Loading summary
Transcript59 lines
- [00:00]
A
Foreigners.
- [00:07]
B
To quote our JAPA authors, artificial intelligence or AI is everywhere these days. You may have used AI to create a social media post or make your own grocery list, or maybe you use AI to write your progress notes or assist in generating a differential diagnosis. AI is transforming the world around us and one provocative question persists. How is AI transforming healthcare? Joe and Martine are joining us from the AAPA conference in Denver today where there has been a lot of talk about AI and its usage. There have been several lectures on AI and House resolutions brought it to the HOD regarding AI. Today we are joined by several guests to help us explore this very question. Marci Farquhar Snow, Amy Simone and Sheil Singh, the authors of the JAPA article Artificial Intelligence and Cardiovascular Practice. In this article the authors explored how AI is affecting their clinical practice and we have a guest co host, Ben, a data scientist focused in healthcare and emerging AI expert. Before we begin our discussion of the article, Ben, can you help us define some key terms like machine learning, linear regression and predictive analytics?
- [01:15]
C
Absolutely. Thank you so much Kim. Very excited to be here as a part of this conversation. So you mentioned three different terms here as a whole. I do want to say overall that if you are a little bit hesitant about any of the definitions, I say that is perfectly acceptable. There isn't a true great definition out there for some of these different terms as the capabilities just continue to emerge as a whole. About machine learning in particular, or ML, you may see that abbreviation thrown around from a variety of places is a type of artificial intelligence where computers learn patterns from data and then use those patterns to make predictions or decisions without being explicitly programmed. So for example, you might see around given electronic health records are now very popular, of course that machine learning model trained on thousands of those might be able to flag a case where somebody may be at high risk for a heart attack. As an example there machine learning has been around for a fair bit, but there has been and overall interest over the past year for machine learning. Linear regression in particular is a statistical method used in both traditional analysis and some machine learning models to understand the relationship between variables. So what that means in a plain language. An example I like to give is when you're converting from Celsius to Fahrenheit in terms of just the temperature as a whole, this is a good example of a linear regression. So the formula if you ever try to convert from Celsius to Fahrenheit is 9/5 times the Celsius temperature plus 32. So this is your standard type of formula for linear regression. And all that really matters whenever I give an example of a model moving forward or you hear us talking about models is that when you receive that Celsius input, say, for example, if it's 30 degrees Celsius, there are some changes that go on to that number based off of information we've learned. So in this situation, you effectively multiply it by two, so that becomes that 30 becomes a 60 and then you add on 32 to it to get 92ish degrees. And that's roughly the degrees that is in Fahrenheit for 30 degrees Celsius. And so when we talk about these other AI models, you can just think about there's a level of change to that input that can be far more complex. But at the end of the day, it's similar to some information we found. Make some change and you have some new idea of some new value at the end of the day. And where this really becomes helpful as a whole is moving into the field of predictive analytics, where we can use current and historical data to make informed predictions about future outcomes. So, for example, I mentioned before about how an example would be that you could find out you could flag somebody for being a higher risk of a heart attack based upon information within EHRs. Another one could be based off of echocardiograms and different lab values that you might be able to flag potential risks at the end of the day to be able to provide that greater insight as a resource, as a whole. Not that these are ever going to be fully perfect, but any level of insight, information we can get to act sooner, that's the real value that's often being explored in the healthcare setting. With that, I'll pass it back to you, Kim.
- [04:23]
B
Thanks, Ben. That was a great introduction. Now let's get to know our guests. Marcy, Amy and Sheila. Can you tell us about your experience, how your experience led you to writing this article?
- [04:33]
D
This is Marci. I'll start off by saying that, you know, I've worked in various cardiovascular areas, both inpatient and outpatient, for several decades, it seems like, and the specialties are varied among preventative to electrophysiology, more interventional cardiac surgery. And with that, I've recognized that AI has been using been used in those various areas in different ways, from diagnostics to the predictive analysis like Ben had mentioned. And so the ability to incorporate AI into our clinical practice not only bolsters the efficiency that we can provide the patient with personalized care, but it also enhances our ability for decision making. So by writing this article, we wanted to show that as AI evolves that advanced practice providers need to understand general AI terms as they're being used in our practice. And I'm glad that Ben talked about there's no true definition because to be quite honest, when I started writing this article, I did a search of what do these different topics, machine learning, generative AI, what do they really mean? What we try to do with this article is to define these terms. It gives some examples how they can be used, and to explore what are the benefits and limitations so we can apply them to our clinical practice. So I'm going to ask Amy and Sheil how you feel about AI in general and how it applies to your clinical practice.
- [06:29]
A
Oh, well, thanks, Marcy. Thanks for handing it over to me. First and foremost, I'd like to thank Kim, Martine and Joe. Thank you for the invitation to participate today. I'm thrilled to be here and chat with you all about AI. So, you know, I think just overarchingly, innovation, progress, advancements, you know, all these things we've seen, they're fueled by disruption. And I think none of us at this point would argue that AI has become a daily disruption in both our everyday lives and also our professional lives. And I'm by no means am I inferring that disruption is a negative thing. Right. It's fuel. I first became interested in AI through my role directing a large valve center here in Atlanta, Georgia, and to participate in an informed way about incorporating AI into our practice. I needed to understand and the kind of the lingo, if you will, and be able to speak the language. So I began to learn on my own, to kind of find my own resources and educate myself. But it was our goal collectively as authors of this article, to have a comprehensive yet concise resource and tool for apps to learn the language. Right. So they can not only understand kind of what's happening in terms of the current applications for AI, but participate in the evolution of AI in clinical medicine and healthcare by recognizing opportunities to utilize and refine AI in everyday clinical practice to, again, better patient experience, better clinical outcomes, make work streams more efficient, and, you know, practitioners more productive at the end of the day. So just a few things that got me kind of on this road to.
