all right next up is uh Brian Krueger one of my my partners and uh and someone I love to call and and ask for advice so he's going to talk about uh artificial intelligence spine surgery and you're going to hear a lot more about about the unit and how it how he incorporates it into his practice all right thanks Zach and thank you all for coming today you know this being our first time doing this it's quite a quite a great turnout and uh so far the talks have been been quite awesome um it's going to be hard to kind of follow all that up um something that I've gotten involved in in the last year and a half is uh this idea of ai's role in deformity correction surgery and you know when you hear the term AI or artificial intelligence especially when it comes to uh its role in in surgery I think our our minds go straight to thinking about robots and uh and robots performing surgery you know as a surgeon like oh am I gonna have robots take my job you know like and you know this is just kind of the mental image that we get but I hope this talk kind of shows more uh the reality of the situation we're probably hundreds of years away from this but who knows um some initial thoughts uh you know you can look at you know if if if the the idea of AI intimidates you you just a background I I I studied biology in college I went to med school I don't I don't have any background in computer technology uh software technology I used to say I'm I'm pretty computer illiterate uh so you know it kind of kind of makes me feel good when the vice president United said says AI is kind of a fancy thing you know first of all it's it's two letters means artificial intelligence well we're all on the same page that's where we all start with that statement okay and so then we can all kind of grow together and so you know if it's intimidating to you you just got to start somewhere you know Elon Musk is a a pretty prominent figure uh in our culture uh he's got a lot of uh pretty pretty rational opinions I would say uh one one he's he's a little bit nervous about the the the blowing up of AI and he says I'm increasingly inclined to think that there should be some sort of regulatory oversight maybe at the national and international level just to make sure we don't do something very foolish I mean we're with artificial intelligence we're summoning the demon and he's he's he's very concerned about uh us creating some sort of Extinction event for uh for for Humanity um jinny redti she's the uh she's a former CEO of IBM I think she put it the best when she said some people call this artificial intelligence but the reality is that this technology will enhance us so instead of artificial intelligence I think we'll augment our intelligence and so I think that's that's the way we should view this is is augmenting our intelligence augmenting our capabilities within the the world of spine and spine research for that matter so to give an overview of what I'm going to talk about there are a lot of challenges with adult spinal deformity surgery uh we'll talk about these challenges we'll talk about the potential that a has in spinal deformity surgery and then you know we all got to start somewhere and today we're going to start with the basic principles of machine learning and then we'll finish with some case examples where we use these readily available predictive models so I have no Financial disclosures but after watching David's talk I kind of want some um but uh just to give a little background uh in 2020 metronic Acquired metrea and we we often times throw around the term metrea but they're kind of moving to this unid platform as they call it uh and it's designed to support surgeon workflow in in pre-operative planning and kind of create these personal personalized implant solutions for for surgery um and so we'll talk a little bit about that today um the case examples and then and then the the a lot of the presentation does does have to do with elonics unit cap capabilities so the challenges of spinal deformity surgery you know there are so many variables that must be considered when planning a surgery and everybody's treatment algorithms can vary significantly I mean just the the the the three presenters before me we all kind of view problems in a slightly different way you know when you think about the surgical pathology patient signs and symptoms their entire skeletal anatomy and Alignment um each segment mobility and Anatomy the different biomechanical forces depending on their body habitus uh patients bone density and composition uh a patient's desired goals how much they're they want to go through their comorbidities and limitations and then your own experience as a surgeon I I remember Grand rounds I gave a a grand rounds on cervical deformity and uh a neurologist was like well can't you just can't you just design a study and test these things and uh I I just thought they was so ignorant when it came to uh to spine surgery and and no it's you know up until I discovered this technology I was kind of disenfranchised by the whole academic world because it it's it seemed almost impossible to isolate these variables and study them effectively This Is Spinal deformity it there's some parameters and things but a lot of it is just somewhat of an art form but I think with the buding of this artificial intelligence technology we are going to be able to study all these variables and how they relate to each other so what is the potential so AI seeks to replicate the experience of human intelligence it's able to learn from immense data sets and you know all this stuff uh uh Chris Ames out at UCSF he he's kind of a he's he's definitely an early adopter in this and I remember being being in residency hearing him talk and he was talking about all this Mass data they were collecting and as a resident I'm like well who's going to uh you know synthesize that data and dissect that data who's going to mine all that data it's us the residents right so I was not very excited about hearing about all this Mass data collecting um but but with mass data you know this was back before AI was kind of a buzzword the AI analyzes all of it because it's humans aren't capable of analyzing this much data so AI seeks to learn from just immense data sets help make decisions and provide recommendations and then it it can adapt to new data readily um so there's various aspects