Dr. Chris Tucker:

Welcome to the Arthroscopy Journal Podcast. I'm Dr. Chris Tucker from the Walter Reed National Military Medical Center and the podcast's founding editor. Today on the podcast, we are discussing artificial intelligence and machine learning. I'm excited and honored to be joined for the discussion by a thought leader and a pioneer in this field, Dr. Prem Ramkumar from Long Beach Lakewood Orthopedic Institute in Long Beach, California.

Dr. Ramkumar has published extensively on this topic, and we will touch on concepts presented in multiple of his articles with an emphasis on several recent publications in the Arthroscopy Journal, including his original research paper titled Clinical and Research Medical Applications of Artificial Intelligence from the May 2021 issue, and also his editorial commentary titled Machine Learning Is Just a Statistical Technique, Not a Mystical Methodology or Peer Review Panacea recently published in the March 2023 issue. Prem, congrats on your work and welcome to the podcast.

Dr. Prem Ramkumar:

Chris, thank you so much for having me on the podcast. I'm really excited to be here and share what little I know about this to hopefully make it more understandable and more potable for our audience.

Dr. Chris Tucker:

Well, the little you know is exponentially more than most of us. So let's get right into it. We all know artificial intelligence and machine learning are extremely hot topics. They're poised to completely revolutionize and transform medicine. But just the mention of those terms, I think, conjures up some different ideas and different emotions for people, likely originating from a wide variety of perspectives and also different levels of understanding of just what they are and what they can do.

I'd like to first have you help get us all on kind of a fairly level playing field by discussing some broad concepts and principles, and then diving into the specifics of your papers and some more focused findings as far as the pearls and pitfalls and the applications of AI. So to start us off, can you just define for us what exactly is artificial intelligence, and how does machine learning derive from that? Just how does this whole thing work?

Dr. Prem Ramkumar:

For sure. So artificial intelligence, also known as AI, is basically an umbrella term. And to dumb things down, because I like keeping things super simple, it's basically using the help of a computer to automate your tasks. And there's various types of AI-related subfields such as machine learning. Now, even regression modeling, which we're all used to in all the papers that we read for the last 40 years, is still considered a subset of machine learning, which therefore means it's a subset of AI, but it's not as scary or intimidating as it is now.

But artificial intelligence right now, from a historical perspective, is not a new concept. It was coined at Dartmouth, I think, at least 35, 40 years ago, where basically a professor postulated that we're going to get to a point where computers can automate a lot of tasks for us that are redundant and we don't necessarily need to all be doing manually.

Now, why is this exploding right now? Well, there's really two reasons. In order for AI to actually be implemented and be successful, you need, one, a lot of data, and two, you need really, really fast and accessible computer processors. Both things in the last five to 10 years have occurred, especially in medicine. We've been aggregating data from the patient-reported outcomes as we've digitized our medical records. We've put them into the Word documents, PDFs, EMRs, various formats of digitization.

So now, as you'll notice, in the last 10, 15 years, that was a big shift in medicine, "Oh my God, we're going from paper charts all the way to the digitization of it." And once it's digitized, it's fungible, it's accessible, it's analyzable. So that's the first thing, first domino. Right? And the second thing is that there's something called Moore's Law. And I might butcher it, but it basically is the concept that computer processor speed will double every year. And that's basically shown right now up until 2023 that our computer processors are really fast.

It's kind of the reason why every year, Apple gouges our money for another grand for the newest iPhone, because things are just faster. They're just better. They're just a little bit sharper. And it seems trivial to the average user, but for people who compute at the highest level, these two things converging is why we are here where we are today. Now, AI, from a broad perspective, let's dumb it down to two separate areas. There's generative AI and non-generative AI.

Generative AI is using AI to create things. So most people kind of got freaked out maybe seven or eight years ago, maybe even earlier, if I recall, when Gmail was trying to finish your sentence in an email or was trying to finish what you were going to type. That is a subset of AI-related work called natural language processing, where it's proposing a generated text for you. So it's letting you decide whether or not you want to accept that word, that phrase, that finishing of the sentence. And because you are choosing whether or not you want to accept it or not, we refer to that as supervised. You are supervising that decision.

