Could artificial intelligence save the health service?

“A conversation between two humans gets pretty complicated pretty quickly,” explains Professor Oliver Lemon, in discussing the problems troubling Alana.

That fact is probably well known by anyone who has attempted to learn a foreign language. A conversation may start off on basic terms, but, when it comes to using a shaky grasp of language to ask for help abroad, an encounter can quickly get confusing.

But for Alana, the challenge is a different one, because Alana is not a person. Alana is an advanced form of Artificial Intelligence (AI) software, who can understand and respond to human conversation, and her understanding of language is still developing.

Professor Lemon, Director of the Interaction Lab in Heriot-Watt University’s School of Mathematical and Computer Sciences, spoke to Holyrood shortly after receiving the news that Alana has made it to the final of the international Amazon Alexa Prize, a $3.5m university challenge to advance human-computer interaction.

But while Alana is one of the most advanced forms of conversational AI in the world, her understanding is not yet fluent.

Lemon said: “Conversational AI is a growing area, because people realised that there are situations where you don’t want to have a keyboard or a mouse, or even a touch screen with lots of apps on it. It’s similar to the Star Trek computer, you just want to be able to talk in a natural way, to interact with information and services.”

And, as Lemon explains, the opportunities are huge. “For example, in trying to reduce social isolation, someone could talk to this conversational AI, which has its own personality, as well as a lot of facts at its finger tips. It can tell jokes and play games, but beyond that it could actually remind you or prompt you for things, like telling you, ‘you haven’t called your daughter today’ or it could even sense when she has put the kettle on in her house and tell you, ‘it looks like she is about to have a cup of tea, why not give her a call?’”

But just how closely does a discussion with an AI bot resemble normal human conversation?

He said: “Some conversations go really well and they do seem quite natural, but in other cases, maybe the human says something very long and complicated, and we aren’t really able to process that properly, and it can quite quickly fall over.

“But we are not trying to pretend to be human. For example, our bot has the whole of Wikipedia available to it, so it’s kind of like spending time with a know-it-all. Also, it reads the news every night and automatically summarises all the news if you want. So, it’s similar to talking to a friend in the pub – you can say, ‘hi, how are you doing’ and have a bit of chit-chat, but it is also very information heavy if you want it to be. Sometimes it works pretty well. Our average call length with our system is just over two minutes long, but ten per cent of our conversations are nearly ten minutes long, so some people really talk to it, which means it is entertaining and engaging, though I don’t want to give the impression all the problems are solved.”

There has certainly been a growing sense of hype around the potential for AI to change the way we deliver healthcare. But how much can it really do?

Dr Ann Wales is Programme Manager, Knowledge and Decision Support, at the Digital Health and Care Institute, an innovation centre working to combine world-leading industry and academic expertise with advances in service, business and technology.

She describes AI as having the potential to bring a profound transformation to health. She told Holyrood: “AI means technology that emulates performance by learning, appearing to understand complex content and engaging in natural, human seeming dialogues. But it may be helpful to distinguish between two types of AI. One is data driven, so that’s machine learning, deep learning, using neural networks and natural language processing and so on. These all work to detect new patterns in data, using non-statistical methods. The other type is what I would call knowledge-based, where the technology uses decision trees, usually based on guidelines, or protocols, and those decision trees emulate human cognitive processing.

by Sam Trendall

Artificial intelligence could save lives of 22,000 cancer patients a year, Theresa May predicts

“The second type – knowledge systems, which are also known as expert systems – are probably a bit more progressed in health care than the data driven approaches, and by and large it is the data driven approaches that are getting so much media attention at the moment.”

Wales adds: “There are huge opportunities right across the health and wellbeing pathway, from risk detection, prevention and early intervention, though to triage, diagnosis and treatment. It can be used in new ways in engaging people in health and wellbeing, and improving population health. These areas can all benefit from AI.”

