Transcription:
Artificial Intelligence and Machine Learning For Spine Surgery

Melanie Cole (Host): Welcome to Better Edge, a Northwestern Medicine Podcast for physicians. I'm Melanie Cole. And I invite you to join us as we explore artificial intelligence and machine learning, how it's advancing orthopedic surgery, the research in this field, the clinical implications and what these advancements could mean for the future of orthopedic health. Joining me is Dr. Srikanth Divi. He's an Assistant Professor of Orthopedic Surgery and Neurologic Surgery at Northwestern Medicine. Dr. Divi, it's a pleasure to have you with us today. What a fascinating topic, this is. I'm so glad to have you here. Most people have heard the words like artificial intelligence, deep learning, machine learning. Can you tell us generally what that means for the purpose of this discussion?

Srikanth Divi, MD (Guest): Sure. And first of all, thank you for having me Melanie. This is a very interesting topic and something that for me recently, I've delved into. To start from the top and kind of break down the definitions, I thinks really helps people understand what we're talking about. So, artificial intelligence is the overall overarching term that describes any ability by for example, a machine to replicate human-like intelligence. And so that's why you hear artificial intelligence in many different categories throughout your daily life, whether that's shopping on the internet via Amazon's algorithms or other websites, whether it's your like self-driving car, using AI derived algorithms to monitor traffic around you, keep you in your lane or, I mean, some even more sinister applications more recently.

For example, drone warfare, that's really taken off in the last 10 years. And it's leveled the playing field for smaller countries to defend themselves and et cetera. So it's really being applied in many different areas and pretty soon we'll be just talking about this as just an integral part of our everyday life. AI will just become a background sort of way of machines and humans kind of interact. There'll be guiding algorithms that either are designed for efficiency or safety, or no matter what it is in each field. In healthcare artificial intelligence is really at this time based to help provide clinician support. The goal is not to take over a physician's role in health care, because I think we are the ultimate experts in processing all the patient variables and providing a logical and guided outcome for the patient. But there are instances where computers can operate better than humans, and processing raw amounts of data, and giving kind of accurate outputs. And in certain instances that can help the clinician provide better care for the patient.

So, I think that's where the role of AI is within healthcare. And another term, that people hear a lot is machine learning and machine learning is actually a subset within artificial intelligence. There's also another term deep learning and those are related, and we can talk about that also, but machine learning just simply put it's used interchangeably with AI, but it's not. It's actually a subset. And it's like traditional statistics in the sense that it helps describe relationship between variables. In our healthcare applications, it would be kind of describing relationship between clinical variables and possibly certain outcomes. Traditional statistics, that's used in, you know, retrospective or prospective clinical research helps infer the relationship between those variables. So what that means is, we're trying to determine using statistics, whether a certain outcome is plausible and many clinicians, may be familiar with the P value less than 0.05. So, what that's saying is that with this statistical model, we are accepting, that 95 out of a hundred times that this average value or this conference interval, with the model that we've created is going to be accurate about 95 out of a hundred times. And so for us, that's like a arbitrarily set, but acceptable model.

And machine learning is a little bit different in the sense that it gives us a way of predicting the output variables. It really kind of is a different way of thinking about the set of data that we have. We don't know how the computer processes these variables using the algorithms that we create. So in that sense, it creates a black box. And for some clinicians, it maybe a little bit hard to understand. Because with traditional statistics, you see the equations, you see the mathematics, and you could say that's the inferential model. In machine learning, you have a complex algorithm, and there's different, kinds of models that are created. But basically the computer is able to process raw amounts of data, lots of it and accurately predict an outcome. The more data you have, thousands and thousands of thousands of patients with thousands of variables, the more accurate it becomes. So, we may not understand how the computer got to the answer, but it's going to accurately get to the answer. And so that's one way of thinking about it. It's the predictive model.

Host: Wow, very comprehensive answer. And thank you for telling us how artificial intelligence and machine learning really relate to statistics. So, how have these techniques been applied in orthopedic medicine? What have they sought to explore Dr. Divi that can help to improve the quality of patient care, enable cost effectiveness, reduced readmissions or mortality rates? How is it looking to provide a set of tools to augment and extend the effectiveness of the ortho field?