That obviously means that there's such a thing as unsupervised models. We'll get into examples of that a little bit later, but we're still in the generative category. So because it's producing information, it's reading how you type, it's reading how the message is going, and it's creating text, that's considered generative.

Now, non-generative AI, I want you to basically think of that as holding a mirror up to your database. So any research paper we've ever read that says, "Smoking isn't a great risk factor for any surgery in orthopedics, rotator cuff surgery, arthroplasty," whatever it is, that's usually the last figure or last table in a regression model that says, "These are the risk factors that predispose failure."

Now, what machine learning or non-generative AI does is it basically tells you what that data set has already told you, but instead of telling it to you after the fact, it's basically moving it ahead of the question to say, "If you are a smoker and if you are morbidly obese, your chance of a successful outcome is this." So it's basically helping you predict what you already know is a risk factor, but not looking retrospectively but looking prospectively. So it's helping you make a shared decision with that.

So why do we care? I think we're going to head towards a generative AI topic a little bit later, but from a non-generative subject, the way I see it, the most valuable aspect of the non-generative AI space is it really is a communication tool that helps us convey to administrators, to insurers exactly what we're doing. What you and I both know is going to actually go down to what this patient's successful likelihood of success is after a proposed operation.

Dr. Chris Tucker:

Yeah. I think that's a fantastic intro. The terms generative and non-generative, I don't think I had seen those before in many orthopedic publications, certainly not in the articles we reviewed for this talk. So I think that certainly helps divide it up nicely for me, like you said, from kind of a large, 30,000-foot view, so to speak.

So at least when I read your studies and your reflective editorial summaries and such, I kind of boiled it down in the broadest sense to two large arenas in which AI and machine learning could potentially be applied in our healthcare field. First, clinical practice, and then second, research applications. I don't know if that's a way that you also think about breaking it down, but that's kind of what I [inaudible 00:08:25] on from application-wise.

Dr. Prem Ramkumar:

Totally. I mean, I have definitely evolved on how I think about this subject. The first study I ever wrote with my coauthors was basically trying to do what I talked about earlier, communicate the difficulty of an operation that everyone in the room that is holding the scalpel or involved in the surgery knows is not going to necessarily have a great outcome, but it's the right thing to do.

So, for example, in residency, we treated a lot of periprosthetic joint infections or failures after arthroplasty. We don't really have much choice, but we do it and we all know that the outcome is not going to be as great, and we know it's a very expensive endeavor, but we want to be able to communicate to administrators, to insurers, these surgeries should be compensated a little bit more. And we know this going in, it's shown in the data, but we're not able to basically reflect that.

And so, what we had initially created was a patient-specific payment model. And again, I'm arthroplasty as well as sports medicine-trained, so my head is a little bit in both spaces. As we all know, I know Arthroscopy has published papers about the devaluation of cuff repair surgery, but also arthroplasty, hip and knee is really taking a big hit year after year. And so, that's also front of mind for me, but specifically being able to communicate the value of our work before the incision, because we are in the business of elective surgery, for the most part, in arthroplasty and sports medicine spaces.

So the way I see is machine learning really flips that question to let all stakeholders be aware of the fact that, "Hey, it's going to require more resources. The outcome isn't going to be done, but we're the only ones that are going to be able to handle this operation. I want you to know this going in, not just suit, but also patient, but also everyone else in the room."

Dr. Chris Tucker:

I think diving a little deeper into that clinical realm which we're covering, I kind of surmised three general areas that it is applicable, and you have just nicely described what you labeled as value-based metrics, so communicating the value of potential care rendered to the patient for surgery. You also referenced application for actual clinical practice. So you talked about ideally removing that time-intensive administrative burden and then also telehealth applications. Can you just touch on those two?

Dr. Prem Ramkumar:

Yeah, for sure. So telehealth applications, it's kind of basically along the lines of being able to detect any kind of physical exam maneuvers, what's normal, what's not normal, and you can use what's called computer vision. It's like a cousin of AI in a way, where you could see what the camera is looking at and basically detect, "Oh, they don't have much abduction. That could be a problem. We should maybe triage this to a specialist instead of just having the first-line primary care physical therapist be the one addressing this."

