Chris Morris on healthcare cybersecurity

One of the principal analysts at South, Central and West Commissioning Support Unit talks to Lyn Whitfield about the evolution of risk stratification in the NHS, and what the future holds for healthcare analytics, in the run-up to this year’s EHI Live.

It’s not exactly news that the NHS is under enormous pressure. Or that some individuals make much more use of emergency and acute services than others.

So, what if it was possible to relieve some of the pressure by identifying those individuals, sorting out their care, and making sure they didn’t need so many expensive trips to hospital?  That, broadly, was the idea behind risk stratification.

Chris Morris, principal analyst at South Central and West Commissioning Support Unit, explains: “Risk stratification is about using data to identify people at the highest risk of a particular outcome.

“The early models were about identifying emergency admissions to hospital. Emergency admissions are expensive, at least sometimes avoidable, and often a not particularly great outcome for patients when they happen, so there were good reasons for making that the first area of focus.”

The great pyramid of Kaiser

Risk stratification is not a new idea. It arrived in the UK about 15 years ago, when the British Medical Journal published a series of papers about the work of US health management organisation Kaiser Permanente.

These helped to popularise the now-familiar Kaiser triangle or pyramid; a graphic representation of the population that rises from a broad base of healthy people making little or no use of services to a peak of highly complex patients.

They also helped to popularise the idea that by focusing the right services on the right groups, and ‘case finding’ individuals for intensive support, it should be possible to ‘move people down the pyramid’; delivering a more appropriate service while saving money in the process.

Tools for doing this arrived in the UK a couple of years later, when the King’s Fund, New York University and Health Dialog launched Patients at Risk of Re-hospitalisation, or PARR.

More sophisticated tools, using a wider range of predictive data, pulled from multiple datasets, and aimed at finding patients at risk of admission as well as re-admission, followed. However, while the idea behind risk stratification is simple, implementing it has proved to be anything but.

Dicing with data

One issue lies with the tools themselves, and what is known as their positive predictive validity. Put simply, tools that identify a lot of people as being at risk of a particular outcome may pick-up a lot of people who won’t in fact, experience that outcome.

And vice versa (tools that identify just a few people as being at risk of a particular outcome are more likely to be right about that; but to miss a lot of others). Another issue is what is known as ‘impactability’; is it, in fact, possible, to ‘move people down the pyramid’?

Morris says this is a big issue in the UK because “while it feels intuitively right that you can identify these people, and potentially avoid some of their costs and adverse outcomes, there is an argument that a lot of the admissions experienced by people at the top of the pyramid are not, in fact, avoidable.”

Even if project has a good outcome, it’s not always clear why. For example, many areas of the UK that have used risk stratification and case finding have used them to recruit patients onto ‘virtual wards’. If GPs also have ‘admission rights’, then is it the model, or the ward, that has made a difference?

Updating the toolkit

So, this is a complex area. But that’s no reason to give up. Johns Hopkins University in the US has developed a suite of population health management tools called Adjusted Clinical Groups, or ACG.

Morris has been working with the university, and previously with the UK company Sollis, to refine and develop these tools, ‘recalibrating’ them to take account of changes in clinical coding, and exploring whether new indicators, such as smoking status or BMI, could make them more accurate.

As part of this work, five new predictive models have been developed. Morris outlines how one of them focuses on people in the top 20% of an area’s costs for three, consecutive six-month periods. Emergency hospital admissions will be just a percentage of these costs, so the model is finding people who experience a lot of other interventions as well.

“You find many more true positives – that is, people who actually experience the outcome that is predicted – than you do with models that just look at hospital admissions,” Morris explains. “Even at the top end of the predictive models for emergency admission, you find a lot of false positives – just because it’s a relatively rare outcome across a population as a whole.

“That means the probability of correctly predicting an emergency admission, that also happens to be preventable, is quite low. Diffuse costs are much more common – so you have a better chance of finding and addressing them.”

Really understanding pathways

ACG is about more than risk stratification, though. The tools are designed to let commissioners understand how different factors affect the health of their populations and how this will impact on the cost and performance of healthcare systems.

For example, Morris says, his CSU has been doing some work on what he calls ‘clinical clusters’; how patients with one condition, such as diabetes, are likely to experience one or more additional conditions, such as coronary heart disease and hypertension.

Which, of course, has a significant impact on their use of health and care services. Morris argues that this is something commissioners will need to keep this in mind as they get into the Five Year Forward View agenda of redesigning patient pathways and creating new, integrated services for their populations.

“Lots of people want to redesign their diabetes pathway, or their COPD pathway, but we need to be thinking about more than one condition,” he says.

“Otherwise, you can imagine getting the pathway right for every condition that a patient might have, and still failing them, because they would still be having multiple points of contact with services that duplicated each other along those different pathways.”

Data goes mainstream

This is complex, sophisticated work to be tackling, when the NHS is under immediate financial pressure, hospitals are missing targets, and the health and care system is being told to get ready for a winter that may bring a flu outbreak with it.

Yet Morris says the Five Year Forward View, and the subsequent development of sustainability and transformation plans and accountable care systems, has generated a whole new level of interest in using data, modelling and analytics in healthcare.

“People are really starting to think about this,” he says. “Nobody thinks it is going to be easy, trying to do redesign at a time of great financial difficulty, but with the arrival of STPs and new models of care, people realise they need to take these steps.

“So, suddenly we are pushing at an open door.” In the longer term, he adds, more sophisticated analysis tools will enable commissioners to do the kind of intelligent commissioning that they have always wanted to do.

“If you go to all the effort of collecting population data, and you do all the information governance and everything that comes with doing that, you don’t just want to be using it for risk of emergency admissions,” he says. “You want to be using it for lots of other things that you want to do.

“So, where we are moving to is not collecting data for risk stratification, and then using it for commissioning, but collecting information for commissioning, and then using it for risk stratification, among lots of other things. Risk stratification will just be part of the toolkit.”