AI in Healthcare

Adam Kobeissi: Hello and welcome to our new series of podcasts on AI for industry. I'm Adam Kobeissi. I'm your host today, and as well as being an AI enthusiast generally, I've got the great pleasure of running one of our practices here in CGI, some of whom you may have heard of should you have listened in to some of our previous pods.

However, this time we're taking a bit of a change in tack. Whilst the tech world has been talking about advancements in AI, Gen AI and who's going to win the race between the hyperscalers, we thought it might be a bit more insightful to have a discussion about how AI is being used practically across some of the different industries we work with. We're going to be joined by a range of different guests to give their perspectives.

Today, our focus is going to be on AI for healthcare, and our guest is none other than Dr Ameet Bakhai, a renowned cardiologist and healthcare innovator working with the Royal Free London. With over two decades of experience in his field, he's been at the forefront of cardiovascular research and patient care, pioneering new treatments and technologies that have transformed countless lives. Dr Bakhai’s work aims to bridge the gap between medicine and technology, which has earnt him global recognition and numerous awards. In today's episode, we're going to delve a little bit into his journey, discuss the latest advancements in AI for healthcare, and uncover how these technologies are transforming the way we diagnose, treat, and manage diseases.

Perhaps we could start with then, you could just share a little bit about your journey and experiences with AI in your professional career so far.

Ameet Bakhai: Thank you. So, my day job is a cardiologist, at one of the North London trusts at Royal Free. I'm also the R&D director here, responsible for one of the themes, which is called Barnet and Chase Farm. and I've been doing cardiovascular research for over 20 years, and I happen to have gained a new position at the Trust called the Chief Research Infomatics Officer.

And one of the reasons for that is because we've been using data to gain insights, for so many years that we've created something called clinical pathways. and these are, dashboards that actually help us understand what's going on with patients coming into our Trust with new onset heart dysfunction, using a term dysfunction, because I don't like using the term heart failure.

Now, failure is a term that we all use, but your heart hasn't failed. You're still alive. So, and yet, everyone knows it as heart failure. And just as an example there. So we code the term heart failure with the word failure in it. And of course, an AI tool is going to learn forever that this patient's got heart failure. And what does it mean. If you actually ask an AI “What does the word heart failure mean?”, the definition should be: the heart has stopped working. Well, that's not true in any of these patients, they're all alive to have the definition of heart failure, otherwise they’d be called dead. So imagine the sorts of difficulties that we're creating for AI in a way, and it's going to have to navigate around our human nuances, because even I'm struggling to use the word heart failure, I'd much rather have patients with challenged heart function, because I'm not an expert in heart failure, I’m an expert in heart recovery. So interesting times, right?

Adam: Yeah, absolutely. And I suppose that sort of parallels the research you've been going through over the last 20 years. I'm guessing the use of the word failure was because from a cardiovascular perspective, a lot of the patients you were seeing were at the end of that healthcare cycle. They'd already reached this sort of end of the journey around where their heart performance was, and they were looking for support from somebody like you to fix their heart.

Whereas, your point about dysfunction in the learning we can use from a data perspective is actually, that journey starts much earlier in the process. The heart starts to have issues earlier in the process. People's health deteriorates and then it ends up with failure, in that mind and that talk about the pathway, perhaps you could give us a bit of a view.

Ameet: You and I are going to continuously interrupt each other on this podcast, we're going to spark ideas off each other, right? And the first thing you've just literally given me one of those moments of epiphany, which is quite rare, which is that maybe the most powerful use of AI in healthcare is to move the dial from reactive medicine, which is where we spend 80% of our life at the moment, as cardiologist doctors anything, primary care, etc. to proactive prevention. And maybe the whole reason we haven't been to go there before is because our predictive power, what's going to happen down the road, has been so poor. And I'm mindful of this because I've just seen the NHS use the mining tool to find six times more patients with condition X.

And actually when we created an automatic detection pathway using blood tests, we started finding five times more patients with heart failure. I'll use the term heart failure for now, suspected heart failure. But we start immediately. The day we switched it on, we started finding five times more people with suspected heart failure. And of course, our manpower utilisation had to expand quite rapidly.

