The Doctors’ Lounge
Where scalpels meet systems — and physicians say what they really think.
Co-hosted by Anish Koka, MD & Anthony DiGiorgio, DO. Candid talks on healthcare policy, reform, physician autonomy & patient care.
The Doctors’ Lounge
Dan Donoho on Surgical Data Science, Regulation, and the Humanoid OR
•
The Doctor's Lounge
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Episode Summary
Anthony DiGiorgio and Anish Koka sit down with Dr. Dan Donoho, an NIH-funded neurosurgeon-scientist at Children's National Hospital and George Washington University, and founder of the Surgical Data Science Collective and the Foundation for Digital Neurosurgery, for a wide-ranging conversation on AI in medicine. Donoho explains why surgeons get almost no structured feedback after a case, walks through a randomized controlled trial showing AI-coached medical students outperformed those coached by humans on simulated tumor resections, and argues that most regulatory "barriers" to deploying AI in healthcare are cultural rather than statutory. The conversation moves through the limits of tactile sensation in surgical AI, why he expects innovation to come from small agile practices rather than large insurance-pay hospitals, his family's data science lineage, humanoid robots performing laparoscopic surgery, the anti-data-center backlash, and ongoing surgical AI trials in Ethiopia and Tanzania.
Chapter Markers
00:16 Introduction and guest background
01:16 The Surgical Data Science Collective and the problem of lost surgical knowledge
03:26 How AI could replace the mentor's voice in a surgeon's head
06:50 Tactile feedback and the limits of video-based AI in surgery
08:47 The RCT: AI coaching vs. human coaching in medical students
11:48 What's the gold standard for grading surgical performance
14:42 Who validates the AI — the EKG problem and patient outcomes as the real judge
17:19 Humanoid robots, self-driving cars, and low-hanging AI fruit in clinics
20:27 Regulatory barriers to AI in EHRs — culture vs. statute
22:09 Doctronic, AI prescribing, and licensure
25:12 Where Dan's interest in data science started
26:39 His father, statistician David Donoho, and the rise of empiricism
28:09 The "don't go into medicine" narrative and why it's wrong
30:26 COVID as a case study in bureaucratic drag on technology adoption
33:27 Should the FDA regulate surgical AI products
35:50 Andromeda Surgical, humanoid robots in the OR, and the cost gap
38:27 The anti-data-center movement and the history of sabotage
39:54 Global AI competition and China's Kimi K3 model
42:41 How to measure whether an AI model is actually better
44:34 Model size vs. training — why prompting changes performance
45:46 What the surgical community currently measures, and why it's the wrong thing
47:23 The Tanzania surgical AI trial and other real-world deployments
49:03 What actually scares Dan about AI
51:21 Will AI democratize power or concentrate it
52:15 Closing questions — what's next, including the Ethiopia partnership
Co-Host Handle
@anish_koka and @drdigiorgio
Show Handle
@drsloungepod
Subscribe Links
Spotify: https://open.spotify.com/show/44vw8eirsKKnjgNIrdDvrR
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-doctors-lounge/id1832097658
YouTube: https://www.youtube.com/@TheDoctorsLoungePod
Dr. Daniel Donoho Website & X
Website: https://www.surgicalvideo.io/
🔗 Connect with the Hosts:
Anthony DiGiorgio: welcome to the Doctor's Lounge. Anthony DiGiorgio here with Anish Koka and our guest today, Dan Donoho. Dr. Donoho is an NIH-funded neurosurgeon scientist at Children's National Hospital and George Washington University in Washington, D.C. He leads AI initiatives throughout surgery, including founding the Surgical Data Science Collective and the Foundation for Digital Neurosurgery. He's been awarded several competitive grants, most recently, including funding from the American Board of Medical Specialties to create. Innovative approaches to surgical training. Today we're going to be talking AI, all things AI, robots taking over the world. Doctor Donoho thanks for being on the doctor's lounge.
Dan: Thanks, Tony and anish it's a lot of fun to be here with you guys. I think ⁓ longtime listener, first time caller. So ⁓ very excited to see you. I know the conversations can be wide ranging, so we're open to it all. Let's do it.
Anthony DiGiorgio: Right. All right. So tell me about the surgical data science initiative. What are you doing with AI? are you just focusing on surgery, all of medicine, all of healthcare, world domination? Where are we going with it?
