AI in cardiovascular medicine: where are we now and what is its future?

In this episode of the Healthy Heart Trust podcast, Albert spoke with Dr. Kevin O’Gallagher about his groundbreaking research on heart failure. As a consultant cardiologist, Kevin is focused on using AI to analyze electronic health records and identify patients at risk of heart failure earlier. We discussed how natural language processing helps us extract valuable insights from clinical data, the challenges of ensuring data security and health equity, and the future of AI in clinical practice. Kevin’s optimism about using technology to enhance patient care truly resonates, and he believes it’s essential to maintain the human element in medicine while harnessing AI’s potential.

Transcript

SPEAKER_00 0:15

So, hello, and welcome to the uh latest episode of the Healthy Heart Trust Podcast. I'm Albert Farrow, I'm chair of the Healthy Heart Trust. This is episode seven. And I'm delighted to be joined today by my colleague here at King's College London, Dr. Kevin O'Gallagher. Kevin, welcome to you.

SPEAKER_01 0:36

Thank you very much. It's uh fantastic to be here today.

SPEAKER_00 0:39

Great to have you. So, Kevin, you're a cardiologist and you're based at King's College Hospital and at KCL, where I work as well. Tell us a bit about what you do.

SPEAKER_01 0:49

Well, thank you very much. I am, as you say, a consultant cardiologist, uh, primarily performing research. So I spend probably 80% of my time on clinical research projects through King's College London, where I'm titled a clinician scientist and funded by the Medical Research Council. That research focuses on a condition called heart failure, which is where the heart's pumping function isn't efficient enough to supply the rest of the body, and typically leads people to present with breathlessness and fluid in the lungs or fluid in the legs. And we use the data in the electronic health records, so in the hospital's computers, to try and identify where patients may have all the features of the disease, but may not be diagnosed with the disease to see whether we can improve those detection rates for patients.

SPEAKER_00 1:41

Aaron Powell So it sounds very uh relevant to the mission of the LVR Trust, in other words, preventing or detecting heart disease early. So, yes, you you said a bit about heart failure, and I I guess we should just talk a little bit about that just for the benefit of the listeners. It sounds like a terrible thing, doesn't it? Heart failure that your heart's failing. And actually it can it can have a bad prognosis. It's quite varied in its severity, and some people that do have severe heart failure, and some people can live with it for quite a while, but it's not a good thing to have.

SPEAKER_01 2:13

Yeah, I think you're entirely right. It's a it's an unfortunate and quite old-fashioned term. I think if we came up with or discovered heart failure today, we wouldn't call it heart failure. It is really just refers to inefficiency of the heart as a muscular pump. That can be either in what we call systolic function or a failure to squeeze strongly enough. Also, the aspect that I'm interested in, which is where you have diastolic dysfunction, so a failure to relax the heart. And in ways we don't fully understand, that failure of relaxation can cause just as severe symptoms of breathlessness and just as severe outcomes as can a failure of the pumping, squeezing function of the heart. Now, you're entirely right, there's a vast range of severity, and most patients don't have this scary recurrent hospitalizations or life-threatening problems from heart failure, which gives us an opportunity to find those patients with the earlier, less severe stages of the disease, or even patients who are at risk of developing heart failure and get involved in those patients early to hopefully prevent them from reaching those more severe stages of the disease.

SPEAKER_00 3:23

Thank you. Yeah, so it's all about early detection. And you mentioned that you're you're using AI artificial intelligence. It's something that is very topical at the moment. A lot of people know about it, have heard about AI, may even be using it. They might be using the uh generative AI type models for their internet searches or to help them with writing essays or articles or letters or whatever. So I think generative AI, ChatGPT, Copilot, you know, all of those similar things are things that a lot of people will be have some familiar familiarity with. But tell us about how you're using AI in in your research.

