User beware: How the FDP could increase health inequalities

The NHS Federated Data Platform promises data integration, but a closer look reveals potential pitfalls in perpetuating biases. By Andi Orlowski and members of the Chief Analytical Officers Network.

The NHS has one of the most valuable data troves in the world. If that data could be brought together in a way that was easy to analyse, it would be the foundation of much beneficial research and service improvement.

At a more basic level, as Sir Muir Gray correctly lamented in HSJ recently, without some form of national data environment, “we have no means of knowing how much was spent on people with respiratory disease or any other condition.”

In this context, the arrival of the Federated Data Platform programme is as welcome an innovation as Sir Muir suggests.

The FDP apparently does exactly what is needed in bringing NHS data sources together in a meaningful way. Yet, rather than inspiring confidence that the NHS can now effectively mine its data, NHS England appears to be doubling down on the FDP’s role as a tool rather than a depository.

By focusing on the add-ons rather than the core proposition, NHSE risks creating confusion, raising concern and mistrust.

A perfect example of this danger came during recent demonstrations by Palantir and NHSE of how the FDP can be used to audiences of healthcare leaders. A range of influential people were left coo-ing at the smooth functionality which allowed the easy combination of data sets. The (relatively few) NHS data analysts in the audiences, of which I was one, were left with a deep sense of unease.

Let me explain why.

At first glance, the FDP’s “Population Health Campaign Manager” looks great. It is a simple “plug and play” tool that walks the user through each step of the process. But beware… because that process could easily turn out to be a simple three-step journey to harm.

Step 1 – Bake in inequalities

You start by selecting your cohort. Simply click the group you want to analyse (diabetics, diabetics with hypertension, etc) and then those that are coded to those diseases.

Hey presto, you have your target list – and that is where the trouble starts.

There is a lot of bias baked into this data. There is, for example, much more information on white middle class people in NHS databases, so just using coded data leads to inequalities. The people selected may end up being offered interventions before others who need them much more badly.

Analysts often use proxies taken from other data sets to try and address this bias. However, at a national level, the FDP only contains secondary care data. The standard offer does not include primary care data, a great source of nuanced data with lots of potential useful alternative codes to add in to address inequalities.

Step 2 – Reinforce the inequalities

Next, there is a list of proven interventions for you to choose from. You can also select from a number of published papers (using the abstracts) to evidence your decision.

But, again, take care!

First, interpreting evidence is often complex. Results are rarely unambiguous and always depend on the context in which an intervention is implemented.

For example, how relevant is the intervention you have just selected? Maybe the people you want to reach are already being served in this way or have been in the past. Maybe the intervention failed them, or the take up was low.

What is more, not all interventions are equal and nor is all published research. The FDP offers no clear weighting for the interventions it offers up. You can easily pick an intervention completely ignorant of the large systematic literature review that suggests it does not work.

Step 3 – Deepen the inequalities

So, you have your cohort, you have chosen your intervention and have evidenced why it was a good choice. Now to contact the cohort to get the intervention going.

The FDP provides information on patient contact preferences, and you can choose “call, text or letter” (but not email apparently!?).

But such messages out of the blue have a very poor record of reaching vulnerable groups, whose relationship with the service is often not the one of trust and understanding enjoyed by those lucky enough to have more stable lives.

There is also no way to let the healthcare provider who will be delivering the intervention, or others who will be dealing with its consequences, know this contact had been sent.

Once you have messaged your cohort, you can monitor which people have taken up your intervention and see how successful you have been.

And the answer may be “very successful” in terms of take up. But remember all those biases that were built in to that intervention – and then consider how the data that will flow from these interventions may deepen those biases even further as the results are coded.

The FDP offers cost benefit analysis for your intervention. This is great when looking at the cost of an intervention in isolation, but you also have to be aware the analysis will be massively oversimplified.

To get a meaningful result, you need to consider whether, for example, large-scale interventions that work for many may not be applicable for more vulnerable groups. Do you exclude “non-responders” from your analysis (again likely the most vulnerable)? The FDP does not offer any alerts warning you to take care when conducting your analysis.

Each step in a PHM campaign needs much more careful consideration than the FDP tool encourages. With the right engagement and nuanced application none of the issues addressed above are insurmountable, but it is worrying that this tool has been developed without some scrutiny, or even awareness that such an application was in scope for the FDP.

The tool must be tested and refined together with expert analysts and the communities who live and breathe these issues. Until then it should be used with considerable caution.