BoneChat #47 - AI applications for orthopedics discussion led by Alex Ravitz, COO @ ClinicalAI

Welcome to Bone Chat Discussion number 47. We're going to dig into AI applications for orthopedics led by the COO at Clinical AI. And we are So I'd like to welcome Alex Ravitz.

Ravitz. Yep, that's exactly right.

Ravitz. Yep. And I'll put Alex's information into the chat so everybody can see. And thanks to member Dave McClain for bringing us clinical AI. And so I'm just going to hand the mic over to Alex and welcome aboard.

Wonderful.

Well, you know, Tiger, thank you for having me. I'm going to go ahead and share my screen just so you guys can see my presentation. Is everyone able to see my screen?

Yes.

Wonderful. Well, as Tiger said, my name is Alex Ravitz. I am the Chief Operating Officer for clinical AI and a company called VisionMed. We are a technology company that specializes in applications of artificial intelligence, specifically within the healthcare and life science spaces. Our flagship platform is called Maya, which stands for Medical AI Assistant. And Maya has a broad range of functionalities from clinical trial prescreening to data extraction, all the way to our newest breakthrough, the ability to observe and analyze videos and live feeds of surgical and nonsurgical medical procedures. To give you all some background on kind of Hopkins, back in 2020, under a company called RoboTrials, we were originally more focused on robotic process automation, but over the last 2 years, we saw AI rapidly advancing. And our team pivoted to create a separate company called Clinical AI to commercialize a suite of AI products. Along the way, you know, we've had many opportunities to develop customized solutions in the healthcare space, you know, including the technology used to analyze videos of medical procedures, which we are now commercializing under a subsidiary called VisionMed. This next slide just shows kind of the executive team. Dave is also part of our team. He's just not on the C-level, but we have a lot of experience in, you know, clinical research and AI software development. So this just talks a little bit about Maya. Essentially, you know, Maya is powered by the most cutting-edge AI technology available. The background technology is able to pass all 3 steps of the USMLE, which are the exams required to become a licensed physician within the US. Uh, the VisionMed technology was initially developed for the ophthalmology space, uh, to analyze the 13 different types of retinal surgery.

Alex, can you hear me?

Yes. Sorry, your question?

Yeah, your, your bandwidth, your bandwidth is sort of garbled, so I don't know if you, uh, sorry, give me, give me a second.

Let me just try to close out of like Firefox. Maybe that'll help. Um, Apologies, give me one second. Is that any better? So I'll share my screen again.

Yes, that's better.

That's smooth. Okay, how is it now? You can still clearly hear me? Good, good. Okay, sorry, Firefox is probably just taking up most of the bandwidth. Anyways, like I was saying, so basically, you know, the vision methodology was initially developed for the ophthalmology space to essentially analyze the 13 Is it cutting out again? Or—

Yeah, it's a little garbled. Yeah. Can everybody hear the garbledness?

Yeah. Got it. Um, okay.

Uh, talk slower.

Talk a little bit slower. Okay. I'll try to just be a little bit slower.

Or you can dial in from your phone. Um, if it, if it continues, that might be—

Yeah, I just wouldn't, I would, I wouldn't be able to, I guess I could share my screen. You know what? Give me, give me one second. Let me try to do that. I apologize for this. Um, let me try to dial in with my phone and then I'll share my screen from my computer.

Okay, anyone have fun plans for the weekend?

Oktoberfest.

I was very bummed to miss out Uh, from— for Eurospine, I, I received several Oktoberfest videos that look like a lot of fun. Are you going to Oktoberfest? Oktoberfest, the real one?

No, I've been there quite a few times, but I'm going to a little Bavarian redneck town called Helen, Georgia. Oh, I've always wanted to go there, and they have really a pretty good, you know, sort of a wannabe, um, I mean, it was just a little nowhere one-horse town that rescued itself in the '70s by literally making itself over to look like a Bavarian village.

Wow.

They've got it. They've got enough authenticity to get a really good schnitzel, really good beer, and it sits on the banks of the Upper Chattahoochee, so it's got that sort of Bavarian woodsy fairies in the bushes kind of feel, you know. Cool, if you like that sort of thing.

Sounds like Fredericksburg west of Austin. Yeah, same thing. You back, Alex? Oh, not yet.

Or, uh, have you guys ever heard of Leavenworth, Washington?

I've heard of that.

That's the Christmas town, right?

Yes.

Yeah.

And yeah, very European. They do a good Oktoberfest and just surrounded by the mountains or hills of the Olympic Mountain range up there.

Very pretty.

I've always wanted to go. I was supposed to be on a flight this morning to Tampa, Florida. That didn't happen.

Oh yeah.

I don't think there's going to be much going to Tampa.

You need a seaplane.

Yeah, yeah, we're going— I'm going to Orlando tomorrow. I'm going to take my son to Universal Studios, so should be light.

You should be no lines, hopefully.

Yeah, right.

If you ever get a chance to go to Munich to the Octoberfest, go to the Lowenbrau tent and sit with the Australians.

Yeah, if your liver can handle that.

I'm 100% in. Actually, the Australians were the ones that were sending me the video, some, some of the videos.

So there you go.

Yeah, funny. Yeah, I've been to Oktoberfest in Munich. It's— it is— those tents are enormous, and every, every beer has their own tent. It is so strange, and you can only drink inside the tents.

