Welcome to this AOSSM webinar titled Artificial Intelligence-Driven Clinical Efficiency, a Guide for Orthopedic Surgeons. Thank you for joining us. Here are the disclosures of the faculty and organizers for tonight's webinar and the learning objectives. CME credit is available for this activity. You'll receive further details at the end of the program. I am now pleased to introduce tonight's moderator and presenter, Dr. Katherine Koiner. Dr. Koiner is an Associate Professor of Orthopedic Surgery at UConn Health and the Director of the Women's Center for Motion and Performance. She is a board-certified orthopedic surgeon with a subspecialty certification in orthopedic sports medicine. Dr. Koiner serves as the Director of the UConn Sports Medicine Fellowship as well as the UConn Huskies team physician. I will now pass the program over to Dr. Koiner. Thanks very much. Hi, everyone. I'm definitely excited to be here and excited to moderate today's webinar on AI-Driven Clinical Efficiency, a Guide for Orthopedic Surgeons. Hope you guys enjoyed our hype videos that Liz and I did, but we definitely have a great lineup of presentations. We'll be covering everything from using AI to streamline patient education for preauthorization letters and even how AI can help writing those time-consuming letters of recommendation. We probably all come from different starting points regarding AI, and some of you are just beginning to explore how AI can help you with clinical workflows, while others may already have integrated it very thoroughly into your practice. Our goal today is really to offer you a few new ideas on how to implement AI, whether you're a beginner or an advanced user. For those of you that are already using AI to boost your efficiency in your clinic, please feel free to share your experiences in the chat or save them for the discussion and question and answer session at the end, because we would definitely love to hear from you and your insights and learn from you as well. Before we dive in, let me take a moment to introduce our expert speakers for today. Dr. Elizabeth Scott is a sports medicine and hip preservation surgeon at Duke University. She completed her residency at the University of Iowa and fellowship at Boston Children's Hospital before starting practice two years ago. She was also the recipient of ISHA's Traveling Hip Preservation Fellowship in 2023. She herself is an avid fencer and equestrian and enjoys caring for both Duke fencing as well as local fencing schools in central North Carolina. Dr. Josh Harris attended The Ohio State University for both medical school and residency. He then attended a sports medicine fellowship at Rush University and he joined Houston Methodist in 2013. He is an associate professor at Houston Methodist. He is also an associate editor for the Journal of Arthroscopy. He is director of the research for the orthopedic surgery residency at Houston Methodist. So let's start by handing things over to Dr. Harris, who will be giving us an overview of AI's impact in orthopedics and how to use it to develop website content and patient education. Dr. Harris, the floor is all yours. Great. Thank you, Kathy. I'm going to share my screen and can you guys see that okay? All right, perfect. And so I'm going to get started tonight. And first off, before I actually begin, I want to really thank Ashley and AOSSM, Kathy and Liz for the invitation to speak tonight. It's a pretty awesome topic. I'm excited to present on it. It's a unique topic. It's not just hip preservation surgery and CAMs and labral tears. So it's unique. So really looking forward tonight. And again, I just want to say thank you. And so I'm going to do the introduction and then also a little description on websites and patient education. And so I'll roll these two presentations into one for about the next 15 minutes. And so this one is AI and orthopedics and overview. Here are my disclosures, none of which are relevant to tonight's topic. And so at the end of 2020, for some reason that I'm actually still unaware of, I was asked to buy the arthroscopy journal to write an editorial on machine learning and AI in orthopedic surgery. And I had zero experience or expertise on this topic. I really knew very little about it. So I looked it up on Amazon and I read the first book that I could find by Eric Topol, where he contended that AI could make healthcare human again by taking care of the medicine part and allowing the doctor patient relationship to come back. So the book was really quite good. It piqued my interest to study it more. And so I decided to take an online six month course at MIT, which is largely considered the home of AI to learn more about it. I liked it so much. I took three more courses over the next two years. And so that two years total to learn more about machine learning and artificial intelligence, I was truly a novice then. And despite what I've learned since, there's just so much out there. I still feel like I'm a novice. I still feel like I'm learning stuff every day, every week. And a bit of the imposter syndrome certainly exists for me even speaking tonight during this webinar. And so even in just the past couple of months, since we've been setting this up, I've already learned so much from my co-presenters tonight, Drs. Kointer and Scott. And so this webinar hopefully will provide something new to everyone on the webinar. They'll find it valuable for their practices, including myself. And so prior to that editorial a couple of years ago, whenever I heard the phrase AI, this is what I thought of. So just movies of intelligent machines, not human, yet seemingly sentient. They weren't aligned with humanity. They attempted to take over their human creators and the world ended. And so although I would really love for the webinar tonight to be just a review of great movies over the years, the topic is AI and orthopedics. And so first, I think before you really get into the details, some basic definitions are necessary. The AI that we're talking about tonight, it's not AGI. AGI was the previous slide. That's the Skynet. That's how. AGI is a subset of AGI. That's artificial general intelligence. And AGI is actually the holy grail of AI to many. And it's basically doing everything that a human can do, thinking, learning, solving problems, adapting, reasoning like humans. And that's not what we're talking about tonight. There are hundreds of podcasts, videos, webinars that are already out there currently discussing if and when that moment actually happens, what we're going to do when that happens. Tonight is more for narrow AI. That's what we're talking about. It's simply a machine or a computer system that can perform very specific tasks using algorithms, data models. And most of us now have heard of and use chat GPT to variable degrees. And that's simply a large language model that uses natural language processing to answer questions and prompt. And so the AI topic, as you can see here, it's clearly popular. On PubMed over the years, it's just exponentially gone up in how much we're studying this. But it's not a new topic. In 1950, Alan Turing introduced the, quote unquote, Turing test. And that's just to see if a machine can convince an examiner that it's human. And then in 1956, six years later at the Dartmouth Conference, which was the birthplace of AI, John McCarthy actually coined the phrase AI or artificial intelligence. So this isn't new. The AI idea has actually been around for decades. And since its inception nearly 70 years ago and still today, AI safety is, I think, one of the most important issues to discuss. And in my intro talk here, I think safety is one of the things that I'll really emphasize the most. But again, it's not AGI. It's not how, like you can see here in this movie, but more pertinent to tonight's webinar, it's on patient safety. And so for safety, I think several issues need to be considered. Data privacy, not just for your patients, but also for yourself and the security of that data is incredibly important. It must comply