And so, it's elements of recognizing, again, fungible data, anything that's digitized and being able to make sense of it. And that's an example of one way we can do that. Really, where I've really evolved is... The big issue is that we cannot be replacing physician tasks with AI, because we have so much expertise. There's so much nuance we've accumulated. And being able to convert what we've learned in residency and fellowship to something that's digitized, that's not a worthwhile endeavor.

And from a self-preservation standpoint, it's not the wisest move. But honestly, it's hard to recapitulate into ones and zeros what we do. But what we should do from that telemedicine standpoint is really be able to help nonexperts check the boxes to triage that to an expert.

Dr. Chris Tucker:

Yeah. I agree. I mean, you documented in one of your papers that the administrative burden currently is nearing 40% of healthcare costs, and it's really taking surgeons away from focusing on these patient-facing activities. So I think this is where AI could be a force multiplier, not a substitute for clinical judgment or physician-patient interactions.

So I think hopefully that alleviates some of the fears and apprehensions that folks not familiar with AI may have, which I suspect some do. I mean, I probably had a little bit of apprehension too before I was more familiar with it. And then the telehealth applications, like you said, anything that can improve patient engagement, compliance rates, remote monitoring, I mean, all of that is kind of a world that can be tapped into, I think, with this technology.

Dr. Prem Ramkumar:

We just need to use everything we can to make medicine much more human. Eric Topol has an amazing book on this, on how AI can help restore our humanity in medicine, so we're not facing our backs to patients when we're talking to them and documenting or we're not getting caught behind. And basically, we just need to be able to leverage all these tasks for these burdens, these things that are not up to the level of our expertise, and where others have basically hired someone to do that job for us.

There's a reason we're struggling. It's because we have such a disorganized, heterogeneous insurer system. And I think I read a stat that says there's eight FTEs per physician. And obviously, I don't want to take anyone out of a job, but a lot of these tasks can be automated. We can really decrease administrative burden surrounding not just doctors, but also the insurance companies.

Dr. Chris Tucker:

Yeah. So after reviewing nicely some of those clinical applications, which, I admit talking to you now, gets me even more excited about how potentially powerful and beneficial they can be, I did want to discuss some of the research applications for AI. Can you just talk to us about the handling of big data and then some actual examples of application in the literature on topics like athlete injury prevention, imaging interpretation, and patient-reported outcomes are all things I've seen in your studies?

Dr. Prem Ramkumar:

Yeah, for sure. I mean, every aspect of not just medicine, but sports health, performance is all moving towards analytics. And I think we should also get with the times and say, "A patient with this preexisting condition when they show up to training camp or before they start up their service in the military is predisposed to XYZ. Their runaway risk is this."

And again, these are just probabilities, so everyone's on the same page about what we think about a certain patient before we operate on them or before we treat them. Things that are known, we can spell out into a numerical form that gets everyone on the same page and swarm to that patient when needed about it. It's, again, the similar concept of triaging the appropriate person. And honestly, I know it's a very in-vogue thing, but I can't think of anything much more equitable than being able to have a patient-specific approach to every single medical condition, and that's from an analytics perspective at least on how we handle that.

Dr. Chris Tucker:

I just wanted to get a little better explanation of how AI and/or machine learning is actually analyzing big data and what kind of predictive models and such are being generated in what applications. Is it athlete injury prevention? I've seen some of your work on MRI interpretations and such. I mean, how does that all happen?

Dr. Prem Ramkumar:

Gotcha. Okay. So there's a lot of great authors out there that are doing amazing stuff. I think Kyle Kunze is one of those people that's basically done a really nice job of taking existing data sets and, again, holding up that mirror to what our data looks like but flipping the question. And that again is more under the machine learning topic to show us what we know.

And I can understand how some people may be frustrated by saying, "Yes, this isn't like changing what we knew, but it's a very valuable predictive model," because again, administrators and others need to know what the combination of different data points looks like into a risk calculator. So one study that we were able to do, I think it was also with Ben Nwachukwu and Riley Williams, and I think Kyle was on one of them, but similar theme, is where we took multiple different data points from MRI findings of bone marrow edema, around cartilage lesions, and we combined it with SF mental health scores and we combined it with their meniscus status evaluated intraoperatively. And these are all representative examples.