Luckily, we'd already done this at one hospital, learned from it, and then a second hospital we were ready and expecting that, and it was exactly as we predicted. So the idea that AI moves us from reactive medicine to prevention, that's quite unique. I don't think I've heard that anywhere before. Have you Adam?

Adam: I think it's certainly, an area of exploration across a number of different, healthcare paths. If you think about it from a perspective of, if we use cancer as an example, which is slightly outside of your field, but, the viewpoint on trying to be able to recognise when a patient has cancer sooner in the process is certainly something AI has been helping with within the health care arena for the last ten or so years from a data analytics and predictive perspective.

However, more recently and I saw this in the news, only last week, I can't remember the health care trust, actually, I need to perhaps look it up, but I think it was a project here in the UK where using again bloods, they were able to see what the chances were of a lady having breast cancer or a recurrence of breast cancer should you have had it before, before it came back, with a 100% success rate. And that prediction and prevention, therefore, of being able to see whether that's coming back is absolutely going to transform people's lives. It's going to take them out of hospitals, it's going to keep them all healthy and ultimately it's going to reduce the strain on the healthcare system. Which here in the UK particularly has been being quite a challenge.

Ameet: It is. And you've raised so many issues. Just in that one statement, you've said, first of all, breast cancer, because I'm R&D director, we get involved in everything. So we're quite lucky. Breast cancer has evolved in its treatment strategies in so many ways. So first of all, do we just do surgery? Do we do surgery with chemotherapy that shrinks the tumour?

Do we do surgery with chemotherapy with another technology that highlights our ability to track the cancer cells, because we're looking at the receptors on there and we find those receptors and we can target a hormonal therapy. Or do we help the combatting immune system to become much more augmented and powerful? Do we bring the army into the police force as well and actually get them to locate the cancer much more powerfully.

And then later on, do we look at the genetics of that patient, and do we actually gene target that therapy specifically to that cancer by putting a laser guided sight on it, so we know where the missiles need to strike. The way that we are attacking breast cancer has changed so dramatically, in which order we do things, probably in the future will only be with the support of AI tools, that we can actually work out the permutations, because this one patient, should they have diabetes or should they have a weakened heart, or should they have other conditions or not? We'll have so many permutations that we can go down, that we won't be able to amass enough information to choose the right strategy without the support of these digital tools.

Adam: Which is fantastic if you think about that future where, all permutations and all possibilities could be, analysed, very, very quickly through the power of AI. You're talking about a Utopian kind of health care future, which I think is fantastic. But it also presents an interesting kind of dynamic about where we are today and how, maybe not just the NHS, but healthcare more generally needs to think about some of the challenges on adoption for AI. So as we talked about, I think prior to this pod, a lot of the healthcare research is done within its fields. Cardiology looks at the heart, physiotherapy looks muscles, and therefore the research and the, AI data analysis that's being performed is still being performed within those pathways, within those segments.

What you've just described is the ability to almost create one healthcare AI to rule them all. Pardon the pun, but the ability for a patient to be optimised across every pathway and every diagnostic region all at the same time, is that the future that you see?

Ameet: So the biggest problem in healthcare isn't that we don't know what the right thing to do, isn't that we've got pockets of excellence. It's the variants in practice between the leading edge and the lagging edge. Wherever it is, whether it's going for a blood test, whether it's interpreting that blood test, whether it's recognising the blood test compared to your previous blood test and graphing that trend, whatever it is, even amongst doctors and nurses and pharmacists and physiotherapists, there is going to be a variance in who you meet and why you meet them, and what their level of intellect and insight and capability is compared to somebody else.

That variance and the silo mentality: I just focus on this component compartment of health care because that's where I feel the strongest, most experienced versus a holistic individual that says actually I'm pretty good in a lot of areas, or I'm used to being in an multi-disciplinary team setting where I have to think about the kidney, I have to think about diabetes, I have to think about liver. I have to think about the psychology. I have to think about depression and so forth, and patient adherence and their social factors and their well-being, and their frailty scores at the same time as the heart. Those are the two things that are going to really evolve.

