Predicting risk of hospital admission and extracting GP data David Osborne Senior Public Health Information Analyst NHS Croydon
Overview Risk stratification Predictive algorithms Extraction of GP data Evaluation
Risk stratification Disease management Supported self-care Prevention and wellness promotion Very high relative risk 0.5% 10% of emergency admissions High relative risk % 25% of emergency admissions Moderate relative risk % 25% of emergency admissions Low relative risk % 40% of emergency admissions Case management Disease management Supported self-care Prevention and wellness promotion
Frequently-admitted patients Source: Department of Health
Frequently-admitted patients Source: Department of Health
Emerging risk Source: Department of Health
Techniques to find high risk patients Clinical knowledge Referral by clinicians Threshold modelling Set of criteria e.g. aged >65 having 2+ emergency admissions in last year Predictive modelling Use historical data to quantify future risk of admission
PARR++ In-patient Records PARR++ Algorithm Patient NHS No.Risk Score XXXXX170.2% XXXXX234.9% XXXXX394.0% Linked date ???
PARR++
Combined Model GP Records A&E Records Out-patient Records In-patient Records Combined Predictive Model Algorithm Patient NHS No.Risk Score XXXXX170.2% XXXXX234.9% XXXXX394.0% Linked date ???
Combined Model
UK context SPARRA ( PRISM ( id=770&pid=33635) id=770&pid=33635 PARR++ and Combined Model (
Use of risk scores for case management: Croydon virtual wards Quarterly list of top 1000 highest risk patients Sent to virtual ward team (community matrons and ward clerks) Discussion between community matron and patient’s GP Patient admitted to virtual ward Regular contacts with community matron When risk score reduces, patient considered for discharge
GP data required for running combined model List of registered patients Read code history for last 2 years Disease registers for long term conditions
Extraction methods Manual extract using MIQUEST Companies provide service to do regular extractions e.g. Apollo, GraphNet, HeathAnalytics General Practice Extraction Service (GPES) being set up by NHS IC (
Information governance Conditions for processing ‘sensitive personal data’: ‘The processing is necessary for medical purposes, and is undertaken by a health professional or by someone who is subject to an equivalent duty of confidentiality.’ (Data Protection Act, 1998) ‘Retraceably pseudonymised data may be considered as information on individuals which are indirectly identifiable. … In that case, although data protection rules apply, the risks at stake for the individuals with regard to the processing of such indirectly identifiable information will most often be low, so that the application of these rules will justifiably be more flexible than if information on directly identifiable individuals were processed.’ (European Data Protection Working Party, 2007)
Obtaining support of GPs Get support of PEC and LMC Show GPs the benefits and provide them with the results More likely to respond if someone they know approaches them direct
Example: Croydon agreement Specific uses agreed for aggregated data including public health purposes Data on individual patients can be used only to identify patients for case management Additional uses agreed with LMC on case-by-case basis
Case management PSA target to reduce emergency admissions by 5% between 2004 and 2008 used as a lever for change 1 Good evidence that case management improves quality of care but majority of studies show no effect on reducing emergency admissions or bed days Evaluation is complicated by regression to the mean 1 Predictive risk project literature review, Kings Fund, 2005
Regression to the mean
Evaluation Simple before versus after analysis misleading Start of intervention
Evaluation If tracking patients over time, compare with a control group
Croydon examples Comparison with a historic control group Change (before/after) in virtual ward group Change (before/after) in control group Results of paired t-test Emergency admissions (p=0.99, 95% CI -1.41, 1.43) A&E attendances (p=0.46, 95% CI -0.96, 2.14) Average length of stay (p=0.85, 95% CI -4.96, 4.10) Emergency bed days (p=0.67)
Croydon examples Modelled rates by risk score by month compared with actual rates for patients on virtual wards
Croydon examples Trend in average number of emergency admissions for top 100 patients in each year
Future developments PARR and Combined Model to be revised Predicting risk that is ‘impactable’ Interventions further down the risk pyramid e.g. health coaching Other applications of risk stratification Resource allocation, performance management
Further information Kings Fund NHS Networks pages Nuffield Trust GPES Evidence David Osborne, Senior Public Health Information Analyst, NHS Croydon