Again, there were, I think, 36 variables that were going into this, but as you can imagine, there are multiple different variables that... I know, Chris, you think about when we all evaluate patients in a clinic. We don't even realize we're doing this. We're taking so many different parameters into our brain. But we've essentially combined this into a model to come up with a risk calculation, and then that's something... I'm not the first or the last one necessarily to be doing it, but many other authors have done a really fine job at condensing all of what we do from a clinical standpoint.

But again, the fear I have is, why are we making it easier to do or to basically approximate our job, to approximate our reasoning? It's good up to a certain point of communicating the risk of a surgery and the resources required for a surgery, but I don't think it's good to say, "Hey, if you plug in these variables, we're going to get to basically a point where we can predict exactly how things are going to go down." And that could be a problem for multiple reasons.

One, if it's accepted too quickly, people are going to think, "Oh, we can actually replace doctors," which, if you know, obviously that's dumb, but at that point, we're the only ones that's saying that's dumb and that makes us look just self-preservational. But two, if you actually look at the models, they're only reaching accuracy of anywhere between 75 to 81%, which, by most standards, is fair to good.

And for a patient, obviously, you don't want to be 75 to 81% right when it comes to diagnosing and treating and managing a patient. That's obviously unacceptable for us and for any individual, I would think. So the big issue that we need to really focus on is, again, using the model to automate nonclinical tasks. And so, I'll ask you, Chris, when you are interviewing someone for either your fellowship or your office, do you prefer to have an in-person interview?

Dr. Chris Tucker:

I do. Yeah.

Dr. Prem Ramkumar:

You do? I think most people, post-pandemic, do. And why do you think that is?

Dr. Chris Tucker:

I mean, I think it's the same principle as when you're evaluating a patient in the exam room.

Dr. Prem Ramkumar:

Yup.

Dr. Chris Tucker:

I think there are so many intangibles that you can pick up on that are hard to describe. But like you said, you're gathering data points so rapidly and our brains are processing it faster than we really even recognize. You can get a feel for personality and motivation and other intangibles that you just can't pick up if you're not in person.

Dr. Prem Ramkumar:

Absolutely, Chris. And I don't think we will ever get to a point where we can truly digitize that. So there's a reason why we still interview people in person, and that reason is because not everything that can be counted counts and not everything... I've screwed up that quote, but not everything that counts can be counted and not everything that can be counted counts.

And the reason is, we're not going to get close enough to say how good a patient is because we have to actually look at them and see, do they seem like a compliant patient? Do they seem like they understand what it's like to take care of a post-op wound, what it's like to have limited restrictions? It's all the things you just talked about, Chris, because there's a reason we, at the end of the day, go down to the in-person visit.

It's because it gives us a lot more information that we'll never be able to codify into an Excel spreadsheet and we'll never be able to spit into an algorithm. So as long as we have that patient-facing element, we're going to be fine. And so, we shouldn't even bother wasting time trying to make up any effort trying to get 75% of the picture, because we're missing that critical 20 to 25%.

Dr. Chris Tucker:

So I think that's a-

Dr. Prem Ramkumar:

Yeah. Sorry, go ahead.

Dr. Chris Tucker:

No. I was going to say I think that's a great segue into my next question, which you nicely discussed some of the limitations of AI already. But I've also read your commentary about the black box phenomenon, and I was hoping you could just describe for us what that is and potentially what that can mean for those using AI and the dangers of using something that I potentially don't understand how it works.

Dr. Prem Ramkumar:

Yeah. I think that's a good point. I think that was something from a non-generative AI, machine learning perspective a lot of people exaggerated a bit, because when you're putting things into a model, you're not seeing the mechanics of how it works and you're getting an output and you, for the most part, agree with it. But if you try to peek under the hood, oftentimes we say, "Oh, you can't. It's just a model. It is what it is." But that's not necessarily true. There's a lot of different things you could use.

One thing is called Shapley plots, where it's almost like looking at what the model in a snapshot in time evaluated. And so, it is still telling you what elements stood out to that model in a period of time. So I still think you have a pretty good peek under the hood. And then when we looked at an implant identifier with my colleagues, shout-out to Mac Karnuta and Sergio Navarro, they were able to look at these implants that we were trying to identify. They were about to undergo revision hip or knee replacement.

And we were able to ask the model, "What area of the X-ray are you looking at that's making you say that this is a Stryker or a Zimmer implant?" And it will focus on a collar, the shape of the stem, various metrics. So I'm not a huge worrier about the black box phenomenon. I get it, but it doesn't take a genius to figure out why a model isn't reflecting reality. It's either your data doesn't reflect reality or you don't know how to peek under the hood in an appropriate way that gets you at the answer.

Dr. Chris Tucker:

So for those researchers listening who are interested in either using this more or getting involved in it, I have read some critiques of AI, is the phenomenon of basically garbage in is garbage out. So the quality of the output and the conclusions aren't dependent only on using these algorithms, but also rely heavily on the input data itself. So could you just explain for us what constitutes a quality data set or, maybe the flip of that, what are the pitfalls leading to low-quality data sets that researchers should avoid?

Dr. Prem Ramkumar:

Yeah. This is kind of what I was getting at with the non-generative machine learning stuff. It's taking a mirror at your data set. It's just telling you how good your data is. And I don't know, and I wouldn't claim to be an expert of what a great data set is, but obviously you want to have... Depends on the question. If you are looking at 500 osteotomy surgeries and you're trying to report the follow-up of it, it's not a great thing if you have 75% loss to follow-up. That's a basic element.

And smarter minds and me, statistical-oriented people, will say there's a certain threshold by which loss to follow-up is or isn't acceptable in my experience, I think. Depending on what time point, 20% can or cannot be acceptable to some. And so, yes, I completely agree with the garbage in, garbage out phenomenon. You're only as good as your data. So everyone's freaking out or worrying about the ethics of AI. I totally get it when it comes to generative AI, and I'm excited to talk about that. But the non-generative AI, honestly just, it's so fatiguing to me.

And Arthroscopy was kind enough to also publish a paper that I wrote with Karnuta and also Dr. Riley Williams about how there's meaningless applications being used for machine learning, where you see a lot of people... I don't know what it is. They either are getting things past reviewers who aren't acclimated or they're not understanding exactly the breadth of what they're doing, or maybe I understand that it's a publish-or-perish world for some and everyone's trying to get their PubMed citations up.

But right now, what is very troubling to me is you're seeing people publish data using data sets that have already been harvested hundreds of times that are basically coming out with a risk model for everything. And that data was just as poor as it was when you published it 10 years ago, except you're attaching the word "machine learning" to it. And again, like I said, it's just moving the answer to the front of the question, and it's basically...

And I have a huge problem with that because, one, you didn't need machine learning to tell me smoking was bad for a rotator cuff repair. Two, did that model give you any unique insights at all? I would like to think a study we published that was a reflection of the data set. It was a unique finding that the biggest factor between MRI findings of bone edema, meniscus status, age of the patient, sex of the patient, I found it illuminating that the biggest thing was the preoperative mental health component of a single patient-reported outcome.

And that was the biggest thing, because you should really be spending time talking to these patients, these cartilage patients from a social perspective to understand what they have going on in their life. That's the biggest thing, and people just kind of get transfixed on the imaging. And Dr. Williams has really led the way on that work using his database. But I think I'm kind of digressing here. But really, I'm very troubled by the fact that a lot of people are publishing machine learning research, and it's not moving the needle. It's taking an existing database, and they're getting a lot past reviewers.

Dr. Chris Tucker:

I appreciate that kind of in-vogue-ness, if that's a word, of AI and machine learning right now. They're nice buzzwords to kind of get a publication highlighted more than it may have been otherwise if you were using regression analysis. Nobody wants to hear about, like you said, the same large database that came up with the same conclusions. But if you attach something new and fancy to it, it gets some traction.

But also a concern, I think, is the external validity of these predictive models, because it is really just modeling off of data that's collected potentially at a single institution and potentially by a single surgeon. And so, when you're looking at these models, which, as you said, even the most accurate of them is only getting 75, 80% predictive, the external validity of that potentially open-source available algorithm, if I take that model and I plug it into my patients and I input all the data and it kicks out a number of, A, 69% chance of success with this hip scope, I don't know how valid that is, because it's not my practice that generated that model in the first place.

Dr. Prem Ramkumar:

Totally. But what are you essentially saying, Chris? You're basically saying that we want external validity of a predictive model, but it would make much more sense if we went upstream and we focused on the data. And so, those data-in and data-out people, if you just had built that model off of a multicenter database, that would be a lot better. That would be a lot more helpful.

But going around externally validating with multiple other institutions, that's valuable for some, not valuable for others. Depends on the question. For outcomes, totally, it could be valuable, if you're trying to unleash it into the world and say, "This is great." But I'll tell you an example. In The Journal of Arthroplasty, we essentially had an implant identifier that was based off of photos of X-rays. Do I need to externally validate that? Absolutely not. But the reviewers still asked me to do it, so I did it, and it was dumb because that's what I was told to do.

But anyone who understands what this field is, they really know that you don't need to externally validate a photo of an implant. So it depends on the question. But what you're saying with outcomes, you totally need to validate and validate externally, but the ideal move is to do what MOON and MARS has been doing for years, where they basically have been aggregating multicenter, very high-quality data. And so, it's not that exciting to me to talk about machine learning from that perspective, because I think I've said this before, but it's like changing...

Imagine changing every title that says Machine Learning Predicts XYZ. Do we ever publish a paper that says, "P-value shows statistically different outcomes between these two groups"? No. It's literally a statistical model. So it's a reflection of your data. And so, from a machine learning, non-generative standpoint, I'm not that excited about that.

Dr. Chris Tucker:

So let's try and talk about something maybe a little more exciting. You've mentioned a few times already that you wanted to dive into the generative AI potential. What do you see potentially on the horizon with that in our orthopedic practice?

Dr. Prem Ramkumar:

From a research standpoint, I'm horrified, because right now, I would give our readership of orthopedic surgeons on AI models... Appropriately so. We weren't trained in this. They didn't spend a year going to a computer science master's program like I did, or Cody has trained himself and taken other courses, and Kyle. They've all really just spent extra time to build this up. We shouldn't be experts in this. But if we're going to publish in this, we should have knowledgeable reviewers look at this content.

But what really makes me nervous is there's, like we've talked about ad nauseam, a lot of bad data sets out there. Obviously, administrative data sets, they've been reported to have 20 to 30% errors, and that's a huge problem. But again, we're working with the best that we have. Now, it's becoming slowly and slowly accepted in nonmedical spheres or sectors to generate data, to create fake data based off of existing trends. It's like forecasting. Right?

Now, imagine a research fellow who's trying to get into medical school. He or she wants to publish in the AI space because it's in vogue, like we've already talked about. And their attending, who they need a letter of recommendation from, drops this database that they've already had. It's a unique procedure, and there's a loss to follow-up rate of 40%. Unacceptable. It's unfortunately dropped down to poor level of research. Even journals aren't biting on this anymore.

What that research fellow could do very easily today is use a generative AI model to create dummy data that represents the rest of the 60% that was there, say AI was used to support the generation of this data, create multiple simulations, fill out the blanks, and then publish a model. Would it probably be that much different? Maybe, maybe not. Depends on the question, but it's horrifying because it's almost like indoctrinating and accepting fake news. That's how I view it.

It's basically permitting the concept of making stuff up. That's where we're headed. And how emblematic is it that this is where we're probably going to end up going in the research world? And I think that's what makes me really nervous about people not necessarily recognizing that that could occur, because no one actually goes through the lines of code. The number one thing that I always ask for whenever I review for any journal is I want to see the software, at least publicly available on GitHub or wherever repository they want to use, so we can actually make sure that it's not crazy. Nothing crazy is happening.

And so, for me, that's an example of generative AI in the research world. And to my knowledge, there's only one paper that's ever been published in arthroplasty and sports medicine, again, sorry, that's my bias, that's used generative AI, and it came out of Mayo. And I've given my buddy Cody Wyles a lot of grief about this, but he basically used X-rays that were essentially generated to help improve the model. And so, the foundation of that has augmented... And I don't really have a major problem with it, because again, the question is based on imaging data, and that's not necessarily changing. That's a pretty static, obvious variable.

But when it comes to patient-reported outcomes and reporting the fidelity of how well human beings are doing, not just recognizing what implant you need to call a vendor for to do a revision operation, that's a whole different ball game. And I'm super worried from a research perspective we're going to head there, and we're just going to miss it.

