Predict and survive
- 29 March 2010
 |
|
Predictive risk modelling is based on using |
Anyone looking into a crystal ball to divine the future of the NHS right now would expect to see little more than a fog of uncertainty, perhaps with a giant £ sign wreathed in mist. There might, however, be one small area of clarity about future activity and it might be labelled “predictive risk modelling”.
Predictive risk modelling is an idea that has been around for some time. It is based on using mathematical models to predict events – such as which people in a given population are most likely to be admitted to hospital in the next 12 months.
The theory runs that once you have identified those at risk, you can target them with preventive services; saving money and improving lives. And it looks set to be a growth area, since it hits several of the quality, innovation, productivity and prevention (QIPP) agenda objectives.
Proponents argue that it has the potential to improve quality and productivity by targeting services at those who need them most. It promotes innovation by requiring commissioners and health professionals to rethink their services. Finally, using predictive modelling is preventative, because it aims to identify those at risk of an adverse event and stop it happening.
There are more concrete reasons to see predictive risk modelling as one of the up and coming ideas of 2010-11. In the wake of last week’s Budget, the Department of Health identified £2.7bn of efficiency savings from improving the care and treatment of people with long term conditions.
Among the measures it highlighted was reducing emergency admissions. There are also several new tools in the pipeline to predict not just patients at risk of emergency admission but also those at risk of being admitted to nursing and residential homes.
Trusts are also experimenting with predictive risk modelling for those most likely not to attend their hospital appointments. Meanwhile, a large research programme is about to get underway to examine whether predictive risk modelling really delivers in practice.
History lessons
The names most associated with predictive risk modelling in the UK are Dr Geraint Lewis – a one-time public health doctor in Croydon and now senior research fellow at the Nuffield Trust – the King’s Fund and Health Dialog, now part of Bupa.
It was Dr Lewis who introduced predictive risk modelling in the early noughties. He developed an algorithm using Hospital Episode Statistics data to predict which patients on a GP’s list were most likely to be admitted to hospital in the next 12 months. He then worked with community nursing colleagues to develop a “virtual ward” with targeted nursing interventions that would help to keep them out of hospital.
The DH liked the idea and commissioned the King’s Fund, New York University and Health Dialog to develop two predictive models. The first was the Patients At Risk of Readmission case finding tool (PARR), which uses inpatient data to predict the likelihood of readmission.
It came with a front end that makes the algorithm usable, although even Dr Lewis admits it is “clunky”. It’s now getting out of date – it uses HRG3 definitions and the maintenance contract has not been renewed.
In 2006, the group developed the combined predictive model, which uses data from in-patient episodes, out-patients, A&E and general practice to predict which people in a population are most likely to need emergency admission to hospital the following year.
Although potentially more useful, the DH did not commission the front end, leaving it to health communities to develop or commission their own. Many of them have done so.
Ian Manovel, associate director of analytics at Bupa Health Dialog, one of a number of commercial players in this space, says: “We estimate that 80% of PCTs are now using predictive modelling. We are getting more and more queries from PCTs and SHAs and there is a real buzz around it.”
He has worked not just with individual PCTs but with entire health communities to develop more refined models for predicting events and tools to make them usable.
For example, Health Solutions Wales (the equivalent of England’s NHS Information Centre) is set to roll out its PRISM tool across Wales on 1 April 1, following a successful pilot scheme.
Working in Wales
It works like this: Health Solutions Wales takes data from GP practices and the acute sector, strips out the demographic data and runs the predictive model on this data.
The analysis is then fed back to PRISM, a web-based service to which GPs can log on and marry up demographic data with the analysis, allowing them to identify not only which of their patients are at risk of emergency hospital admission in the next 12 months but at how much risk they are. They can then target cohorts as they see fit.
Leo Lewis, a project director in Carmarthenshire on secondment from HSW, says not all doctors will take it up immediately. “The way we are engaging with our users is to give them a tool to identify cohorts of patients that they might be able to work with differently.”
Gareth John, head of information products at HSW, says this is not a static model. He hopes to integrate data from A&E as well as from social care in due course. It’s not just about lists of patients, either. Users can benchmark their population and practice against each other through the system, for example.
“Potentially it could be used for horizon scanning – picking up things that will become issues for patients in future. It is going to take a culture shift to accept that sort of thinking,” he adds.
Working in England
Another Bupa Health Dialog project is also about to come to fruition with the launch, in the West Midlands in mid-April, of a new Population Health tool kit. “It uses the combined predictive model but with a number of financial risks built in,” says Manovel. “Instead of measuring the risk of being admitted to hospital in the next 12 months, it ranks the costs if that risk becomes a reality.”
Other large scale developments are taking place in London and across Devon. Mr Manovel argues that there is a real benefit to working with a commercial partner in these projects. “PCT informatics departments are frantically trying to pump out management reports,” he says. “They do not necessarily have skill set to do the kind of statistical work needed to develop a reliable model or the time to develop the front end tool.”
One city that has developed its own model and tool is Bristol, where Andy Kinnear, head of Avon IM&T consortium explains: “We are using our GP extraction service to carry out risk stratification. We take Secondary Uses Service data feeds and GP feeds from primary care to build an index of patients. Our killer USP is a dashboard for users that brings in the risk stratification.”
He reckons that this tool can pinpoint 85% of the people who are at risk of being admitted to hospital in the next 12 months. “Of course, the holy grail is working out what you do with them,” says Kinnear.
The multi-million pound question
Indeed it is, and so far there is scant evidence that service redesign based on predictive modelling actually does what is claimed for it: reduce hospital admissions. Lots of small scale projects to implement virtual wards, in particular, have been evaluated positively. But Dr Lewis knows that is not the same thing.
“We know that they are intuitive and patients seem to love them, but we do not know whether they actually work,” he says. The Nuffield Trust is about to start on a research project that will look into this systematically.
The Nuffield Trust has also been commissioned by the DH to conduct a feasibility study of building predictive models for social care that would allow PCTs and local authorities to predict which people are at risk of being admitted to a nursing or residential care home.
Dr Lewis says: “Emergency admission lent itself to predictive modelling because it is undesirable, expensive, recorded in routine data, and potentially avoidable. I would argue that admission to a nursing or residential home meets the same criteria. Similarly, admission to an intensive care unit or an extended length of stay.”
Bupa Health Dialog is looking at building in community data from the RiO clinical system, community mental health data and social care data. In addition to allowing PCTs and practice based commissioning consortia to refine and target their interventions, this could also shed light on the relative costs of secondary and primary care.
Manovel says: “At the moment we can identify costs in secondary care but we cannot identify what it would cost to treat the same person in the community. Intuitively, it would be less but we do not know. This will allow us to make the rigorous comparison.”
This all matters desperately in the current financial climate but Manovel says: “Actually I couldn’t give a stuff about the money. The single most important thing about this is that it is an ethical issue.
“There are people who are disabled and society should care for them. We are failing to look after them properly and waiting for them to suffer trauma and end up in the back of an ambulance going to A&E. Using this data enables us to invest in looking after them better.”
Links
Latest from the Nuffield Trust on predictive risk modelling
