis revealed key insights such as multi-morbidity being the norm in the population, multi-morbidity not being distributed uniformly, that multi-morbidity was the biggest cost driver and predictor of activity and future risk and that there were significant variations in case mix between GP practices. The ACG System automatically assigns a six-level (low to very high) simplified morbidity category termed a Resource Utilisation Band (RUB). Calculating RUBs by GP practice not only confirmed this case-mix variation between GP practices in Slough (which had long been suspected by GPs) but also quantified it to facilitate targeted action.

We used the data to highlight and quantify the degree of co-morbidity and multi-morbidity that exists within the population. People with common chronic conditions like asthma, chronic obstructive pulmonary disease (COPD) and diabetes rarely experience a chronic condition in isolation. In Slough, for example, only 21% of people have diabetes and no other chronic condition. More than half of the people with COPD had three other chronic conditions to manage.

Clinicians were keen to understand which conditions had the most significant impact on future risk. SCW and JHU listed the number of people with each of nine common chronic conditions and the mean probability of an emergency admission for people with that disease. To illustrate how risk increases with co-morbidity rather than a single condition, we also provided the bottom risk scores for patients with two combinations of diseases. This proved to be a seminal moment in Slough CCG’s appreciation that they should be focussing on the support of patients with multi-morbidity rather than just managing single diseases.

In terms of commonality in different risk groups (figure opposite), clinicians were keen to identify any commonality between cohorts of people most at risk of an emergency admission, those predicted to be the highest cost patients in the coming year, and those flagged as ‘frail’. The degree of overlap was much lower than the CCG expected with only 10% of patients featuring in all three categories. There were significant numbers of people who only feature in one of the three cohorts.

Based on the findings, a decision was made to run a pilot ‘complex case management service’ at Slough CCG to support around 750 primary care patients with one of four specified multi-morbidity combinations identified through our analysis. After 18 months, we measured a 19% reduction in unplanned hospitalisation compared to the expected rate based on historical data. The scheme was subsequently rolled out across two neighbouring CCGs covering a population of 450,000.

To support this approach, traditional case finding techniques that centred on identifying people most at risk of an emergency admission or people with a particular disease should become more sophisticated. They need to be able to identify people that fit multiple criteria including certain co-morbidity combinations. Even more importantly, there needs to be an alignment of case finding techniques with the types of care programmes that exist within a health community and case finding techniques need to be based on the admission criteria to these programmes.

How We Work with Partners in Lot 5

Specifically within this Lot, we will continue working with organisations such as Johns Hopkins University, Baltimore and their US-originated Adjusted Clinical Groups (ACG) system, which SCW adapted for use in UK health and social care settings. We also work with the academic sector on interpretation and advanced analytics; with Cerner and their HealtheIntent programme for delivering real-time support tools in clinical settings; and with Beautiful Information, the first NHS/private partnership of its kind providing unique real-time information to NHS trusts to help them plan and resource clinical services to meet hourly fluctuations in patient demand.

 

Case Studies

SCW conducted a comprehensive IG compliance review for the Symphony PACS Vanguard data set in South Somerset. Because of our scale, we were able to use an independent IG team that had not been involved in the initial development project, and whose role was to critically review and confirm legal compliance for the project. They were also tasked to provide clarity on information ownership, access, safeguards and privacy impact assessment processes. The ultimate aim of the review was to establish the legal basis on which GPs would share patient data for use in the Symphony database; how SCW would link GP data with SUS and Social Care data; and whether underlying data flows, information sharing agreements and privacy impact assessments were appropriate. The final report confirmed compliance and recommended additional work and a regular review programme to ensure the system remained compliant as the project continued and extended in scope. 

Subsequently, working in conjunction with the national New Models of Care (NMOC) IG lead, we supported Somerset as an IG test site for NHS Digital’s DARS process. In December 2016, Dr Geraint Lewis from NHS Digital visited Somerset to run a workshop with key stakeholders. In January 2017, we supported Somerset to complete an NHS Digital IG Audit where Symphony data was a key area of scrutiny. As part of that process, we facilitated a full-day workshop in Leeds between NHS Digital, NHS England’s NMOC team, and representatives from Somerset and SCW to validate the application and associated IG plans. 