with HIPAA. It must comply with insurance coverage and for employers. Also, I think one of the big things that you have to consider is misdiagnosis, inaccurate diagnosis, which founds treatment. And I think one of the things I'm going to go over the next several slides are AI hallucinations, where it gives you an answer. It gives you a good looking answer, but the answer is completely wrong. And if you use something like that, that's invalid, if it's not reliable, you can make some bad choices. And so the lack of transparency as well, I think, is a big thing. They come up with these answers, but we don't know really where the answers come from. And that's really founded from the statistics within machine learning, you know, the quote unquote black box, it may lead to poor treatment decisions. And I think eventually we'll get to a point in our clinical practice where we become very reliant upon this and hopefully not over reliant. Hopefully we still have human based clinical judgment where we use AI as a supplement and not a replacement. And also AI can be biased. These models are founded from the internet. They're founded from humans. And so the training data that we use, if it's bias data for age, sex, race, among several other things. And if that doesn't match up with your testing data, you can potentially make bad decisions. And so this is an example of the hallucinations that I was talking about. A couple of months ago, my hospital actually had some issues with my contract and how they were billing and coding a hamstring repair surgery that I do. And a hamstring repair code is CPT 27385. And if you actually ask chat GPT, how many WRVUs, how many work RVUs for a hamstring repair, it actually tells you that it's actually, it's a patellar tendon repair. And it tells you that it's basically 10 and a half WRVUs. And so I reply back, well, that's hamstring repair, not patellar tendon repair. And they're like, you're correct. I apologize for the confusion. And actually then describes hamstring repair. So had I just trusted that I would have made the wrong choice. And then also I'm like, well, 10 and a half RVUs doesn't sound right. And so it actually then tells me like, well, for 27385, it's actually 12 RVUs. Well, I'm like, well, that certainly isn't right either. So I consulted the codex from the academy and it's actually seven work RVUs. And I'm like, well, let's take this a little bit further. So that's a little bit concerning to me. So I did sciatic neuralysis, which I frequently will do with a hamstring repair. So it actually said that it's ulnar nerve transposition at the elbow. I'm like, well, that's not right. And I know it's not seven RVUs. So, oh, you're absolutely right again. And I apologize for the mistake. It actually says that it's now sciatic nerve and it says, oh, it's actually 10.2. I'm like, well, that's not right either. It's actually eight. I'm like, okay, well, I'm now concerned that chat GPT is just giving me any wrong answers that it's hallucinating. So then I do ischial brisectomy, which is 27062. And it says that it's removal of a hip prosthesis. And so it's clearly wrong. And it also states that it's 18 and a half RVUs. It's not, and I correct it. And then it says that it's nine. I'm like, well, that's also not right. It's actually almost six. And so if you follow those hallucinations, you can potentially make bad decisions. So you have to always trust your clinical judgment and have a little bit of that validity and reliability in the back of your mind when you're using chat BT for something like this. And so in your clinical care, in your clinical practice, it can hugely help with making the right diagnosis, personalizing and customizing your treatment to the patients that literally like right at your door. And the cool thing, and this was the reason why they actually asked me to write that editorial four years ago. It's to see if prediction of outcomes can be done face to face with a patient that's in your office. And this is a study from Rush from where I did my fellowship by Shane Noam. In that study, they actually came up with an equation where you use these eight variables to potentially predict the outcome of your patient that's in the office. And if you just substitute some numbers in there, you can actually see that you can predict the probability of achieving a clinically significant improvement. It's about 77 and a half percent. And each variable can tell you if it's for or against that actual prediction. And so it's crazy to think that just four years ago when this topic was introduced, where we've come since then, a lot has actually been done since. And beyond just prediction of outcomes, this is what we're going to talk about the rest of the hour with the other two presenters. But you can actually use AI to actually scribe your notes. You can do remote monitoring and wearables. You can write letters for your clinical practice. You can utilize social media and you can adjust your website, which is what I'm going to talk about actually in the next talk. And so not just clinical practice, but you can also use it for research too. You can come up with ideas, write grants. You can do your bio sketch. You can help with IRB, systematic reviews, meta-analyses. You can do lit reviews for projects. You can help writing manuscripts, but you can't write manuscripts. It can suggest journals. It can suggest institutional collaborations. And this is a paper that Brian Cole, also at Rush, wrote in arthroscopy about a year and a half ago. And this was a really cool paper. It's one page and you can actually see the text of the paper that's right there. But if you actually look at the highlighted yellow section, this is what it says. It says everything you've read up to this point was automatically generated by CHAT-GPT with the following input. And it's basically just write a page long editorial for the journal that provides insight on AI and CHAT-GPT on what it will do in the near future and how we need to handle this as authors, reviewers, and editors in journals and still maintain scientific ethics and quality. And so really cool idea, but you kind of, the punchline was at the end of the article and it said that he didn't actually write anything. And so because of that article, that journal among several other journals in orthopedics, they came up with declaration statements on how you're allowed to use AI in writing papers. And really, I think this is the way to go about doing it. You can't use it to write your whole paper. That's just wrong. But you can use it to improve readability and especially for English language translation and for some simple things that don't change the actual content of the science and the quality of the science. And so for education, you can use it as well. You can help prepare for tests. You can help teach. You can help write letters of recommendation. You can prepare for cases. You can actually study individual topics. And I think that one of the cool things was just a couple of years ago, we actually saw that CHAT-GPT could pass USMLE Step 1, Step 2, and Step 3. And Step 1 is actually largely regarded as one of the most challenging academic tests that exists on earth and actually passed the test. And so what an amazing achievement that a machine could do. And so AI still can't quite do everything. And so it's obviously, it's a new technology. You really need to trust the validity and the reliability. I've tried to do this for x-rays in the office for actually measuring alpha angles, lateral center edge angles for hip preservation. And it's not there yet. Also cost. I mean, most stuff that you do nowadays, you can actually accomplish largely for free or at a really low cost. CHAT-GPT for the actual subscription model is 20 bucks. But actually just 12 days ago, NVIDIA actually finally released their new hardware-based AI for just a little over a cool half million dollars for one. And the crazy thing with that AI is it was actually using training data of 72 petaflops. And that's a million gigabytes per