So I didn't know that you can only drink inside the tents.

Yeah, it looks like, looks like a giant fair, you know, looks like the Missouri State Fair when you walk in. And then where's the beer? And then you see these massive white tents, and like, that's like the Lowenbrau tent. And you go inside, there's 5,000 drunk people singing.

That is awesome.

But then when you have to go to the bathroom, you have to go outside the tent, and there's so much beer being eliminated that it looks like the LA River, man.

Yep. You got to step over the drunk people that are passed out.

I mean, we've all been to Academy at, uh, in New Orleans. That's pretty much the same thing. AOS New Orleans.

Kind of. Maybe less Steins.

Uh, Tiger, I, I had a quick question actually. Uh, if I email you just the slides, would you be able to share your screen and I can just tell you when to go to the next slide?

I can try, yeah.

Okay, I, I will.

Why don't you email to— why don't you email to Dave?

Uh, Dave, you, you should have actually the slides. Do you, do you still have them or do you need to send you another copy?

Or, I mean, you had it on your computer. All you had to do was dial in on your phone too.

I did it. It kicked me out. It let me join for a second and then it kicked me out after like a minute for some reason. It just said goodbye and then it hung up on me.

Oh, at least it said goodbye.

So unfortunately that was a bit of an issue. Sorry, David.

You want to just go without slides? That's the other option.

Sure, I could do that. Unfortunately, there is a demo that I did want to kind of show you guys, but that's okay. I could— maybe I could try to pull it up later without talking about it. But the audio quality is pretty good just talking like this.

Yes. You have a website?

Yeah, the website's in the chat, I think, if you can see it.

Yes, the— unfortunately, the slides won't be in that, but I'm happy to— if anyone would like to get a copy of them, I'm happy to send them.

Well, we'll distribute the slides afterwards. Yeah, the website's clinicalai.us. All right, it's all yours, Alex. Take it away.

Okay. Yeah, so like I said, you know, kind of we initially developed the software for the ophthalmology space, and essentially, you know, it is able to differentiate all types of retinal surgery with no input from the users. Maya is also able to create a procedure note for any type of medical procedure without the need for ambient dictation or document analysis. As I'm sure many of you are aware, procedure notes are required for any procedures so that other physicians understand what was done and why it was done. Maya can automate this process for physicians, saving valuable time, you know, for already overworked people. You know, one of the surgeons that we work with specifically, he completes roughly 6 surgeries on average per day, and it takes him roughly 20 minutes to create a procedure note. This software saves him an average of— he told us about 2 hours per day. You know, obviously, this time savings will vary depending on what type of surgery you are specifically doing, but we expect the software to be able to save physicians between 5 and 30 minutes per procedure, you know, and then you would just take that number of procedures they do per day. Normally I would have a nice little graph that kind of shows you that the volume of surgical procedures is increasing exponentially, and these surgeries are becoming increasingly more complex with time. Manual note-taking is a very tedious process for already overworked physicians, and Maya is able to give insights observations, recommendations, and the ability to pull up reference clips in real time as a procedure is being performed. So this is— well, this is currently a capability of the software. In order for us to con— basically include it in a commercial product, we are going through the process of getting 510 FDA approval. The procedure notes, because that is generated after the surgery, does not need that same approval. Um, again, this is kind of normally where I would show you an actual demo of the software, uh, using a saved video. Um, I could actually— if anyone would like to see it, I'm going to include the YouTube link in the chat right now. Um, so if anyone does want to watch this video, um, feel free to go ahead and see it. Um, it is basically, um, it is an inferior mesenteric artery IMA division using the Da Vinci, uh, basically robotic surgical system. This software can be applied to basically any type of medical procedure. It is not discipline-specific. You know, we are able to basically use this for any kind of device or surgery or non-surgical medical procedure. You know, basically if you do watch the video, at the end of this 2-minute video, the AI is going to show you an example of a procedure note. But I'm actually gonna share my screen again. Hopefully I don't cut out too much. I do kind of just want to show you what that is. And this is the video itself right here. But I will briefly— I don't think I have this pulled up. Sorry, give me one second just to pull it up. And I will show you kind of what an example output looks like.

So while he's doing that, let me try to put it in layman's terms. So this software will watch any video. Anything. It'll watch kitchen cooking. It'll watch retinal surgery. It'll watch, and it'll annotate. It'll tell you exactly what's going on. This ligament was released. It took 15 seconds. Then they went over to this artery. They did this. And it's, it's incredible. And, and we're trying to figure out what the applications may be in orthopedics. Asher's got his hand up.

Hey, is this a, uh, uh, basically like a ChatGPT wrapper, or did you actually develop your own like artificial intelligence, like, uh, like OCR, some kind of like novel tech behind it?

Yeah, so that's a great question. Uh, am I cutting out at all? Are you guys able to hear me clearly while I'm sharing my screen?

No, you're good, and we can see the notes.