And actually we can't keep up with the research. Who can keep up with 20,000 papers just coming out in cardiology a year versus 50,000 papers coming out in cancer a year? Who's ever going to be able to keep up with that? Right. So maybe AI is the glue that might help make those little connections again. And reaffirm us.

Adam: Makes perfect sense. And I guess, again, sort of leads us into some of the advancements around what's going on in AI. You talked about that plethora of information and who could possibly keep up with it from a from a human perspective? Do you think the advancements in generative AI are going to revolutionise that aspect of healthcare, is that an opportunity for us to be able to bring that virtual assistant, that digital health assistant to a range of consultative areas much quicker through the use of generative AI? And are you seeing that being operated anywhere in health care currently?

Ameet: So we're looking at generative AI to reduce the workload, for example, in coding. This patient came in for this operation, but they also had a background of diabetes. They also had a background of immune disease and cardiovascular disease, and so they are a more complicated hip replacement than the person who didn't come in with those, and we need to code for that activity. And that's why they stayed a day longer, etc. These are safe places that we're testing generative AI at the moment. But this word testing, really important. When we move to use a new medication, it has been rigorously tested. First of all, it's been tested on healthy human beings to see what dose levels we can tolerate and whether people get side effects.

Then it's been tested on a that's called a phase one trial, and it's been tested on some early disease patients to see, making sure there's nothing unexpected going on and we're not worsening the disease and actually improving significantly and getting the dose right. And then you have a large-scale phase three study, a pivotal landmark trial, and that takes 4 or 5 years to test. And then when we validate it, we've got about 7 to 10 years of patent life left on that drug while we try and adopt it. But it's so rigorously and safely done. Speeded up only during Covid for vaccines, it was contracted down to one year because the whole human race worked on one technology and one problem in one way, and that was the fastest we've ever managed to get it.

And some great lessons to be learned on research and research agility, during Covid. But the way that things get tested is really safe. We always try and do no harm first, and then minimise harm and then do significant changes. But with generative AI tools coming in, we find that there isn't a framework to really evaluate it.

And we also don't always know what the status quo of the problem before was, well articulated, to see the benefit, once we intervene and put a new tool into it. So taking a step back, do we have sufficient framework that we know how to evaluate a digital technology, which can't be a placebo control technology? So that's a question back to you, Adam. Is the industry thinking like that properly, do we know how to evaluate, be transparent, do we know how to judge? Do we know how to see what's significant impact and is it cost effective?

Adam: I think it's a fantastic question. That's certainly one that's being explored by both technology organisations and private organisations within their different sphere. The whole idea behind: are we ready for AI at the first instance? What is our AI readiness, as a business, and do we know how to control it, Is one aspect to it. The other is then the security and risk and compliance aspects of what am I doing with that data, where might that data end up, whose data is it, who does it belong to? All of these things are being considered.

But I find it quite an interesting dynamic, particularly in health care. You talked about the research of drugs, and particularly the Covid vaccines, which as we know, were supported by a number of AI, data analytics, testing capabilities to be able to test what we thought were the likely outcomes of certain vaccines against certain proteins, within the Covid disease, and that was all supported by AI. And then we licensed the ability for those drugs to be injected into humans within a year.  And yet at the same time, we have lots of conversations around AI within the NHS and the blockers seem to be, well, we're concerned about data privacy, we're concerned about the sharing of the data. And I find it an interesting dynamic of have we actually thought about speaking as a group to our users, to our patients, to our communities, about what data they would be willing to share or not, in the light of being able to get better health care outcomes? Because I think often industry takes a blanket view, either based on a position of not understanding, we don't know what the outcome is, therefore, we're just going to take a risk averse approach, or it takes an approach based on the understanding of a government or a regulatory body that hasn't necessarily engaged with its audience.