petaflop. And that's 72 million gigabytes of training data. And that's just for the data itself. And it can actually process 64 terabytes per second. That is 64,000 gigabytes a second. It's just crazy to actually think about what these machines are actually doing. Hence, the reason open AI is now publicly valuated at about $157 billion. One of the most profitable companies that exists on earth. And so I think all of that being said, surgeons are really stubborn. We're resistant to change. And so for something like this, the best thing to do is just practice. Use it a little bit. Once you get a little bit of experience and expertise with it, you'll start to actually use it a little bit more and more. And the other ones beyond that of ChatGPT, which we'll talk about tonight, the things like the BARD, which is now Gemini, Anthropx Claude, Twitter, or X, which has Grok, which really uses social media, Microsoft's Copilot, Metasllama. They all have different unique skills, drawbacks, limitations, things that are good and bad. And I think just a little bit of practice, you'll learn what works best for you and your clinical practice with your research, with your education, and really find how you can integrate this into your practice. Because I think it's something you'll use it. I think you'll find value in it, and it can actually help a ton. And so in general, AI really has already transformed medicine. It's transformed orthopedic surgery, and it'll only do it further in the future. It'll improve our diagnosis. It'll personalize or customize our treatment. And ultimately, that's going to benefit our patients. It should improve outcomes. It should be used, though, really as a supplement and not a replacement for our clinical practice. And I think it's not just dozens. It's not just a few. There are hundreds of examples in how you can integrate this into your day-to-day workflow. And so that concludes the AI introduction talk. And what I'll do for the next few minutes is talk about how you can use AI to streamline your website content, so your personal website, and how you can personalize, customize the patient education for patients that you're actually seeing in the office. And so same disclosures as before. And what should a website really do? A lot of us have institutional websites, but many of us also have a personal website that is customized to just your practice. And so really, quite simply, it should clearly, concisely communicate what your expertise is and the services that you provide. It should be really easy to navigate on your desktop, on your phone. It should be like that for patients. It should be like that for referring physicians. You should be able to schedule appointments. You should be able to provide education for what it is you do and for any patients that you operate on for their post-op recovery. It could really cover your common procedures, your conditions, and you should be able to provide some reviews and testimonials on what you do. And so I think this is a really good example of how you can use ChatGPT and AI for your website. And so I asked it, what are five things that my website, joshuahairsmd.com, does well, and what does it actually do poorly? And so the five things that I found were really number one, and this was, I actually did this multiple times, is search engine optimization. It should target keywords. I'm a hip arthroscopist, and so it should really target hip arthroscopy Houston and the headings, the articles, titles, the actual metadata within the actual website and the blogs. This is how patients and other doctors will find you, and this is really important. Next is the actual appointment. Like if you want patients in the door, they've got to have an easy, efficient way to go about doing that. If it's cumbersome, if it's burdensome, they won't actually come see you. They'll go elsewhere. The frequently asked questions section, I think, is also really valuable. Patients may not use handouts. They may not use the FAQ, but when they do have it, it really improves the efficiency of not just you, but especially your office staff. The website should also be really visually appealing, and that front page, that's your first impression. That's where you get one chance to do that, and if it's too much, if it's too much to navigate, they'll go somewhere else. If it's not enough, they'll go somewhere else, and so you're really kind of seeing a little bit of a Goldilocks effect. The social media, I'm not a big social media guy, but patients are, and they will actually look at all the different social media sites that are out there. Having social media is actually really, really helpful. Those were five things. I'm like, well, I guess my website itself isn't all that great. I'm like, but is there anything else? They found four more things, and so I guess my website wasn't actually very good. The limited visual engagement, and you'll see on the next slide what I had. It wasn't great. I had limited visuals. I really only had three pictures. I had a football player, a runner, and a dancer, and there was actually no picture of me. I'm a 44-year-old orthopedic surgeon, but I look like I can barely even rent a car, so when patients actually show up in the office, they're oftentimes surprised to see what I look like and what my actual calendar age really is, and so it's always entertaining to have that conversation. Beyond that, the content depth actually really matters. A lot of patients come in, and they try to be really educated, and so if you have high-quality education available to them, it makes your job a ton easier. Load speed also matters, and so if your website's actually really slow, after they try to go through a couple of different clicks, if it's really slow to load, they're not going to keep looking. They're going to actually find somewhere else, or they'll close it down, and that's a less chance they'll actually get in your door. And then a call to action, a declarative call to action, actually really does matter quite a bit, so having something that actually says, come here for an appointment or book an appointment here is actually different from just appointment, and so this was my actual website, and this was the actual front page. It was literally just my name, a phone number, book online, and a picture of a football player, so it's not that great. It's not horrible, but it clearly could be better. For the mobile version, this is what you see if you do it on your phone, and so more than half of patients will actually use their phone and not an actual desktop to actually come find their doctor. So there's no tabs at the top and so even that could actually be better. And this is what I think better actually looks like. This was the best doctor website that I could find. This was my fellowship mentor from 11 years ago. This is Brian Cole's website. That's a QR code where you can actually look at his website. This is what it looks like on the front screen. So it's got his name, it's got a cell phone number, 3,000 plus patient reviews, 24,000 surgeries, 25 years of experience, 2,000 lectures, 1,000 articles. And so his expertise is like screaming at you from just this actual picture of him, which is him giving a TED Talk. And so it's actually pretty amazing. If you actually watch the video of just what it would look like if you're actually going through his website, it's not just him, it shows his institution, it shows what he's doing, some of the patients, who he's worked with before in the past, some of the news media, what he does for his podcast. If you scroll through, you see testimonials, you'll see his expertise, where he is, what his credentials are, what services he'll provide, a supplement that he does, a podcast that he does, his YouTube videos. And so obviously from all of that, there's only one Brian Cole out there and not everyone has all of those credentials behind their name. But really advertise, show what you have to offer for patients. And if you can actually show these things from a visually appealing