Our Symphony data analysts and in-house data management leads have worked with NHS Digital to develop a DARS-compliant framework for integrated data. These proposals went for NHS Digital IGUARD approval in March 2017 and SCW was the first integrated data proposal in the country to receive approval. NHS Digital has subsequently shared our IG solution widely.

“I've just been at the Vanguard PACS IG event where NHS Digital used our diagram as an example of good practice!” Jeremy Martin, Symphony Programme Director, Yeovil District Hospital NHS Foundation Trust.

SCW has worked closely with academic colleagues at the Johns Hopkins Bloomberg School of Public Health to recalibrate their ACG system for use with UK data. An initial recalibration was undertaken, which demonstrated existing U.S. model weightings worked well on UK data, but that the new co-created weightings worked even better. We also co-developed a new UK model to predict the risk of emergency admission, to align the ACG tools with UK priorities. 

In the most recent recalibration, we co-developed five new models for use in the UK, including models to predict Emergency Bed Days, Total Bed Days, Urgent Care Contacts, and Persistent High Costs. This range of predictive models supports a level of sophistication for bespoke case finding that is not available through most other risk prediction tools. Our joint work in this area has shown that a range of predictive models is required when supporting cohort identification programmes that seek to stratify beyond the historically typical top 1-2% of high-risk patients (risk of a non-elective admission). This is particularly relevant for most new models of care. Our continued collaboration with academic colleagues at Johns Hopkins ensures we have the flexibility to develop new models and markers as the UK market evolves and the need arises.

Partner Profile: “John Hopkins University and SCW have a special relationship that goes back eight years. We have done significant development work together including recalibration of the predictive models in the software and developing new analytical approaches to support the work of CCGs and more recently to support the population health management activities of STPs. We are grateful for the support SCW has given with our educational programme, speaking at various national and international conferences and contributing to our webinar series. We value the on-going relationship and look forward to continuing to work with SCW on developing not only the ACG System but an approach to population health management that supports the work of Health Systems.”  Alan Thompson, Director of User Support, UK Team Leader, Johns Hopkins ACG® System

As part of our evaluation of the South Somerset Vanguard, we worked with the University of York Centre for Health Economics to incorporate ‘level of deprivation’ as a matching variable to identify cases and controls.  This analysis enabled us to assess the extent to which the intervention groups are more or less likely to be deprived than similar patients. The findings indicated that, for both the Complex Care Hubs and Enhanced Primary Care Teams the STP were developing, patients identified for interventions were likely to be more deprived than similar cases. This suggested that there was no evidence that interventions would have a negative impact on health inequality

West Hampshire CCG required insight that would support practices to improve dementia diagnosis rates as the CCG was not meeting the national prevalence target. We used HHRA data to identify patients who had been diagnosed with dementia in secondary care by using a Read Code that is not included in QOF definitions. We co-developed a trajectory for achieving the national target and the number of additional patients needed to be diagnosed, with suggestions of where to identify these patients. This allowed the CCG to target their resources in a more effective way using more timely information compared with the national QOF data, which only provides annual GP-level data – and providing improved support for patients.

We use the bow tie method to analyse and visualise healthcare processes from the patient’s perspective. Patients can be represented as a chronological timeline of their medical events. This process mining (e.g. a clinical pathway is the process of care management for a given disease) involves the exploitation of large datasets in new and innovative ways. The outputs and visualisations that result from such analyses allow us to build a clear picture of the process and clear insight into the risk factors and incidents contributing to a specific event and the longer-term outcomes of the event. The illustration to the right shows pacemaker implantation as a specific event, with a wide range of inputs and outcomes linked to it covering the whole population experiencing the procedure.