And I would argue health care is one of those areas where most patients would probably be willing to share more data than they are today, if the outcome was: I was going to live longer, have a healthier life, and be able to ensure my kids had a great future. I say that from a kind of utopian view, but I think you're absolutely right that the work needs to be done to make sure that the framework is as agile as our ability to invest in new medicines, in new pathways and new treatments, because the technology's coming like a train, whether you want it or not, we've got to learn how we can control it for the best outcomes for everybody. Right?

Ameet: So on the one hand, it's coming like a rocket and we've got to utilise it, and it helped us find millions of permutations for how do we attack this virus, which piece of the virus do we take on, which piece of the virus do we put our drug motifs around or use our mechanism's to try and identify the virus etc.  So we could simulate a whole range of warfare and games by which to understand what we were doing, which is fighting the Covid virus.

And so that was phenomenal power. Nobody gets harmed in simulations. We're just simulations, and actually, the best simulations give us the direction of travel. Fantastic place to be, finding new user drugs for specific diseases, etc. looking at large sets of data and social media to say, where are these young people, where is this pocket of epidemic coming from, who's getting infected, who's not getting infected, etc. all of those hotspots, where's the worst amount of deaths? Where do we need to focus our vaccines on first? What's the order of which we need to invite patients to have their vaccine, if we do have vaccine, what's the order of which we shield them or let them come back into the public environment when they're ready to come out, etc.  All of that technology was dependent on our AI benefits and tools. However, the opposite is we become so dependent on it that when something goes wrong with it, can we actually cope.

So there's always this schizophrenic side in healthcare that at the end of the day, it's human beings. Should human beings and data and digital be absolutely intertwined, or can human beings still function, irrelevant of the technology? Or what if the technology is telling us the wrong thing to do, where's our intuition against the power of, the post office scandal, etc.? No, no, no, data has got to be right. The software is going to be right. The algorithms got to be right. And no we're trying to tell you when we don't really believe it from the floor. In healthcare, that would be an immediate, if we were seeing people being harmed, Postmasters being harmed, if we were seeing patients being harmed, literally within ten patients, we'd be raising an alert, going: I don't care what you think. Tell me exactly what's going on. Right. Which is crazy, right, that we didn't think of it like health care.

So one of the things we've become quite proficient at, and Royal Free at the present time is we do something, one of the roles I have is I'm deputy chief, so clinical safety officer, what that means is that we hazard test any digital software coming in that may potentially relate to patient harm. And we actually always ask the ‘what if’ question: what if this went wrong? what if this human being accidentally made this mistake? what if the training wasn't done? what if a bad actor was coming in, etc.? what could they do to manipulate this pathway or this data, etc.? One of the interesting things about this flow software. We've been doing flow simulation for so many years. UCL is always, leading edge on that. The F.A.R. Institute has always been the first to actually demonstrate, in real time, we can actually track where the ambulances are when they're picking up the patient, when they've actually lit up. And actually, we might be able to and this is the sort of thing we did almost 40 years ago when we decided, hospitals, when we started training, we said we need to deploy a lot of our decision making in the ambulance.

So about 20 years ago, we decided, they'll be ECG machines in ambulances, and the ambulance crew will be trained to see if the ECG had elevation. They will bypass a hospital that doesn't have a cath lab and go directly to a hospital with a cath lab that could receive that patient to go directly into the cath lab to open up that artery.

We actually had a national target that we meet, which is called ‘call to needle time’. So call to needle time, used to be less than 90 minutes. Then we moved it to call to balloon time. Actually having a balloon in your artery, opening that artery up so that you actually saved from a heart attack set that had to be less than 90 minutes. And that's the sort of data that tells us, actually, London is the safest place in the world.  I think even, and I say world, I'm just thinking definitely in Europe, if not the world, to have a heart attack, because within 4 to 5 minutes, an ambulance will arrive. Within 25 minutes, you'll be at the A&E door or bypass that and go straight to the cath lab door and you'll be on the angiogram table, with a sheath going through either your wrist, or your groin, opening that artery up. And this was all due to flow management. Now, the way we do flow management for heart attacks, we don't seem to do it for all the other things that clog up the system. And that's because of scalability, can we really create that scale for chest infections.