Yeah, wonderful. Um, so yes, so basically, you know, we work with a number of different AI providers in order to kind of utilize their language models. Uh, what we do is we train those language models to do specific tasks. So for example, uh, this software specifically, uh, is actually powered, um, by— it's not ChatGPT, but it's the same AI that powers ChatGPT, um, if that makes sense. We have a developer version of the software which is not available to the general public, um, but we do have contracts essentially with OpenAI Google, Microsoft, and we're currently working to get a contract signed with X, Twitter. They are going to be entering this space in the near future. Supposedly they are going to have a supercomputer that powers the AI that is supposed to make their AI a little bit better than their competitors. We'll see how that works. But essentially, you know, what we do is we find which AI is able to be best utilized for each specific task that we need to do. And that is the AI that we pair with our technology. The benefit of that is that, you know, you basically get an AI that a company has spent hundreds of millions to billions of dollars investing in. You know, as a smaller organization, we do not have the capacity to basically match what OpenAI is able to create. You know, they just have billions of dollars in funding and unfortunately we don't. But, you know, the benefit of that is that as that AI advances, as they come out with newer versions of it, the software on the user's end will automatically improve. It becomes faster. It becomes— I don't want to say more accurate. All the accuracy actually comes from us training the AI to do specific things, as you can kind of see right here. Does that answer your question a little bit, Asher?

Yeah, I guess it makes sense. So when you're training the agent, where are you getting your actual dataset? Is it like ported view med videos or is it something else?

Yeah, that's a great question. Essentially, you know, we have a large library of surgical procedure videos and that is kind of what we use to train the AI. So for example, right here, We have a couple terabytes of data that was provided by our retinal surgeon partner, and that is kind of how we train the AI. So we feed it the full procedures, we feed it reference clips of specific things being done in those procedures, and then we feed the AI the corresponding op notes. And we essentially train the AI on those procedures so that it understands what's happening. It's able to accurately output this information and it's able to identify what is being done so that it can pull up these reference clips in real time. So for example, right, if you were a surgeon and you asked Maya, you know, let's say this is plugged into, they have a microscope with a video feed. Let's say that Maya is currently live within your system. You would say, "Hi, hey Maya, pull up a video of an epiretinal membrane peel." And Maya will pull up a video of either if we have that surgeon-specific videos or a different surgeon performing that same procedure. And they can look at that video in real time, kind of as a little refresher. Maybe they see something that's abnormal and they don't know what to compare it to. Let's say you're doing a colonoscopy, for example, you see an unusual lesion. Hey, Maya, you know, can you pull up something similar to this? Maya will pull up reference photos and say, you know, this patient was diagnosed with X. You know, it might be a similar diagnosis for the patient you're currently working with. Does that make sense?

Yep, makes sense.

Very cool. Wonderful. So you can see here, this is kind of— we basically trained the AI in the format of this retinal surgeon's procedure notes. So we basically kind of mimicked what he did. But all of this is generated by the AI. The AI is not taking any kind of audio dictation. It does have that feature. But for these videos, because we're using these kind of saved YouTube videos, so we can kind of demo this to people. There is no audio dictation coming from the surgeons, nor is the AI analyzing any kind of document. So it's not finding this information from a document that was provided by the surgery. The AI just kind of watches the surgery with no sound, understands what's happening, and basically creates this note for this surgery. We do obviously kind of work with surgeons to verify that this information is accurate. The surgeons that we are working with have told us that the AI is able to output information, you know, kind of at the resident physician level. You can kind of see here this is for nephro— And I have a number of other examples as well so you can see here what it might look like for macular hole surgery but like I said all of this is completely generated by the AI and I've, I have a whole number of examples and again I'm happy to share this with anyone if you would like to watch these videos they are publicly available on YouTube. But yeah, this is just for ophthalmology specifically. But this can be deployed for any kind of procedure. I see a couple people have raised their hands.

Yeah, so how you, you had mentioned whenever you want to use it for a different procedure, you need to use the term feed it information on that procedure. Is that a fairly hands-on approach? So if you wanted to use it for a knee surgery, you need to compile 10, 20 videos, pretty clear videos, I'm sure, of what a complete total knee replacement looks like and then tell the AI, all right, this is them resecting the femur. This is them resecting the femur again in this video. This is them doing it again in this video. And then the AI gets smarter with every video. Or is it, does the AI know enough to see it once and then from then on it's good to go and it can start identifying those steps?

Yeah, so all of that training is done on the background and the basically by us before we deploy it to our clients. But the answer is yes, we generally will feed it usually gigabytes of different kinds of videos, lots and lots of different procedures. So what we need the AI to do is be able to differentiate on its own what kind of procedure is being performed. So it needs to understand when you're doing You know, unfortunately, I'm not a surgeon, so I don't know a lot of the actual surgeries themselves, but I'll use the kind of ophthalmology examples, right? It needs to know the difference between an epiretinal membrane peel and a macular hole surgery on its own without a human telling it what's being done. So that is something that we do on the background. We basically feed it a very large database of data. It learns from that database, and then it's able to kind of watch new videos and be able to understand what's happening. We do pair those videos with the corresponding procedure note so the AI does understand what's happening at every stage. And then what we'll usually do, like I said, is give it reference clips. So those are shorter clips, usually like 10 to 15 seconds of something that's very specific that's happening. And that way, when that happens again, the AI understands that that's what's being done. It does take, you know, probably, you know, it took well over a year to kind of build up the AI to where it is now. Training it to do new types of surgery probably take— we can do within 2 months. It's really just about making sure that we partner with the right surgeon who can go in and help validate the results. If the AI is— sometimes AI will hallucinate. We need to basically understand why it's hallucinating and how we can go in and fix that. But generally, you know, in order for us to keep the AI HIPAA compliant, the AI is not going to learn from the procedures that are being done live. That's how we're able to keep things compliant. So unfortunately, like, you know, if you were to use the AI in a live surgery, it's not going to remember what you fed it. It only remembers what we feed it before we deploy it. Does that make sense?