- [08:05]
E
Begin with and I'll just hop on. This is shield to say that, you know, I am a PhD student and I'm at a point in my research career where I'm constantly absorbing new information on how to conduct research rigorously and responsibly. And so when it comes to AI, it's a very new field for me. And so I was really excited to be invited to help Marcy and Amy develop this manuscript and just learn a little bit more about how I can use AI in my daily life, but also use it responsibly.
- [08:34]
F
Well, thank you again all of you for being here. And I just, I really enjoyed this article. And if you hear some background noise, you'll. You'll notice that I'm at the AAPA conference where, like Kim said, there has been a lot of discussion about AI and its use. So I want to talk a little bit about your article and you, you let us off with a case. A 78 year old black woman that presented to the emergency department with chest pain and shortness of breath. They obtained an echocardiogram that revealed aortic stenosis. Further testing ruled out any like acute myocardial infarction or any significant coronary artery disease. She was discharged and instructed to follow up with her pcp. The patient was from a rural community and was referred to a follow up with her pcp, but she had a. She was referred to get a comprehensive valve program that was basically 100 miles away. The center offers an echo surveillance application as quality initiative using predictive anal analytics. Excuse me. This application assessed that she was likely to rapidly progress to critical aortic stenosis. The PCP was notified and an app called the patient to discuss the need for additional workup, including a CT scan. In interpreting the ct, a predictive modeling application identified risk for a transcatheter aortic valve replacement and the patient ultimately underwent a surgical aortic valve replacement. This is truly a powerful case. So what I want to know is what are some of the ways that AI augmented this patient's care?
- [10:10]
A
Oh, thanks, Joe. I'll take that one. So this is Amy. I wrote this case study. So it's, you know, near and dear to my heart, but I think this is a really great case that that brings the power of AI today into very fine focus. And this is not science fiction or something we're looking at down the road, but applications of AI that are referenced in this, in this case study are within our reach, at our fingertips today. So very exciting. So just in terms of ways that AI augmented this patient's care, few things, few thoughts around this. So number one, I would say AI helped to mitigate disparities and improve her access to care. So just like this patient who resides in a rural community 100 miles away from a comprehensive valve center, this also applies to patients living in urban areas without access to care. I think AI applications can certainly increase reach and ensure patients who need treatment for various disorders, diseases and ailments are sought out with proper representation. Right. I think AI absolutely applied this, this very powerful predictive analytic component to diagnostic testing and imaging. Within this case study, also within our reach today. This patient was swiftly referred to do a valve center and AI aided in these operational efficiencies.
- [11:29]
C
Right.
- [11:30]
A
Timing is everything. But I mean, truly many patients with heart failure, coronary disease, valvular heart disease, and so on, so forth are at an increased risk of decompensation, which can in turn prompt urgent hospitalization or perhaps intervention as well. All of which are costly, yes, but impact quality of life and clinical outcomes for patients. So AI in this example also provided very personalized educational resources and tools. Right. That were tailored to her literacy level AI to help to generate this individual kind of shared decision making toolkit. I absolutely love this. So that we as practitioners and providers can ensure our patients are participating in their care journey and treatment decisions fully informed. So knowledge creates this element in this environment of psychological safety and where patients wishes and goals can be incorporated into discussions appropriately. And ultimately AI certainly was, was here to strive to mitigate complications. I saw this example firsthand in my practice caring for structural heart disease patients. And treatment plans were changed or tailored for optimal clinical quality and outcomes based on predictive modeling that we utilized at the VAL center for where I was employed. This was absolutely tremendous and really just kind of astounding if you think about the impact of that downstream and how quality and outcomes were affected. I think AI has been shown just in this brief case study to really have multiple touch points to this patient across her care continuum. And I think it's a great example of things, again, not science fiction. These are capabilities within our reach today. Thanks so much for the question.
- [13:10]
F
Well, and thanks for your obvious excitement in this, in this realm. I really appreciate that as clinicians we're responsible for learning a lot of guidelines and however, patients don't always present by the book. So what are some examples of how machine learning algorithms can assist us in identifying the best therapy for our patients?
- [13:31]
D
I'll answer this. You know, machine learning sets up algorithms, but it doesn't learn to give actual management strategies that are tailored. But AI itself is very powerful as a basic predictive tool and it gives people a general understanding of where they are. So for example, there's a lot of tools that cardiology uses for cardiac risk and these algorithms and scorings have been around for many years, and they are used to evaluate if a patient, say, presents with chest pain. And you're not quite sure what kind of pain is it, Is it cardiac? Is it non cardiac? Is it possible, you know, cardiac? You're not sure. So there's different scoring systems that can be used. You know, you can do it on paper, you can download an app that, that will do it with a certain kind of machine learning where it plugs into an algorithm. So you have the TIMI score, the heart score, the Edax, to name a few. And with those, you're able to refer a patient to treatment by putting into a low risk, intermediate risk, high risk, and decide with shared decision making with the patient if they proceed with medical management, interventional or surgical management. So the machine learning is a powerful tool that can analyze automatically the data that the electronic medical record provides, and that leads to efficiency. We're all looking for ways to save time to prevent, you know, cumbersome workload, to prevent burnout. And this is another way that our apps, whether you're a novice, a student, or you're unfamiliar with cardiac practice, you can use these tools. Another way that machine learning is helpful with heart failure, with deep learning, which is a subset of machine learning, you can integrate multiple data sets, sets or data sources. And by using these scoring tools and integrating some other ways, such as EKG findings or an echocardiogram, you can now network with those other technology to incorporate them into the heart failure scoring tool. And then that way it can do that predictive analysis. So these are some tools that are currently being used. But also there's a future of making more integration with this deep learning. Anything that you want to add to this, Amy?
