I am a rising fellow at Special Surgery for Sports Medicine. So this is kind of just an overview of where we stand with AI today because it's a pretty big buzzword that I think needs to be brought down to planet Earth. So a couple of questions. First what is AI and why should we care? Talk about the lab that I helped create in residency. Talk about its core team and output and the four verticals that we basically view the importance of it and future directions. So simplified AI is essentially computer automation. It's using pattern recognition and it's patterned based on pre-existing data. Most people try to analyze AI but the work that Eric Mockney here and everyone else out there are doing, collecting data is the most important thing. So putting the cart before the horse and focusing on the AI is not the move. So data is key and so there's a lot of problems that can happen when you don't have great data and then you apply AI. So Amazon essentially built an AI tool to help improve their employment and the problem is they looked at historical data and if they're not employing many women historically that carries forth that bias in the recruiting process and they realized that they had a lot of problems with that. And we think disparities are, I mean in the tech world they think disparities are bad when only 40% are women but especially in orthopedics we know how difficult we have it but that was a huge problem in that world. And so it carries on into when you have bad data out there, especially in the internet, when Google itself made a program when you use natural language processing to say what happens if you say two Muslims walk into versus two Christians, two Sikhs, two Jews and found that if you have bad data out there that's what constructs a narrative and so we need to essentially be careful with what we do. That being said, there have been a lot of people out there that are trying to regulate AI in terms of an ethical standpoint but the takeaway is ethics of AI is the same ethics that you guys are all applying to data collection and management so let's not overthink it. So in terms of the lab that we created, it basically started off with trying to find a way in a perfect world for arthroplasty what surgeons should be compensated. And so we basically built an early model saying that if you take into account a patient's comorbidities, minor, major, moderate, major and extreme, that's the amount that should be reimbursed. Obviously that's not happening. By 2024, the cost of an injection for knee arthritis is going to be the same as a total knee. So couldn't have done it without my mentor, Dr. Krebs, who helped get us the supercomputer and my co-authors and this led to various publications, patents and products. And so the first vertical is big data where we care about large databases, social media, teleconferencing. And so we basically wanted to prove first that machine learning models outperform regression in most cases when the setting is very rich data. And so rich data essentially includes a high volume or high quality. We didn't have high quality data but we had a lot of public information in the world of NHL and MLB data. So we just wanted to show that machine learning models in that setting outperform regression. So the next step here is basically looking at association of if you have a strong registry like the cartilage registry at special surgery, what turns out to influence outcomes. And unsurprisingly to those that do that, mental health is probably the biggest key. In terms of value-based care, this was an article we did that basically tried to reiterate and show that you can use these models, again, to say what is a sustainable infrastructure when you're communicating these to your C-suite and administrators. Obviously a hip fracture case is not, we all know in this room, is not an elective procedure compared to a total joint. But CMS had initially considered trying to make bundled payment models. So when you show this table to administrators and so forth, it basically shows that in a sick patient, it's a very expensive model. For now, that's been scrapped. Small data, what does that mean? That includes all the data that we carry around with us in our phones and especially in remote patient monitoring like Apple Watches and so forth. We explored this area. There's a company out there that I helped create called Focus Motion that basically looked at how a patient recovers after surgery. You basically combine passive data of what is existing from the sensors including their sleep data, their, the number of steps, their, and then also active data such as their pain, their COO score, and then we were able to solicit them to do their exercises and show their compliance. And so we're trying to change the paradigm in order to decrease in-office visits and get richer data from remote patient monitoring. Imaging. So this is a use case for AI. And so long story short, in the setting of joint replacement, you can use it to identify implants which will help us before revision arthroplasty, and this basically shows that if you have good images, you'll get 100% accuracy. And then similarly, you can ask it what are the features that tell you what implant it is and it can show that this is the area of interest to help delineate the specific implants as a teaching tool. Did the same thing in the total hip world, showed the same answer. Use CAT scan data with the special surgery group to basically help predict once a patient's indicated for surgery, for a hip scope, do the preoperative radiographic indices actually matter? And the short answer was no. And then if you have a patient who is about to undergo their first primary hip scope as a shared decision-making tool, what is their future risk of ongoing into revision arthroplasty, sorry, revision arthroscopy, total hip, hip resurfacing, and PAO, obviously helps give that percentage when you're sharing that information with the patient. MRI, we've started to start looking at bone marrow edema to help predict that in concert with patient board outcomes to help give patients their risk of success. So the question is, taking it a higher level perspective, should we have hope or hype? It's interesting, these are all the rejection comments I received when I started trying to put this out in the literature. Basically everyone's saying it's not appropriate. It has nothing to do with orthopedics, why should we care? And then all of a sudden, I don't know what happened in December 2019, people started having this interest in it, like AJSM invited, JBJ asked, Yellow Journal, and all of a sudden people are just very interested in this hot topic and I can't really figure out what turned. But the problem is that we have right now, these are all the publications that have come out from 2018 to 2021. The red represents original research for people that are actually doing the technical work and the blue represents people that are just talking about the research. So for every two articles that are put out actually doing the work in AI, there's one that's discussing it, which is a problem. And then so, there's a lot of bad literature out there that makes the world a little bit more confused about what we're actually discussing. So how to spot bad AI research so that the work that, you know, Dr. Mockney and Spiker are doing is meaningful is that first, you have to make sure that when you're reviewing this article, you need to make sure that it actually needed AI. Second, you need to make sure that just using it answers a previously answered question and you shouldn't really be equating machine learning to AI because that's a subset and they're not the same thing. Fourth, if you have a title that reads ML or AI shows, proves, predicts, that's equivalent to saying p-value shows differences. We don't write papers based on that. That's a statistical technique. Fifth, you should provide the full open source code when you're publishing. And sixth, shouldn't apply these regression-based guidelines called tripod guidelines to evaluate research. And seventh, and most importantly, you probably shouldn't be releasing prediction models out in the world without external validation. So why are we excited, though? The degree to which computers are actually capable of automating tasks with speed, accuracy, and anticipation has actually exceeded our expectations and there's an opportunity to automate redundant tasks. So if you take a step back and think about it, the task that need automation in our day-to-day lives is clinical medicine where we are, it takes 17 clicks to order inpatient Tylenol. That's way too much. There's a heavy documentation burden. So why are we essentially fearing it? Is it a threat to our self-determination? Are we going to get replaced by these robots? Is there a burnout? Well, three different principles of autonomy, competence, and relatedness in terms of motivation, this is essentially an illusion. And so the takeaway is we need to initiate others to welcome AI as an adjunctive tool to clinical medicine. We need to leverage AI to decrease our administrative burden from documentation to personnel and humanize medicine so we can actually focus on the practice of medicine. And I'll ignore that for now because we're running out of time. This is just kind of beating a dead horse that we basically need to focus on using AI to replace administrative burden. And so we need to first rein in AI research, leverage AI to automate the administrative gluttony, and then develop meaningful applications on the back end without marketing it on the front end. And so what we're working on now are CAT scan data, and we're excited about that. So thank you. Thanks.

artificial intelligence

data collection

biases in AI algorithms

healthcare

evaluation of AI research