So, first off, thank you for having me here today. Let's get started. All right. So, is machine learning always the way? I think the first thing we need to do is talk a little bit about what is machine learning. At least for the next six minutes, machine learning is going to mean a computer process for prediction. So this does not include like everything in artificial intelligence. People often refer to these as like nesting dolls, where AI is the big nesting doll, and then within that you have natural language processing, visual AI, and machine learning is one component. And so that is what we're going to talk about here. All right. So, I'm going to change the question, when is machine learning the way? You know, just because you have a hammer, it is not always hammer time. All right. So, when to consider, when to avoid. So, when to consider, if you're interested in prediction, purely in prediction, machine learning can be really valuable. When to avoid, if you're interested in just understanding is there an underlying association here and making inferences and building, machine learning may not be the way. So, if you're going to have a continuing program and this is a model you're going to build and you're going to supervise it and you're going to continue on and on, machine learning good. If it's just a discrete one-off project, you want to do it and then leave, machine learning may not be what you're looking for. You have a very large data set where you can split it into training and testing and all this stuff and take really advantage of the algorithms that machine learning offers, yes. If you don't have a very large data set or even if you have a relatively moderate data set but there weren't many events, so let's say you did 3,000 ACL reconstructions but you didn't have that many failures, that's probably not ideal for machine learning. And then finally, when to consider common variables, right? If the idea of machine learning is you're going to build a model that other people can use so it would be helpful if you've included variables that other people will have. I would avoid machine learning if you've got a unique or costly variable that no one else is collecting because great, you built a model that was useful for the past for you but it's not going to be useful in the future. So why talk about machine learning and insurance? And here I want to say we're going to briefly touch on two ideas. One is the use of insurance data for academic purposes and the other is the use of insurance data by insurance companies. All right, so let's go back to when to consider machine learning, right? Interested in prediction. Well, insurance data is going to have utilization, complications, costs, so that matches up nicely. Continuing program, right? Providers have existing and continuing relationship with beneficiaries, providers, purchasers, so that seems to tee up. Very large data sets, right? Insurance companies have very large data sets. And then finally, common variables, right? When we fill out claims forms and things like that, it's a standard set of variables that they're always getting. So all these things sort of line up nicely to apply machine learning algorithms to insurance data and to be used by insurance companies. All right, so here's a quick academic example, decided to pick on our own projects. And so, you know, multiligament knee injuries, I mean, this is like, right, a devastating, very challenging clinical area. The other thing is no one center is going to have that many. So this teed up nicely, I think, for sort of big data type project. And so when we initially did it, this was the title, Machine Learning Approach to Identify Risk Factors for Post-Traumatic Osteoarthritis After Multiligament Knee Reconstruction. So, you know, we were excited. And then the question is, well, was this really good? All right, so let's go back to when to consider using machine learning. Were we interested in prediction for prediction's sake? Not really. Were we going to have a continuing program? No. You know, did we have a very large data set? Yeah, relatively. And then common variables? Yeah. So we benefited through the peer review process, and then ultimately machine learning was taken out because it really was not the most appropriate thing here. And then just leveraging insurance claims data to identify risk factors. So machine learning is not always the way. And so non-academic uses of machine learning. So early on, a lot of this in the insurance world, and still is, about fraud detection. Machine learning is very good for fraud detection and even beyond insurance. And so here are just a few examples. But more recently, and I think we don't really have a sense of how much this is going on, is the following, where insurance companies are going to use this big data and machine learning algorithms to try to predict outcomes, and then steer patients to specific surgeons. So you know, is it all about the machine learning, or is it all about the data? I would argue it's probably more about the data than the algorithms, at least at present. So how can we prepare for a machine learning future? Consider what variables the models are going to use. Include all diagnosis codes and be specific. Ensure the codes are linked with the procedure. This hopefully will allow us to take control of the expected number of complications. You know, we're going to be judged on the ratio of the observed to expected complications. I would argue we actually don't have that much control over the number of observed complications. Right? Presumably, when you go back on Monday, you're going to do the best you can do. The hospital setup is already there, right? There's a limited amount of control we have over the observed number of complications, but the expected we have a lot of control over. And I think this is important from a societal level, because we want to control the expected number of complications not by lemon dropping or something like that. We want to control it because we actually documented things appropriately, had good models, so the expected number of complications was accurate. A little bit more. So I think we should think about what do we offer, right? What should models predict? If you go back to the news story about this insurance company and what they're, they're just predicting readmission rates, amount of money spent, they're not necessarily predicting like how good the patient did. And so, you know, we should be thinking about what do we offer and making sure that we actually record that. Because if it's not recorded, no one's ever going to be able to predict it. And then who monitors these clinical prediction models, right? And I think there's probably a role for societies and registries to sort of take a larger role with that. And then should models be open and transparent? I mean, obviously, I think we would all say yes, but it's a little bit more complicated. And so making sure from the front end that we are, that we are trying to enforce that is going to be helpful. And then finally, you know, these models can be passive, but we can't. So thank you. Thank you.

machine learning

prediction

insurance industry

data

algorithms