Can we create that scale for COPD infection? Can we create that scale for an older frail patient that's fallen down, etc. and so those are where we're struggling. And I'm hoping that AI and the flow management will identify a better way of using our resources properly. It's all about the energy required to resolve that issue that's occurred to a patient and our ability to motivate and mobilise ourselves to meet that appropriately, without putting strain on a different part of the system. If all of a sudden I created a fall centre, am I going to put pressure somewhere else on the system, etc.?

Adam: Yeah, it makes perfect sense. I think we talked about this previously where, some of the work you're doing around, using blood test analysis is great, but actually it just means that you need a thousand more blood tests. And, how do you make sure that those things happen? From my perspective, I think this is an interesting interlink between the physical and the digital world, where actually maybe we need to start thinking about how the AI pathways that we're creating can also be linked into, better health integration into the wider health ecosystem. So actually, does that blood test need to be done in a hospital, or could it be done in a pharmacy, for example. And, we're using the technology to provide that information back to the cardiologist in a much better way.

Ameet: I’m going to interrupt you there, this is really great. This is where you as a group, CGI consultants to government, industry and so forth. This is where you think much more widely than we will. We're used to the patient comes into hospital, they get a blood test for detect heart failure, heart attacks, etc. but actually you going wider and going actually ‘is that really the best place?’ when actually if we'd have put these areas here, diagnostic centres here, or motivated pharmacists to be doing front line testing for diabetes, cholesterol, atrial fibrillation. Or drones will arrive at the house right now. Isn't that a better way of doing things? Yesterday at London Tech Week, there was something called Medicube X, and that was a little pod where you could walk into and literally you get most of the things that I would do in a cardiology console, you get an ECG, you get a pulse oximeter, you get an age detector, you get a heart rate monitor, weight and blood pressure and all those things.

And you actually get a risk score of your heart age compared to your biological age. And so we use that as an opening gambit. And here people were it's a bit like your camera, in a Tesco's, you go have your passport photos done, actually here go and have your cardiovascular and age, detection done in this cube and that cube will give you that result in eight minutes. And that was the NHS health check. And that's again one of the programs that the UK is absolutely famous for in the world. We were the first people in the world, and thanks to Gordon Brown actually, and that actually set up this program, where we said, actually there are 7 million people between the ages of 40 and 74 that haven't been to a doctor, haven't had health checked yet. We don't know their blood pressure, diabetes, etc. and they cardiovascular risk score. And actually we could upstream give, them aspirin, statins, Ace inhibitors, managed diabetes to stop them having a heart attack. So let's employ this technology called NHS health checks and anyone between the age of 40 and 74 should be invited for that by their GP.

Interestingly, we put £600 million aside to do that over five years. And actually GP's were so well-motivated to do that. Majority of that was done in 18 months. We target so much earlier, when you put a clear focus and clear resources, the NHS and health care system is really motivated to gear up, but of course, we kick the can down the road because we found so many more people with heart attack potential because they had high cholesterol, obesity, high blood pressure, potential for diabetes, prediabetes or early diabetes, that we didn't know how to cope. And now, of course, luckily drugs have arrived that can reverse diabetes. Or we catch you at the prediabetes stage and stop it going to diabetes. Amazing stuff.

Adam: And I guess for you, working within health care, that's always going to be the challenge, the balance between what can I solve today and what could I do in the future. But, we've been looking at different ways that we can help patients through the, health care pathways. If we look at, other ways that we might be able to help them, particularly dealing with dissatisfaction, health care, other areas where you see that your patients are dissatisfied, we could do better. And where could maybe AI help improve that satisfaction?

Ameet: That's a really interesting point, Adam. We've talked about AI being a help in finding new drugs, the diseases predict things, etc. but actually one of the biggest areas, and the NHS is amazing at capturing complaints. For example, one of the areas that we see is that communication failure is such a large component, is usually the second commonest cause of patient dissatisfaction.