SPEAKER_01 4:09

Absolutely. And as you mentioned, there are different types of AI, different categories of AI. The category that we are using is what's called natural language processing, which in essence is the ability of the computer to understand human language. And we use advanced search function to dive into the electronic health records of the hospital, dive into the computer notes, which thankfully most hospitals now have, and we pull out any references to heart failure or the symptoms of heart failure, and then all of the relevant clinical information associated with that. Now you might say, well, that's just a little bit like opening a PDF and pressing control F and searching for heart failure. But with the AI, the computer can understand both the tense. So is this a present reference, is it a past reference or a future reference? Who does it refer to? Are we talking about the patient themselves, or are we talking about the family history, which is also recorded in the notes? And also whether it's a positive or a negative mention. So are we is the doctor saying, or or clinician, are they saying this patient does have heart failure? Or are they saying this patient does not have heart failure? And all of those added bits of information are really important to mean that when we pull out our data set, that we can be confident that we have a list of patients who truly have the condition. We can then extract a whole range of information, test results, both structured test results, by which we mean the numbers of blood results, for example, but also unstructured data, including descriptions of symptoms, written descriptions of what something might look like on a scan, that those unstructured bits of data don't typically make it into a data set. And so that's another advantage of the AI. So we then get a data set of patients with the disease, and we can analyze that and look for features and hidden aspects that may not be obvious to a standard bit of research. What we find is that what was always suspected but never proven, which was that the vast majority of patients with heart failure with preserved ejection fractions, that diastolic dysfunction, failure of relaxation, the majority of those patients have not been diagnosed with a disease, even though you can see that they have all the clinical features of the disease. That didn't matter 10 years ago when there weren't any effective treatments for it. But now it does matter because we've got two classes of drugs which have been proven to improve outcomes in these patients. So it's really important that we now take what we've been doing, which is clinical research, and identify ways in which we can apply it prospectively in clinical practice to make a difference for our patients. And that's the next step of our project.

SPEAKER_00 7:01

And to be clear about this, the database that you're using comprises thousands of patients. So you've got a huge array of data there that you can look for patterns and so on, is that right?

SPEAKER_01 7:12

Absolutely. So the we can apply the work to the electronic health records within King's College Hospital, which has over a million individual patients, and we when we I extract patients with heart failure, we get several thousand patients. So absolutely, there is great insight to be achieved by being able to use data at scale. We're also lucky in that within King's Health Partners, which is the power of King's College Hospital, Guys and St. Thomas's Trust, which includes four very large hospitals, they all use the same electronic health record setup. So with appropriate collaboration, appropriate regulatory aspects in place, you can actually perform quite large-scale research.

SPEAKER_00 7:55

So just to make it clear, you're you're looking for patterns of data that will predict the occurrence of heart failure with preserved ejection fraction. And do those patterns apply to groups of people, or are you able to predict in an individual whether an individual is going to be at risk of developing this type of heart failure?

SPEAKER_01 8:19

Both. So we can look for patterns in groups of patients, and what we have tried to do with some success currently and are working on for the future is to what we call cluster groups of patients to see whether this large syndrome of heart failure with preserved direction fraction can be categorized further down into closer groupings, because that then gives you what we would call mechanistic understanding. How does the disease happen? Does it happen differently in different patients? And we think it does. On the individual patient level, what we are working towards is the ability to give an individual a percentage likelihood of having the disease, so that you can make those predictions with also some context. So you're not just saying this patient may or may not have it, you're saying, well, we can be s pretty certain if they've got a 95% likelihood that this is the disease, versus someone who has a lower likelihood. That allows you then to apply thresholds and then say anyone above this threshold, we should definitely get them to see a cardiologist. That's the idea. That's the potential for this type of work.

SPEAKER_00 9:27

Sounds fascinating. This is obviously a a research thing at the moment, Kevin, and and you've obviously done a lot of work in that. How how far do you think we are from implementing this or a similar initiative clinically? Is it is it the sort of thing that doctors are going to be able to use in their clinics in the near future? W what are your predictions? I know this might be a crystal ball sort of thing, but but what do you think?