of AI there's natural language processing robotics computer vision but machine learning I think is the kind of the the the key aspect here that we're implementing in our field so this is a quote from that paper I just showed the opportunities are Limitless in what questions may be answered beyond the scope of our current knowledge base and human limitations spine surgeons have historically relied on clinical judgment and experience in addition to retrospective cohort studies based on linear and logistic regression to educate patients and inform their decision-making so kind of this primitive model into just Limitless potential I think this is going to Revolution ize our ability to conduct research in spinal deformity um you know we typically will try to uh think of what variables affect patients but artificial intelligence will be able to uh look at other relationships that we're not thinking about and able to analyze this massive amount of data this helps reduce user bias and we can you know analyze everything without leaving anything out so if you can imagine all the variables you look at in a spinal deformity patient age uh sex bone density body habitus BMI uh alignment you know you have all this data and then when you look at outcomes over the long term AI is will will be able to tease out what matters and what doesn't so kind of in general about the potential artificial intelligence has the potential to be all of our collective experience so that in the future future surgeons any experience level can use that to guide their decision- making so machine learning there are three uh paradigms of machine learning and I think this is just a great place to start if you have no background uh in this stuff the first is supervised learning so supervised learning it uses labeled data sets to train algorithms to predict outcomes and recognize patterns so what that means is that the data is labeled and that it contains examples of both inputs and correct outputs let's say for example you feed into the machine pictures of various trees and you label those trees this one's an oak tree this one's a maple tree this one's a Arbor and it learns from the input then you start sending it data that's not labeled and it has already determined which characteristics make it this tree which characteristics make it this tree and it spits out kind of an answer so here you can see rectangles circles triangles hexagons these shapes all have a label once the machine kind of learns this pattern you can send in unlabeled data and it can spit out answers so the application in spine surgery would be show the algorithm uh a fulllength scol x-ray and the Machine that is kind of already learned will be able to calculate bone density it'll be able to calculate lumbar lordosis pi and just kind of do it for you the next aspect of machine learning is unsupervised learning so this is where the models are given unlabeled data and they're allowed to kind of discover patterns and insights without any explicit guidance or instruction it kind of makes up its own rules and can kind of structure the information based on similarities differences and so kind of using the same example you have a bunch of shapes now it doesn't know what these shapes are but it can kind of spit out that these are similar these are similar these are similar so when you apply that to spine surgery you could have a whole bunch of patients who have undergone a one level Fusion or a T10 into the pelvis and have certain outcomes and so that machine learning could potentially take all these patients and their outcomes and group them turns out that patients with lower bone density have higher rates of Hardware failure you know we know that through our research but this is what they could take on a whole another level patients over the age of 65 who happen to be female who happen to have a a dexa score of negative 1.5 and and uh and and carry their body mass anteriorly uh are are going to get 50% of the time pjk you know whatever the example is but they'll be able to kind of pull these variables together and find out what are problematic combinations and what are ideal combinations uh of those variables and then reinforcement learning is is very similar to the way humans learn and it takes actions in a dynamic environment through trial and error to maximize the collective rewards so this is where you provide positive or negative feedback and then the uh algorithm can kind of learn from that so this is very important I think in the development of this technology you know when I show the you know this technology and and and what I have been using you know the so far the outcomes of the AI may not be all that impressive but every time you utilize the application you are giving it more information so that 10 years down the road you've got all these surgeons pumping in all this data that's being mined by whoever created the algorithm and the algorithm gets better over time so just to kind of recap if you don't learn anything at all about AI machine learning just just know that there's there's three three aspects supervised learning unsup supervis and reinforcement so now if the whole world of AI and machine learning and how it applies to spine seems foreign and and and too much you know you just got to start somewhere and this is a great place to start so the unid platform is metronics AI platform and I think uh David's slide looking at the different companies and how they've evolved in their market cap right now it seems like metronic and globous new vasive are really kind of leading the charge in in some of this newer technology I know theuse synthes is like a third and they've got stuff coming out but this really is kind of a first mover Advantage as you start collecting this data and building these algorithms so the unit platform is more than just AI it starts with this kind of scheme of biomedical Engineers where you see a patient you come up with the plan and then you send your plan to these engineers what they will do is they'll utilize their software to kind of chop up the scolia X-ray and make the adjustments that you said you would say for example I'm going to do an L5 S1 they well they they'll they'll chop the image at that add 15 degrees and they'll kind of create this pre-plan um and then they utilize the algorithms to predict the outcome and uh on the next slide