- [16:37]
A
No, I think you summarized that absolutely beautifully. No. I wonder if Sheil has any thoughts from her perspective.
- [16:45]
E
No, I agree. Marcie. Got it all.
- [16:50]
F
Thank you, guys. Yeah, I mean, like, that touches on so many, like, hot topics. I feel like within the realm of medicine, you mentioned, like, burnout, which is another huge topic, and how AI is helping to reduce burnout. I think it's great. So before our next question, though, Ben, can you help describe some of the different subsets of machine learning, like deep learning, natural language processing, and generative AI, and can you provide maybe some examples of them?
- [17:18]
C
I would be happy to. So there's a wide field of artificial intelligence, machine learning, and some of the topics we talked about are just some of those subfields as a whole and what those different subfields will become as all this. This does change over time, as new capabilities emerge as a whole. One of the fields that has really blown up within the past few years is the field of deep learning. And so the idea, when somebody says deep learning, I gave the example earlier of what a linear regression is. In this particular type of instance, imagine a model that has millions of different parameters and interactions in such a space and time. It's actually modeled after the neural networks similar to a human brain in that space and time. And so the idea here is that with more complex information, you have a lot more little chances to find those different patterns, because things like language are incredibly complex in terms of the insights we provide. Like, the tone of my voice right now is providing some level of information as you're listening to it. And these larger models can capture those individual components to help build a model in that situation. A good example of a deep learning model is actually a ChatGPT. ChatGPT, the engine behind it is a deep learning model. It's a specific type called a large language model. But there's a large amount of those that are being used to power all these different tools today. Natural language processing as a field is specifically where a computer can handle human conversation and human language that's both spoken and written. So natural language processing really does mean the processing of human text in conversation. So a great example of this is a spam filter on your email that is a form of natural language processing. If you ever try to hit Control F on a document and find a keyword, that's an example of natural natural language processing, as well as all the different large language models and other tools out there that also work in this space. On the last one, this is really what's blown up. The overall field is generative AI. AI as a whole is not new, but generative AI as a subfield is what has emerged over the past couple of years. And so this is a form of AI that can create new content like text or images or audio or other different fields based upon previous information. It literally creates something new in that process. So going Back to the ChatGPT example, not that I have a preferred relationship with OpenAI or anything like that. It's just it's the most popular tool today. Use ChatGPT. You can ask it to create an image or some audio or those other filters or other content in that type of situation. And so those are some examples as a whole, as a part of this process. But again, AI is not new. What's really changed here is that as the amount of computing power that we've gained and the amounts of data that have now become available thanks to the Internet and all those technological innovations, there's now the ability to find those patterns which then can be applied in different specific settings.
- [20:23]
F
Thanks so much for that, Ben. You know, I think this highlights my naivety in AI basically, and just the fact that like, not even thinking about the daily use that we, that we all are using AI just in like those spam filters and the control F, like I never even would have thought about that being an AI system, but it totally is.
- [20:43]
C
Yeah, actually the Control F is not AI. That one is actually just a form of natural language processing. That is an AI. In that situation, it gets really complicated in terms of definitions there. Everything changes all the time. But spam, emails, filtering, that is AI. That is AI.
- [20:59]
A
Gotcha.
- [21:00]
F
Thank you. Thank you. All right, to our authors. How have you seen these different subsets of machine learning used in cardiology?
- [21:10]
D
Well, just to start off, cardiology really has embraced AI and these different models. And there's quite a few examples that I'll go into. And then Amy's going to share some examples. Examples as well as shield. You know, machine learning examples will be the cardiac risk scores. You know, those have been around for a long time. There's hospital readmission risk scores also in clinical trials, you know, trying to screen for eligibility. That's also been used extensively in heart failure. An example I can give from a recent study is cardiac death or sudden cardiac death has been a major cause of death in Kenya and there's so many rural areas and access to care is very poor. So what a researcher out of University of Texas has done is he's applied an AI model that has made it more accessible to use EKG interpretation to detect if somebody has left ventricular systolic dysfunction, since echo technology is not as readily available in a lot of these areas. And so this is a really nice way of using both machine learning, but then also this deep learning where you're able to tie it in with some other methods. The other example for deep learning would be Apple watches. People are using that. And I'm sure those of us working in cardiology have seen this patients coming in and saying their Apple watch, notify them that they have atrial fibrillation. You know, it's a good thing that they are able to recognize. But then what do you do with the data? So that's another issue that I think we'll be talking about later. And then another example would be, as Ben was talking about clinical trials, you could use it for predictive models to see what outcomes are. I think Amy has some other examples that she's going to share.
- [23:29]
A
Great. Thanks, Marcie. So Ben gave fantastic definitions in terms of the, the two summary. You know, the, the two, the two types of AI that I'm going to discuss. So first I'm going to start with some natural natural language processing, which is the application that has the ability to comprehend human language in either spoken or written format. So this is what I try to think of as a two way street. Right, because it could work both ways. For example, there are some echo surveillance tools out there, just like the patient in our case study had applied to her echocardiogram. And certain applications that utilize NLP can quote, unquote, read echo reports that are dictated by the reading physician. So they're not only looking at discrete fields in terms of aortic valve mean gradient or aortic valve area or so on and so forth. Not just discrete fields, but also looking at the language that's dictated by the reading physician and then interpreting that language. Another example of NLP can be also used in clinical documentation improvement, or CDI to identify missing or unclear information in medical records. AI can suggest more accurate medical coding or language to ensure documentation supports billing and the big one, regulatory compliance. So NLP can also generate very dedicated and tailored patient education tools for the literacy level that we talked about with our patient in the study and also the language spoken by the patient. And just as we referenced, you know, AI can create specific images and graphics and also create treatment plans or kind of visit summaries based on patient data by listening and interpreting language, human language. So those are some examples of how NLP or natural language processing can be utilized. So generative AI, which we heard a little bit about, it's, it's super dynamic, right? A dynamic system capable of analyzing, analyzing data, making predictions. Now that is where we had the patient on a case study. Generative AI was applied to her diagnostics and it was found that she had a complication. Right. And that if she had a transcatheter aortic valve replacement procedure, this potential complication would have potentially led to a negative clinical outcome. So her treatment plan was altered based on this, based on this predictive modeling. Very, very interesting. And again, real world, real world. So another sort of example of generative AI that's being utilized in cardiac care now is the use of AI generated personalized 3D heart models which our patient had in sort of a two dimensional fashion in that, in that software. But this really aids and as we talked about pre surgical planning and, and even planning prior to interventional procedures like tavr. So really, really interesting because it takes a specific patient's unique cardiac anatomy and can really apply that clinically into planning complex procedures and, and, you know, makes it easier to communicate with patients. For example, that patient in our study, she really wanted to have a transcatheter, maybe a tavern. But, you know, once you kind of gather this data and can show the patients like this, this patient that, you know, she did not have proper anatomy that would be conducive to a successful TAVA procedure, and she needed to have a surgical aortic valve replacement instead. I think that really aids in patient satisfaction. Patient experience kind of builds that trust and rapport with the multidisciplinary heart teams. So, and AI was kind of a partner through, through all of that. Those are my thoughts about those two types of AI.
- [27:16]
D
Amy, before we go on to Jill, I just want to add that the power that AI gives, it's a somewhat more consistent tool. So when you're talking about interpretation of EKGs and echo, we know in clinical practice we have cardiologists that are interpreting the results, and sometimes there could be a little bit of variance on how they interpret the findings. But with this machine learning, it is usually a more consistent manner. And that can somewhat leverage the benefits versus the potential negative of AI not interpreting appropriately. Don't you agree, Amy?
- [28:07]
A
Yes, perfectly said.
- [28:09]
D
So, Shio, can you talk about how generative AI helps in research?
- [28:14]
E
Yeah, of course. So generative AI and natural language processing are becoming a huge asset in the research space too, especially if you know how to use it creatively. So let's say you're screening for eligibility criteria from a massive data set like a national registry that used to take hours or even days of manual work. And now, with the right prompt, generative AI can help you sift through it in just mere minutes. And so it's like having a, a super fast reader on your team. But here's the catch. You have to be thoughtful about how you ask. So there are two ways to approach a task. You can either ask the AI to directly find a specific trait or criteria, or you can ask it to write code that will do the filtering for you. Think of it like this. If you had a really smart friend who reads fast, they might still miss something just because they skim. But if that same friend wrote a program that looks for all instances of, say, a particular diagnosis, they're going to be much more precise. So just thinking about the best way to leverage these tools to get the most accurate responses in research.
- [29:23]
D
Yeah, I agree Shiel, that the consistency and how things are being viewed or interpreted really is powerful.
- [29:33]
A
Wonderful.
- [29:33]
G
Thank you so much for your input. And with those real life examples in cardiology, being a PA in cardiology, I can say that this is a huge help and advancement in technology, especially in the heart failure world. We use remote patient monitoring and all the systems that are in place are really helping us with those patients. And access and Marci, you touch on access in remote areas and I see how this can help a lot with EKG readings, echocardiograms. We don't have specialist care cardiologists everywhere. So this is all great information here and I'm learning a lot from all of you in the article. You also and if you hear some background noise, by the way, I'm also at AAPA here with Joe, so that's why we are having a lot of background background noise. You also describe in your article rule based expert systems in your and these systems mimic our decision making processes using if then then statements. This technology was used to create a heart failure mobile phone based telemonitoring system that generated alerts and care suggestions based on the patient's inputs like weight, heart rate and symptoms. So it sounds like the potential applications of rule based systems are huge. Can you share a few more examples from the world of cardiology?
- [30:47]
D
Specifically, I'll start off saying that rule based systems have been around for a long time and they're very helpful because it follows an algorithm that if something happens then you're going to recommend this. Things such as risk based scores is very straightforward thinking where you're going to get that basic understanding of what their risk is. Is it 5%, 15% risk of having cardiac disease within 10 years or so? And so the rule based systems use these guidelines and then they're going to provide you with a way you can communicate with the patient. Are they low risk, intermediate, high risk? And as I mentioned before, this is important when you're trying to determine with that shared decision making are they going to receive medical management, interventional or surgical management? And we know that studies change all the time about, you know, are you low risk, you taking the cath lab or do you just have watchful waiting, you know, treating with medical management until things progress? You know, what is the amount of blockage in their coronary arteries before you decide you're going to do a bloomer stent? And so these rule based systems are giving you a general understanding. They don't learn. So in a complex patient, you still need that provider that's going to use their unique comorbidities, their multiple diagnosis to consider before you make that decision making. And so for example, a patient may do a risk score for should they be started on a statin or not. There's apps that will tell you if you're at a high risk and it may say start a high dose statin, but it may not consider that they have a history of liver disease. And so you're going to, you know, need to consider other things before you prescribe that management. Some other examples in the field of electrophysiology with afib, there's also risk scores, there's many different risk scores to determine if you are going to start anticoagulation, are you going to start some antiarrhythmic therapy. And then you talked about the heart failure. But there's also pacemaker rule based systems that will detect if you have vtach svt and then you know, the pacemaker nurses or apps monitoring it will contact the patient to decide what their next step will be. So those are some ways I think that they are very helpful and I wanted to see if Amy had anything to share about this. Sure.
- [34:02]
A
No, Marcie, that was, that was excellent. And essentially I'm going to mimic my much of what you said because you know, risk based score, they're incredibly helpful. And as for example in the structural heart world or the surgical, you know, cardiothoracic world, you, you utilize an STS score, the Society of Thoracic Surgeons, you know, risk to try and again, risk, stratified patients, are they low, intermediate, high risk, perhaps prohibitive. But you know what I always tell students that I'm, you know, working with or, or something like this or new apps to the field, that this is merely a jumping off point.
- [34:35]
B
Right.
- [34:35]
A
It gives you a nice kind of guideline about if this, then this, okay. And this is so this is a calculator and it, you know, does exactly what it's supposed to. Now the STS score, for example, does not take certain things into account when it's, you know, doing its calculation and spitting out the score. And that projected risk of such things like oh, various blood dyscrasia, hostile chest frailty metrics, cognitive decline or anything that about your patient. So yes, they are extremely helpful for a jumping off point. But nothing can ever replace the patient in front of you and it's the provider, the clinician, putting your hands on a patient, talking with the patient. The eyeball test is kind Of a more of a little bit of a kind of flippant way to say it, but really nothing can replace the provider aspect in that multidisciplinary team kind of coming together to discuss patients and then involving the patients in the care plan.
- [35:30]
C
Of course.
- [35:30]
A
But so really I'm just kind of dovetailing on what you said, Marcy. Couldn't agree more. So very useful, but jumping off point again. But nothing can replace that clinician insight that, that humanistic element, if you will, to help really guide patients along the right care pathway.
- [35:47]
D
You know, I'll just say looking in my library bookshelf right now, one of my most favorite and helpful books that I use starting off in my clinical practice was Collins Differential Diagnosis. And I love that because it had algorithms for different symptoms and it was easy to understand. They were rule based where you said if you had certain decision making pathways, you would do that. So that's kind of the way our brains work.
- [36:22]
G
That's great. And I'm happy to hear that the human. We will still need the clinicians expertise with those machines that we are not, we are not going to solely depend on them because I know, we know a lot of people are worried about that, how AI are going to replace our jobs as clinicians. So robots are being, are beginning to become more common in healthcare settings. So how have robotics and AI being integrated in the healthcare setting? And is there also potential to use this technology within clinical education? And also probably Amy could chime in specifically into PA education.
- [37:01]
A
Oh, Martine, this is a great question. I like this a lot. Okay, so in terms of robotics and AI and how they've been integrated in the healthcare setting, I think this answer is pretty amazing. Right. So for example, robotic surgery, right. If you think about systems like the da Vinci surgical system, they assist in minimally invasive procedures. Many cardiothoracic surgery procedures like robotic mitral valve repair, robotic coronary artery ByPass grafting, robotic ASD or PFO closure or repair, or even robotic removal of some cardiac tumors. Pretty incredible stuff. But robotic surgery also used very commonly in urologic, gynecologic and also some general surgeries. So we're definitely seeing, seeing that that trend kind of proceed and continue on an upward trajectory. I think we touched on another important point being the medical imaging and utilizing AI to help interpret diagnostic imaging. So X rays, CT scans, MRIs to detect certain conditions like cancer, fracture, strokes, lung disease. That extremely high accuracy. Also very useful in these kind of incidentalomas. You know, incidental findings that are seen on diagnostic imaging performed for some other reason. So AI tools can really flag early signs of, say, breast cancer in mammography before radiologists can actually detect any sort of abnormality with their, with their naked eye, if you will.
- [38:38]
D
We have quite a few referrals to cardiology services based on these incidental findings.
- [38:45]
A
Absolutely. And one of the classics, classics that we see time and time again is, you know, a patient needs to have a cardiac clearance for something routine like a cardiac colonoscopy or a knee replacement or something, and they end up finding something significant in terms of their cardiac history, valvular heart disease or coronary disease, what have you.
- [39:06]
C
So.
- [39:06]
A
Yes, exactly right, Marcy. Let's see other ways that I think robotics and AI are making their way kind of integrating every day into healthcare clinical decision support. We talked a lot about that kind of along the lines of our case study from earlier in the podcast here. AI analyzing patient data to assist in not only the diagnosis, but also treatment planning and risk prediction. You can see this in a non cardiac situation such as predicting sepsis onset before clinical signs appear. A lot of process automation in terms of operational efficiencies like billing and claims and data entry, reducing staff burden and errors I think is one of the main goals there. So yes, increase efficiency, productivity and ideally reduce errors. We see a lot of virtual health assistance chatbots who are providing this kind of 247 patient support, not only helping with medical queries but also helping with those. Again, kind of part of that administrative piece like the appointment scheduling, monitoring chronic conditions is something else that is really taking off AI and drug development, precision therapies, really accelerating drug discovery and clinical trial design by predicting therapeutic targets more of efficiently and effectively than more than traditional method has historically done. And something I find extremely interesting is this continuous monitoring and predictive care piece. And Marcy, you touched on the patient who has the, the Apple watch and says, well, it's telling me I'm an afed, but you know, we're seeing more and more of these wearables and implantable devices powered by AI or working in partnership in conjunction with AI, providing this real time health monitoring, predicting some sort of arrhythmia, like we said, a crisis or even a decompensation kind of before it happens. You see some of those pulmonary artery sensors that are really attuned to this and kind of give these warning signs as they start to see some indicators that a patient might be coming fluid overloaded and headed toward decompensated heart failure. So really interesting there in terms of how it's AI is rearing its head in our everyday clinical practice. But Martin, you asked a really important question about, you know, do you think there might be a potential to utilize that technology within education for advanced practice providers? And yes, I mean, I absolutely. I don't think we're far away from this model. And this extends to all types of education where a student's kind of foundational knowledge base is assessed by AI and then a tailored or personalized education plan is formulated and then, you know, various time points where patients, patient where a student's progress can be evaluated kind of over time and then that consistency in terms of how does a learner learn the best, what's the best modality to kind of not only absorb but retain information and to be able to pull it back out. You know, there are different types of learners and I think AI can really help establish what is best for each individual person. So that was a little bit of a long winded answer, but I super nerd out about this kind of stuff. So thanks for the question.