We are currently using this approach for the sepsis pathway, using ‘suspicion of sepsis’ emergency admission as a specific event to better understand the risk factors for admission with infection and the outcomes for those admitted.  For example, which patients go on to develop sepsis, survive sepsis, be treated for another condition, and so on.  We are using national data to develop the overarching model and testing regional variations using subsets of this data. This supports local patient safety collaboratives to identify critical risk factors that can be monitored or tracked and to develop practical interventions that reduce the risk of admission.

Critically, we can use the risk factors we identified through the bow tie analysis to build and refine our predictive models for specific conditions. By applying the analysis to data from TFEs, we can see the most significant risk factors and use this intelligence to develop meaningful and refined predictive models that reflect real life. This type of analysis also provides useful visualisations of health and care processes that are intuitive and provide a useful starting point for discussions between health and care system partners when considering new models of care.

SCW worked with national clinical leads and colleagues at Imperial College Health Partners to develop a ‘suspicion of sepsis’ dashboard using statistical process control (SPC) methodologies. Through this dashboard we provide SPC charts for key patient outcomes including mortality and length of stay, plus insight into those areas tackling the issue most successfully. The use of SPC allowed for an objective and statistically robust view, highlighting where ‘special cause’ variation was evident in key outcomes. Use of this analysis also allowed SCW to support NHS Improvement and NHS England to understand the capability of the system to achieve improvements in sepsis outcomes across England, and provide estimates of impact at a national scale. 

“SCW has done some terrific work within an extremely short time frame to advance the national, regional and organisational understanding of patients admitted to hospital with infection. This work is pivotal to both understanding the scale of the condition, its costs and being a starting point to developing a true operational definition of sepsis. Furthermore, it enables us to track improvement over time, a critical component of all improvement strategies; and to ascertain what the key characteristics and interventions associated with better outcomes in all cause infection and sepsis. I have no hesitation in commending the incredible work that has been accomplished already and am very excited to see what is to come and the collaboration that this data will enable in the future across the entire care pathway.” Matt Inada-Kim, National Clinical Advisor Sepsis and Deterioration.

We have worked closely with Johns Hopkins University (JHU) in adapting, validating and improving a range of predictive algorithms for use on the UK population. The original US ACG system contained two models to predict any admission in the next six and next 12 months.  We worked with JHU in 2013 to recalibrate these models for us with the UK population and to extend the range of models by developing models for emergency admissions, long length of stay admissions, total cost and drug cost. We supported a further recalibration in 2016 to refine the existing models and develop four additional models for persistent high users, emergency bed days, all bed days, and urgent care contacts (A&E attendances or emergency admissions). Both recalibrations were robust in their use of large data sets covering a population of half a million and using a portion of the data to test the effectiveness of the models, splitting the data into a “training” and “validation” datasets. The newly developed algorithms were tested to ascertain their external reliability and validity against the validation data. The extensive published evaluations of the ACS models we use include:

  • Development and validation of models:  Lemke KW, Weiner JP, Clark JM. (2012) Development and validation of a model for predicting inpatient hospitalization. Med Care 50:131-139. http://www.ncbi.nlm.nih.gov/pubmed/22002640
  • Tailoring models to chronic care programs: Murphy SM, Castro HK, Sylvia M. (2011) Predictive modeling in practice: improving the participant identification process for care management programs using condition-specific cut points. Popul Health Manag 14:205-210. http://www.ncbi.nlm.nih.gov/pubmed/21241172
  • Measuring outcomes: Murphy SM, McGready J, Griswold ME, Sylvia ML. (2013) A method for estimating cost savings for population health management programs. Health Serv Res 48:582-602. http://www.ncbi.nlm.nih.gov/pubmed/22924661
  • Using Propensity Scoring: Segal JB, Griswold M, Achy-Brou A, Herbert R, Bass EB, Dy SM, Millman AE, Wu AW, Frangakis CE. (2007) Using propensity scores subclassification to estimate effects of longitudinal treatments: an example using a new diabetes medication. Med Care 45:S149. http://www.ncbi.nlm.nih.gov/pubmed/17909374