Dan: definitely ⁓ we're going to Mars, but first we got a couple things to fix on this planet. ⁓ I think ⁓ lot the excitement around this stems from our real experiences as doctors and as surgeons, particularly. as an example, ⁓ I did a surgery today, you know, a couple cases today, did a pituitary case today. and during this case, I have, my mentors ⁓ continual running sort of ⁓ talk in my head as I'm operating. You know, ⁓ do do that, do This, don't do that, etc. ⁓ But the is at the end of that case, I get no feedback whatsoever. ⁓ I today wild. I had ⁓ a film crew, ⁓ five students, ⁓ the operating room staff, know, fellow, all this stuff. And I do this every case, I kind of look around and I realize, like There's nobody in this operating room who's seeing the same thing. There's nobody in the operating room who has the whole picture, certainly not even the surgeon. how we improve in that environment? How do do the best for our patients and maybe most importantly for our specialty and our field? ⁓ How do actually save the knowledge that we're generating during that surgery? Because I definitely learned I grew as a surgeon, as I try to do in every single case. How do we those learnings, but maybe even more importantly, how do we transmit them? ⁓ that key insight there's a broken piece in surgery. ⁓ Where the routine day-to-day learning and teaching doesn't make it out of the operating room. And think at the Surgical Data Science Collective, and basically in all of my work, is around this one core problem ⁓ of do we share and transmit and learn ⁓ as as we can from these incredible experiences that we're privileged to have as surgeons. ⁓ And do we get the insights from our mentors to live on and hopefully to benefit many surgeons who they've never met? We've tried a lot of things in the that maybe haven't worked and I'm really excited about the promise of the level of understanding we can get from tools like artificial intelligence from helping, use those tools to actually connect people around extremely meaningful work.
Anthony DiGiorgio: what does the AI do? How is AI going to? I have the same thing, right? I'm finishing a case and my mentors are in my mind. A lot of times I think I act a little bit too much like my program director in the OR. His voice even comes out. ⁓ so how AI gonna that? Is it gonna watch the videos? Is it gonna give us continual feedback? Is it gonna be like a life coach when we're done? What what are your thoughts?
Dan: Yeah, I think it's an iterative process, right? And so we have to start from where we can get the most benefit and balance that against risk and potential harm because I think as we've seen from a lot of these technologies, they are powerful. And that means they do have the potential to have missteps or, out and out ⁓ rampage. and ⁓ I think when we think about really important high-risk domains like surgery and technical crafts, ⁓ we have to be measured in our approach. ⁓ that being said, there are other areas where we can crap let things fly a little bit more. ⁓ But As far as surgery, I think we're really focused on understanding really important operative decisions ⁓ helping surgeons optimize their performance. So we start in the post-operative and pre-operative settings because those are the lowest risk areas where the surgeon ⁓ an intermediary between the information from the system and the patient. And start simply by ⁓ cataloging understanding patterns. And it doesn't sound very exciting, know, it's extremely profound once you can actually do that. You understand: hey, this is what, you know, ⁓ we have thousands of cases right in this catalog. And so what have thousands of surgeons done in this similar scenario is a core question, right? So we can break down these sort of game tapes, create, ⁓ an understanding of what the actual actions were of surgeons, and then feed them back to our, you know, our surgeons and our research team to understand ⁓ exactly mattered during that surgery, what drove that outcome. What were the other possibilities or ways you could have played that particular game? ⁓ And then it into the preoperative world of, you're Going into this case, what should you know about the last five cases that you did, the last 500, or the best five cases that anyone's ever done for this kind of operation, for this kind of patient in this scenario? ⁓ And I that's really those are the most important, you know, terrains to start in. ⁓ Meanwhile, ⁓ does escape us that the exact same system capable of doing these things ⁓ could provide profound insight during surgery. but I will say that the way to provide that information ⁓ is not settled. It is not exactly. Exactly the best way to mix human and artificial intelligence during a very time-sensitive performance. the common thing if you're, coaching elite performers, right, ⁓ is ⁓ you'll tell a of stuff potentially before surgery, before an before a game. But during the game, I think really good coaches are really effective and minimalistic at what they say. They're only communicating what players need to hear. How do we get an AI to do I think ⁓ is not 100% subtled. we're very excited to be at the forefront of efforts that are trying to figure that out and trying to figure out how to bring that into training, into early practice, and to really, get the best out of everyone who's trying to master their craft.
Anthony DiGiorgio: So AI is essentially watching the video, I assume you're plugging in pre-operative imaging ⁓ ⁓ a history. ⁓ there's a lot in surgery that I don't think AI can capture. Like it can't really capture like the the tactile feel, right, that the tissue gives us. at least not yet. and I so much of that is just, the instinct of of how does the tissue feel. I don't know, is the video capture technology three-dimensional enough where AI can pick up depth? Like it seems, AI coaching me through a surgery, yeah, maybe saying, you've made one, two extra movements here versus there versus AI, like truly getting that same instinct that I as a surgeon as I feel the tissue planes, So I'm putting a pedicle screw in, there's a there's a certain feel to it that tells me that it's good or as I'm dissecting tissue off certain tissue planes have a very tactile feel. Is AI gonna get to that point too, or are we really just looking at, yes, you could have made a different cut here or took one too many extra steps there.