Dr. Chris Tucker:

Well, I think you've nicely highlighted a new nightmare for researchers that they may not have even thought they had, but now they do.

Dr. Prem Ramkumar:

Or a new dream for a whole bunch of computer science majors trying to pivot into med school, because no one's going to know.

Dr. Chris Tucker:

So you talked about that's one of your fears. I kind of wanted have a little fun here and ask you a little bit of an outside-the-box question. When you think about artificial intelligence, I think anybody in our age or older probably thinks about the Terminator movie series and Skynet and these machine-driven takeover apocalypse.

So, I mean, on a scale of one to 10, where we're talking about one being a minor statistical error that gets inadvertently published, or like you said, this generatively derived data that's not even real, so to speak, but it contributes to maybe a false conclusion. If that's at the low end of the scale and at the high end of the scale is this total Terminator-driven apocalypse, I mean, just how risky is the use or misuse of AI in medicine?

Dr. Prem Ramkumar:

Super risky, when it comes to generative. I don't know if you guys are aware of who Lil Miquela is. Have you heard that name?

Dr. Chris Tucker:

Nah.

Dr. Prem Ramkumar:

So it's a completely AI-generated social media influencer made by a company that markets this fake person, this fake persona online, and this fake account generates $10 million a year. Okay? $10 million based on a fake, AI-generated person that doesn't exist in this world. This person has 3.6 million followers on TikTok and 2.7 million on Instagram. And anytime this fake AI account highlights a CHANEL purse in their fake post, that generates about $10,000 per post.

So I think we are headed from a generative standpoint for a real moral crisis here, because we all know if you follow the money, that's where trends are headed. And if we start accepting this en masse socially accepted fake news situation... Because that's what AI is. We've already showed that fake news kind of bothers us, but we're not really clamping down on it. But now, with generative AI, we have the ability to produce it at mass scale. We are 100% going to lose touch with all reality. 100%. That's what makes me worried.

Can you imagine? Let's talk about a little bit sports-related stuff. We all know that after a certain quarterback for the Jets went down, there was a lot of marketing in the aftermath of what implant system was used for that celebrity to get his Achilles back. Right? Everyone saw the ads. Everyone saw that person go around. Now, imagine if another celebrity got injured and that celebrity athlete didn't get that specific implant, but another implant company decided to say, "This is a simulation of what this celebrity surgery would look like with our implants," and then was able to create the specific technique guide for that specific patient's likeness or anatomy and post it on Instagram.

And although it's fake, because again, with HIPAA, we can't really disclose whether it's true or not, we can't really tell what's real or not on social or the internet or anything like that, that company is going to make a lot of money based off this fake surgery that never occurred on this celebrity that no one can ever talk about whether that's true or not. So that, to me, is like a whole new level of technology enabling industry to just go crazy when it comes to direct-to-consumer marketing. And that's where we're headed.

Dr. Chris Tucker:

Yeah. I think to have left and right limits on our own practice and just community practice in general, I think it comes full circle to where we started our conversation talking about how we have to aim as a profession to not have AI be used in place of us or replace us, but to allow us to continue the patient-facing contact.

And, I mean, I think as long as we can continue to emphasize the importance of that personal interaction of us with our patient, I think the smart consumer will recognize that the rest of it is kind of window dressing or distractions and, like you said, direct-to-consumer marketing. But I mean, in the end, they're going to need a human to do their surgery that's interactive.

Dr. Prem Ramkumar:

Yeah. I'm not really worried about doctors getting replaced, because like we talked about earlier, any codified data, the value of an in-person interview, the value of all that stuff, even if someone tried to replace doctors, they truly could not, because they're not capturing a lot of the stuff that we're capturing. And I don't think even the best covert narc orthopedic surgeon who wants to go and feed a company all the elements that makes the perfect indication for a cuff repair or a SLAP, whatever it is, they're not going to be able to necessarily capture it that easily.

I don't see that being a real threat for replacing physicians, but I do see patients getting further and further confused about what's true and what's not. I think there's going to be a lot of confusion about how much a doctor actually makes. And again, back to the arthroplasty world, for total joint, the surgeon's actual fee is 6% of the cost. So right now, the narrative is such that patients think that we are greedy physicians that just profit off of the patients.

