Question 1
Laura: Hello Tom! Ok starting with the basics then, what is actually meant by the term AI and in what ways do we already use it in day-to-day life?
Tom: That’s a harder question than you realise, because, it’s actually a grey area. There isn’t a clear definition that constitutes AI. AI is a series of techniques that allows computers to perform in a way that is analogous to human performance. If you are actually doing AI you don’t usually want to use this term because it can confuse people. There are different techniques under the umbrella of AI and no one definition. One example of an AI technique is Large Language Models, these are very clever text prediction engines, such as Chat GPT. You can use them to do something like write an essay on a topic of your choice. The problem with this AI is it is plausible about things that are not necessarily correct, but they are still plausible. When Artificial Intelligence works smoothly, computers are able to spot, for example, cats in photographs at lightning-fast speeds. When it goes wrong, they can confuse images e.g. when an image of a turtle was misidentified as a rifle – The Verge. In this way image-detecting software are open to exploitation and programmers need to pay attention to what they are teaching the computer. Image-detecting software is used in apps like Plant net or Birda, these apps identify a species and name it, working on the probability of what that image might be, then selecting the most likely one. AI makes mistakes, on average they get things right, but when they make a mistake, we may struggle to empathise or understand it because its way of “thinking” is so different to ours, and its mistakes will be unlike those a human might have made.
Question 2
L: How will AI techniques transform the patient care that you, I, or our loved ones might receive in hospital or within GP surgery, and local community services?
T: Well AI can be used to augment human performance. The concern would be that AI tried to replace humans. Humans have strengths that AI doesn’t have and AI has strengths that humans don’t. It’s about combining the two. There’s a famous piece of psychological research where people are in a room playing basketball, one team is wearing white t-shirts and the other team black t-shirts. You are asked to watch and count how many times the team in the white t-shirts pass the ball to each other and in doing so you miss a person dressed as a gorilla walk through the room. When we focus on tasks intensely, we can miss other information of significance. Human brains can throw away information, that is not relevant, and it is missed e.g. being task focused. AI has the opportunity to avoid being task focused. There is an AI product that has been used retrospectively to look at CT scans for incidental findings. AI systems exist that can look at chest x-rays using image classification and look at suggested features. The ideal is humans and AI working together, this is more powerful, and both together are less likely to miss things.
Question 3
L: How will the use of AI techniques change the day-to-day roles of healthcare staff, and how do you think staff will respond to these changes?
T: Depends how it is used, many clinical tasks can’t be delegated to AI, I mean AI can’t help get someone dressed or washed. We live in a time of information overload – we have access to more information regarding patients and it’s impossible for humans to synthesise this information in the short time available. It’s very much about AI supporting the role of humans. Any AI system looking to replace people is likely to fail. Artificial General Intelligence is AI that has human intelligence; large language models are not AGI, and we are a long way off. The changes have to be done in combination with people and questions such as who, how, why, and in what way is it going to fit into the workflow, often this question is avoided.
Question 4
L: What role do AI techniques have for those of us involved in healthcare improvement and how are you already using it in the Improvement Academy?
T: We’ve used it in our Shared CAIRE simulation using AI in clinical consultations and what we found was that people’s acceptance or rejection of AI is not age dependent, if it doesn’t assist people they won’t use it. So, it’s really important that we engage with people who will use it. Locally, we’ve looked at AI to help with qualitative analysis (inferencing services) and when shown to an expert in qualitative analysis, he said the AI did a reasonable first pass. It is not replacing humans but trying to assist them.
Question 5
L: Can you tell me a bit more about the Shared CAIRE project?
T: Shared CAIRE is a project run between the Improvement Academy and the University of York (Assuring Autonomy International Programme), funded by The MPS Foundation. We are assessing how AI used in clinical consultations can help (or hinder) the interaction between patients and clinicians, and particularly how clinicians respond to the different models of AI used. We have been doing this in consultation with patient experts, generating realistic simulations of healthcare scenarios with professional actors and real clinicians, to test how the AI is used and how the users feel about it. We believe that by testing in simulation in this way, we can generate recommendations that will help AI be useful and avoid the loss of the patient voice.
Question 6
L: What is your view about the increasing use of AI techniques in healthcare, do you have any concerns? Or do you foresee that it will make healthcare better for everyone?
T: Done right it will be useful. Less confident we can get by with minimal staff because front line staff are fundamental to our working capability in a way that AI is not. The classic argument is that IT and NHS are fairly broken, and people ask “why bother investing in AI if we can’t fix our computers?” I don’t fully accept that argument, we need to repair broken computers and we need to invest in AI. AI can help us to improve efficiency, quality, and safety. We also need to be aware that AI is good at interpolation, where it encounters situations close to its training data, and bad at extrapolation where it encounters something new. Humans are generally much better than AI at handling novel situations.
Question 7
L: How can people find out more about the work the IA are involved in, in relation to AI and healthcare improvement?
T: We have recently presented our paper Development and Translation of Human-AI Interaction Models into Working Prototypes for Clinical Decision-making at DIS 2024, for which we received an honourable mention. We have also previously published a paper with the team at the University of York (Clinicians risk becoming “liability sinks” for artificial intelligence).
Analysis from Shared CAIRE is currently underway, and results will be shared soon. Further information can also be found:
