The Power of AI in Medicine

Podcast Transcript

Neuro Pathways Podcast Series

Release Date: March 1, 2024

Expiration Date: March 1, 2025

Estimated Time of Completion: 29 minutes

The Power of AI in Medicine

John Morren, MD

Description

Each podcast in the Neurological Institute series provides a brief, review of management strategies related to the topic.

Learning Objectives

• Review up to date and clinically pertinent topics related to neurological disease
• Discuss advances in the field of neurological diseases
• Describe options for the treatment and care of various neurological disease

Target Audience

Physicians and Advanced Practice providers in Family Practice, Internal Medicine & Subspecialties, Neurology, Nursing, Pediatrics, Psychology/Psychiatry, Radiology as well as Professors, Researchers, and Students.

ACCREDITATION

In support of improving patient care, Cleveland Clinic Center for Continuing Education is jointly accredited by the Accreditation Council for Continuing Medical Education (ACCME), the Accreditation Council for Pharmacy Education (ACPE), and the American Nurses Credentialing Center (ANCC), to provide continuing education for the healthcare team.

CREDIT DESIGNATION

• American Medical Association (AMA)

Cleveland Clinic Center for Continuing Education designates this enduring material for a maximum of 0.50 AMA PRA Category 1 Credits™. Physicians should claim only the credit commensurate with the extent of their participation in the activity.

Participants claiming CME credit from this activity may submit the credit hours to the American Osteopathic Association for Category 2 credit.

• American Nurses Credentialing Center (ANCC)

Cleveland Clinic Center for Continuing Education designates this enduring material for a maximum of 0.50 ANCC contact hours.

• Certificate of Participation

A certificate of participation will be provided to other health care professionals for requesting credits in accordance with their professional boards and/or associations.

• American Board of Surgery (ABS)

Successful completion of this CME activity enables the learner to earn credit toward the CME requirements of the American Board of Surgery’s Continuous Certification program. It is the CME activity provider's responsibility to submit learner completion information to ACCME for the purpose of granting ABS credit.

Credit will be reported within 30 days of claiming credit.

Podcast Series Director

Imad Najm, MD

Epilepsy Center

Additional Planner/Reviewer

Cindy Willis, DNP

Faculty

John Morren, MD

Neuromuscular Center

Host

Glen Stevens, DO, PhD

Cleveland Clinic Brain Tumor and Neuro-Oncology Center

Agenda

The Power of AI in Medicine

John Morren, MD

Disclosures

In accordance with the Standards for Integrity and Independence issued by the Accreditation Council for Continuing Medical Education (ACCME), The Cleveland Clinic Center for Continuing Education mitigates all relevant conflicts of interest to ensure CME activities are free of commercial bias.

The following faculty have indicated that they may have a relationship, which in the context of their presentation(s), could be perceived as a potential conflict of interest:

The following faculty have indicated they have no relationship which, in the context of their presentation(s), could be perceived as a potential conflict of interest: Cindy Willis, DNP, John Morren, MD

CME Disclaimer

The information in this educational activity is provided for general medical education purposes only and is not meant to substitute for the independent medical judgment of a physician relative to diagnostic and treatment options of a specific patient's medical condition. The viewpoints expressed in this CME activity are those of the authors/faculty. They do not represent an endorsement by The Cleveland Clinic Foundation. In no event will The Cleveland Clinic Foundation be liable for any decision made or action taken in reliance upon the information provided through this CME activity.

HOW TO OBTAIN AMA PRA Category 1 Credits™, ANCC Contact Hours, OR CERTIFICATE OF PARTICIPATION:

Go to: Neuro Pathways Podcast March, 1 2024 to log into myCME and begin the activity evaluation and print your certificate. If you need assistance, contact the CME office at myCME@ccf.org 

Copyright © 2024 The Cleveland Clinic Foundation. All Rights Reserved.

Introduction: Neuro Pathways, a Cleveland Clinic podcast exploring the latest research discoveries and clinical advances in the fields of neurology, neurosurgery, neuro-rehab, and psychiatry.

Glen Stevens, DO, PhD: The brain remains one of medicine's greatest mysteries. With artificial intelligence as an invaluable assistant, physicians, surgeons, researchers are gaining an unprecedented window into how this enormously complex organ works. In this episode of Neuro Pathways, we're exploring the intersection of neuroscience and artificial intelligence and harnessing the power of AI to safely and ethically advance the field.

I'm your host, Glen Stevens, neurologist/neuro oncologist in Cleveland Clinic's Neurological Institute. I'm very pleased have Dr. John Morren join me for today's conversation. Dr. Morren is a neurologist in the Neuromuscular Center within Cleveland Clinic's Neurological Institute and serves on the board of directors of the American Association of Neuromuscular and Electrodiagnostic Medicine, where he leads their AI task force. John, welcome to Neuro Pathways.

John Morren, MD: Thank you for having me, Dr. Stevens.

Glen Stevens, DO, PhD: It's my pleasure. And I know you, but for those in the audience that don't know you, tell us a little bit about your background, how you made your way to the Cleveland Clinic, and what you do here.

John Morren, MD: Sure. Well, I'm originally from the tropical island of Trinidad and Tobago. That's the southernmost island in the Caribbean belt. So, I'm a little bit outside of my natural habitat, so to speak, but that gave me a very interesting experience growing up. That being said, I grew up in a village and my path to medicine was somewhat atypical.

I remember thinking I would never really want to deal with a lot of the complexities that come with medicine, but also had some serious doubts about medicine because of things like body fluids and things that would be not something I thought was appealing. Now, I got sick as a child and during my teenage years, I depended on doctors to really make a difference in my life.

I remember one time looking at them and saying, "Wow, you must go home feeling very satisfied with your day's work." And at that point, something kind of clicked. I was like, "It's really probably not a good excuse for me telling myself those are the reasons why I shouldn't get into medicine." I had a somewhat humble background. I knew if there was a little bit of a meritocracy, if I work hard enough, I could get into med school. At that time, my love was mostly electronics.

I would kind of scavenge things that people would throw away in terms of the electronic equipment and computers and try to rebuild them. And I thought I was heading in the direction of engineering, maybe electrical engineering if I did make it to college. First in my family to do so, but then I said, if I do take this path of medicine, I mightn't be able to abandon my love for computers, wires, and circuits. So, needless to say, in med school, I gravitated towards neurology. And I guess we were going to talk about how the love for computers didn't go away either.

Glen Stevens, DO, PhD: Well, excellent. I know so much more about you than I actually knew before. I love it. I love it. Well, it makes sense why you're interested in AI now. Of course, my children, I think, would say that I am a Luddite, but I don't think that that's really true.

For those that don't know what a Luddite is, the Luddites are the group of people in England that at the turn of the century broke up all the cotton mills and those types of things because they thought technology was going to do something poor, but I was thinking about how much AI do I really use?

And I know that you're going to define this for us, but I was thinking, I look at my phone and I turn it on, and it uses my face to engage, and the infrared dots know what I look like. And when I'm typing an email to somebody, it's correcting my spelling. I don't even have to do it. And when I'm on a trip, somebody's telling me where to go or one of these electronic things. So, I guess we're all using a lot more AI than we think that we're using.

John Morren, MD: That's absolutely right. We see it all around us, and there are just growing areas of application that really doesn't require us to sign up per se.

Glen Stevens, DO, PhD: So, we're going to go forward today without fear. We're not going to have any fear that this is going to replace anybody. We're going to go forward with a better understanding of how it's going to help everybody. So, in order to get everybody on the same page, we want you to define some AI buzzwords for us. But just tell us, number one, what AI is. Tell us things such as machine learning, deep learning, LLM, which I use as well. Not very much, but a little bit. But why don't you define these terms for us so everybody's on the same page?

John Morren, MD: Sure. No, these are definitely terms gaining traction in our daily vocabulary these days. So, AI could be defined simply as a computer system that performs tasks commonly associated with intelligent beings. Another way of looking at it is they're computer systems endowed with intellectual processes very characteristic of humans, such as the ability to reason, to discover meaning, to generalize or learn from past experiences.

A lot of this has been work that come out of pioneers in the 1940s, 1950s. So this is not a new concept, but a lot of the acceleration in the AI space happened over the last half a decade or so because of the increasing ability of computing power and namely the development of GPUs, graphics processing units, which really allow for parallel computing and really put a lot of power into the hands of these computer scientists and really dealt with the rate limiting step in AI development.

You'll hear the word machine learning a lot, and I think if we look at that word, it has the word learning in it. And believe it or not, that's what these computers are doing fairly autonomously. Before, a lot of automation with computing was based on rules, so we call it rules-based algorithms, if this, then that. So, with machine learning, you have these machines improving at tasks with repetition or with experience devoid of explicit programming, which is pretty neat.

So, machine learning is a subset of artificial intelligence, and then a subset of machine learning is deep learning. So, it's deeper. So, the word is kind of self-explanatory. But what deep also means is what's really low down in a set of layers. So basically, deep learning is machine learning using an artificial neural network. And we'll talk more about that, and that's what gets me excited as a neurologist because a nerve cell has a cell body and axons that goes away from the cell body, and you have multiple dendrites feeding into it.

So, when Rosenblatt and colleagues developed the concept of a perceptron, this kind of artificial neuron in the computer space in the 1940s, 1950s, they were using the same concept of this piece of our natural biology, nerves. We have input that's being fed into a cell body that's summated, so dendrites feeding into cell body with a summation to cell body executes a function. When a certain threshold is exceeded, then there's an output down the axon.

And if you arrange multiple artificial neurons in a layer and then you have multiple layers on top of each other, arborizing in a very similar way that our cerebral cortex does it, then you have an amazingly powerful computing potential. So, when you have multiple artificial neural networks layered like that, you have a very powerful AI machinery and it can do a lot of complex tasks. So, that's what deep learning is, and a lot of the deep learning algorithms are able to do a lot of amazing human-like complex tasks.

Glen Stevens, DO, PhD: And then the large language model?

John Morren, MD: So now that we understand what deep learning is, LLMs or large language models, are DL or deep learning algorithms that could perform a variety of natural language processing tasks. So, NLP tasks as they're called. So, they could recognize, they could summarize, they can translate, predict, and generate content using very large data sets. And we're in the era of big data, and that's why these things are so much more possible. So, the core of the algorithm predicts the next word in sentences, and that's the oversimplification of the process. But this is basically what drives popular applications, like ChatGPT.

Glen Stevens, DO, PhD: I didn't even realize I was involved in LLM until I looked at my phone today and I started typing and it was predicting the next word. I mean, that's the same thing, right?

John Morren, MD: Right. In some ways, yes, but it's definitely more complex than that. LLMs can be used to actually build protein models, for example. So, it's a simplistic way of explaining LLMs, but we could build on that as we do the discussion.

Glen Stevens, DO, PhD: AI seems built for neurologists. I know the podcast is for a large group of people, but based on what you sort of described on the neuron and how all of this was developed, it seems like an excellent fit for neurologists. Have neurologists adapted this? Should we be adapting it? Should we really just love this stuff? Should we be touting it and at the front of the line, or where should we be?

John Morren, MD: Well, as you know, Dr. Stevens, I think we should be owning this a lot more than we do because the engineers, the folks who are building these models, are looking at basically the object of our regard and passion, the nervous system, and using it to add power to these computer systems. So even if we'd look at how we've evolved in our understanding of learning and memory, at one point we thought we were laying down these physical substrate or chemical substrate called engrams, and then we learned that, well, actually it's not that simple.

There are a lot of learning and memory occurs due to a lot of dynamic synaptic activity. But there's still a lot about the brain and cognition and learning that we still don't completely understand and what the computer scientists call a black box phenomenon. And that's exactly what's going on right now with this revolution and AI using deep learning. The artificial brain, so to speak, it's producing results that are very powerful beyond what the scientists can explain.

So, you'll hear the term explainability. A lot of these models, they perform so well, but there's poor explainability or we are solely peeling the layers of mystery around them as we speak. 

And another thing you could look at, there's some parallel here between some of our more advanced learning theories. And this might be a little bit in combination of neurosciences and education theory. Traditionally, there was a lot of emphasis on cognitivism in our learning process, but more recently we realized that more meaningful learning happens with making meaning out of experiences. And that's exactly what these models do, they allow for a better understanding when data is repeatedly put through the algorithm, so the machines get better and better with repetition. And that's what we do as human beings. 

And there's another modern learning theory called connectivism, and that also runs parallel with what's happening in this space. So, I find it very exciting to see these parallels, and it's really one of the major reasons why I think we should own it a lot more than we do as neurologists and neuroscientists.

Glen Stevens, DO, PhD: And how is AI going to help us in clinical care? I know there's always the fear that it will replace us. I don't personally have that fear. It's really to augment what we're doing. But how do we use it to support clinical care? And how is it being used currently? What do you see in the future?

John Morren, MD: I want to make a comment on the use of your word augment, and I think that's very key and important. I think when I say AI, most of the times I am making reference to augmented intelligence rather than artificial. So, the A for augmented rather than artificial. And that really speaks to the partnership that's needed with human intelligence to make AI truly powerful and less dangerous.

But to your question about AI for clinical care, there are multiple ways this is applied. I think most of the audience, when they think of AI, they think of ChatGPT and similar large language models like Google's Bard or Claude AI or Microsoft Bing AI with Copilot, these are all free available web-based LLM applications are hugely popular, and AI has really shot into the public consciousness after the release of ChatGPT in the fall of 2022.

But for clinical care, there actually was equivalent to ChatGPT for clinicians, and that's called Med-PaLM. And that's a product from Google. And Med-PaLM 2 is out, and you probably would've, some folks may have heard about how it's performed way above 80% in the USMLE and similar medical exams. And there's a similar large language model type AI technologies used in radiology, there's something called Annalise.ai, and you could go and put in your CT brain and have this software read it for you and does a decent job, and you'll see how it performs. It's really exciting, and again, it's to augment not to replace radiologists. 

Isaac Kohane at Harvard talks about how he uses even ChatGPT responsibly to crack some hard diagnostic cases. I would certainly encourage folks to read his book, The AI Revolution.

There's a lot of work in medical scribing, so DeepScribe and Dragon, DAX and Augmedix, they use ambient AI technology. So, we can probably get our notes done accurately, quickly, and liberate ourselves from the keyboard and the computer screen so we could spend some more meaningful time with our patients. 

And then you could use AI for a number of administrative purposes in the medical office, using something called Consensus, which we could talk more about.

You could use it to optimize your patient portal messages to make sure that you're speaking to the patient at the grade six or seven reading level as the NIH recommends for a lot of the medical explanations and things. 

And even here for clinical care, we're doing a lot with an AI in Medicine grant here neuromuscular to improve things like EMG. As you would imagine, patients have been asking for a long time, how can we use technology to lessen the pain and the discomfort that comes with needle electrode examination? And I think this is going to be a big part of how we really revolutionized that.

Glen Stevens, DO, PhD: So, John, I started in an era where it was just pen to paper writing the notes. So, things have definitely progressed. But it just sort of seemed like one day ChatGTP was out there and it was free. What was the thinking? Why did they just release it free? Is there something behind it? Do you understand the story, why it was released, what the intent was?

John Morren, MD: Yeah. I think it was a very strategic move. I mean, we could talk about how OpenAI started up as a nonprofit and now they're a for-profit. I mean, there could be a business side to the explanation, but I think it was very strategic because I think the developers knew that there was a public fear about AI being a dangerous thing, but they wanted to change the public consciousness about AI as something that's actually truly powerful, almost scary powerful, but very useful.

I think there has been a lot of transformation in the perspective around AI after ChatGPT was released. There continues to be a lot of doubts at time about whether it's being used responsibly and there are guardrails that need to be put in place.

For example, when ChatGPT was released, within two months of its release, it had over 100 million active users. So that's pretty amazing. And it got to almost two billion. So that's with a B, monthly website hits for that ChatGPT website by the summer of 2023. So, it certainly allowed the public to become more familiar with AI technologies and how powerful they are.

Glen Stevens, DO, PhD: Yeah. So, using AI in medical education and training, are you using it currently with your fellows and residents, or how is it being used nationally?

John Morren, MD: There's a lot of applications for AI in medical education training. And yes, I am. I could break this up into how AI can be leveraged to be helpful for the program and how it could be helpful to the trainee. 

From the program side, there've been applications where you could use a recruitment chatbot on your program website to attract really good candidates. And that's been proven to be true. You could actually use it to predict the ranked and matriculated applicants, which is something if you're in program management is almost like a nirvana experience when you could see that happen. It could be used to detect implicit biases and application materials, and that's very important for the obvious reasons. It could categorize narrative feedback from faculty into particular milestones. So that automates an otherwise very tedious process. 

From the trainee standpoint, they've used AI integration to automate and improve case logging, to optimize communication simulations and other simulations to improve proficiency, to help prepare for board and in-service exams as a learning coach, because a lot of these LLM models are performing much better than a PGY-5 level trainee. It's exciting. So, there's that. And one of the examples I had shown in a recent talk was how you could use it for qualitative research in education. So, a lot of times you interview trainees or learners and you have these long transcripts, and how do you digest that to get themes? Well, LLMs are really awesome at doing that in a few seconds. It allows for qualitative research to move forward in a way it hasn't done in education for a while.

Glen Stevens, DO, PhD: So, you mentioned a little bit about guardrails. Obviously with the good is the potential for misuse. Talk about that a little bit. Limitations, how can we control it, how can we manage it? I mean, you hear a lot of concerns about academic misuse.

John Morren, MD: Yes, with great power comes great responsibility, and that's true for anything that's truly innovative like this. I think the first thing I should make very clear, even though we have easy access to powerful AI tools like ChatGPT, we should never enter sensitive information, patient identifiers, protected health information into a chat on those kind of publicly available sources because we don't have security of information for that, and that's a HIPAA violation.

But I think the most popular limitation has been hallucinations. And as neurologists and neurosciences, I think we should use the word confabulations with these models because the output from these tools can come out with a very confident tone. And if you're not careful, you would buy it wholesale and apply it and you'll be just compounding an error.

So, this has been much better with more recent iterations of these tools, but it's ultimately the responsibility of the user to double check. A lot of the newer versions for example, in Google Bard, it will flag sentences that it's outputting with a color code and says, "This, you might want to check" or you could click on it and it'll do a Google search to verify it. So, it's kind of auto-correcting itself in a lot of ways.

And you could always ask the model, "Where did you get that from?" Or, "Quote your source." And it's able to do that very well in most applications. But I think it's kind of like using dictation software. For those who use, liked I use Dragon dictation a lot, it's great, it automates things, but when I dictate into a patient's chart, I'm ultimately responsible to check that, to make sure that it captured what I said. So, it's the same kind of thing, you have to proofread before you deploy it as final. You got to confirm. So, I say you must trust but verify.

Glen Stevens, DO, PhD: Well, I think that's the concern, right? Is where does that ultimately end? Who is the verifier? Who is the person responsible to make sure that these things are done correctly? I think that's the difficulty, right? I mean, you as a person can do it for yourself, but as an institution, who checks on the institution use of these?

John Morren, MD: Yeah. And that's part of the legal ethics. So, if there is a patient harm that occurs or there's a breach of patient privacy, where's the locus of responsibility for that? So again, this is kind of still gray and a bit opaque. There are other aspects like the humanistic ethics. If you're using this to generate responses to patients through like patient portals, can you erode compassion, authenticity and the current recommendation best practices to actually disclose use of AI generated content in any of the healthcare applications.

There's bias training into these models as well, especially some of the early models. So, we have to be aware of the potential to have underrepresented groups be at a disadvantage when these technologies are applied to them. I don't know if you saw, there was an article with the New York Times suing OpenAI and Microsoft for copyright infringement. So there's that whole copyright thing.