SPEAKER_01 9:52

I think it is something that we will be using in the near future. And I say we loosely as the medical community within the world. Naturally, in this technological revolution, there are going to be people who press ahead, and there are going to be countries who employ this type of technology sooner than others. The technology I think will develop such that we'll be able to do this very quickly. Thankfully, in the UK, we've got a very sensible health regulatory authority, the MHRA, who make sure that any AI technology that's going to be used in clinical practice is appropriately checks and balances to go through first. So what's important for us all to understand is that an AI software is classed like a met as a medical device, in the same way that the stents that we put in patients' heart arteries or the heart valve that a surgeon will put in at a valve replacement are medical devices. And so there's going to be a great set of checks and balances to go through before these bits of software and algorithms are used in clinical practice. I think it's really good that there won't be a rush ahead of speed to put these in practice because I think our patients quite rightly demand that we know these things are safe before we use them. There may be a time where the MHRA has to use AI itself to judge these algorithms because there's going to be a real head of steam built up and the speed of development is going to increase exponentially over the coming decade. But for now, yes, we're on the right track to use it in clinical practice. There are regulatory pathways to go through, and I think that's really appropriate.

SPEAKER_00 11:41

So so um this all sounds fascinating, Kevin, and obviously it's i it it it has real promise there to predict patients who are going to develop this type of heart failure. What about the the risks? Are there any risks that you might be incorrectly identifying people who are not at risk or or any other risks of doing this kind of research, do you think?

SPEAKER_01 12:04

Kevin It's a really good and a really important question. The risk, because no algorithm is ever 100% accurate, is of course that there is a chance you identify a patient and you've labelled them as possibly having FEF PEF who turns out not to have it. And that would be a really big problem if the algorithm was the be all and end-all, and it was defining labelling and defining all the future management thereafter. But of course, when we put this into prospective testing, the end point is going to be someone's going to go and see a cardiologist. So we're identifying people who we think need to see the expert who can then make the diagnosis. So I think that's a really important fail-safe. I think it brings a broader concept across all AI implementation, is that we have to have fail-safes in place. We need to make sure that we trust the technology, that it performs as it should do, so we can say to our patients, yes, this is the right thing to use for you, just like any other medical device, as I said. Broader risks, I think things that are important to talk about are, you know, data security. If patients' data is being used in the generation of AI models, if it's being processed via AI models, we need to make sure that all the regulatory and safety aspects are in place for data security. In the NHS, in the UK, we've got a national data opt-out, which is important for anyone like me who does work with the electronic health records, in that patients can prospectively sign up to say, I do not wish my data to be used in this type of research. And that then allows us to know actually, do not do not access that patient's data. So that's important things, an important component. Other broad risks if we link in with health equity, is we need to make sure that our algorithms work equally well in all aspects of the population and particularly ethnicity, particularly social deprivation. So we know that across cardiology and indeed across a wide range of medical conditions, that often our scoring systems don't perform as well in non-white ethnicity and perform less well the more deprived area a person comes from. That's a risk in AI models. I see it as an opportunity. We can utilize and harness the power of AI to say we know this is a problem currently in clinical medicine. Let's use AI to solve that. And that's something that within my research group, every bit of work we do, we make sure that we're addressing the problems of ethnicity and social deprivation and the variation in care that can occur from that.

SPEAKER_00 14:48

That's obviously a really important point about health equity. It's an issue not just in cardiology but across the whole of medicine, really. And I think, as you say, AI may provide a real opportunity to address that. So we might turn to some more general applications of AI in cardiovascular disease. Now I know, Kevin, your research is in this particular area, hef-pef heart failure with preserved ejection fraction. But I think it's fair to say that the applications are more diverse than that. It th there's many areas in cardiovascular medicine where AI may have a a useful role in diagnosis and in planning treatment and so on. Are you able to say something about other areas in cardiovascular medicine where AI might be useful?

