Risk Stratification and Model Development: Potential of “new” data and Predictive Modelling Stephen Sutch, MAppSc, BSc. Doctoral Student Johns Hopkins Bloomberg School of Public Health Baltimore, Maryland 21205 USA [email protected]Presented at Nuffield Trust 13 June 2012
21
Embed
Stephen Sutch: Risk stratification and model development: Potential of new data and predictive modelling
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Risk Stratification and Model Development: Potential of “new” data and Predictive Modelling
Stephen Sutch, MAppSc, BSc. Doctoral Student
Johns Hopkins Bloomberg School of Public Health Baltimore, Maryland 21205 USA
• Risk stratification of whole population • Improving the use of clinical data in predictive
modelling – Use of other data, Rx, Labs, frailty ….
• Build models for specific purposes/outcomes • Classification and Predictive Modelling, contextual
information
Copyright 2008 Johns Hopkins University 3
Working Definitions
• Case mix / risk adjustment (RA) - taking health status / risk into consideration for health care finance, payment, provider performance assessment and patient outcome monitoring.
• Predictive modeling (PM) - prospective (or concurrent) application of risk measures and statistical technique to identify “high risk” individuals who would likely benefit from care management interventions.
3
Copyright 2008 Johns Hopkins University 4
Needs Assessment
Quality
Improvement
Payment/ Finance
The risk measurement pyramid
Case- Management
Disease Management Practice
Resource Management
High Disease Burden
Single High Impact Disease
Users
Users & Non-Users
Management Applications
Population Segment
Copyright 2008 Johns Hopkins University 5
Using Predictive Modeling to Assign Persons Within the Care Management Pyramid
5%
Level 3 High risk
with multiple chronic illness
15% Level 2
Moderate risk patients with single chronic
illness or risk factors
80% Level 1 Low risk
Intensive Case and Disease Management
Health Coaching and Lifestyle Management
Health Education and Promotion
Copyright 2008 Johns Hopkins University 6
Purposes of Predictive modeling
• Clinical prediction - Individual patient, to improve clinical decision-making
• Population predictive models - Groups of patients, to forecast healthcare trends and identify candidates for healthcare interventions (e.g. DM programs)
Copyright 2008 Johns Hopkins University 7
Key non statistical considerations for model selection if it is to be used administratively
• Transparency – How easily can the model be understood and
explained?
• Clinical Texture – Does the system make sense to clinicians?
• Flexibility – Does the system support a range of applications?
• Customisable – Adjusts to local data, new models easy to derive and
validate?
Copyright 2008 Johns Hopkins University 8
Prior High Cost Year-1 (Prior Use)
Predicted High Risk
Year-2 (Using Year-1
Data)
Actual High Cost
Year-2 Not High Risk
High Risk, Current Costs Low, Future Costs High
Value of Predictive Modeling Population of Persons Across Two Year Period
Copyright 2008 Johns Hopkins University 9
Data
• Secondary Care – Acute Hospitals, Inpatient, Outpatient,
– Mental Health, Rehabilitation, Community care
– Diagnoses, Procedures
• Primary Care – Attendances, Diagnoses, Prescribing
– Labs, Examinations, Findings, Dispensing
• Patient Data – Risk factors, lifestyle factors, Health Status, Rx
Possession, Self Care
Copyright 2008 Johns Hopkins University 10
Distribution of READ Codes: Illustration Drugs 39%
Findings 23%
Procedures 17%
Administration 11%
Clinical findings 8%
Other 2%
Copyright 2008 Johns Hopkins University 11
GP diagnosis Coding and Drug prescribing
Diagnosis coding & drug prescribing by GP
PCT data US data
Prevalence Diags/Drugs Prevalence Diags/Drugs
Asthma 8.69% 3.60% 0.71% 4.38%
Dx + Rx Dx Only Rx Only
9.77% 2.67% 1.48% 5.63%
Dx + Rx Dx Only Rx Only
Congestive Heart Failure 2.52% 0.18% 0.05% 2.29%
Dx + Rx Dx Only Rx Only
1.85% 0.30% 0.85% 0.70%
Dx + Rx Dx Only Rx Only
Depression 6.23% 1.36% 0.25% 4.62%
Dx + Rx Dx Only Rx Only
10.38% 1.28% 0.66% 8.43%
Dx + Rx Dx Only Rx Only
Diabetes 3.91% 0.60% 3.25% 0.06%
Dx + Rx Dx Only Rx Only
5.45% 2.77% 2.23% 0.44%
Dx + Rx Dx Only Rx Only
Hyperlipidemia 5.32% 1.28% 0.22% 3.82%
Dx + Rx Dx Only Rx Only
14.87% 5.23% 6.85% 2.78%
Dx + Rx Dx Only Rx Only
Hypertension 13.09% 4.53% 0.45% 8.11%
Dx + Rx Dx Only Rx Only
18.95% 8.78% 6.05% 4.12%
Dx + Rx Dx Only Rx Only
Copyright 2008 Johns Hopkins University 12
Stratifying Whole Populations
• Multimorbidity – Understanding and measuring
• Classification of health need – Stratification of disease popultions
• Multiple purposes • Validation on whole populations
– Generalisable?
Copyright 2008 Johns Hopkins University 13
Co-Morbidity is key – Multiple morbidities encountered in UK GP practices
Average consultation in elderly involves someone with 1.9 QOF diseases and 6.7 chronic diseases using ACG/EDC chronic disease designations
Source: Salisbury et al. From GPRD data, 488 practices 2005-2008