I'll tell you about those algorithms once the uh kind of plan is finalized they create this patient specific implant and the reason for this as was Dave was talking about earlier is a way to sell it um but I do think you know these these pre uh pre-fabricated implants are kind of nice and I'll show some examples of why because when you get into surgery and you've got these big plans on correcting sagittal balance it's really hard to kind of check yourself intraop and and say are you doing exactly what what you said you were going to do and then after surgery there's posttop analytics on did you do what you said you were going to do and then did the uh that the rest of the spine changed the way we predicted so here are the three algorithms that the unid platform has and uh the one I'm going to talk about the most here is this adult de deformity so this algorithm is supposed to predict that at six months posttop what the thoracic kyphosis will be and what the pelvic tilt will be how important pelvic tilt is and thoracic kosis I mean pjk is our you know our gasta our Holy Grail trying to figure this out um in the ped's world uh their algorithm looks at lumbar lordosis and pelvic Tilt at 12 months postop now these are scoliosis CES cases where they're usually not fusing down to the sacrum and then they also have an algorithm looking at D gen cases and at 6 months looking at lumbar lordosis of the upper non-instrumented segment and then sacral slope or or pelvic tilt so this team of biomedical Engineers kind of has this generic understanding of of alignment but you can really work with them as much as you want and let them know what's important to you are you just sticking with the basics do you have different ideas about what you think proper alignment is and it's a very personalized EXP experience for the surgeon so I'm going to move on to some cases and you know these are these are these are fine cases they're not like the best outcomes but um I think they really illustrate the technology and kind of what we were able to do with them so this first case is a uh a 50-year-old woman with a history of l45 fusion and she presented with you know complaints of of sagittal deformity inability to stand up straight inability to walk any significant distance just overall the quality life severely diminished so you know there's just a one level Fusion it's not a completely fixed deformity but my plan here was an L2 to4 L lift and then an L5 S1 a lift as a stage one and then from the back L2 to 23 34 and 51 grade 2 osteotomies with T10 to the pelvis so I send that to the biomedical engineer team and this is what they turn out they basically measure every end plate and they kind of come up with an alignment and just a backtrack these these algorithms are only built on alignment they're not taking into consideration anything else they're not considering patients's age they're not considering bone quality they're not considering the fet anatomy or the the flexibility of the curve it's literally just looking at alignment which I think is so inadequate and I've I've kind of talked to them about you know we're collecting all this data up front and and and surgeons are using this I mean we should really be collecting every aspect of the of the uh any any data that we can collect on these patients we should be getting after it immediately you know if you're going to have a team of biomedical Engineers working on this I mean there you can you can you can hand them a dexas score um you can hand them the patients uh sex male or female you can hand them the age just simple simple simple data that is all readily available so going going back to this uh this first case so here's here's the pre-operative plan and just kind of you know from a bird's eye view looks nice and you could see the pre op measurements the plans measurements and kind of the perceived correction at every level it kind of gives you uh some advice on what angle of implant to use and then again it pre manufactures the rod so on that kind of pre-manufactured Rod I I do want to say one thing when I look at the hospital that I do these surgeries at this cost of this whole platform and the pre-manufactured rod chews up about 2third of the hospital's margin on these cases so they're not really telling me I can't do it yet but you know it's kind of always in our best interest to try to create value and keep cost down and so one one of the future possibilities here and and and I've talked to the people at metronic is is not actually giving you a pre-manufactured rod doing all of this but then giving you a sterile template that you can bend your own Rod to and that wouldn't be very hard and that that cuts the cost by about 75% um and makes and at the hospital where I'm using this it would make the cost of all of this about the price of a pedicle screw or two um which I think would would really increase adoption so postoperatively corrected in the coronal plane uh corrected in the sagittal plane the thoracic hyos is was a little bit higher than uh it had predicted but then again this kind of goes back to inputting this data into the into the algorithm for future future algorithms one one thing I really like this for is you know um when you're doing a coronal correction a lot a lot of times people don't have it's not the worst sagittal deformity you've ever seen and you can find yourself as you're loosening up the spine the patient sinking into that Jackson t uh you you you can find yourself overcorrecting in the sagittal plane quite often uh so I think this is is kind of nice if you really want to loosen up the spine give you kind of an end uh an end vision for for the sagittal correction now here's the uh kind of the uh the the three Monon posttop and already seeing some increased kyphosis and I don't have the uh the six-month posttop here I haven't had it in the she but she straightened back up so you know what does the thoracic spine do in that first year you know we're maybe not collecting all of this you know we think maybe it's just going to tip over we do you know it's been it's been shown that people actually as their muscles improve they start to stand up taller um so kind of measuring that I think over the time would be interesting here's another case a 56 year old man he had a 5year history uh with complaints uh with his deformity inability to stand up