At a public health level, experts are working on ways to use the data-driven approaches described by Wales to identify population trends, allowing decision-makers to pick up evidence of risk then take action. Meanwhile, in a hospital setting, AI software could allow a radiologist to detect problems far earlier.

Wales said: “It is also useful for filtering out unnecessary or less urgent requests for imaging and other diagnostics. A lot of our waiting time problems at present come from requests for tests or specialist referral which could be managed in other ways, but AI, because of its power in data processing and modelling, can help with that, so that only essential cases are sent to an expert, and the less important ones can be triaged elsewhere.”

Sam Deere is the founder of Welbot, a new Leith-based start-up which aims to encourage behaviour change. Still under development, Welbot will sit on someone’s desktop, or on a mobile app, providing nudges or prompts to encourage the user to do things to improve their health.

Deere told Holyrood: “The idea we came up with was based around behaviour change in the workplace. If you look at almost any issue facing society, it has its roots in behaviour, and particularly the big chronic issues of our age – things like diabetes, cancer, obesity – these are traceable back to things like unhealthy eating, not getting enough activity, smoking, all of these things. We wanted to look after your wellbeing – covering everything from your behaviour, your nutrition, physical wellbeing and mental wellbeing – by drawing it all together and then nudging you to take positive behavioural change, which over time should have some pretty serious effects.”

The nudges presented by the software are tailored to the user, taking account of both what motivates them, as well as any relevant conditions they may have. Doing that requires data, not just about the person in question, but also about their workspace.

“The idea is that the more we know about you, the more we can understand what you might like, and what you might be more amenable to doing,” Deere said. “So imagine you want to ensure everyone gets 30 minutes of activity a day – generic solutions, like a poster on a wall, have a very small effect. But if you begin to understand the specifics about a person, for example, they support a particular football team, or in the morning, they are more willing to engage in an outgoing activity – if you can model the context of someone’s life – then you should be able to provide a programme of behaviour change which has a greater long-term effect.”

But while technologies such as Welbot and Alana may be at an advanced stage of development, AI technology has yet to sweep across the health service as a whole.

Dr Julian Huppert is the Director of the Intellectual Forum at Jesus College, Cambridge, as well as vice-chair of the Cambridgeshire and Peterborough NHS Clinical Commissioning Group, and one of the Independent Reviewers of DeepMind Health. While he agrees that AI will have huge long-term effects on the health service, he is more sceptical about its current reach.

“There have been a lot of nice stories,” he said “but only a very limited amount of genuine influence in healthcare. In the longer term, increasingly, almost all diagnosis will be machine aided at the very least, if not machine done, and that raises really interesting questions of the role of the clinician.

“Increasingly, we have clinicians who know a lot of information. But if you have machines that are likely to be better at diagnosis than any of them, then it seems to me that the role of a doctor will become much more holistic, much more about understanding the factors that drive a human, and communicating with them, rather than being the person that recognises that three particular symptoms are most commonly associated with a condition.

“There will be a residual role for people who are extremely expert, who would be the ones who develop the new machine learning, who spot things a machine can’t, but I think it means the role of general practice will become much more important, rather than the kind of specialisation we see increasingly in hospital medicine.”

And so, while the potential is clear, it appears the reality for system-wide change may still be more distant on the horizon. But, given the challenges facing the health service, is it realistic to see AI as a possible solution? To what extent could it take the strain off?

Huppert said: “I think there will be some areas where it relieves pressure, but it is also worth saying that if you discover new conditions, if you can identify more things, there is also the potential for that to lead to significantly increased costs for the health service. If you look back at the history of the NHS, when it was founded, there was a belief it would be cheaper every year because people would be healthier, and what actually happened is that we can do much more – we can detect things earlier and start to intervene – which is brilliant, but it is much cheaper if people just die. I am not advocating that as a policy but actually, it is cheaper for people to die of these conditions than to treat them.

“So if you imagine AI at the best it could be, it will identify a whole load of new treatable things, which we will treat at great expense and could actually increase pressure on the health service, while simultaneously also helping people to live longer, which is obviously a good thing.”