Dr. Divi: As of right now, I'd say in the past two to three years, there's been sort of an explosion of research, looking at these specific answers using machine learning in particular, Machine learning is able to aggregate thousands upon thousands of patients and more variables to accurately come up with an outcome. And nowadays it's being applied to let's say, large institutional databases. So for example, a center like Northwestern, that has many hospitals under it with lots of patients and you can have accurate outcomes that are captured within the system. Ostensibly patients would stay within the system and not go outside of it. So you can track outcomes more accurately for that one year, or two years around surgery, you have a lot more access to patient data, for example, their preoperative lab values, postoperative lab values, any other medical complications they had, any other studies that they underwent. All these things sort of add up the painting a full picture of the patient. In the past, data and this kind of patient tracking was sort of all over the place. There were some nationwide administrative databases that may have been able to track patients across state lines or nationally, but the data integrity just wasn't there.

So, in recent years, more people have been trying to apply this with large institutional databases or statewide databases, which may be more accurate and certain national databases that have the ability to track patients across the country. So having this accurate set of data, you can, then apply these algorithms to identify the outcomes of interest. So obviously, for patient care, and understanding which patients need more health utilization basically, is very important because that tells us as surgeons, how we can counsel them preoperatively.

For example, we all have a, gestalt about a patient, when we see them in clinic, they either pass the eyeball test or not. We may not know why. And maybe that's a way of saying as a surgeon, we're applying our own kind of innate machine learning because our brain is subconsciously adding up all these variables and all our experiences in the past and say, oh, this patient may have an adverse outcome and maybe it's like, their bone quality is not good enough for surgery or that they're anemic or a combination of factors and they need other sort of optimization before considering surgery. For example, a concrete machine learning model is where that can sort of clarify for surgeons, what exactly needs to be optimized and, whether patients are at higher risk for wound complications after surgery or reoperation, or other medical complications, such as a heart attack, stroke, blood clots, and pulmonary embolism.

So I think it can be a very useful support tool if we can get models that are highly accurate. One other thing I think I should mention is that let's say if we create a model that is very accurate for the Northwestern system, for patients, within the Northwestern system. It may not be externally valid, meaning it may not apply to let's say a hospital system in the east coast or west coast, just because the patient population is different and that there are other variables that we may not know.

So, in the end, we can have an internally kind of valid tool but to make it globally applicable, we need to have external validation. And so that's another area that's being kind of worked on as sort of a more generalizable model.

Host: Dr. Divi, as you're telling us the ways that Northwestern Medicine is innovating in the AI space as it relates to orthopedic and spine health, I'd like you to expand a little more on the ways you're using it, but I'd like you to tell us about your research that you recently published in a Surgeon's Guide to Understanding Artificial Intelligence and Machine Learning Studies in Orthopedic Surgery.

Dr. Divi: So the article that we recently published is a review article that was aimed at helping surgeons understand how to read an artificial intelligence or machine learning study. Because these studies are so new and have a lot of foreign language essentially for clinicians, we wanted to break down at the simplest terms, what each term means, and also how to apply this for their own practice. So, not only is it important to understand what the study is doing, but to identify whether it's appropriate, because we are all used to you know, sort of traditional statistics and clinical research.

We can apply our own intuition to determine whether this article or the study was valid. And, their methodology was valid. People may not know how to do that for AI or ML studies. And that was the goal of our review article was to break that down into simplest terms. Within, Northwestern Orthopedic Surgery here and, Spine Surgery, we're really pushing forward with applying machine learning models, as well as natural language processing models.

And then finally we want to start using computer vision models and I'll explain a little bit for the latter to which people may have not heard of. Natural language processing is using a lot of unstructured free text notes. In our electronic health record, we have endless amounts of those, whether it's progress notes or therapy notes, or nursing notes, operative notes is very important for surgeons. All notes are basically pieces of data and without someone going through and organizing them, they're hard to use for research purposes. Natural language processing kind of takes that out, it reads all this documentation, tries to use complex algorithms using analyzing sentence structure to then identify or predict outcomes of interest.

