Can AI fix the NHS?
A quick post today, in response to Tony Blair’s widely-reported idea that AI can solve funding problems in the UK’s National Health Service (NHS)1. Is this feasible? Has AI, and presumably by “AI” he means ML, reached a stage of maturity where it can be relied on in a sector where safety is so important?
Reading the Tony Blair Institute for Global Change’s report on this topic, the focus is very much on data. In particular, they propose setting up a “national data trust” where all health data from NHS patients will be made available for research and development purposes. One idea is that this mass of data will attract investment and commercial research, with access fees giving a further boost to NHS funding. Though probably the most important idea is that AI tools trained on this data will reduce costs through things like automated screening and early diagnosis.
Access to data is important. Currently it’s pretty hard, at least in the UK, to get hold of medical data. Generally you need to know the clinician who collected it, you need to persuade them that you can do something useful with it, and you need to convince the NHS ethics committee that you should be allowed access to it. This last point particularly grates with me, having filled in a 70 page NHS ethics form in the past for precisely this purpose.
So, making data available, and large quantities of data at that, is highly desirable. Yet, there seems to be a notable gap in the plan somewhere between providing data and getting dependable AI tools back. This “magic happens” presumably lies in training and evaluating ML models — but if you’ve read this Substack for long, you’ll know that doing so is less than straightforward, and full of pitfalls. And these pitfalls are particularly worrying in the context of medicine, where decisions made by AI tools have a profound influence on people’s lives. Oh, and there’s also the small matter of organising and cleaning the data, which could be time consuming, to say the least.
Big data generally means big models, and big models generally mean deep learning. But deep learning also means a lack of transparency, and a lack of transparency is a troublesome thing for medicine. So yes, the AI tools may actually be good at screening, diagnostics, prognostics, and other phases of the healthcare decision pathway, but we’re unlikely to know how it reached these decisions. And this means that clinicians can’t explain these decisions to patients — without at least looking at the data and seeing whether they agree with it — and this takes away many of the advantages of automation. So, in practice, the kind of efficiencies that Tony Blair imagines may be difficult to achieve.
And this is even moreso for generative AI. In theory, we could use all this lovely data to train our large language models, and then allow clinicians and patients to query these models for advice. But in practice, the lack of transparency inherent in these models, and their well-known deficiencies in terms of dependability and fairness, may again make this difficult to achieve.
Nevertheless, a context where a “national data trust” may be really useful is within rare diseases. There are lots of these, and their rareness makes it challenging, and often impossible, to collect meaningful data sets that can be used to to do things like train ML models. If (and this is a sizeable if, given the need to gain informed consent from patients) the trust is able to collect data from all the patients with a rare disease in a country with a population of ~66 million people, then it seems likely that reasonably-sized data sets may be achievable. And this could really stimulate research into treatment and cures, and perhaps even AI tools.
So, whilst I think bringing together health data in one centrally-accessible repository is potentially a very good thing, I think we also need to address the limitations of current ML models and the issues surrounding ML practice before this can be fully leveraged in the way that Tony Blair imagines.
For anyone unfamiliar with the NHS, it’s the UK’s public health service. In a country where private healthcare is still relatively rare, the NHS really dominates the UK’s health sector. Currently it’s looking a bit ragged around the edges though, due in part to the UK’s rapidly aging population, and in part to other more economically and politically-tinged issues.