So, so it's not continuing to strengthen its knowledge whenever it's deployed. It's just, it's still going off of whatever knowledge it was deployed with.

It, we could deploy it in a way that it learns over time. But if it's fed any kind of patient information, it would not be HIPAA compliant. For sure. That's so, so right now the way we have it deploy is to essentially wipe its own memory. It will analyze, it'll output these notes, it will talk to you, but then it has to basically wipe its own memory. I'm sure there are ways around that, but right now that's kind of how it works.

That makes sense. I would be interested to see like on the knee, on the knee side still, a primary knee versus a revision knee. There's going to be a lot of similar steps at the beginning and the AI needs to differentiate pretty early on. But it's fascinating stuff and it's— yes, it's the technology that's out there today.

If anyone is interested in actually seeing a demo of the AI using like maybe a video that you have or you've completed previously, I would be happy to feed that to the AI and show you a live demo or generate like a YouTube video which you could see a later time. If anyone is interested in that, feel free to shoot me an email. All I would need is just a copy of that video or a YouTube video, a link or something like that.

Yeah, thank you.

Absolutely. Yeah.

I have an oddly unique perspective on this from very close people that have been on the orthopedic side as well as the retina surgery side. Retina surgery and their note-taking and ophthalmologists are very different. They're very unique, particularly in retina surgery. They require a lot more notes because they see the patient on an emergency basis and they're transferred over to somebody else, right? So the note-taking for surgery, I don't see it being— if you've seen an orthopedic surgeon's notes, it's like ORIF proximal tibia and that's it. Um, uh, if anybody— so, but I see a lot of value on this in the clinic, on the clinic standpoint. That's where, um, physicians in particular have— they get buried in notes. And, um, uh, but I also see a value in this in— on the internal medicine side when you're identifying scar tissue and the time it takes to clear out the scar tissue and so on and so forth. Just trying to dial you in on some expectations from the orthopedic side, but clinically, this could be extraordinary.

Yeah, absolutely. You know, you know, obviously, as someone who's not a surgeon, we're not familiar with, you know, every single type of procedure note, how that kind of looks. But, you know, we are really good at kind of applying the AI to do different things and to customize it to our clients' needs. So if someone came to us and they did have— let's say they saw the value in the software and wanted it to be applied a little bit differently than kind of what I've explained today, that is absolutely something that we would be able to do. We're very flexible, you know, with technology like this. But thank you, Liz. I appreciate the insights. Mark.

So, Alex, you're not keeping any PHI, it sounds like. You're— after every procedure, you supposedly wipe it?

Yes, that is kind of how we're able to keep everything HIPAA compliant is by essentially having the AI wipe its own memory. So it processes the data locally. And basically, you know, we have a partnership agreement, for example, at least with this technology, with OpenAI. And that partnership agreement says that they are not going to use this information to train the AI, nor will it be saved, nor can it be viewed by any other human. So even when, um, let's say that we do deploy this to a client, we ourselves are not actually able to see what people use our AI for. We can see that it was called, we can see how much information was processed with it, but we can't actually see the procedure videos and we can't actually see the outputs. The only people who have access to that information are going to be the users themselves. And then who they choose to share that information with is completely up to them.

So if a user were a clinical research organization, they could keep the PHI.

Yes, exactly.

And okay. So I just, I recently just wrote an article on the use of ambient AI scribes versus AI-assisted templated scribes. Which are physician-controlled. And so, you know, this is an area of kind of a certainly a strong interest to me. I would like to see the video, you know, see how you guys are using machine learning and deep learning and so forth. Who's your machine learning Berkeley PhD expert on board? Who's that person? Or do you have one?

Yeah, so generally, unfortunately he wasn't able to join today, but the lead software developer for essentially all of our software is our CEO. His name is Hamza Hassan. Um, he is kind of the one who at least initially does a lot of the work. I'm unfortunately not a software developer. I work more on the operations side. Um, my job mainly is to kind of be the interface between our software development team and our clients. Um, if you did have any specific technical questions, I would be happy to put you in contact with him and he would be able to give you a much better answer than I could.

Yeah, I might. And also, if down the road you have a need for completely wiping out PHI from existing recordings, I have another UC Berkeley person who has a particularly interesting software that can remove PHI from audio, video at the meta level. And most sanitizing software that you can buy commercially doesn't do that. It's still kind of still there. And we've all heard about computers that have been wiped that really still forensically can be analyzed. And so it's a very similar thing with PHI. Anyway, I'm gonna ask for—

That's very interesting. Yeah, that's actually very interesting. If you don't mind, maybe if you don't mind, let's just send me an email. My email's alex@clinicalai.us. I would be happy to set up a time to speak with you more on that. Chat, schedule a demo, and maybe if—

yeah, I think, I think that would be a useful connection for you at some point.

Yeah, that would be greatly appreciated. Thank you, Mark. You're welcome.

Alex, from a—

uh, anyone else have any questions?

Yeah, um, I'm sure you guys have thought of this, from a legal perspective, uh, let's say that a surgeon encounters from a legal perspective, um, the data that's collected is inaccurate, and, um, and they're challenged where the data says X and the surgeon says Y. In that perspective, um, how, how do you guys plan to handle that? Because AI, again, you said it hallucinates, and we've all seen it, um, If somebody doesn't actually review their notes, what would they do?