So, there are deep learning models associated with this. And it's extremely interesting, and we've, recently submitted abstracts to the North American Spine Society, looking at lumbar microdiscectomy data. We're pushing forward with analyzing our cervical fusion data. So we have lots of free text information that we're analyzing to see how this is predictive of certain outcomes and a recent article that was published, shows another application that may be interesting too, for example, providers and health system administrators. A group out of Columbia looked at operative notes and was able to use natural language processing to see if anything in the notes could generate for example, billing codes. So just by training a model to read the note and see what was done, they were then able to kind of accurately generate billing codes. So in the future, this is how something can be developed that can automate certain aspects.

Obviously you want to do this with high accuracy. So, before it was ever implemented, it would have to be double and triple checked. But, that's one example. And then computer vision is another area that's very interesting. The cardiology space here and radiology departments here have been looking into this and their applications are looking at whether it's chest CTs that identify lung nodules, or looking at cardiac echoes, to identify other variables of interest. Within orthopedics, this could be a very large, area of exploration. Orthopedics is traditionally very reliant on imaging and there are interesting ways we can use computer vision to help improve outcomes. For example, if you're able to create a 3D model from just a 2D x-ray, AP and lateral, x-ray using deep learning that may save a patient from needing a CT scan preoperatively, or an MRI with just as much accuracy interoperatively.

So that's one area of interesting application. And, there's also ability for an algorithm to read an MRI and identify any pathology that might be missed by the clinician. For example, if it's tumor or, infection or other things that may not be as obvious, to pick those up. So that's how that could be used as a support tool.

Host: Fascinating. Absolutely fascinating. So, I'd like you to tell us about the clinical implications of your findings and how you feel that these studies will influence spine health and the orthopedic field research and treatment options for patients in the future. How will this care model improve the way patients receive their care?

Dr. Divi: I think in the near future, the most immediate kind of applications and impact will be allowing the surgeon to provide individualized care, through a support tool. And so what that means is, deciding whether a patient needs further optimization or identifying whether they're candidates for surgery. That I see is kind of the more immediate application. In addition, as far as, whether this may affect how payers guide care is a little bit harder for me to know. I think there needs to be a lot more data, needs to be a lot more validation before I think we can make broad sweeping changes.

This is very new, in terms of like research in this field. I'd say in the last three to four years, there's been an explosion of this research, but I think in the next four to five years to come, there'll be even more. And once we have reliable answers that could be replicated, then maybe that will kind of shift how payers think. But for now I think, we can use these tools to help individualize patient care within a system and help improve outcomes, decrease 90 day readmissions by identifying, preoperative factors that may have previously been missed. That I think is where the power of this kind of research is going to come in to play.

Host: I agree with you. So interesting, Dr. Divi, as we wrap up, can you leave us with one parting piece of information for providers listening and how you think this will change medicine for patients, for providers, and maybe even result in a paradigm shift towards precision orthopedic medicine?

Dr. Divi: Yeah. I think providers across all different disciplines, should be aware of artificial intelligence and its applications. They shouldn't kind of sweep it under the rug, but they kind of embrace it as part of their daily life. It doesn't mean that our jobs are going to become obsolete in any way. It's just going to be another tool or technology that we use on a daily basis. I think within orthopedics, there's a vast opportunity for improvement in, for example, instrumentation, or techniques, that as I previously mentioned with computer vision, for example, patient specific instrumentation. There's also a huge opportunity to improve clinical care perioperatively. So, many different ways of looking at this and, I think, the next five to 10 years, it will be very exciting to see where this takes us.

Host: I agree with you. Thank you so much, Dr. Divi, what a fascinating episode this was. Thank you for sharing your expertise and insights into artificial intelligence and machine learning as it relates to orthopedic medicine. To refer your patient or for more information, please visit our website at breakthroughsforphysicians.nm.org/ortho to get connected with one of our providers.

And that concludes this episode of Better Edge, a Northwestern Medicine podcast for physicians. For more updates on the latest medical advancements, breakthroughs and research, please follow us on your social channels. I'm Melanie Cole.