website, you'll get a lot more patients in the door. So I think your website really does matter. And so to kind of switch to the last topic, this is for patient education, and specifically it's personalized patient education. And so I have a few different patient education handouts that I give patients, and these are generic things. And so the two most common ones that I'll typically provide patients are the one you see on the left and the right here. It's hip impingement and hamstring tear. And this is actually something that I give to pretty much every patient that comes in for either hip impingement or a hamstring tear, they get this. But if you actually take that and you upload the PDF of that document to specific patients. And so let's say you've got John Doe who comes in your door, who's had a year of hip pain, and that hip pain has prevented him from playing football at the Ohio State University. And unfortunately, we just lost a few nights ago. So that's a little embarrassing. But despite PT, despite Celebrex, despite some injections in his hip, he has cam impingement and a labral tear, but no arthritis. So what can you do if you upload that and personalize it to him? Well, it can literally give John Doe his entire clinical scenario, nonsurgical versus surgical treatment, and what he should really do in this situation. So this actually can customize to him in the door, you can load this up on your computer actually in the room. And if you look at what the goal is, he's like, well, I want to play football, and also prevent further damage down the road. And since he's tried all of the nonsurgical treatments, then surgical treatment might be the right answer. If you do the same thing with hamstring tear, you upload that document. And let's say I'm in Houston, Texas. And so let's say you've got a 50 year old male runner who was bull riding at the Houston rodeo, and he has a full thickness hamstring tear with a little bit of retraction, he loves running. And based on the patient education handout, should he have surgery or not? So it summarizes his clinical presentation, what nonsurgical versus surgical treatment would provide. And it generally recommends for his tear for what he would like to do, it generally recommends surgery. And so I think that's a really cool, personalized, customized approach to clinical care. And so but if you just do a generic one, so let's say, can you give me a one page handout on hip arthritis and its impact on hip arthroscopy, a common clinical scenario that we see in hip preservation, it actually tells the patient, it tells you what hip arthritis is, what hip arthroscopy can do, and what it may not be able to do. And if you look at the starred section right there, it says hip arthroscopy may not be able to address the broader issue of cartilage loss, making it less effective in hip arthroscopy cases. And so for patients that have this moderate to severe arthritis, hip arthroscopy is less likely to be effective in a hip replacement may be the better option. So that can hugely help your patients before they even get in the door. But even once they're in the door can help you guys make a shared decision making approach to optimize your clinical approach. And so I think in conclusion for your website, AI can really create a website or it can just optimize a user friendly experience with clear and concise information that communicates your expertise and the services that you provide on both desktop and mobile platforms. And it can also customize the education you give your patients by tailoring that information to the patient that's there face to face with you so that you can ultimately improve their outcomes and their satisfaction. So thanks again to AOSSM, Ashley, Kathy, Liz, I will now stop sharing my screen, and I'll turn it over to Liz for the next part of our webinar tonight. So thanks again. All right, thank you, Josh. That was awesome. Let me share my screen here. Perfect. Okay. So for the next 15 minutes, what we are going to talk about are some other very practical kind of hands on ways that AI can enhance your clinical practice. So what we're going to talk about are some of the different dictation service options, how they're sort of grouped into categories, ways you can think about them. So there are a lot of a growing number of different options for that. Talk about some of the kind of pearls and then potential pitfalls or problems you can run into with the AI in your clinical practice come in its current state. Some of its limitations. We'll also talk about how you can use AI in letter writing in your clinical practice. So things like letters for appeals and denials, my experience even just with op note dictation, and then some other personal ways that I use AI. Of course, this lecture will probably be out of date in some ways in another even three to six months, because there are a growing number of different services out there for AI, specifically for clinical dictation. So I literally a week ago, I googled AI for clinic note taking for physicians, and these all came up within the first page of Google. Some of them are more popular than others. They are again, all constantly changing all, you know, pushing to be to have the latest greatest service that's different and distinguished from the others. Certainly, we cannot go into any of them in detail. So instead, what I want to do is just talk about some of the overall differences and kind of concepts with them. So in general, most of these services, you can kind of divide them into one of three different things, the sort of lowest rung, the most basic way that you might use AI and clinical note taking is really just straight up transcription. So instead of a human, taking your oral dictation and turning it into a note, now it's AI doing it. So exactly what you say exactly what you give it is exactly what you get out. Most people kind of understand this, that makes the most sense. The next sort of step up is going to be AI that uses natural language processing. So it's taking your audio dictation, and instead of having a bunch of run on sentences or really rough draft, it's going to create more of a finished product by extracting the relevant information and then generating more of a structured note in the sort of format that you prefer. So it can actually identify entities like medications, symptoms, procedures, and in a sense is a little bit like transcription with more of a brain. So natural language processing relies on machine learning algorithms, linguistic rules, statistical models, again to sort of analyze and interpret your language. And then the one that most people get really excited about that's still kind of constantly improving is ambient listening technology. So that's where you're going to take the app on your phone or your iPad, you take it into the room with you, you talk to the patient, have a normal conversation with them, do your exam, and then the AI turns that into a structured note, taking all the dialogue and turning it into a clinical note without you having to do a lot of additional oral dictation. So that's a more sort of direct way to capture what the patient's saying, your discussion with them. Many services allow then you to come out of the room and then dictate a little bit more to state your assessment and plan and things like that. But the overall idea is that you're having to do less work afterwards and it's doing all of the legwork of coming up with a note just based on what happens in the room. So then there's some pros and cons, some similarities and differences, of course, then between a human scribe and an AI service. So many surgeons who've been around a while have human scribes, you're very used to how that works. They know that it involves a significant effort to train them, but they also tend to be very high quality once trained. Of course, a human scribe can get sick, they do it for a few years and then graduate and go to medical school or they retire. And they're, of course, limited by the fact that they can't be in two places at once. So if you and your mid-level provider or resident or fellow are all seeing patients at the same time in clinic, only one of you gets the human scribe, of course. And then there's a lag time sometimes between patients while they finish up the note. On the other hand, the AI service, the idea behind it is, yes, it can be cost-saving versus a human scribe. There is still training time associated with it, both for you figuring out what you need to sort of say to the AI to get the note you want, as well as training on the AI's end to collect data and get feedback on what you like and don't like in your notes. Of course, AI cannot graduate, retire, or get sick. It's usable by everyone at the same time. But as we'll talk about on the next slide, it is still somewhat limited by what I like to call garbage in, garbage out. So what that means is that there can be some issues with AI, particularly if you don't take the time to really understand what you say and then what comes out and figure out how to correct that to get what you're looking for in a note. You can have some issues depending on the service, integrating it with the sort of standard templates that you're sort of used to and the way we all sort of trained and practiced to have this very specific way that we want our notes done. You can also sometimes run into issues with, for instance, very specific authorization requirements for insurance. You need to have XYZ things in the note very clearly listed. And if you say transition into ambient listening technology but don't take the time to make sure those components are still in your note, then yep, you're going to run into a whole bunch of authorization issues until you get that kind of hammered out. Just like a human, they can mishear you as well. So they're not, they're looking at you in the room, they're listening to you in the room, which means arthroscopy can still sound a whole lot like arthroplasty and you can end up with some very odd mistakes in your notes. Similarly, often the way we talk to patients about diagnosis, about planning, about imaging is very different than the medical language that we usually want in our notes, which means if you're just relying on the ambient listening technology to create a note for you, typically it's, again, it's going to take the language you say in the room and that's what it's going to make the note from, which means that if you don't take the time to then clearly state what you want in terms of your read of the MRI or the assessment and plan, then again, you're going to get something that's a little bit off from what you're used to. And then as we all know, sometimes patients just say really weird things. They say sensitive things that we know intuitively, they just don't, you wouldn't actually want to put that in the note, but AI does not know that, it's not a human, and so you do need to really make sure you're scanning through sometimes the narrative portions of the notes to make sure that those sorts of sensitive things are taken out. So overall then it's really important to understand that AI is still not replacing your need to review notes carefully for errors. Just like you review a medical student's note or just like you review a learner's note or any other transcribed note, you still need to catch those errors and fix them, but it can save you time, particularly if you do not have a human transcribing for you and you're off doing your notes on your own, it can be a really big time saver. Oops. So other ways that AI can help you in clinic, I really like to talk about specifically uses of ChatGPT, largely because that's a free service, which means for those of us in large medical practices where it can take a long time to get new software approved, ChatGPT, you can use tomorrow. So here are some three different examples of how I've used ChatGPT in helping me speed up letter writing for clinical practice. I love using it for all kinds of appeals for insurance. For instance, here, this top one is a letter of medical necessity, had a patient good gluteus medius tendon repair denied after we'd already done her surgery. They covered part of the surgery, but not another section. I gave it two sentences and it produced a really nice letter that I added one or two extra things to and sent it off and it ended up getting approved. It's also really nice for physical therapy. A lot of us are used to maybe having batch templates for these really common things like PT denials, but it still takes a lot of time to go in and change all the details to make it personalized for the patient. AI is really, really good at doing that. So this was a second prompt as a patient that had prior surgery and was doing PT, but the number of PT visits ran out. I plugged in just a couple details of her diagnosis and that it's a her and that the PT said that they are, you know, she's doing well and should continue to have PT. And it wrote me again, a very nice personalized letter that we were able to send off to the insurance company. Of course, you need to be careful that you're not putting in a bunch of HIPAA details in there. I always use like a star star star for their name so that it's clear that so that I can then, you know, identify where those places are and later plug their actual name and information. And I usually do like to use the gender though, just so I don't have to change a bunch of pronouns. And then the last example is I even like it for writing school notes. Again, it's something I have a template for on the computer, but especially for those patients where they start to ask for a bunch of really specific things that you don't normally write in the note. You could have your nurse, you know, sit there for 15 minutes trying to write out some complex thing and make it sound good, but you put one sentence in the chat GPT and it will spit out a really, a really nice note. That's like very well written and very impressive. And, you know, literally takes you five seconds. So those are just three different ways that I love using chat GPT in my practice. I think it saves me and my staff a lot of time and energy. And again, kind of makes you sound really smart and look good when you have a really, really nice writing. Here's a, an example where maybe we're still, you know, working, work in progress is for op notes. This is not something necessarily we all need to do. You know, I got my dot templates. We use Epic where I am and I'm, you know, I don't have to change too much, but I decided to play around with chat GPT and see what it could make as far as a operative note for me. So I asked it for a operative report for a standard hip scope. I told it I do inter portal capsulotomy and I use these three portals and it wrote something that was pretty good. It was an op note. Indications were spot on. It got, you know, very similar sentences to exactly what I use in my own notes. Patients failed conservative treatment, physical therapy, injections, yada, yada. And the majority of the note was, you know, off to the right start. And then we got towards the end and I realized, wait a minute, it's made some very critical, very important mistakes. So for one, it, it didn't do a capsular repair to me that is below the standard of repair or standard of care for hip arthroscopy. Number two, I noticed it kept traction on the entire case, both intracompartmental and extracompartmental work. And the very last thing that was done at the end of the case, according to chat GPT is the hip joints taken out of traction. That's a pretty important thing that you would not want to have incorrect in your op note. Of course, then all it took was a couple of sentences for me to say, Hey, chat GPT capsulotomy, I repair that. Please put that in and then make sure that the traction comes off as soon as the intracompartmental work is done. And it wrote something that was, which much more in line with what I would actually use. So who knows, maybe there is maybe a role here for using chat GPT for op notes, at least to produce a template for something, something I might continue to play around with. But it's just something I wanted to put here to maybe stimulate some ideas in your brain again, about how you can use this in your practice. Overall, my piece of advice when it comes to chat GPT and clinic, again, I really think it shines when it comes to writing notes or writing something that you just don't know how to write. So it's a great service for things like administrative assistants who might be newer, who don't have a lot of experience or don't really know what you're looking for. And you say, you know, help me write a note for this patient, for this thing. I just tell them, Hey, you know, put it in chat GPT. You don't have to have any idea how to write it. Just, just start with something in chat GPT, let it give you some kind of draft and then work, work from there. Cause really, you know, often it's, it's just hard for us to stare at a blank screen and get started with a draft for something. And you ended up procrastinating and it's now a month later and you still haven't done it. And the patient's asking, you know, for this letter for the eighth time, and just, just try chat PT or another service to help you. And you'll be really surprised at how, how much quicker you can finish having something in front of you to work from. And then a last kind of fun thing is just how do I also use chat GPT in my own kind of life outside the operating room. I work with a lot of local fencing schools and I do some sports performance education and sports medicine education for them. A chat GPT is great at making things funny and making things have a interesting tone or a tone that's different than how you would normally write. So for instance, I've had it write me some funny things to help high schoolers stay engaged and learn some sports performance topics in a way that I certainly could not write on my own. It's also super great for, I have a little side gig that is some personal training for clients and it's awesome for rewriting workouts for me. If they are say going on vacation and have different equipment, I'll say, Hey, rewrite this workout, except take out all the machines and put in resistance bands or weights and it redoes the whole thing in two seconds and saves me a lot of time. So there are a lot of ways you can use it. If you get creative and start exploring. I also love chat GPT for travel. I'm going to Finland in January and it helped me pick out hotels, helped me come up with a nice little itinerary. I'll let you know in January how that goes, but again, it's just a nice way to get an idea going for something and modify as you go. And again, I think you'll be really pleased if you start to just experiment with ChatGPT, it can really increase the efficiency of your practice. And I really do believe is going to be more and more involved in our everyday life as surgeons outside the OR over the next few years. So that's what I have for you. We're going to turn it back to Kathy in a second here and keep going with our webinar. Thanks, Liz, that was great. Well, Liz and Josh, you did an amazing job getting everything sort of explained from that perspective. And now we're going to shift gears a little bit and we're going to dive into two areas where AI can make a huge difference in our daily professional lives. And that's writing letters of recommendation and crafting sort of the perfect professional email. And it's definitely all about the prompts. And I really think of prompts as having a conversation with a friend. You want to be specific, detailed when explaining what you need. A good approach is to provide the background examples or reference materials before asking for more complex questions. And I think of AI prompts like recipes with five main ingredients, task, instructions, context, parameters, and input. And for example, you know, there's a simple AI prompt that you may just need the task component and you put in and you get out what you want. But to really make your AI prompts more effective, you need to understand each of these components. But instead of a recipe, let's make a orthopedic analogy. And you can think of the prompt like a preference card we use in the OR. I don't know about you, but my preference cards are never right, but hopefully, you know, AI will help that. But so a task is really what you're asking AI to do. It's like telling a scrub tech that you're doing an ACL reconstruction versus a rotator cuff repair. The instructions are the specific directions you give to AI. It's like telling the circulating nurse that you do rotator cuff in the beach chair position versus the lateral position. And then the context, it provides the background information, ensuring that the AI's response is aligned with your expectations. You should tell your first assist that not only does this ACL, you know, is an ACL, but it has a root tear as well as a ramp tear. So you'll need the extra equipment. And then the parameters, these are the settings of your sort of toolkit. And this is, they dictate how the AI processes the information or the complexity of the responses. You need to specify the level of detail. You want to say you want a three-page response or you want a bulleted list. And then the input is the raw material that you provide. And this is the information data or the examples that AI use to generate its responses. And AI really makes letters of recommendation faster and easier. It ensures that they're personalized and they're high quality. I'm sure we all follow very similar processes when writing letters of recommendation, but now I'd advocate that adding AI at the drafting step. And I'll show you three ways that you can sort of implement AI to quickly generate high quality drafts based on the details you provide, saving you a ton of time. And then once the AI drafts the letter, you review it, make adjustments, and then submit it. So the first method is just like dipping your toe into the water. You're taking a small step to explore the possibilities with really not diving in too quickly. And this slide really outlines how AI can be used to refine your letters of recommendation after you've generated the draft. The purpose is to ensure that the letter is grammatically correct, it's clear, and it maintains a professional tone. And the process is really simple. You just submit your letter of recommendation draft into ChatGPT for review. You request the ChatGPT, check it for grammar, clarity, and tone. And then AI will highlight sort of similar to what Josh did for his website, the areas that need improvement, ensuring the final letters meet your highest standards and will save you valuable time. And we've all been on the other side reading a letter of recommendation and thinking, did they really just write that? You can almost hear HR sighing as every time they see this writer's letter's name pop up. So over the next three or four slides, we'll cover how AI can help detect and eliminate gender and racial bias in your letters so we can avoid those cringeworthy moments and more importantly, really create more equitable, unbiased recommendations for our students and trainees. While AI tools like this can assist you in creating and refining your letters, it's essential to remember that accountability, similar to the research that Josh talks about, lies with the person writing or submitting the letter. You're ultimately responsible for the content and accuracy of this letter. In this editorial, the authors highlight that technology should be used as supplementary tool and AI can support recognizing and managing the implicit bias, but it's not a replacement for human judgment and accountability in crafting professional recommendations because also similarly, as Josh mentioned, AI in and of itself, depending on the machine learning model, can be biased. This study analyzed 171 general surgery resident applicants and the results showed that AI and computer algorithms successfully identified patterns and gender bias even when factoring in clerkship grades and looking at letters from