The predictive models we use provide GPs and other primary care clinicians with patients at risk of a range of adverse outcomes such as risk of emergency admission, risk of high cost due to multi-morbidity, risk of high pharmacy costs, risk of being high cost for the next three six-months periods (without regression to the mean). With our dataset of integrated primary and secondary care data covering the whole population (and other settings of care where available), we are able to further segment and stratify ‘at-risk’ groups based on parameters given to us by clinicians and system leaders. For example, where they have clinical and financial evidence that an intervention is having an impact and/or that patients identified by this further stratification are amenable to different forms of preventive care. We provide the following examples of identifying people amenable to preventive care:

In East Berkshire: We identified multi-morbid patients with certain combinations of long-term conditions who GPs stated were amenable to treatment in primary care rather than hospital-based specialist care. Three CCGs have now commissioned primary care-based multi-morbidity clinics that resulted in a 19% decrease in emergency care costs for these patients in the first 18-months of the scheme. We are able to identify the patients suitable for this service by deploying an algorithm that is based on the LTCs the patients have and their risk of incurring high cost in the coming year, while omitting patients nearing the end of their life.

In Gloucestershire: We identified people with polypharmacy to facilitate medication reviews that would improve safety (stopping unnecessary medication) and reduce costs. GP practices are reporting significant impactability with these types of patients, and both in-year and recurring cost savings of tens of thousands of pounds resulting from medication reviews. With an integrated primary care and secondary care dataset that contains prescription information, we can create algorithms to identify patients of interest for medication reviews based on characteristics the GPs define. For example, one GP practice reviews all patients with multi-morbidity and polypharmacy issues as defined by the number of unique drugs rather than the number of prescriptions issued.

We supported this Vanguard to create a new care programme to identify, track and support at-risk patients, using a combination of ACG outputs combined with clinical review. This is a GP-led project using our ACG tool to identify cohorts of patients based on their own specific criteria. They place these patients on a register that can be identified and monitored within the tool for an increasing risk of admission, at which point they are placed on an MDT register for discussion between health and care providers. Their risk is then monitored to identify whether interventions aid reductions in A&E admissions and long-term hospital stays. We are now supporting system partners to understand the profile of instances that were eventually taken on as part of the caseload, to assess the risk level of these patients.

Working with our partner Experian, we applied Mosaic social indicators at household level to our linked and risk-stratified dataset to generate insight into costs and settings of care. We observed that some indicators (e.g. patients living at home alone) have as much effect on health and social care costs as some clinical diagnoses (e.g. COPD). Our partner the University of York Centre for Health Economics verified these observations and local health communities are working to introduce social factors into their service redesign programmes.  For example, working with the voluntary sector, an ‘End Loneliness in Mendip’ website has recently been created. We translated results using mapping software to create more engaging ways to present data for non-analytical audiences and to demonstrate impacts more clearly for clinicians and other decision makers. As a further step in Somerset, Patient Activation Scores are being collected and linked to the data set, as are measures for generic health (EQ-5D) and mental wellbeing (Warwick-Edinbury). Again working with the University of York, we are using regression analysis to understand the effect of these characteristics and indicators on the per-patient cost of health and social care in Somerset to inform health planning.

Our partners JHU have extensive experience of these approaches. They recently worked with the Canadian Institute for Health Information to apply a deprivation index to case-mix data to further stratify the population and explore issues of impactability. The key components of the index are both material (% population without high-school graduation, employment ratio, average income), and social (% families headed by a single parent, % population living alone, % population separated, divorced or widowed). 

An alternative approach is to use individual-level information. We support the further work required in the UK on recording this information consistently and linking with non-health data sets. Our partners JHU have developed models in the US and Canada which take account of individual level demographic data, for example, employment data.  Examples include models such as the ‘County Health Rankings Model’ to identify social, economic and behavioural factors that contribute to wellness, health and length of life, and their published development work on ‘Risk Adjustment for Socioeconomic Status or Other Sociodemographic Factors’.

