At a time when the NHS continues to face substantial financial and demand pressures the use of predictive analytics to plan services and allow patients to be treated sooner, could help alleviate some of the strain being placed on our health system.

Predictive analytics use technology and statistical methods to trawl through huge amounts of historical patient data and information to try and establish patterns and predict future demands. In a nutshell, they help to predict outcomes for individual patients and populations.

One way of looking at the current situation in the NHS is to say that it’s a victim of its own success: public health has improved and we are all living longer. However, healthcare costs are increasing. The underlying growth in costs in the health sector is outstripping growth in costs for other businesses by broadly double and, unfortunately, this shows no sign of slowing.

Just look at diabetes; the cost of diabetes drugs has almost doubled in the last decade, while the number of diabetics has grown by 65%.

Predictive analytics can be used to help target treatments and resources as efficiently as possible, stripping out waste and improving patient outcomes.

Predictions for individuals and populations

On an individual patient basis, there are already a number of tools we can use to help drive patient behavioural change, like QRISK, a prediction algorithm for cardiovascular disease that uses risk factors like age, blood pressure, smoking status, weight and family history to predict a patient’s chances of developing heart disease in ten years’ time, for example.

Once the patient knows their level of risk they can start working on ways to reduce their chances of developing heart disease. QRISK can also help the clinician judge which patients will benefit most from taking particular drugs to lower their cholesterol or blood pressure.

This sort of advanced modelling is now widely used in general practice. On a larger scale, there is all sorts of information that affects how populations interact with the health service.

Take the weather. If we know a cold snap is coming, we can assume that our emergency departments and out-of-hours services are probably going to be busier, while pharmacists are going to have to work their supply chain so they have more Paracetamol and cold relief products in stock.

However, using predictive analytics we create algorithms which continuously analyse what actually happened during similar weather events in the past, and then make sure more staff are on duty to be ready for peaks in demand, or stand staff down at times when the data shows that they won’t be very busy.

The next wave

The next wave of predictive analysis will come around machine learning - where huge volumes of data are uploaded into a powerful analytical engine which then identifies relationships between different data items. Rather than being taught those relationships; the machine is then ‘learning’ – the richer the data, the more it learns.

Although still in its infancy, the likes of IBM and its Watson computer system, and Google’s Artificial Intelligence arm DeepMind are really interesting.  My experience is that these kind of systems only really work if you take the time to preload them with knowledge that you’ve already got, so the starting point is often to feed in years of existing scientific research.

For example, we might know that patients with diabetes are best treated with two particular drugs, because that’s what the scientific research says. We could then apply that research to a sample of the population to find out if the treatment is more effective in certain subsets of the population than it is with others. You can then refine who gets which drugs because we can work out the chance of their diabetes getting better or developing side effects. That can then allow drugs to be targeted more efficiently.

The more data you have, the more accurate predictive analysis becomes.

Enhancing accuracy

At the moment, we can look at patients’ medical records, but in the future we’ll be able to look into their genetic makeup too, which will bring another dimension to the data. Your genes pre-programme your body as to how it will respond to medicine. For example, some people have a gene that means some common drugs used to treat tuberculosis simply won’t work on them because they can’t metabolise them, and so they will need a different type of treatment. Without genetic profiling, the treatment for TB was trial and error. The potential here for improving oncology outcomes is massive.

By adding more information, the more accurate the prediction becomes. Unfortunately (or maybe fortunately), it is unlikely machines will ever be able to predict everything – random environmental and behavioural events are part of what make us human. There are still challenges that need to be overcome for predictive analytics to be widely used and which have slowed development down, but none of them are insurmountable.

Overcoming challenges

First and foremost predictive analytics need to be proven to be safe and effective, much in the same way that a new drug does. That involves a level of academic rigour that is time consuming. New drugs take five to eight years to come to market. These predictive algorithms are treated just like drugs because they can affect a patient’s medical outcome in the same way that a drug can.

The drugs regulator MHRA puts algorithms through an approval process similar to  medical devices like pacemakers, which includes an academic validation, to make sure the implementation is safe. The next challenge is access to data.

Technical access isn’t the real problem (although huge processing power is often needed); the issues tend to be around complexity of privacy regulations and who is the controller of the data. The NHS produces billions of pieces of data every day. Each GP practice, hospital trust and ambulance service has data that they are legally responsible for, and data scientists can only access that data if they’ve been given permission.

Thanks to Fiona Caldecott’s review of how information about patients is shared across the health and care system, it is becoming clearer how these challenges will be addressed however previous mistakes around Care.data have made this a complex and unpredictable process.

Milestones

Over the next decade I think there will be three main developments for predictive analysis in the NHS:

Currently – we see better use of data for health service planning and face-to-face consultations between GPs and patients. QCancer, QRisk, EFi Scores, QAdmission are just a few data derived tools used for population risk stratification and in consultations to improve the quality of care within the NHS – all approved by NICE.

Within three years – we’ll see increasing use of predictive tools for the patient, which will include far more self-help. Patients will be expected to help manage their blood pressure or diabetes, because there simply isn’t the resource in the NHS to do that. A GP or another service will review their results once a month or so and give feedback. There is a growing realisation that the NHS can’t do everything and doctors must work differently. Going forward, they’ll be increasingly focused on patients with complex social, psychological, medical needs who need face-to-face contact.

Within five years – we will see more mainstream use of AI (artificial intelligence) in healthcare in the UK. We must make sure they are safe and regulated, and there will also be a huge cultural shift to go through and it’s likely that for the forseable future, AI will augment rather than replace the doctor… but who knows!

It’s clear that data driven healthcare has the potential to make an important contribution to helping the NHS become more efficient and therefore more accessible for patients. The next decade will see it being applied in almost every aspect of the service, as the quantity and quality of data available gives ever greater insight into the health of the nation and the resources needed to serve the public effectively.