Yeah, so, you know, one thing that we do stress to our clients, right, an AI is not a physician. An AI should not be treated as a physician. Any work that an AI does, any output that it creates, needs to be checked by someone who was kind of responsible for that data. So the surgeons themselves, the physicians, they do have to confirm that all of the information the AI outputted was correct. You know, one thing, depending on the specific device, if the AI— if the device has a frame of reference for size, the AI does a good job of being able to estimate essentially the sizes of different things that it's viewing. If that device does not have that, so for example, in retinal surgery, there is no frame of reference for size in these microscopes. The AI is going to basically estimate, and I can actually show you an example of this. So if I scroll up, so you can see here, right, the AI said a 25-gauge valve trocar was introduced, right, or a 3.5mm posterior to the inferior temporal limbus was identified, right? This is the— basically the way that it generates these numbers is by using what is standard in these procedures. But just because that is what's standard does not mean that this is the actual size of this, I guess, limbus, right?

Identify the mark.

You do need to check this information. If the AI does have a frame of reference, and many medical devices have that, the AI does do a good job of that. But otherwise, it's going to assume based on prior surgical precedence what these numbers might be. So that definitely is something that needs to be confirmed. Retinal surgery, it's unfortunate just because the devices don't have that frame of reference, but that is something that needs to be checked. Otherwise, what we can do is if a client requested it, any specific number like this, we can replace with X. So it would say an X-gauge Valtrocar was introduced, right? And then they would have to go in and just kind of replace that.

So rather than— what I understand, rather than go into a little booth and dictate the case, all they need to do when it takes 10, 20 minutes, they just read this and just sign at the bottom, yeah, that's correct.

Yeah, but I'm thinking from a perspective. So I would actually incorporate photos like they take in sports medicine cases. They actually take photos of the tears and so on and so forth. So I would actually allow it to choose to opt in pictures because they're already doing this. And then you have documentation of select photos that you can add to the notes.

But yes, so one feature that we do want to include, like I mentioned, which it does have this feature, is the ability to talk to the AI while the procedure is being done. So if you want to tell the AI a 20 let's say a 30-gauge trocar, right, was inserted, right? Then the AI would know and it would use that information. Same with if you do want to have, let's say, the AI take a snapshot, take a brief recording. Hey, Maya, record the last 5 seconds and save it. Hey, Maya, you know, take a snapshot every 5 seconds for the next 30 seconds, right? The AI is able to do that so long as you tell the AI to do that. We could make it an automatic feature, If a client, you know, let's say that they did want that for every surgery, they want that to be standard, we absolutely could incorporate that. Otherwise, it's generally going to be very customizable by the user themselves how they want the AI to operate.

Makes sense. Hey, Alex, what's the hardware needed to do this in the operating room?

So basically, you know, our goal right now is to kind of integrate the software with the medical devices themselves. So ideally, let's say that, you know, you're an ultrasound manufacturer, an endoscope manufacturer, the AI would basically come incorporated into the tower. So you would not need to bring in any kind of external device so long as you had an internet connection. That is really all you would need to have. That and just, you know, you need to tell—

it's like an API. It's like an API sharing data with something that is already in the operating room.

Got it. Yep.

And then just kind of telling the AI where you want this information to be saved, right? If you do want to have it saved locally to, let's say, a specific computer, you want it saved to a cloud-based kind of database, that would be something that we would help, you know, basically our clients set up. Got it. Did someone else have a question? I saw someone raise their hand a minute ago. Asher, was that you? Did you have a question or did I answer it? Guess not. That's your point.

Yeah. So this is, I think this is where everything is going with AI. You know, everybody's trying to figure out, is it going to take over the world? It's, you know, ChatGPT is this monster. It's sort of a general AI. But I think where everything is going is what a lot of people call agents. So every company will have an agent. Every surgical procedure will have an agent. Every, they're just gonna be these little helpful agents everywhere you go. And yeah, go ahead.

Have you ever heard Kevin Kelly talk about it?

Yeah.

Oh yeah. I'm a big Kevin Kelly fan.

Yeah.

I think the way he puts it is, is really interesting. So, you know, there's a kind of checking stuff. He kind of describes it as, um, it's like a really smart intern.

Yeah.

It's great, but you still wanna check its work. Um, and then the other thing he said, which I thought was fascinating, he heard him give the example of cameras. He said the idea that there's going to be kind of one uniform AI that will do everything is not right. So if you think back to the camera, there was the original kind of like Box Brownie style camera. Then you've got cameras for sports, you've got infrared cameras, you've got underwater cameras, you'll have kind of—

GoPro.

Yeah.

Yeah.

Subspecialization of AI in every imaginable realm. And I guess it's who gets to those niches first and ties them up.

Yeah, that's a great way to put it. Yeah. That's one of the reasons why we call the software Maya, right? It's a medical AI assistant. Software, you know, we design software not to replace humans in, you know, basically you can't replace a human with an AI. You could help automate some of the work, but it's not meant to replace people. It's meant to be a tool that allows you to do your jobs easier, that saves you time, and that just kind of overall, overall improves your ability to kind of do these tasks, right? That's all we're trying to do is, is just make everyone's lives easier through the application of AI.

So, so what else could Maya do other than just watch surgery? Because there's, you know, orthopedics, there's a ton of, there's pre-op planning and diagnoses and then post-op. There's a lot. Everybody brainstorm that while Aster's asking a question.

I was gonna kind of build off what you were just asking, Tiger. Have you guys thought of adopting this or expanding it to more of like perioperative, like management where you— and this may not be appropriate for ophthalmology, but like for orthopedics, there's a ton of like, you know, surgical requirements, paperwork, you know, non-surgical treatments, preoperative planning, and then being able to use that to both like manage an actual practice, but also then automatically fight prior authorization with the insurance companies.

Yeah, that's a— that's actually a great suggestion. I'll give you guys, let you guys in on a little secret. Generally, the way that we build technology is that we have clients come to us and they ask us, hey, you know, I would like to have an AI that's able to do this. And so we build them custom solutions and then we essentially turn those custom solutions into more general products that we can sell to anyone. So this particular application was actually the suggestion of a surgeon that we were working with. We said, this is a great idea, let's build it, and then we can kind of sell it to anyone else. That is kind of how we built this. But any other suggestion like that, that is basically how we build our software, right? Unfortunately, like I said, you know, I'm not a physician, neither is the CEO. You know, we are mainly experienced in kind of building software, software deployment, things like that. So when people come to us with like ideas like that, hey, you know, this would be a great use of the AI, we're always open to ideas like that. We love kind of building out demos and proof of concepts for different ways that AI can be used.

So it's interesting. Start with the customer's needs first. That makes sense. Go ahead, Alisa.

Sorry, I'm bombarding you with questions. I've been thinking about this because I've seen loved ones just get bombarded with notes, spending an entire day on the weekends taking notes or And so my question is, Cerner is working on this. I know Epic's working on something like this. I know that quite a few companies over in Palo Alto are working on something like this. How are you protecting— do you have IP? Are you protecting it or is it just like the wild, wild west?

We do have a patent in review. It has not been kind of formally I guess, accepted yet. We are planning to, though, patent kind of the overall technology that we're using. I will say, you know, what we've been able to do is, you know, we don't know of any other company right now that, that kind of does what we do. There are a number of imaging, AI imaging companies out there that are able to analyze like static images. Not many are able to do video. I do know that like GI Genius, for example, is able to identify lesions, uh, like let's say you're doing a colonoscopy, right? It is able to identify them. It doesn't create a write-up, but the AI is able to recognize like those patterns. Being able to have an AI generate like an op note, kind of give recommendations and pull up reference clips. Like I said, as far as we know, nobody has that ability yet. Um, exactly how that's gonna turn out in the next 6 months to 2 years, remains to be seen.

So there's a company called CarePix that does this for diabetic lesions, and they've been doing it for several years. And they literally scan the patient, or the patient can go in and scan their own diabetic lesion at their house, and CarePix will generate a note. They'll also suggest the implant or whatever is needed for the patient, as well as do a dictation note. So I mean, I would, I would seriously consider, this is just my own perspective, hiring a patent attorney for the company that just writes and defends patents. And they're very hard to find ones that actually write and defend internationally because there's a lot of stuff out there. And CarePix in particular has been on the market for about 6 years.

Oh, yes. Yeah. Like I said, that is a little bit about what kind of differentiates us is like I said, there are a number of AI imaging technologies that do use those static images like what CarePix does. It's just they're not able to analyze those kind of full videos in terms of working with a patent attorney.

Because the patients— yeah. So and the patents themselves can be broad enough. They can say videos or images. It depends on what the patent is. So definitely looking, because this is fantastic. And whoever comes to the market with this the best, um, with the most marketing money, will win.

Yeah, I'll definitely look into them a little bit more.

I think it's a, I think it's a space race. Whoever can get to scale first is going to win. I'm not sure the patents even matter. Um, you know, it's, if you get enough customers And I mean, it's like beta versus VHS, you know, whoever has the most customers wins.

It might be the second one. It might be the second best because all the companies will go after the first one.

Exactly.

Exactly. Yes. Yeah, we're— it's, it's a little bit of a wild west right now, but we are, you know, we are working with a patent attorney. So hopefully—

what does your company look like with funding and kind of sales force in the future?

Yeah, so I could touch on that a little bit. You know, right now Clinical AI is basically kind of the parent company. We were able to, very fortunately for us, grow organically. The help that we received from Johns Hopkins in the beginning, basically, you know, connecting us with industry leaders, potential clients, that really did help us to be able to grow. Without the need for any kind of funding. We've been able to essentially just kind of grow our business organically. For VisionMed specifically, because this is going to need 510 approval for those recommendations, for the ability to pull up those reference clips, we are looking to take on a Series A round of funding. You know, we could basically consider that we did all of the initial work to build this the way it is. Clinically AI has invested a significant amount of money, you know, but We estimate that each trial for each medical device will probably cost around $3 million. And, you know, it just makes more sense for us at this time to try to bring on private equity or venture funding. So that is something that we are actively pursuing. We're looking to raise essentially about $20 million in a Series A round of funding. Okay.

And Mark?