Dan: Yeah, I think humans operating do rely on a variety of senses. But as a thought experiment, imagine what would happen if you told your patient you're gonna go into the operating room with the blindfold on, right?
Anthony DiGiorgio: Wait, you don't tell your patients that?
Dan: I I don't routinely tell my patients that. But we all wear one or two pairs of gloves. We diminish tactile sensation really, incredibly. so I don't want to say it's unimportant. I think it's really important, but I wonder, particularly for MIS and and endoscopic cases, ⁓ how much of that is sort ⁓ able to be parsed from video. we're very excited to, ⁓ systems that have some physical intelligence, are still, ⁓ I think, a little bit of a ways away, there are some exciting new models ⁓ that have come out in the last few months. That I think show the promise of where this technology can go. if it can understand and create physical relationships and physically intelligent AI, ⁓ think the promise of a technology that understands physical properties in the world ⁓ is coming. I don't want to discount that there's a lot that goes into an operation. ⁓ but I think All of these problems, ⁓ are whittled away. And, what would have said, was absolutely impossible. There's no way we could get this kind of performance out of a thing ⁓ four years ago, five years ago, eight years ago, ten years ago. ⁓ Each of those barriers is, being in certain ways. I don't want to be overly optimistic because I do that there's a lot, I see this every day, there's a lot of difficulty in bringing the promises of this technology. To bear on the real world, is extremely must ⁓ messy and sort of human-centered in very important ways. But I think the raw technological barriers, sort of the things that are solvable with compute, ⁓ the things that ⁓ could potentially understood, I think we're going to see more and more of those falling away and things that we thought, ⁓ there's no that the system could understand this, or there's no way that this property could be measured. I think, the answer from, a lot of people is like, okay, hold my beer, right? We're gonna go figure this out and we'll report back. So that's where I'd leave that particular question.
Anish Koka MD: clarify for our audience, where we are ⁓ based on some of the trials you've done. ⁓ there's trial ⁓ you actually did a randomized trial with ⁓ Canadian medical students, is that correct? Can you ⁓
Dan: Yeah, yeah. So there's there's a lot of trials that have been done, particularly in education, that tell us that, particularly when it comes to teaching and learning, ⁓ where, we would have said that there's, there's no way if an AI can't even feel how pedicle screw goes in, there's no way he can teach a good pedicle screw, right? Or can't feel a brain tumor section. I think that in ⁓ particular scenarios, which were, in simulation, to be clear, we're not, turning medical students loose on the ⁓ unsuspecting populace, or at least yet. we found that we can't know human performance until we measure it. So this case, measuring the performance of humans as teachers. And once we measure that, you can be surprised. ⁓
Anish Koka MD: So what so what exactly s so ⁓ but tell me what exactly was the trial? How what was the randomized control trial? What was the design of it?
Dan: Yeah, so the the trial there were sever actually there's a number of these efforts that were done over the years. But essentially in this in this crossover design, the students were ⁓ randomized to either receive feedback from ⁓ a human right, or to receive several of different degrees of feedback over the years, but sort of culminating in ⁓ a full that could give them real meaningful feedback about their surgical performance in a brain tumor resection task. So you I mean, you guys know this, you do this all the time, but for our audience, ⁓ removing a tumor is one of the
Anish Koka MD: Okay.
Dan: Riskiest things, that we can do in all of surgery. And ⁓ I think learning how do that is often left until sometimes the later years, even of neurosurgical training, of ⁓ being, the who's actually doing the difficult part of the case. ⁓ And ⁓ so in scenario, were able to take a group of medical students measure their force, their bleeding risk, and many, many other parameters in terms of how effectively, ⁓ how riskily, and how efficiently they were these simulated tumors. ⁓ And either try to coach them and improve their performance with humans or with AI. And surprisingly, we not only found that AI led to more ⁓ performance gains in the students in this actual task, but that when we switched students from getting AI to getting human coaching, they got worse. And this is ⁓ unexpected and maybe controversial. But ⁓ and similarly, when we switch students from human to AI feedback, they again improved much more quickly, ⁓ as though received perhaps little performance improvement at all. So I think is one example of a nice piece of work that shows that one of the domains that we thought ⁓ really would be as you said, what can AI teach me, right? How can it do all of this stuff? ⁓ I think there's a real that we have to measure the benchmark ⁓ of human Performance before we assume this. So, you know, I think very, very recently, Ethan Go had this nice letter with this sort of laundry list of all-star co-authors in Nature Medicine ⁓ talking about the need for human benchmarking of human performance tasks. And secondly, something that I completely agree with of measuring meaningful outcomes, not just looking at individual care processes when we're adjudicating whether AI works or not, but actually looking on whether it improves the ultimate outcomes that we really care about ⁓ terms of patient care. I think the more and more of those two things that we do, the better off we'll.
