Following the Government’s announcement in 2010 that hospitals in England would not be reimbursed for some emergency hospital readmissions occurring within 30 days of discharge, the interest in finding ways of preventing these became greater than ever. The Nuffield Trust developed an algorithm for predicting which patients are most at risk of short term readmission, designed to be used by acute hospitals.
Ian Blunt, Nuffield Trust, outlines the motivation behind the development of PARR-30, a predictive model for readmission within 30 days of discharge
Approximately eight per cent of patients discharged from hospital are readmitted within 30 days, costing the NHS an estimated £2.2billion a year.
Although NHS hospitals can use a model for predicting readmission (called PARR), it makes predictions over a 12 month period rather than 30 days. PARR was primarily designed for use by primary care trusts (PCTs) or community services to identify patients suitable for case management, rather than for hospital clinicians preparing patients for discharge.
A number of hospital-based models to predict risk of short-term readmission (within 30 days of discharge) have been developed in other countries. A well-known example is the LACE index, developed in Canada, which was built using routine electronic information then converted into an index that hospital clinicians calculate manually when they are preparing patients for discharge.
Patients readmitted to hospital within a month of discharge cost the NHS £2.2billion a year, which is why this project is so important
There has been interest in the NHS in applying LACE in English hospitals but its reliability cannot be guaranteed when used with different groups of patients to the one it was developed for.
The PARR-30 algorithm was developed using English NHS information – a 10 per cent sample of all NHS hospital admissions in England over one year. Like LACE, it uses a small number of variables to allow risk scores to be calculated quickly and simply.
We intentionally selected variables that we believed would easily translate to information available from patients’ notes or from the patients themselves. We also explored the estimated costs of readmission to commissioners.
The percentage of inpatients identified as high risk (that is, at a risk score threshold of 0.5 or more), who were subsequently readmitted within 30 days was 59.2 per cent (with a 95 per cent confidence interval of 58.0 per cent to 60.5 per cent).
This proportion increases as the risk threshold is increased, demonstrating that patients predicted to have a higher risk of admission by the PARR-30 model were readmitted more frequently than those predicted to have a lower risk.
PARR-30 performs well when compared with international models. A common performance measure for predictive risk models is the ‘c-statistic’, and the value for PARR-30 is 0.70 (with a 95 per cent confidence interval of 0.69 to 0.70). A recent systematic review of predictive risk models for 30 day readmissions documented c-statistics ranging from 0.50 to 0.72.
In common with other models, the proportion of all people readmitted that are classed as high risk (known as the ‘sensitivity’) is quite low.
However, PARR-30 is proven to identify subgroups of patients that contain a high proportion of patients who will be readmitted within 30 days, and knowledge of the percentage of patients in each risk score band can be useful in determining appropriate resources for patients, with more or different types of resources assigned for patients who are most likely to have a hospital admission.
The full findings of this project were published in an article for BMJ Open in August 2012.
In August 2011, the Department of Health announced that it will not be commissioning a national upgrade of two predictive models: the Patients at Risk of Re-hospitalisation tool (PARR++) and the Combined Predictive Model. Scotland and Wales have their own predictive models called SPARRA and PRISM respectively.
In November 2011 we published guidance to help clinical commissioners choose predictive risk models from an open market, based on factors including: the outcome to be predicted, the cost of the model and its associated software, the availability of data, the accuracy of the predictions, and the preventive intervention to be offered on the basis of predictions.