across different decades when they were written. And the conclusion is clear, AI can be an effective tool in detecting subtle bias, helping us make our letters recommendation both fair and objective. Here's a sample prompt you can use. Essentially, it's asking that AI to review a letter for any signs of gender or racial bias and offer suggestions for improvement. And Ashley has hopefully included in the chat that little PDF that I put together with just some sample prompts as well as examples of other AI responses. And here's an example of gender bias. Bias can often slip into descriptions of candidates affecting how their abilities are perceived. Using AI help catch these subtle differences and ensures fairness in the portrayal of both male and female applicants. Trust me, adding this one step to your letter of recommendation process will not only be a better letter writer, but you'll also probably heighten your awareness of your unconscious bias that can influence your recommendations. Also, I'll note that almost all of my images are created with ChatGPT as well. So at this point, now you've survived dipping your toe in and now it's time to go a little deeper. So we're moving beyond just editing your draft. In this step, you'll provide AI with specific context about the applicant and it helps you craft the entire letter of recommendation. And here's a sample prompt that you can find in the PDF. To start using AI for drafting the letters, you wanna give it to key details about the applicant. You wanna identify your relationship with the applicant, what position they're applying for, and their strengths like clinical leadership and their surgical skills. You'll be really impressed with how personalized these letter reads. It feels like they've been watching your guys' interaction over time, but most importantly, you'll save a ton of time. When drafting these letters, it's really helpful to provide clear details upfront. You wanna start with the tone that you want. Is it formal? Is it supportive? Is it neutral? And then you wanna decide the length. Maybe it's a page or maybe it's just a few paragraphs. You wanna identify your key focus areas like clinical skills or leadership and specify how strong you want the endorsement to be. You wanna include specific examples of the applicant's accomplishments and mention any relative personal qualities like professionalism or work ethic. Sometimes really, for me at least, the average letters are the hardest to write. And this was a prompt that I found useful. And it's basically stating write a general letter without overly emphasizing any outstanding qualities. And you can also find this prompt in the PDF. So now you've accomplished those two ways to write a letter. Now you're ready to dive in head first. So this is a memo from the UConn School of Medicine that I often receive when writing letter of support for a colleague that is up for promotion. I'm sure we all get these, whether it's from our own institution or elsewhere. And they really take a significant amount of time for me. I recognize how important they are both academically as well as potentially financially. So I spent a lot of time on these. And my typical process involves reviewing all of the candidates' supporting materials, matching them with the institution's promotions guidelines that you can see highlighted here on this document, and then adding my own insights and then drafting the letter. And I'm sure all of you follow a sort of similar process. So this is what I call a game changer. And I've just figured this out recently. So you open File Explorer where all of these documents are saved. You then highlight the candidate's CV, personal statement, promotion criteria, teaching responsibilities in our case, and publications. You select all those relative files. You then drag and drop them directly into the prompt box in Chet GPT. And this is what it will look like. And then you input a prompt similar to this on the screen and ask AI to review and organize the information, cross-reference it with the promotion criteria, and draft a comprehensive letter of support. This letter will blow your mind. And the time that you saved is amazing. So here really is how AI is working in the background. It first parses through all of the documents like the CV and personal statement, pulling out the key details. Then it identifies the important themes such as surgical expertise or leadership. And AI makes sure to focus on what's relative to the position that the candidate is applying for and avoids repeating the information as we sometimes do when we write letters. Lastly, it highlights unique qualities that make the candidate stand out. And all of this saves us a ton of time and you also don't miss anything important and you have met all the criteria for that particular institution. So now I think you guys are all ready for a little test drive of your own. But switching gears to how to write professional emails with AI, we've progressed a long way from the sort of AOL days. And AI really saves me time by drafting and reviewing emails. It improves my message by improving my clarity. So it's to the point. It even tailors the language to fit specific needs, whether it's a patient update or professional inquiries. And best of all, it really keeps the tone professional even when I'm tempted to send that passive aggressive email. So you can use AI to streamline a lot of your everyday communication. For example, patient follow-up emails, post-op instructions, or even collaborating with colleagues and research. It's also great for scheduling tasks, handling sort of administrative work and sending out professional responses to conference invites. Here are two examples that I'm sure many of us have dealt with. And the first is an email to the ER about an inappropriate consult. And the second is an email to anesthesia about turnover time. You guys can see those responses in the PDF. Similar to how do you approach a letter of recommendation, you really need to start by defining the purpose of the email. Then you want to provide the key details. You want to specify the tone that you want. You want to define how long the email wants to be. And finally, you review and edit the draft. And then you can also tell chat GPT to say, make it more formal or make it more collegial, make it longer, make it shorter. And then ultimately you can send that message in a lot shorter period of time. So here are a few examples of how I use AI to simplify my email response. This is one of those dreaded emails from GME. They're looking on how we teach patient handoffs to our fellows, what evaluation tools we use, and if the fellow is competent in sports medicine, there's not a lot of handoffs. So I quickly jotted down my thoughts in shorthand. I fed it into chat GPT and it generated this amazing well-structured, thought out, thorough response. I'm sure longer than anybody else responded to. I copied it, sent it off, and it took me no more than three to five minutes. So AI in my professional life, I've used all of these, created a template for a business plan. I create fellowship evaluations. I even created some epic dot phrases for risk and benefits of surgery or generic sort of templates. I've generated word clouds. I use it for progress note templates, award nominations, writing and editing bios, creating images, talking points for media interviews, even creating the outline for an emergency action plan PowerPoint for the fellows course that I had never presented on before. And every year as program directors, we're at task, we're sort of completing the annual program evaluation. And oftentimes our sports medicine fellowship doesn't change much. And this is where it can really, I figured this out just recently, and you simply enter the prompt and you can have it rewarded the content without changing any of the key information. So basically it's just changing your words to sound differently. So if someone's reviewing it, if you're not repeating the same thing year after year. I also can reuse it