And if you have that kind of level of mistrust that's already sown, imagine if there's no checkrein on fake news and fake marketing and a lot of stuff. It's quite an enabling technology to exacerbating a class warfare issue that is going in medicine that no one wants to talk about.

Dr. Chris Tucker:

Well, like I say most times on my podcast, I think the best research often generates more questions than answers. So along those lines, what do you think is the next most important unanswered question in the field of AI that we should be addressing?

Dr. Prem Ramkumar:

Yeah. I think the most important thing that I've landed on right now, and this is a longtime entrepreneurship project that I had in mind. We started this company called Intelligent Health Analytics Inc. And it was really driven from empathy, believe it or not, because a lot of physicians get really mad at insurance companies who either authorize surgeries that probably don't have that much benefit, or they deny surgeries that really need to get done, or they put them in prior authorization land for weeks, and these patients don't have access to care.

And so, the thing that gets me excited and what I'm working on in my spare time is using these models to help insurers and help digital health partners get the analytics before they make a decision about a musculoskeletal event, like what does the data, what do experts, what do the outcomes show before you decide to rubber-stamp yes or no blindly with your one or two salaried individual working for your insurer that probably isn't up-to-date with the highest level of evidence? And being able to provide these people with at least an analytic model.

It's not perfect, but it's a lot better than where we're at today when you talk about how insurance plans are authorizing or not a lot of musculoskeletal surgeries and injections and therapy. So that's what gets me most excited, is being able to help these insurance companies to figure out how they're going to help us do our job in a more expeditious, more analytic, evidence-based manner.

Dr. Chris Tucker:

Well, I think that's a fantastic endeavor, and I think we've got the right man on the job. Prem, you've provided us some really nice, comprehensive, and insightful introduction to the world of AI, machine learning, and some applications for the field of medicine and specifically orthopedic surgery. Did you have any other closing remarks before we wrap up?

Dr. Prem Ramkumar:

No. I think it's just super important to reiterate that this was a huge team effort. I wouldn't have been able to get any of the aspirations I had off the ground back in 2018 without the support of Viktor Krebs as well as my friends and partners, Josh Woo, Jaret Karnuta, Sergio Navarro, a lot of these people I've worked with over the years, Riley Williams, Ben Nwachukwu. A lot of people helped support me with this.

And it was a long road because we got a lot of rejections in the beginning. People didn't really understand it, and now it's an opposite issue where you see people accepting it a little too quickly and not really understanding it. Yeah. I really hope and think that people who listen to this podcast will have understood how I think about it and what is important, and it's really critical to separate generative from non-generative AI, and then that is where the conversation should go afterwards.

Dr. Chris Tucker:

Prem, I want to congratulate you on all your entire body of work, and thank you for sharing your time and your thoughts with us today. I think you did a wonderful job of digesting this for us and distilling it down to something that we can all understand pretty clearly now, a fairly complicated topic.

Dr. Prem Ramkumar:

Thanks so much, Chris, and thanks to AANA and the whole organization for the support and letting me have the microphone.

Dr. Chris Tucker:

Dr. Ramkumar's original research paper titled Clinical and Research Medical Applications of Artificial Intelligence and his editorial commentary titled Machine Learning Is Just a Statistical Technique, Not a Mystical Methodology or Peer Review Panacea are available in the May 2021 and March 2023 issues, respectively, of the Arthroscopy Journal, which is available online at www.arthroscopyjournal.org.

This concludes this edition of the Arthroscopy Journal Podcast. The views expressed in this podcast do not necessarily represent the views of the Arthroscopy Association or the Arthroscopy Journal. Thank you for listening. Please join us again next time.

Medical Disclaimer:

The information and opinions discussed herein, including but not limited to text, graphics, images, and other material contained in this podcast and its referenced paper are for informational and educational purposes only. No material in this podcast or its referenced paper is intended to be a substitute for professional medical advice, diagnosis or treatment. Specifically, all content and information in this podcast and its referenced paper does not constitute medical advice. Always seek the advice of your physician and/or other qualified health care provider with any questions you may have regarding a medical condition or treatment and before undertaking a new health care regimen, and never disregard professional medical advice or delay in seeking it because of something you were exposed to from this podcast or its referenced paper. The information discussed in this podcast and its referenced paper may not apply to every individual and may cause harm.