I am cycling from London to Brighton to raise money for LSHTM’s scholarship fund. I don’t charge for my articles, so please sponsor me if you can by clicking below
This week, after a nudge a few weeks back from a helpful colleague, I’m going to write about AI. This could be seen as being a bit of a retread, as I’ve written about AI before in a different format, but the issues I raised then remain current and pressing, and, bluntly, my thinking on its use has evolved since I last wrote about it.
I will start by being really honest about AI, I find some elements of it incredibly useful. I use it to tidy up scribbles of notes or make emails clearer. Occasionally I have asked it to review things I have written. For admin jobs I have neither the time nor inclination to do I lean heavily on AI. Like most things, it’s good in moderation and when used effectively,
On this Substack, I have also been really open. I use AI to create images (where I want to avoid searching for a picture or running into copyright issues). I sometimes use it to find evidence that would take me hours to gather, or to pull the data or references I need, and I use Grammarly for some of my sentences.
But I stop short of getting it to write my articles because, as a frustrated journalist, that’s the bit I enjoy, and with a now two-decade career in global health, it is my perspective that I hope people might want to read.
So, to AI, there is a lot of focus on the impact on education, and I will probably come back to that in another post, but for now I want to focus on the potential impact of AI on clinical care in resource-poor settings. This is where I want to raise the alarm about the direction, we are, for understandable reasons, heading.
The road to ruin is, after all, littered with good intentions
The WHO has begun to weigh in on the issue and is increasingly pushing considerations on clinical care’s use of AI from the theoretical to the practical; the shorthand is that AI is coming to clinical care, not only in rich nations but in poor ones too. This will have impacts, so, they warn, it’s time to wake up to this new reality. But what does ’ waking up’ mean in practice?
A common theme I return to in Globally Nursing is planning failures and shortages. Educating nurses is costly and lengthy making shortcuts incredibly tempting, and quick fixes are the refuge of many international development initiatives.
Look at the programmes which try to create health workers on the cheap and over months and not years. From traditional birth attendants supported rather than midwives and/or community health workers educated to plug gaps where no doctors or nurses are available.
It would be wrong to say that the billions poured into traditional birth attendants or community health workers have been entirely wasted, but I think we also need to be realistic about what it can achieve on its own and whether we can build high-quality health care on the backs of a workforce trained for months, not educated for years; I’m willing to bet we can’t.
Where does AI come into this?
Put simply my concern is that there are those who view the coming revolution in AI as a way to shortcut nursing education and fill gaps in knowledge lost through hours in the classroom and clinical practice with the ‘shop floor’ support of a large language model.
Community health worker + AI = Nurse ?
A year ago, I posed a relatively simple question: if AI increasingly supports clinical decisions around the world, whose clinical reality does it base it’s answers on?
AI can produce clinically impressive answers and may offer real value in settings with limited access to specialists. Clinical usefulness, however, depends on whether those answers fit the patient, disease burden, language, available investigations and medicines, referral pathways, workforce and scope of practice in the health system where they are used.
AI is of little use if it advocates the benefit to patient x of a CT scanner that is hundreds of miles away or of lab tests that won’t be available for a decade.
I wrote about AI in clinical care and the ‘elephant in the room’; namely, the lack of data upon which the AI models were being built from countries which were the home of actual elephants (Sub-Saharan Africa and Asia).
There is already work in this area
Put at its most basic, you implement an AI triage assistant in a low-resource hospital setting, and staff use it to assess a child with a fever. Some causes of this child’s fever are universal, but how much weight will an AI model developed in North America place on malaria as a differential? What about Kala-azar? Further along, will the AI suggest possible treatments? Will it weigh up risks and benefits based on context?
Most of the world’s AI models have been developed in healthcare settings without enormous resource constraints, and for older populations. What happens when they are implemented into countries where the reverse is true?
I can already hear the arguments that low-resource settings will have to be supported to develop their own tools, but look at every other development in healthcare. Low-resource settings with limited purchasing power rarely feature at the front of the queue and most commonly occupy the back of the line.
The risk here is that a version of healthcare is grown with less experienced and educated health workers allegedly supported by AI models, and it is seen as the ‘cheat code’ for a rapidly expanding health workforce without really investing in them.
The governance nightmare that this arrangement could prompt is worth considering for a moment. I hear concerns from experienced nurses about what happens when they ‘go against’ the AI advice. In an increasingly litigious environment, who carries the risk?
How confident would a community health worker be in challenging the AI even when they have a hunch that the AI is getting this case wrong?
Furthermore, we could see this combination of community health workers supplemented by AI being seen as a viable shortcut when educating new doctors and nurses seems either too expensive or lengthy.
As with so much in our lives, from our Google searches to our daily management, AI has a role to play, but beware of those who focus on the bright and shiny as a quick win. If the last decade has taught us anything, it is that simple solutions to complex problems fail to deliver what was promised.