Um, thanks, Tiger. So the, um, in the research that I did for the article that I wrote on ambient AI scribes versus other types of scribes and human scribes, most of the Palo Alto types, the UC Berkeley machine learning, uh, really smart guys out there, um, they're not healthcare people. And what I'm seeing happen with AI is the same thing I saw in the early 2000s with the EMRs that came one after the other. Most of them were written by really smart people that had never treated a patient in a clinic. And so much of the workload got offloaded to the clinicians. And, and of course the same thing won't happen here necessarily, but it is that uniquely healthcare clinical surgical perspective kind of parachute it into the IT brain that will really make a difference. Because it's really easy to put out a product that you— that makes you feel special. What's really hard is to put out a product that all of the other people think you are special. And that's, that, that's what creates market dominance. And Alyssa said something about marketing money, and you, you kind of echoed that. It could easily be the second or third best product with the first best marketing team because people that are buying it, they don't know what they don't know.

Yeah, that's, that's a great take.

This all looks amazing to the sort of AI illiterate brain. And most of us are pretty medically literate and we're learning to speak AI. And so It's pretty cool to see it happening in real, in real time.

Yes.

Yeah.

Our, our team is very fortunate that, you know, we did have that experience of coming from Johns Hopkins. You know, our team actually itself was founded in the Department of Radiation Oncology. And so that was kind of our background is that, you know, we were running clinical trials. We had that healthcare experience and we were fortunately able to find the right kind of software developers to be able to kind of build what we envisioned and what we saw as the researchers themselves as being what our needs were. How can we automate our jobs? That's what, that's our original goal was how can we automate our jobs? How can we make things simpler for ourselves? Hopkins really liked that. We basically deployed our software within their internal systems. They really liked it. They were like, hey, are you guys interested in commercializing this technology and selling it to other organizations? And that's really how we got our start. But they were, instrumental in that support of being able to help us launch kind of our suite of products.

Yeah. And the one common thread that I see running across many of these is that they save time.

That's the goal.

That's probably the easiest. I mean, I feel like sometimes when I'm listening to these that I'm kind of getting sucked into some kind of space-time continuum that I don't really understand. And I feel like sometimes I need to maybe a dose of psilocybin or some mushrooms to really, really see what's going on. So maybe we can arrange that sometime, Alex. I don't know.

We'll see. Yeah, I'd be happy to sit down like a one-on-one meeting or with, uh, maybe we take some psychedelics and we'll have a great time.

Yeah, I think that would create the best solution, I think.

Yeah. Next week's phone chat.

Mushrooms. Yeah. Yeah. So, man, I mean, what the opportunity— there's almost too many opportunities, you know, it's, I mean, we haven't even talked about clinical study enrollment, you know, or, you know, physical therapy, you know, watching a patient do physical therapy after surgery. I mean, there's just endless. You just got to pick. I think you really have to pick the right application. That, that adds value, saves time. Um, if David Kaye were here, he would say every hospital has labor problems and we need to figure out a way to, uh, take the burden off of the hospital staff. And this is— this answers that question in a little way.

Yeah, actually, uh, to touch on that just very briefly, I won't go into all of it, um, but our actual best-selling product at Clinical AI is our pre-screening system. That is actually the The original reason why we developed Maya was to be able to compare patient data from an EMR to different protocols in order to make eligibility determinations for large groups of people. So, for example, right, let's say that you're doing, say, just diabetes research, right? And you have 5,000 patients in your EMR who have type 2 diabetes. You're running 5 different diabetes trials. You want to know out of this very large patient population, which ones are gonna be eligible for each of those trials. We developed Maya to be able to do that. So it'll go through all 5,000 patients. It'll tell you, hey, maybe 137 meet the criteria for this trial, 86 meet the criteria for this trial, 215 meet the criteria for this trial, and so on and so forth. And that kind of works twofold. You know, it'll, it allows clinical researchers to dramatically expand that potential patient population. Most of them are limited by kind of the hours they're able to spend, how many people they have,, and they might be able to screen, let's say, 80 patients in a given month. Our software enables them to screen 500, 5,000, as many as they would like. And it helps them to reduce their screen fail rate and basically save time so that their researchers are focusing more on patients who are potentially eligible as of the latest information from the EMR, as opposed to patients who are obviously going to screen fail. You know, they get paid based on enrollment. They only want to enroll the best patients. They only want to schedule screening visits if it's going to lead to an enrollment. And that is kind of how Maya as a pre-screening tool can help them. We also do have a virtual assistant that we can train on different kinds of documentation. For example, like a protocol, and you can ask it questions. You know, you can say, you know, what's— what tissue samples need to be collected on the second visit, right? When should I schedule this patient? Are we allowed to, you know, enroll pregnant women into this trial? Things like that. Is a pregnancy test required?

And so that is another software we deploy.

Deploy for specifically for clinical research. We do a lot of work in clinical research, but they are very, very popular software amongst our clients. And again, the whole idea is that, you know, we're helping to make researchers' jobs easier, physician jobs easier, because there is a big shortage of researchers in the field.

It does sound like it would be incredibly useful for creation of us— creation and simplification of SOPs.

Yes, we also offer that as well. So we, if you go on our website, you can, you can see a number of different applications that we've worked on in the past. I also do have a more general client presentation for people working in clinical research. We have about 10 different technologies. If anyone would like a copy of that, I'd be happy to provide it.

Yeah, please do. We'll put it on our Slack channel. Yes. So is the interface voice always, or I guess it depends, right?