SPEAKER_01 15:48

Yes, indeed. There are the general aspects of AI and the the power in terms of organization and in calendar optimization, which you can we all use in our daily lives when we ask ChatGPT or Claude to organize our calendar on our phone, you can see how that could be used at scale to really make listing, clinic allocations, theatre lists much more efficient. So I think there's potential there. Specific to cardiovascular medicine, well, imaging the ability to look at pictures and extract information from pictures is a real strength of AI. Cardiovascular medicine is very imaging heavy, and by which I mean almost every patient who comes near a cardiologist will end up with an echo scan, an ultrasound of their heart. Many patients will have CT scans of their heart, their heart arteries, and MRI scans of the heart. There will inevitably be hidden things within those images that we currently don't understand or we can't utilize for diagnosis or or uh prognostication that AI technologies will be able to look into and give us new insights. That's over and above just the general approach to using AI in imaging, which is to speed up the analysis of the images. So currently, many scans that are done in the heart are measured manually. So someone will take the moving image, will pause it, will draw a line around the inside of the heart, for example, will draw around a line around the outside of the heart to then measure the thickness of the walls. AI can do that much quicker than a human, and we hope and expect can do it at least as accurate as a human. So there's real scope to speed that type of processing up. In terms of helping us with management, there are some evolving technologies that we have access to in cardiac catheterization labs across the world, where if we are looking at someone's heart arteries through an angiogram where we've popped a tube through the wrist, passed it up the heart arteries and are injecting contrast down, we can measure narrowings, we can take accurate measurements, but also just by looking at the way the contrast flows down the picture, the computer can give us an estimation of the pressure difference across a narrowing, and therefore give us an idea of whether or not we should be putting a stent in. It's what we call a fractional flow reserve, which traditionally required putting a wire down the artery to measure the pressure and therefore infer the flow. Just by putting a wire down an artery, you have an additional percentage risk to the procedure. So obviously, if you can do this without putting the wire down, you can get information, but at a lower risk, which is what we want for our patients. There are also emerging technologies where it will help us predict the outcome of putting a stent in, telling us where best to put our stent, what size of stent, what length of stent, and telling us, well, if you do this, that will give you the best outcome in terms of predicted physiological change. So there are lots of ways in which AI technologies will be applied in the coming years, assuming we can prove they work in the individual cases, assuming we can prove they're safe, and fundamentally that they will make a difference to our patients.

SPEAKER_00 19:21

That at least for now, anything that's AI generated has to have human oversight. Yes. That nothing is going to be left to computers without humans having an input and saying, Yeah, we agree with that. Is that is that right?

SPEAKER_01 19:33

I think that's absolutely right. I think it's absolutely appropriate. And you know, even with current systems which have humans in the loop, we focus a lot on what we call explainability. So we often think of AI as a black box, and you put some numbers in and it spits some numbers out, and no one really knows what goes on inside. I think explainability is really important. We need to know why the computer is coming up with X opinion or X recommendation, if we're then going to sit down with our patients and say, this is what we want to do. Because often, quite rightly, when you're sitting with a patient, you say, Well, I think we should do this course of action. They'll say, Okay, why do you want to do that? And we have to then be able to provide them with the answers.

SPEAKER_00 20:16

Now, what we've talked about is very much around AI assisting doctors with their work. And I think a question that often comes up, and it's that's something that's been discussed endlessly, is whether AI is going to take away jobs. In this particular case, we're talking about AI taking away the doctor-patient relationship. Now, I don't think we're anywhere near that at the moment. Do you see a time where doctors won't be needed or the will or or patients will see the AI doctor without necessarily seeing a human? I know this is kind of in the realms of science fiction, maybe, but do you think that's realistic? Do you think that's ever going to happen?

