Martin Bardsley: Predictive risk 2012: context
1 like471 views
This document provides an overview of predictive risk modelling in healthcare: - Predictive models use existing patient data to identify those at high risk of future health events like hospital admissions. - Around 3% of patients account for nearly half of total healthcare costs. Predictive models aim to identify these high-risk, high-cost patients. - Models have moderate accuracy, correctly identifying 50-75% of readmissions and 5-10% of highest-risk patients. - There is a growing commercial market for predictive risk tools in England as new organizations develop their own models. - Simply having a predictive model is not enough; organizations must choose how to apply models through interventions for high-risk patients.
Martin Bardsley: Predictive risk 2012: context
- 1. Predictive Risk 2012: Context Predictive R 13 June 2012 Martin Bardsley Head of Research Nuffield Trust © Nuffield Trust
- 2. Predictive modelling • BMJ in paper* in 2002 showed Kaiser Permanente in California seemed to provide higher-quality healthcare than the NHS at a lower cost. Kaiser identify high risk people in their population and manage them intensively to avoid admissions • Modelling aims to identify people at risk of future event • Relies on exploiting existing information +ve: systematic; not costly data collections; fit into existing systems -ve: information collected may not be predictive • Use pseudonymous, person-level data • In health sector a number of predictive models are available e.g. PARR++ and the combined model. • *Getting more for their dollar: a comparison of the NHS with California's Kaiser Permanente BMJ 2002;324:135-143 © Nuffield Trust
- 3. Uneven distribution of costs The proportion of total costs spent on patients with category of annual costs (area of shape) with the proportion of all patients in annual cost band (dots) Around 3% of patients are responsible for nearly half the total patient costs © Nuffield Trust
- 4. Predicting admissions in advance Change in average number of emergency bed days Predictive models try to identify people here © Nuffield Trust
- 5. Health and social care event timeline © Nuffield Trust
- 6. Patterns in routine data to identify high-risk people next year © Nuffield Trust
- 7. Distribution of Combined Model risk scores Importance of risk adjustment Very high risk Top 0.5% Top 10% High risk Moderate risk 0.5% - 5% Low risk 5% - 20% 10% - 45% 20% - 100% 45% - 85% 85% - 100% General population WSD participants – receiving telehealth or telecare © Nuffield Trust
- 8. Applications of predictive risk • Case finding for people at high risk of admission seen as increasingly important for people with LTCs and complex conditions • Evaluation and risk adjustment eg WSD • Predicting future costs eg work on resource allocation Related: Scope to make the most of linked data sets in describing care pathways © Nuffield Trust
- 9. Choosing the predictive model bit • What event should we be aiming to predict? • What models and tools are available? • What data do I need and how often? • How often do predictive models need to be run? • How accurate is the model? • How much does it cost? © Nuffield Trust
- 10. (1) Predictive tool = Predictive model + Software platform Inputs Processing Inpatient data Outpatient Tools to Predictive organise model GP data input data Population data Presentation and Outputs Patient lists with risk analysis tools score -Gaps in care -Priority lists Users © Nuffield Trust
- 11. Age distribution and mean risk scores within a diagnostic categories © Nuffield Trust
- 12. Testing for gaps in care © Nuffield Trust
- 13. (2) Key metrics for performance of a model (PPV and sensitivity) 100% Positive predictive value 90% When the model says high risk how often 80% is it right? Sensitivity (Brown) 70% PPV (blue), 60% 50% 40% Sensitivity 30% What proportion of all 20% events will the model detect? 10% 0% 0 10 20 30 40 50 60 70 80 90 100 Threshold value (lower bound of defined 'high risk' group) Pooled 4-site 1k model © Nuffield Trust
- 14. Typical performance of models – predicting events next year Predicting ... How many What proportion of all positives are events are found correct (PPV) (Sensitivity) Readmission based on prior 50%-75% 30-50% admissions eg PARR Admission to hospital from a 20-50% 5%-15% general population As above but just for highest 70-80% 5-10% risk groups (top 10%) Changes in social care use 20-50% 5-15% © Nuffield Trust
- 15. (3) Emerging market in England • August 2011, the Department of Health announced that it had no plans to commission national updates of the latest Patients at Risk of Re-hospitalisation tool (PARR++) or the Combined Predictive Model • Range of new/established commercial organisation developing risk tools • Creation of new commissioning groups and new markets • Increasing ease of accessing GP data • Continuing financial pressures and the search for ways to reduce emergency hospital care. © Nuffield Trust
- 16. Examples of case finding models available (with or without software platforms) SPARRA PARR (++) SPARRA MD Combined Predictive Model PRISM PEONY AHI Risk adjuster LACE ACGs (Johns Hopkins) MARA (Milliman Advanced Risk Adjuster) DxCGs (Verisk) Dr Foster Intelligence SPOKE (Sussex CPM) QResearch models eg QD score LACE RISC Variants on basic admission/readmission predictions: Short term readmissions Social care Condition specific tools Costs © Nuffield Trust
- 17. (4) The model by itself doesn't change anything... Choosing an application • Which people should I target? • What interventions should we use? • Who will use it and how? What clinical staff need to see results? • Will some patients benefit more than others? • When can I expect to see a return on investment? © Nuffield Trust
- 18. Summary • Predictive modelling is a practical case finding tool for identifying high risk patients • Growing market for predictive models – extending beyond simple annual predictions of readmissions • Ability to look at linked data valuable for other analyses • Technical details of model performance is important – but so how is the way the model is implemented • We hope today's conference will help you learn more about peoples’ experience of using these models. © Nuffield Trust
- 19. The day ahead • A review around the UK • Examples of different ways that risk models have been applied in the NHS • A view from outside the UK Germany and US. • Developments in modelling • Open session...share your experiences. © Nuffield Trust
- 20. www.nuffieldtrust.org.uk Sign-up for our newsletter www.nuffieldtrust.org.uk/newsletter/login.aspx Follow us on Twitter (http://twitter.com/NuffieldTrust) © Nuffield Trust © Nuffield Trust