- [42:20]
G
That's great. That's great. And I like how you mentioned incidental findings. So on the opening night here at apa, there was a talk about AI and how a lot of pulmonary nodules were being detected when you ordered all those CT scans of a cardiac score for CAC score and stuff. So yes, we talked about a little bit about how people were a little bit afraid about AI taking our jobs and all of that. And can you think about any limits of the models and applications you have discussed? You can share some of those limits. We kind of touch a little bit on that or if you have any additional information.
- [42:55]
D
Well, I'll start off on this by first saying that the first case study that we gave really showed what is currently being in practice. AI is working in current practice and people are concerned that it's going to take our job. But it's already working. It can enhance our jobs, it could promote efficiency, decrease workload, and as I mentioned before, you know, cut down on some burnout. We know that's a huge issue right now in our professions. If we use it appropriately, it's going to allow us more time for patient interaction. We won't have to fumble around to find the most current guideline. As long as the AI system has the right tools and data plugged into it, we're going to be able to communicate with the patient with better tools and incorporate the comorbidities. This is based on what kind of AI model you're using. Is it more basic machine learning? For predictive tools or is it more generative AI? So on the limits side, I'm just going to say that our jobs may be restructured and just like our jobs have evolved over the years, there's going to be more, well, depending on in the beginning, like when we first started utilizing Electronic Health System, there's going to be more IT positions to get us to plug in the data, to discern what is good data that can be plugged in and whatnot. So it's going to be time consuming. Yes, just like Electronic Health System. And unfortunately a lot of us started off in one system, whether it be Epic, Cerner or another one. And then you changed over when you found that it didn't fit your organization. So there's going to be a learning curve as we've mentioned over and over. You're still going to need a real person. And so that's where I think it may cut down having a real person in some areas, but you still need to have a real person on the other side to discern or interpret did the AI work appropriately? So I think we are afraid of something new and we want to look at the bad side of things before we look at the benefits. And that's where we try to point out these things in the article by using the case studies to show how it shouldn't be something to be feared. So Amy, your thoughts on some of the limits?
- [45:48]
A
Oh gosh, I mean, I think you just summarized that perfectly and I just want to touch on that, that whole concern. Will AI take our jobs? You know, they do it. A robot does my, my, you know, robotic MVR better than, better than I would with less error and more accuracy. No, no, no. But I, I like how you said earlier, Marcy, about AI enhancing our jobs. I think that is absolutely the case. And I'll just say this quickly. I mean I think AI is, is an adjunct to our clinical practice and our lives in general. Right. These applications will always need a human to be the interface and provide that humanistic element. And I think medicine has a very creative side to it. It is an art after all. And this, that kind of creativity and ability to really think outside algorithms and, and recommendations about, you know, rule based and the risk based score system that we humans inherently possess. This we create, we learn, we evolve and that's something that I think will always be necessary and no, will never be outed by AI. I, I just don't, I just don't see that happening. But an adjunct for sure. And I will say that, you know, in hearing. Oh, I have a question for you actually, Marcy, in, in hearing all about this, you know, how AI is being applied today and potentially tomorrow, you know, it's, it feels like the future is here. So successfully integrating this technology, I think you are hinting there are some ethical and cost concerns. And I wonder if you wanted to kind of touch on some of the challenges that you see in applying AI across the care continuum today and tomorrow in terms of ethical and cost considerations.
- [47:34]
D
Well, you know, as I mentioned, there's going to be a learning curve just like there was for integrating electronic health systems. And there's also going to be those ethical concerns when you use a system that is pooling data and it's going to, you know, use these rule based systems or algorithms to possibly create biases based on the pool demographics or the social determinants of health in the database itself itself. So, you know, when you go for a medical appointment or you're admitted to a hospital, they always ask you, you can tell I'm a recipient of patient care. They always ask you, you know, are you going to be able to afford your medicines? Do you know where, you know, like your next meal is all these things. And so is this going to create biases or are they going to class you in a different category? Will the patient information be de identified like you do for clinical trials? So this is where we look at the ethical concerns, but on the upside, you're going to be able to trend patient populations that have similar diagnosis and comorbidities. And this is powerful to predict outcomes regarding the cost benefit, you have to look at the return on investment. You know, AI technology, just like when we had electronic health systems is going to cost some money. And so you need to determine if it's going to enhance your efficiency, thereby reducing the overall organizational cost. And can this improve the revenue and streamline your data management, reduce the errors and maybe speed up your claim submissions? We know it usually takes about three months for things to really roll through to get reimbursed. So can this help? You know, this is going to be difficult for some smaller organization to incorporate at first. And the case study showed that AI could expedite care in this rural center and know if they're concerned about improving patient outcome, then yeah, let's in corporate AI or maybe share costs with your referral organization. So I just want to point out that in 2009, the Hitech act stimulated electronic health adoption. And then later on, the House HR Bill 1331amended the requirements and penalties if you didn't institute electronic health system. And that was per CMS for are you going to get reimbursed? So it remains to be seen AI technology or integrating it into more of a national consistent basis is going to receive any national funding or is it just going to be incorporated in some of the newer equipment and technology that organizations are purchasing? And then last, from the ethical standpoint, I just want to say when we do sports, the calls for high schools, you know, we have them. Do you know, blood pressure, heart rate, you know, can you bend over? No scoliosis, you listen to their heart to screen them to see if they have sudden cardiac death or a predisposition for it. There's been in sports medicine a kind of push to have EKGs done on all athletes, high school athletes, just to screen for this. So if you think about AI ethical standpoint, if we did this, is it going to cause some discrimination from that athlete to not get a skull college scholarship? You know, the athlete may not want to know if it's mandated. So AI itself can also cause some challenges on how we present predictive data. So Amy, any other comments from you?