As an example of our ongoing commitment to developing new measures, we are developing a measure for fragmentation of care, working with JHU to test, adapt and validate US-developed measures for use in the UK. One of the algorithms within the ACG system that is not currently used in the UK looks at how well a patient’s care is co-ordinated based on how many different doctors they see and how many different providers they have received care from. This is an extremely useful marker and one that can potentially support us to identify people who may benefit from an intervention based on improving the co-ordination of their care. The measure of fragmentation of care can be applied as an additional differentiator for high-risk patients, to identify the cohort of patients experiencing care fragmentation for whom improved continuity and coordination would be an effective intervention.

In 2016, SCW won the national tender for formal evaluation of the Symphony PACS Vanguard programme in Somerset. This was a combined bid from SCW (comprising local analysts and staff from the Quality Observatory that we host, based in Sussex) and the University of York Centre for Health Economics, who would lead the quantitative econometric evaluation based on Symphony patient-level linked datasets previously created by SCW. We later collaborated with the South West Academic Health Science Network to develop the delivery model to further support qualitative evaluation.

To execute the analysis of South Somerset’s integrated care solutions (Complex Care Hub, Enhanced Primary Care), our partner the University of York is conducting propensity score matching combined with a ‘difference in differences’ analysis (DiD) to establish detailed patient-level cohort case matching. Methodologically, it is analogous to a controlled before-and-after study and is considered a robust method for testing for an intervention’s effects in the absence of randomisation. Suitable controls are also being introduced to allow capture of the ‘counterfactual’ – that is, what would have happened in the absence of an intervention. This relies on identifying one or more matched cohorts of control patients. The evaluation is ongoing, and we are reporting to the NHSE New Models of Care team on a quarterly basis concerning observations and results from the Vanguard.

We propose two different approaches to meeting this requirement, although we recognise that in the reality of individual call-offs a ‘hybrid offer’ may be required depending on the individual maturity assessment of existing assets and the legacy systems involved.  

  • A ‘system agnostic’ approach: in which SCW co-delivers with any provider of electronic patient records. This would be particularly relevant where we are supporting existing and established interoperability projects, such as the five major programmes we are currently supporting in central and southern England. 
  • A ‘Cerner specific’ approach: in which we work in partnership with our interoperability provider of choice, Cerner, to develop the concepts from a ‘zero base’ position, which may be more appropriate for health systems yet to physically start the interoperability journey.

SCW offers deep strategic and technical expertise combined with a rigorous programme management methodology that we have used to develop integrated care records for several major health systems. We have been central to the development and deployment of five major integrated care record programmes - Connected Care (Berkshire/Frimley), the Hampshire Health Record, Joining up Your Information (Gloucestershire), the Oxfordshire Care Summary, and Connecting Care, (Bristol, North Somerset and South Gloucestershire) covering a total population of more than 5 million. We continually enhance our expertise in this area and our insight into health system requirements and the national ‘direction of travel’ by actively supporting and advising NHS England programmes relating to Interoperability, Target Architecture, Personal Health Records, Cancer digital programmes, and Local Health and Care Records – this has included preparing the business case analysis for the national LHACR programme.   

Working with our partners in University Hospital Southampton, our lead Chief Clinical Information Officer undertook a preliminary analysis of the impact of the use of HHR as an integrated Local Health record in Urgent Care scenarios. The study looked at what impact access to an integrated longitudinal records had in the management of patients with complex care needs in an A&E environment. Our study looked at matched cohorts of patients and used the audit data on the system to compare the outcomes for those patients where clinicians had accessed and used the shared care record for triage and the control group where they had not.

The outcome of the preliminary study showed that compared to the control group the use of HHR to support decision-making resulted in a statistically significant reduction in length of stay by two days. This could translate into potential savings of between £420 and £740 per patient, resulting in substantial savings at whole-system level. Portsmouth University are now developing a full study on the back of this to articulate the full benefits to the health and care system.

Want to see more

Why not browse the entire SCW case study library ?

SCW Case Studies