Anish Koka MD: You it's always hard to get at what the gold standard is ⁓ in many things that that we're doing, especially in when you're trying to push the field forward. So what is the gold standard in ⁓ grading medical students who are attempting to do a resection in a ⁓ cadaver or simulation? that's a simulation ⁓ virtual simulation. Okay, So sorry, so this simulation ⁓
Dan: In in a simulation, a virtual simulation in this case. But we've I've done cadaver work in other studies too.
Anish Koka MD: what's the gold standard? Because you have AI kind of giving a score to the students, correct? And then you have nurse surgery, ⁓ I assume veteran experts like yourself kind of looking at it and giving also giving a score or ⁓ what's the yeah.
Dan: Mm-hmm, mm-hmm. mean I think there there is a scoring component to it, but part of it is actually objectively measuring the task. So this is sort of analogous to stuff we've done in cadaver work where you can actually measure blood loss or whether someone succeeded or failed in, a binary task. in this particular model, they can measure the force applied to tissue, the amount of injury that was caused, the amount of ⁓ hemorrhage that occurred, ⁓ right? Because this is a fully paradigm. But even in the physical world, right, in let's say cadaveric simulators, right, we have those objective measures that we can create. And once we have those benchmarks, right, once we have those real-world data points, we can then model them and predict them using ⁓ that maybe we wouldn't have thought were possible.
Anish Koka MD: fascinating. there were human experts in this I'm sorry, I keep going back to this one trial. I know you did a bunch of them, but it was so interesting to see how how you did this. I would never have have thought I just didn't realize that ⁓ AI was at a point that w it could actually do this. So it it's fascinating to me. I I'm curious if right, okay.
Dan: Yeah, and this was old AI. This this was, you know, really most of this work was done now, years ago. ⁓ So when think about model generations occurring ⁓ over weeks ⁓ or ⁓ right? again, I don't want to be the straight line go up kind of guy, right? I think the world is a little bit more complicated than that, ⁓ but
Anish Koka MD: Right.
Dan: it just shows like something that we do all the time is we, talked about this at a talk I gave at the Society for Robotic Surgery this past weekend, ⁓ it's really important for ⁓ I people in a particular craft, medicine, whatever it is, to have your own internal test for AI, right? Your own reality and sanity checks that mean something to you. so in this one example, there's a mathematician ⁓ who at that time was working at Microsoft and working with a lot of open AI guys, ⁓ and his thing was. I'm gonna ask this particular task of an AI, it's gonna draw a unicorn using a particular kind of software library and a particular kind ⁓ sort of manuscript preparation software. And we're gonna see how good it is, right? Other folks have different tests that they ⁓ And once have one of these things that you can throw at the AI, over and over again from these different systems, you can chart their progress. So in our world, we look at you know its ability to understand surgery, to decompose a surgery into its consistent. Gestures, steps, and phases to predict its outcomes. And I just say over the last three or four months we've been continually surprised to see how much progress has been made. And I think these kinds of benchmarking efforts where we keep the AI honest using things that we really know and understand are really critical. ⁓
Anish Koka MD: Dan, who's who's validating the AI? we've had some type of AI at EKGs, and putting out these automatic interpretations. most of my consults from neurosurgery are because, ⁓ the computerized algorithm says abnormal EKG. And then it's like, well, call Koka. But but ⁓ but so who it is, it is it it it no, no, no, I I like I it. ⁓ I I love it.
Dan: Hmm. Yeah, well it sounds like it's keeping you in business, so why are you biting the hand that feeds?
Anish Koka MD: who's validating these AI in terms of its interpretation of whether or not the trainee or whoever is doing a good job? Like do the experts agree with most of the time, whether it be the model, that initial model, or maybe is it improving now with the current models? and the problem there is that
Dan: Yeah.
Anish Koka MD: I've come to know plenty of neurosurgeons in my short career so far. two ⁓ skull based surgeons, one at ⁓ X place, another trained at Y place. If look at the same video, are both gonna be like, Whoa, what the heck is this guy doing? Like he should be doing this. I mean, what's the gold standard?