- [52:03]
A
I mean, I think that's a really good example that you just utilized just in terms of the potential conflicts that that particular athlete would be faced with. And you know, you summarize that extremely well. So I'll just kind of, I'll be brief with my remarks, but I, I do see several key challenges in applying AI in clinical practice in terms of ethics, mainly because at this point in time there is no overarching regulatory body. And I think some clinicians are concerned, right, you know, about bias, fairness, transparency of data sharing. With that you, there are concerns around privacy and security of your data, informed consent, accountability and liability. You know, if who, who actually is liable if you do screening and it's showed something positive potential or a predisposition. And yeah, and then over reliance on technology and the AI applications can be, can be challenges that are perceived by not only clinicians, but I think everyone who sees AI creeping into their everyday lives. But I will say that I do think addressing these concerns, these specific concerns, requires a very transparent design, very regulated ethical oversight and involvement from all of us who are users of AI application to really give input and insight so that things can progress in a very positive way. And again, that's the human element of artificial intelligence and our current standing in relationship and partnership with it. Sheil, I'm really interested in your thoughts around this, just based on the clinical research angle and perspective that you have what are your thoughts?
- [53:55]
E
Well, I think, you know, thinking about and I might be jumping ahead here but I think research wise one of the main things, one of the first things we do when we start any study is we think about how to protect our human subjects. And so something when it comes to utilizing AI in research is I think about privacy concerns. You know, and it's not always about with AI it's more than just de identifying your data. You are trying to protect your participants from adversarial attacks that might trick models into giving up information that they were never meant to share. And so in research, when it comes to utilizing AI and the cost and the benefits of that, I think administratively it is a great tool that will help with burnout, similar to clinicians. But I just thinking about how we can use it responsibly is my main take on utilizing AI.
- [55:02]
C
I really appreciate your take there and insight there Shield but I would love actually to follow up with one of the comments that Amy mentioned during that section is one of the largest questions within AI is I mentioned earlier that there isn't really a good definition at the moment because oftentimes there's a question of scale of one a larger model versus a smaller model. And we've already talked a little bit about the role that automation may play, which isn't necessarily an AI task either. And one of the things that you mentioned in the article is often oftentimes that there is more power with some of these advanced machine learning models that they may have over the more traditional smaller models, but that power is often dependent on the quality and quantity of that data available. The idea being that there's better data quality, AKA more information within it that leads to more potential interactions with that the advanced models can then find. You talked a little bit about the data sharing. I'd love to hear some more of your thoughts about what does that, what should that look like within healthcare to power these more advanced models?
- [56:05]
E
So I can talk a little bit about data sharing. So you know, as you mentioned, sharing healthcare data at scale means we have to be incredibly careful. Data is a fuel for AI. The better the data, meaning more accurate, more detailed, more diverse, the more sophisticated our models can be. And so with a rich data set, advanced machine learning models can detect patterns that linear models might not be able to. But when it comes to data sharing, we're talking about protecting against things like jailbreaking, where someone tries to manipulate the AI into revealing private data. Training data or even language itself can be used as a tool to bypass restrictions which makes the job of securing AI models way more complicated than just putting up a firewall. So we're in the silica balance. On one hand, you want to build these powerful models that can improve diagnostics and reduce disparities and even predict patient outcomes more accurately. But on the other hand, you've got to respect and protect the people behind the data. So it means we're not only in need of smarter tech, but smarter regulation and better encryption to make sure we're not creating tools that can be reverse engineered in harmful ways. And so essentially, I think data sharing in healthcare has really, really enormous potential. But I also think it has to be done with precision when it comes to privacy and security, at least on the research side, for sure. Marthi, I don't know if you have any other thoughts.