SPEAKER_01 20:58

In the long term, I think many interactions which are currently seen by patients to be not clinically serious or not clinically urgent may well be taken over by a human AI interaction, assuming that the technology develops at the speed and skill that we think it's going to develop. I mean, in many ways, it's already happened. Without AI, we always talk about Dr. Google. And I was sitting talking to a patient today, and they said it's really embarrassing. I checked all this before, and I just want to double check that what Google told me is correct. And that's not AI facilitator, that's just a simple Google search. Yes. So you can see where we would reach a point in which a patient can ask an AI model sensible, simple questions, and can get back sensible, simple answers with the provisor that that model is going to be able to identify life threatening things that need to be seen urgently. I think we're always going to need doctors. Patients are always going to demand doctors. At the end of the day, and we're a very long way from going to a fully AI delivered service whereby all interactions are with an app on your phone, and if you need a procedure, it's actually a robot who does it facilitated by AI technology. I think that's that's a very, very long way off if ever we get there. Would we ever, as a country, accept that? I'm not sure patients would. I think patients want to see a physical doctor and for good reason. And you know, my next 20 years of employment are quite happy that that's the case.

SPEAKER_00 22:35

I think my rate for years of employment are rather less than yours, Kevin. So and I I'm sure I won't see it in my lifetime. Just coming back to the present and and maybe something around what's happening with government policy and the specifically the NHS 10-year plan. And there's various aspects of that that people may be aware of, from treatment more towards prevention, care moving from hospital closer to the community, that sort of thing. AI has a very important place there as well, doesn't it?

SPEAKER_01 23:07

Yes, I think that the NHS tenure plan is has a focus on using data to provide precision medicine to patients with a heavy focus on prevention. I think that's entirely right. I don't think you'll meet find many pe people in the country or within healthcare who'll say that a heavy emphasis on prevention is a bad thing. I think we all support that. The NHS tenure plan also talks about the use of wearable devices, which many of us have with Apple Watches or Fitbits or other brands, which are continually collecting a wide range of data on our physiology, which could potentially be of use in the future to help predict, prevent defined management for us as individuals. I'm generally an optimist. I'm slightly pessimistic as to whether we as an NHS organization will get to where we want to be or where the NHS tenure plan says we want to be in terms of AI implementation within the next 10 years. I'm always cautious to talk about politics. When you grow up in Northern Ireland, you're always told by your mum when you go out, don't talk about religion and don't talk about politics. So I've got no intention to make a political view here, but it's inevitable that with the way governments turn around and turn over the NHS tenure plan as it is now is is subject to change potentially. So there's a slight caution there that will there be the will to implement the NHS tenure plan? I hope so. Because if if we can harness this new technological revolution, push it into healthcare and get the maximum benefits out of it, I think we can have real excellent effects for our patients.

SPEAKER_00 24:50

So Kevin, it's been a a really interesting discussion. And AI is, as I said before, something that's really topical, something that a lot of people are aware of. But is this a flash in the pan? Is this something that's a fad at the moment? Or is this really going to make a difference, do you think?

SPEAKER_01 25:07

I think AI is definitely here to stay. I'm sitting here talking about my research. I'm not the only person in London doing this type of research. Almost everyone who's involved in clinical research has some element of AI work within their research groups. You know, I I do a little bit of work with the European Society of Cardiology as part of their committee planning the ESE Congress, which is the world's largest cardiovascular congress, 30,000 people attend every year. It used to be that data science and AI had a little niche role in the conference. You'd maybe get a handful of abstracts relating to it. This year it's the theme of the Congress. So every different subgroup, heart failure, electrical heart disease, inherited heart disease, structural heart disease, intervention, every one of those themes has an AI slant to it. That's how much it's come in and taken over, and it's here to stay for the future. You know, if you think about in America, it used to, I think 20 years ago it used to be that 10% of practices or hospitals had an electronic health record system. Now 90% or more have. As data and computing is more heavily embedded and ubiquitous in our hospital processes and in our primary care processes, we're in a position to really harness the power of AI.

SPEAKER_00 26:30

So it's it's certainly true, isn't it, that the availability of data and the computing power to harness that data is something that's not really been available before. It's also interesting to note that actually doctors in training that I've seen in the old days, they all wanted to do laboratory science. Most of them now want to do data science. That's uh certainly something that I've noticed in uh in our doctors who want to do research. So, Kevin, I think we'll wrap it up there. Thank you ever so much for joining me today for what I think has been a really good discussion. Thank you very much.

SPEAKER_01 27:03

Thanks so much, it's been really enjoyable.