I think, given that AI and generative AI knowledge language models are actually so good at communication in so many different ways, communicating what your appointment is, trying to decipher the discharge summary in the outpatient letter that's supposed to be between your consultant colleague or your doctor, or your health care professional and your GP. But actually you need it in normal speak, in normal language and even things as simple as, can you at the moment actually email your health care professional and not have an abusive system there that the healthcare professional can't cope? All of those communication areas are things where we're really excited. Actually. That is really the true power of generative AI in large language models. We think it's going to be able to help quite a lot in that space.

Adam: Yeah, and it's a fantastic use case. clearly improving communication, not just in terms of patients receiving, an output from their health care provider, but as you rightly say, being able to actually speak back to the health care system, and particularly in this country, where language is a is a huge barrier. We are a multicultural, diverse country, and the health care system has to deal with a range of different cultures and language every day. But actually all of the communication happens in English. Is that the best way to get urgent information to a patient? Or could we use some of these technologies to translate, describe and give it to them in the language that makes sense to them.

Ameet: You know, that is such an important area. It's important area across the UK, but it's also an important area particularly in London where we've got the most diverse population. There are inequalities that are increasing. The gap in inequality of care is increasing because of the communication barrier. And so if we can break that, we are always proactively in research trying to go, no, no, no, we really want people from a disadvantage situation, people from minorities to be represented in research because we don't know if this medication is going to help them as much as it can help anybody else, or even more hopefully. Maybe with generative AI tools, actually tracking the patient's mobile phone as their NHS number, maybe we can actually provide care rather than use it for the wrong reasons, if that makes sense.

Adam: 100%. I think it's a fantastic view of a future where, health care for everyone, which is what the NHS was founded on, really becomes true, backed by technologies such as generative AI.

I'm conscious that we've taken up an enormous amount of your time this morning, and you're a busy man, and therefore we probably need to wrap this up, but if I could ask you one final question from your point of view, we've talked about the good, the bad, the ugly, I guess around, AI and the health care system this morning

If you were looking to give some advice to a young professional or a researcher moving into the field of AI in health care, what would be the advice that you would give them? What would you be saying that they should be thinking about?

Ameet: Think a young researcher moving into that space, I think the first question is should you be doing it alone or should you be doing that as part of a team? Have mentors, have peers around you so that you've got a critical mass. The worst case scenario is you've got one enthusiast in one location working in an isolated fashion, and they do an idea, and then you find that that idea has been replicated somewhere else or done better. And that whole energy processes has not been as useful. Of course, the exercise was useful, but not the output of that necessarily. So working as a collaborative. The second thing I think we have to ask ourselves is, are we as healthcare professionals, the most mature in our thinking about not just can I use this to do something smart and something clever?

What about the potential that actually it might do harm, or it might change a system that's already reasonably good? Does it need optimising and if so, is it optimisation or are you trying to do it for glory and ego? Because you want to do something slightly new that, is it really a problem that absolutely needs solving, or should we be focusing on the problems that actually need solving, that we need your energy for?

That matching of that energy, where it should be for the patient benefit, for the societal benefit, if you do that, you'll never fail in whatever happens, even if you don't succeed in your outputs on what you thought you were going to do, the lessons you learned are going to be so important. And also have a framework - think about a clinical trial, learn from how we've done things for drugs and devices before you move to diagnostics, digitals and predictive tools and algorithms and AI perse, I think those are probably the three things that I’d stop on.

Adam: Fantastic. Great bit of advice, positive in its thinking and as you say, the nice balance between using new technology and also remembering where we've come from to make sure that we put the patient first, and that ultimately should always be at the heart of everything we do in health care. Look, it just leads me to thank you, Ameet, for your time this morning. It's been absolutely amazing talking to you. I know we could have spoken for another good couple of hours, and maybe we should pick up some of these themes as a as a follow on, depending on any of the feedback we get. Obviously, if you do listen in to this pod and you've got some ideas, feedback or questions, please do forward them through to us and we'll make sure that, Dr Ameet gets a view on those as well. Thank you from me for, all of your insight today. And, thank you for listening.

[END OF AUDIO]