Generally, most people want kind of the voice interface because if you're kind of actively in the middle of a surgery, you don't want to have to walk away, go to a computer and type in questions. So it's going to have two options. One is going to be that Maya will be like Alexa. You say like, hey, hey Maya. Instead of saying like, hey Alexa, hey Maya, you know, pull up a reference clip. Hey Maya, save the last 30-second clip, right? So that would be one way to interface with the software. Some people do not want an AI to be constantly listening to them. I totally respect and understand that. In that situation, you know, we could deploy more of a, like we have a button, right? And you have to click the button and then Maya's not listening to you unless you do that. But again, it's fully customizable depending on what the client themselves want. How they would like the AI to kind of handle the information if they want it to be constantly listening to them.

Interesting. Everybody's just— wheels are turning, trying to— and you can create— this is a little off topic— you could create an agent for a company, right? For X Implant or illuminate. I mean, you could create an agent that just watches emails or traffic or videos and, and basically gives you advice, right?

Oh, yeah, absolutely.

That is— I don't want to say it's simple. But compared to some of the other technology we've developed, it's very straightforward. We have the skills to do that.

Yeah, I think, I think that's where everything's going. I think everybody's gonna have companies are gonna have AIs and people, or agents, I call them agents. And companies will have agents, people have agents, there'll be somebody on the look behind your shoulder, watching what you're doing, giving you advice all the time.

Yeah, like a virtual assistant. Yeah.

Like Maya.

Exactly. That's what we're doing. We're building people virtual assistants for a number of different applications.

So, you know what I call the hand innovation challenge? George Orbay, when he brought out that company, it did one thing. The distal radius fracture was the most common fracture seen in the clinic every month. And that was what he chose, and that's what he brought a solution to, and he blew it up and made more money than he'll ever spend. So with what you guys can do, so many different things, it would be super cool to pick one thing that, you know, is super common and go really deep, really vertical, and you will own the world.

I, I echo that. I mean, that's the— we've talked about this many times, but what kills startups is lack of focus. And what, what startups do really well, like, like Therese, you know, or Orbeez Hand Innovations, is they take one thing and they dominate, they go deep. Um, I once had a boss that said we have too many opportunities. He goes, I feel like a mosquito in a nudist colony. Everything, everything looks good. And, and within 2 years that company went out of business, you know, lack of focus.

Yes, we That is advice we have gotten from a number of our different advisors is to pick kind of one thing and really focus on it. So far, I guess we technically do two things. One is the kind of the clinical research side of things. That's kind of where we got our start. And then after building this, this is kind of our— we're splitting our time kind of 50/50 between the clinical space and essentially the video analysis space. But we're hoping that we're going to have the bandwidth to really work. I mean, the, the pre-screen technology, everything that we made for clinical research, that is kind of already existing tech. We don't really have to do a whole lot of work on it. Whereas, you know, this is going to need a little bit more development time, especially to integrate it with the different devices.

Are you, are you guys getting much traction in the CRO space?

For CROs, funny enough, CROs are actually like the smallest out of all of our clients. So, you know, we, at least for in the clinical research space, you know, we sell to sites directly, CROs and sponsors. We sell to more sponsors than anyone else, and then sites would be second. For whatever reason, we don't sell to very many CROs. We do have a few, but it's our smallest kind of market by far. Interesting.

They're, they're laggards. They'll come around.

They want the sites to purchase the software, right? So they see this as at least the, the pre-screen technology, the virtual assistants, right? They're like, oh, these are tools we're gonna give to our sites. Why should we pay for them? And the answer is, well, the reason that you should pay for them is because if you can help increase your site's enrollments, that's gonna reflect really well for your sponsors. And when you go to sign new contracts, you can say, hey, the reason you should pick us over our competitors is that we have these AI integrated tools. We're able to increase patient enrollment. At the end of the day, that is kind of what the sponsors care about most.

You should, you should definitely speak with the legacy PPD team. You know, PPD was bought by Thermo Fisher. PPD is unusual in that it has its own sites as well. It has its own site-specific business and that business is not doing well and they do need to differentiate. So it might be worth knocking on that door.

That's great advice. I unfortunately don't have any contacts in PPD, but I could try to reach out to them, see if I can find someone on LinkedIn.

DM me. I know some folks there if you want an intro.

Oh, I do.

Appreciate that. Thank you.

Yeah, also, I was hoping that Lawrence Binder was going to be on the call today because he's got a— I mean, he's producing thousands of hours of, of surgical video every week. Uh, he, he has technology where anybody in the world can dial into a case and see 10 cameras and chat, chat with the team in the operating room. And I mean, there's gotta be some synergies between— his company's called V Surgic.

So yeah, Tiger, we actually had a call set up with Lawrence.

Oh, good.

It had to be postponed, but we're trying to reschedule that. So I think that's a great potential.

Yeah, there's gotta be some synergies, Dave.

Yeah.

Yeah.

Well, cool. We're coming up on the hour. People are dropping off. Really appreciate it. That was really interesting, Alex. Thanks.

Well, thank you, Tiger. You know, I appreciate you giving me the opportunity to present. You know, it was great to meet all of you.

Yeah, send me— send me the materials. I'll put them in BoneChat Slack, and I'll also send you a recording of this call so you can play it back if you need to.

Will do. Yep. I'll send you a number of different items.

All right, guys. Good stuff. See you next week.

Thanks, everybody.

Thank you, guys. Appreciate it.