to reformat my CV if I'm looking at a specific publication or research grant. I've used it to edit the fellowship handbook to write a nurse navigator job description. And also I'm so not very creative at all. I use it to generate ideas. I asked it for 10 catchy titles when I was doing a talk on steps for ACL revision. And you can see all of these very unique titles here. I also use it in my personal life. I use it for party planning, creating a shopping list that's organized by aisles to make my shopping more efficient. I've used it, one of my favorites to create custom bedtime stories for my three-year-old. I've used it to appeal a parking ticket. I've used it for travel itinerary and meal prep. And at first I was really frustrated that it didn't include my signature, but if you pay the paid version of ChatGPT, you can go to personalization, you can toggle this button on, and then you can sort of put in your professional details and it'll give you suggestions. This is another one we're running short on time, so I'll sort of flip through this, but there's a voice capability that I actually use in the car and make myself more efficient. So I engage with it hands-free, it saves it for text later. Josh talked about AI hallucinations. I would just really caution you about these citations. This is a publication that's coming soon, but basically we looked at 180 publications and we categorized them in properly cited, improperly cited, and did not exist. And only 50% of them were properly cited and there was actually 32% that didn't exist at all. We don't have a ton of time, but there's lots of AI out there. Copilot is Microsoft and you can really use it with Outlook, Excel, PowerPoint. My two favorite features on this is taking meeting notes and creating action items by just ambient listening, as well as the Summarize feature in Outlook. I'll come out of the OR, there'll be an email string with 18 responses and I hit Summarize and it sort of tells me what I need to know. So this is my best advice. If you find yourself sort of dreading a task or it feels like it's taking too long, ask yourself, can AI handle this? Start discussing it with your peers and you'll discover so many creative people that are already using it. And unlike Allen Iverson, but like Josh, I do believe practice does matter. And I said, make this image less cartoon-like and this is what I got. And ultimately I sort of- What are we talking about? Leave you with this. We're talking about practice, man. So key takeaways, be specific. Then you wanna provide the context. You wanna specify the tone. And finally, don't hesitate for ask for revisions and refining these will really make you excel using sort of different AI. So thank you for that. And now we'll open up to panel discussions with questions and answers. Josh and Liz, was there any other questions that are in the chat there while I was talking? One of the questions that came up, Cathy, was if Spanish language translation can be used among other languages. Yeah, so I saw that question and I actually had a chance to use ChatGPT when that question was asked. And I said, make a ACL patient handout. The one thing that I know that it does do is make it so it's at your choice, seventh grade reading level. And that's very important as we talk about patient education and maybe even our website content to designate the reading level that you want it at. I then translated it to Spanish. However, I don't speak fluid Spanish. So I can't tell you how accurate it is, but I have heard that it is pretty accurate and can do numerous languages. I've had the same experience in my office where I've used Spanish language. I've not used any other languages, but I have no personal credibility to that. One of the guys that actually works in my office, he has actually said for several different items that it appears accurate. And so from a limited anecdotal experience, yes, I think it does work for Spanish. And I think it'll only improve with time. A lot of the words that we use in English, like impingement, for example, doesn't exist in Spanish. And so I think you have to really kind of make sure that someone who has some experience in that specific language can validate what it's showing. Any of the other participants have any questions that you wanna ask live or you can type them in the chat? Josh and Liz, maybe any other sort of, now that we've sort of covered those broad topics, any other things that you use your go-to that has really made your clinic more efficient? Nothing, I think, more than what we've talked about, but the one thing I will really reemphasize, and it's great that your presentation ended with this practice. It's amazing if you look at the quality of the questions and prompts that I asked when I first started. Like, I think my first prompt that I ever did was write a song like Led Zeppelin, but make it funny like Dave Chappelle. And that was the first question that I asked. And so, clearly, I think you can use it for work. You can use it for personal. I know Liz was talking about going to Finland. Kathy gave a bunch of really nice personal examples. I'm also using it. We're going to Patagonia in February. We did the exact same thing. We gave a 14-day itinerary, and it literally spelled everything out for us. And I would have never thought to use it for something like that unless you just practice. You kind of tinker with it. You find what works for you, what doesn't work for you. And just the more you practice, the more you'll find it saves you time, it makes you more efficient, and you actually kind of enjoy using it, as long as you remember in the back of your mind that you have to trust the validity. And if you ever question it, if you're like, oh, this citation doesn't seem right, or maybe that doesn't exist, always validate it with what you know is your gold standard. Yeah, I agree. And I see we just have another question that Laura asked. Are there legal ramifications that you're aware of when using AI? I think the most important thing is that you're not feeding in there a bunch of HIPAA data, right? So you have to be careful in the way you ask that you don't spell out a thousand details about the patient. You keep it in more general broad terms. Again, as I think I brought up, I put like star, star, star, star in for the name, and AI will put that into the note, and you can change it later afterwards, or add in at the end of a letter, some additional details if you need to, once you've taken it off of AI to add in the relevant information. So just being aware that everything you feed in needs to be anonymous as far as the patient is concerned. Josh, through your MIT courses, any other sort of comment on the legal ramifications of that? No, I think that covers it well. There is one thing that I think is relevant, especially to ChatGPT. Remember that that actually is dated. ChatGPT isn't current as of today. So it's actually unaware that SpaceX launched a rocket in Southern Texas yesterday. It only, I think, goes through, to my knowledge, I think it's October of 2023. So it's basically like a year behind what has happened over the past year. So always remember, if you're trying to look for current events, like things that are happening right now, there are other different large language models. The Grok one actually, or Grok 2 actually now, for Twitter slash X, actually does a really good job with being up-to-date. But most of the other ones are dated, actually far back in the past. And so you always gotta remember that. Well, to be respectful of everybody's time, it's 9.02. We really appreciate you guys attending this, and hopefully you got some take-home points. I really wanna just say thank you for Dr. Harris and Dr. Scott for participating tonight, as well as ASSM for hosting this, and the Enduring Education Committee for inviting us for this webinar.

Artificial Intelligence

Clinical Efficiency

Orthopedic Surgeons

AOSSM

Sports Medicine

AI Applications

Patient Education

AI-powered Dictation

Administrative Tasks

AI Limitations