- [57:30]
D
Well, as far as looking at, you know, the research aspect, there's a lot of national registry that cardiology has been using for, you know, over 25 years. The American College of Cardiology has a national cardiovascular data registry, the ncdr, and it tracks various topics in various fields. Atrial fibrillation, chest pain, electrophysiology procedures, valvular replacement and repair, as Amy had mentioned before. And it's really the most comprehensive and largest database that looks at outcome based patient data. And this is very helpful in benchmarking hospitals, health systems practices so they can improve patient outcomes and their quality of care. It gives them insights on what other centers are doing that they can mimic. And I know when I go to conferences, I like talking to people from those institutions that have good results to find out, hey, what are you doing? So based on this, there's analysis, research can be performed and then that benchmarking. So as I used to do clinical drug trials many years ago, as well as entering data into these registries, and at that time everything was handwritten and it's very time consuming. But there's also a margin for errors due to whoever is recording the input. I always like to say garbage in is garbage out. So this way of collecting data in a consistent manner with AI, it's able to extract, you're able to now integrate the data in these registries. And that's really going to revolutionize how we look at health outcomes. So deep learning algorithms where you're able to pull from various sources is going to uncover some of these nuances, these patterns and correlations that previously were hidden. And I know Sheil had talked about you may have some jailbreaking or pick up data that you shouldn't be looking at. And that's where there needs to be some kind of policy making or some rule of de identifying certain data that may cause that discrimination or bias. By pulling this data, you're going to have a larger pool so you can have more precise diagnostics developed. Treatment protocols can be modified and that predictive analysis is going to be a little bit better, more rigorous. So on a whole, it's going to fundamentally change how clinicians approach patient management. You know, again, robots, AI is not going to replace us. It's how we approach the patient management is going to change. We're going to have more tools. So this is really important in complex cases that require you to look at multiple areas and think about. You get this big chart or you know, this big file that you have to sift through. This will help summarize some of the data as well as sharing data between different organization. You know, it just depends on some of those privacy and safeguards. But on a whole, it's going to help extract data or possibly to set up how we do future research protocols. Because we can use large populations, we're going to have a larger end for studies that makes them more rigorous. So we just need to know how is the data going to be shared? Is it going to be used in an appropriate manner now, such as is government or insurance organizations going to use it for other purposes? And one of my examples I can give there is. I know working at a outpatient clinic many years ago, I saw many truck drivers who came in with hypertension and in California and I think in other states too, but the Department of Transportation would take away their license if they had high blood pressure. That was uncontrolled. So would this information be now shared with the insurance organizations or government or their employer? That now is going to jeopardize their job. So is that information inputted into the system going to be accurate? How is it going to be shared? Those are some things that we have to think about. I know that was long winded, but anything else that you want to share, Sheil or Amy?
- [62:57]
A
No, I agree with you, Marcy. I mean, I think as you said, garbage in, garbage out. Really need to look at the accuracy. That's a huge component of this. And then extrapolating various things from what we're putting into applications.
- [63:14]
G
Yeah.
- [63:14]
A
So it is multifaceted.
- [63:15]
E
Yeah, I think just overall being aware of these things, it shouldn't be make people feel intimidated to utilize these tools. It just, I think or hesitant. I think if people are aware, it just makes them. It makes. Provides opportunities to be more careful and prepare some safeguards in the process.
- [63:33]
B
Thank you all. This has been a really eye opening and enlightening conversation. I want to turn to you, Ben. So first, thank you for sharing a lot of good context to help us understand some key concepts and definitions in your role as a data scientist and population health professional. What are some ways that you've seen AI successfully utilized to promote health and wellness? And do you have any words of caution or wisdom for clinicians looking to utilize this powerful technology?
- [64:02]
C
Yeah. Thank you, Kim. This is a good question and I think it's really important to note that when we're talking about AI here now, not just as the many different examples have been given throughout this fantastic and I mean just some truly incredible points covered throughout this discussion that first, it's going to take time to see full adoption across many of the different fields, realms, et cetera, from both the clinician side as well as from the patient side as well. And so I'm going to talk a little bit about some stories I've seen from the patient side as a part of that process, with the disclaimer being that it's very important to note forward that AI is growing in terms of capabilities at a near exponential rate. And so when we're thinking about what the capabilities are today, that will not be an accurate reflection on what the capabilities are a year from now. And just to keep that in mind as we're continuing to test and use these different tools, but one thing I really am hearten to hear stories about is something a term I like to call the AI empowered patient or the individual who's able to then be able to use these different tools to get that insight and information to be able to better access and make decisions for their own care that they couldn't necessarily do before based upon that information. That could be somebody who didn't know how to access into the specific healthcare field in terms of being able to find, to get an appointment or how to utilize the system to somebody who maybe will be looking into how to best handle their symptoms. One example I love that I've heard anecdotally is I was teaching somebody how to use, better use a different specific generative AI tool. And they reached out to me afterwards and said that they've this was in a professional setting, but they said like hey, I've my family has some severe dietary restrictions for us it takes four to five hours a week just to plan out our grocery lists, our shopping lists to make sure that we can keep all of our, all of our food in together and making sure that we don't just eat the same meal again and again. And using these different tools, I've been able to cut down that time to like 15 minutes. And this has saved my family so much time in terms of being able to make sure we get food that we can actually eat as a part of that process. Another example I like to give this one's from Google of so a little disclaimer there that this is from an organization, but they gave a case study recently where there was a father whose son was diagnosed with Alexander's disease, which is a very rare genetic disease. And the father was not in the medical field any way, shape or form, but he was able to start research progress being developed into potentially finding different cures for this disease and reaching out to different medical research labs and to do advanced testing using his son as an example and others across the board to help advance the understanding of such a disease, which he would not have been able to do without using different AI tools. And so when we're thinking about moving forward, taking heart in the idea of kind of this AI empowered patient where more individuals be able to take the steps and actions to improve their lives using that information, these tools available, not just also on the clinician side, reducing that burnout, be able to provide that more realized care that also have been called out as a part of this process. I do wish we could keep going. I mean, I've learned so much from this podcast so far and this has been absolutely fantastic. Yeah.
- [67:33]
F
Thank you guys. Thank you all for being with us today. I mean, this has really been such an interesting topic and I just want to especially thank our guests today and also our wonderful guest host, Ben. Thank you for being here with us. This has really been an amazing podcast. I'm glad that I could be a part of it. While this is our first podcast surrounding AI, we doubt it will be the last. We've learned so much from all of your experience and we'll definitely take it forward with us in practice in this rapidly changing world. And don't forget, to our listeners, you can earn CME by listening to the podcast. To receive your CME credit and access your certificate, just listen to the podcast, then complete the post test and evaluation in AAPA's Learning Central at cme.aapa.org until next time.