straight he had bilateral S1 radiculopathy inability to walk with you know with severe pain quality of life diminished so looking at this it's like how how did this happen how did we get here and this was done by a a non- deformity surgeon in the community so I had to go back to 2018 and find like what was the original scoliosis X-ray and you know if you do some simple measurements there was a a slight um P mismatch and a little bit of hypokyphosis of the thoracic spine but it it wasn't it wasn't a terrible deformity at that time and I think the uh the surgeon got a little overzealous with the correction looking at these cages you know putting in 18 degree cages and 12 degree cages on these levels that were originally relatively flat he immediately uh developed compression fractures at multiple levels and ended up fusing uh kind he's also got some really interesting Anatomy this is a the original surgeon did a T12 to S1 but there's this mobile kind of S1 S2 vertebrae he technically has five segments in the sacrum so you could call it an L6 S1 but the S1 you know the S1 nerve route exited at this level so it was uh I kind of just referred it to it as an S1 S2 level uh but the plan here was uh to do an S1 S2 a lift with an L4 PSO and T T12 to the pelvis he had some he had some pain at the proximal Junction and then given his history of this uh subsidence and compression and and the sacrum kind of being a soft bone I decided to to cement augment the the S2 screws if you will so this was the plan basically going to get correction at that bottom level and then a PSO um I wanted to utilize this software just to kind of get a second set of eyes on all these measurements and kind of put together a pre-operative plan uh rather than going in in the traditional way so here's the posttop uh so I probably got uh you know a little bit less correction from the PSO than the plan had had had had told us but the um the the thoracic kyphosis was less because of the less correction um so you know this this wasn't the most necessary case to use it but for all the reasons just talked about I thought uh it gave gave some good information going into it now this this third case is actually the first deformity that I use the platform for and it it's what really made me a Believer and uh Jake's in the room over here he kind of convinced me he's like let's just pick the hardest case you're going to do and let's just give it a go and see what you can do so this is a 72y old woman who presents with uh you know the same sadal deformity complaints uh but she actually had a normal daxa scan normal bone metabolism Labs uh and as you can see from these films this is like this is a this is a lumbar kyphoscoliosis you know we typically think of deformity as the thoracal lumbar problem but this was such a tight deformity and I don't know I I just I felt ambitious I wanted to uh not just you know necessarily get a little correction anduse in sight to I wanted to to to fix the spine make it make it look the way it should her ribs are sitting on the ilc crest she's for the last 20 years been in so much pain that you know lifestyle was dramatically uh influenced by that so my plan was to kind of just approach it standardly inner bodies uh from L2 to to S1 and uh from the back breaking everything up um and these were like legit gr to taking the the the the ligament completely out inspecting every single nerve route because I knew that I was going to be changing things dramatically I didn't want anything getting hung up or caught um and so when when I did the surgery I broke everything up and we were using our our lanky clips and and it was just like a like a wet noodle com into the to the pre-manufactured rod it was it was one of the most amazing things I've I've kind of witnessed uh and just kind of blew me away that that it was able to work so here you can see the pre-operative plan quite ambitious in the sagittal plan if you look at look at that diagram to the right to be able to in a in an in a 72y old uh loosen everything up but uh indeed indeed we did and it was uh it had me as a believer in this in this technology and what we're able to do so if if we didn't have this pre-planning and the and the pre-manufactured rod like you go into surgery and you want to restore everything you want to break everything up how are you going to know intra op if you kind of did exactly what you did now we do have 36 inch x-rays intraop so with all of these especially with the chronal deformities I'm trying to make sure that we haven't accidentally sent them one way or another uh but this was really powerful now you may see I I used to be a huge believer in these sublaminar bands and in this patient she she was treated with a spinal cord stimulator and I didn't I didn't really take these into consideration but you know further dissecting for sublaminar bands and kind of disrupting the posterior elements and then having a previous spinal cord stimulator where the where the musculature has already been dissected and a laminectomy has already been performed I kind of left a a a very vulnerable midthoracic region so you know over over the next year she she kyos at that level but this was a pretty simple extension up and she completely bounced back to where she was I mean this is one of the happiest patients I have uh and and with their kind of renew renewed Zeal for life are traveling the world and doing all the things that they couldn't do over the last 20 years so um yeah yeah well you know that's what the future holds here but I think you know this is this is problematic because now the algorithm thinks that that if we correct somebody like this we're going to get pjk you know and it's just one patient but there are so many other variables that that resulted in this pjk most of it iatrogenic uh and so you know I I just think that we have such a long way to go with this but as you can see you know AI is is talked about in every industry and it's it's not robots with their own intelligence doing spine surgery it is that more augmented intelligence and if it seems intimidating we are literally on the ground floor and and it's going to be part of the future so I think it's I think it's worthwhile to get involved because that's really how you'll learn um but yeah machine learning and it's our collective experience thank you very much you guys have any questions