This is scraping a lot of data from the internet and it's hard to ascertain whether or not you're infringing copyright. But with the very nature of how this content is generated. You have to be mindful too of the date limitations, right? So for example, ChatGPT, it only goes up to January of 2022. So if you're looking for more recent data, the onus is on you to make sure that you're producing information or making decisions on the most up-to-date information.

Glen Stevens, DO, PhD: There may be an AI version of me doing podcasts.

John Morren, MD: I think that's out there already so, but no need to worry about job security.

Glen Stevens, DO, PhD: Well, that's the question, am I real or am I AI?

John Morren, MD: It'll keep us on our toes.

Glen Stevens, DO, PhD: Yeah, keep you on your toes. I don't want to tell you, but we wrote the intro for today on AI.

John Morren, MD: Right. And that's the reality. But there was a nice little disclaimer at the bottom there, and that's how it should be done. And that's the idea. I think we can't pretend that it doesn't exist and we can't pretend it's not helping us. Transparency is very important.

If somebody said, if you were in an airport and you're going to tell the folks boarding a plane, "Hey, we're deciding not to use autopilot on today's flight," a lot of people will come off the plane. I think we realize that yes, we want a human pilot, definitely people want that, but we know we're better off with a combination. So it is like synergy. The total is more than just the sum of its parts. That's how I see AI+HI, and that's how the brain works as well.

Glen Stevens, DO, PhD: Well, I like that analogy. So, this has gone so rapid in terms of the development of AI, not our podcast, but where are we in five years? It's hard to imagine, right?

John Morren, MD: That's true. Because of the exponential nature of this. I think we're going to see models that have very little hallucinations, for example. There's this new aspect of the technology called retrieval-augmented generation, RAG, and that allows these models to use the most updated information in their output, and it could actually use specific databases to produce the response to inform your decisions. So, if you're running a business, for example, it could use specific information, even proprietary information for your particular company so that it's contextualized before it's applied. And that's immensely valuable. 

I think this is going to be more embedded technology. For example, it'll be in our electronic medical record prompting us to make the right decisions. It'll be in the wearable technologies that we have because we're constantly supplying data just by wearing a smart watch, for example, having a cell phone on our body. It's something that will be truly exponential in its improvement.

Glen Stevens, DO, PhD: So, John, final takeaways you want to leave with the audience?

John Morren, MD: I think it's important that if you haven't kind of embraced the AI revolution that you should. It might start with small experiments. Go on there and use some of those free AI tools, like those large language models so you could understand the power, but also understand the responsibility that needs to accompany it. You should be a change agent and find more use cases for your specific field and subspecialty like I'm doing for neuromuscular and EMG, for example. 

I see AI as a general purpose tool. I think Andrew Ng says it best, it's kind of like electricity. If you ask somebody what is electricity good for? Where do you start? And that's what we should think about AI in terms of its applicability. Just find more and more very useful ways to use it. 

And then you should follow best practice guidelines. As you know, I'm part of that subcommittee with the AANEM, putting out position statements on the responsible use of AI. I think there are a lot out there, even by the World Health Organization. There are some stipulations about how it should be applied to healthcare. So, we just need to be aware of the guardrails. We're a little bit behind because the technology is moving so fast, but the regulatory aspect is catching up and we need to be aware of it so that we are safe, especially as it pertains to the interest of patients.

Glen Stevens, DO, PhD: Well, John, I'm excited to hear that I'm not a Luddite, and I appreciate this fascinating conversation. It's been educational to me. It's something that I think that, as you mentioned, that we need to embrace because it's here and we need to understand it, and really looking forward to how the field continues to evolve. Thanks very much for joining us.

John Morren, MD: Thanks so much for having me.

Conclusion: This concludes this episode of Neuro Pathways. You can find additional podcast episodes on our website, clevelandclinic.org/neuropodcast, or subscribe to the podcast on iTunes, Google Play, Spotify, or wherever you get your podcasts. And don't forget, you can access real-time updates from experts in Cleveland Clinic's Neurological Institute on our Consult QD website. That's @CleClinicMD, all one word. And thank you for listening.