Dan: Yeah, yeah. I mean I think to be honest, ultimately the patient is the judge, not the surgeon. Right. And I think that's really the critical distinction. We can look at all of these operational things of do surgeons agree. And we've done these studies of radar agreement on various judgments that surgeons can make. Those are all great, but they're instrumental, right? Where we really need to go to your point about EKGs, right? Yeah, it's great that we flag abnormal EKGs. Maybe I don't know. I mean, I haven't, I forget what the letters are even of an EKG, right? but what ultimately is does it save your life? ⁓ Right. And so there's lot of really interesting. So Ziad Obermeyer had a ⁓ recent paper around like population level EKG analysis showing that when you look at large scale amounts of of EKGs, you can predict critical events, right? Even long after those EKGs occur. You can diagnose other phenomena that aren't commonly thought of as present in an EKG itself. things ⁓ like upcoming cardiac events and even potentially other things as well. ⁓ And I think that's you know it's one direction that we have to go in. I think to your point about I don't want to lose your point though, which I think is really important, which is a lot of the technology we use today in medicine sucks, right? was made ⁓ decades ⁓ right? doesn't even incorporate the things that we've learned, ⁓ that basic, ⁓ first-year computer science student in could, vibe code up in a night, could probably beat any of those automated EKG algorithms that come pre-installed in your EKG machine. ⁓ Right. We could talk about, the broken of medical innovation, and challenges in regulatory and in all of those things, right? That's your answer for why your EKG sucks. And my personal opinion, it's don't blame the technology. Blame the barriers that are preventing actual real technology from getting to you and getting to your patients at the speed that technology can improve.
Anthony DiGiorgio: I think that's a really good point. And the futurist in me says, Yeah, but probably at some point in the future there's gonna be humanoid robots that can do a craniotomy in the middle of the night better than I can and don't get tired. But I think that's probably a long way off, right? We can we still don't even have driving cars that self-driving cars that work on, most roads in the US without some sort of human input. ⁓ right, 'cause all the Waymo's have a human behind them at some point in that chain. ⁓
Dan: Yeah, one to forty, I think, is the the rumored number.
Anthony DiGiorgio: Yeah. So, yes, maybe we'll get there, but there's AI applications right now that we're really falling short on. Right. So we in a lot of clinics, they will pay a human to go through the charts for the clinic for the neurosurgeon the day before clinic ⁓ and collect all the in a way that can be summarized for that neurosurgeon before they go into the exam room. Right. That is like the the right now AI can do that probably better than that human can. Yet we're still paying a human an extraordinarily large amount of money. typically it's like an advanced practice provider. So someone making, well into six figures gets paid to do this that, the current models of Chat GPT could do right now better than that human can. So why aren't we? I mean, this is like low hanging AI fruit. Why isn't this being done right now?
Dan: I think it is. I think it's just not evenly distributed. And why are our broken businesses in medicine not applying best practices? I mean, I think that's a tale, a tale as old as time. I think the nice thing about, you know, innovation at the pace of software is really are, real-world applications that ⁓ you can, and apply, and they're HIPAA compliant and cybersecurity certified and all of these things ⁓ that do this. But think that ⁓ the broader lesson is deployment of technology into healthcare requires Sustained focus and iteration, working hand in hand with a customer that is a really arduous ⁓ intense process. ⁓ it's a lot easier to sell a consumer SaaS tool, ⁓ You want to be your first vibe coded, one employee, a billion dollar startup, ⁓ right? You're probably not do that in healthcare, ⁓ right? You're probably gonna that in I don't know, some you know absurd consumer application, right? SMB application. but I think there are some reasons. Really positive signs. are entities that are really interested in funding disruptive technology companies in healthcare because ⁓ these exist. And guess I would say, like, ⁓ where do I think that innovation going to be distributed? ⁓ I wouldn't go to an insurance hospital and expect anything like the best in class technology. I actually think you'll find small agile, operators of practices of ASCs that are willing to innovate, potentially smaller, ⁓ health systems that. understand this and sort of imbibe this from the beginning and say, hey, we want to disrupt this. And I think we see real world examples of how that can drive, ⁓ tremendous that it can improve quality of life for healthcare providers. And yeah, happy to happy to go into more details about those things, but I think ⁓ the answer is like in healthcare is really hard, particularly in calcified, insurance pay And ⁓ innovation is going to occur elsewhere. Those hospitals, personal opinion, they all want to be second. But there are places that want to be first, ⁓ we should look to those as our examples.
Anthony DiGiorgio: See, Koka I knew he'd belong. He wants more competition to drive innovation, just like us. I mean, I think that's a great point, right? So ⁓ think the large hospital takes a big risk by deploying something, it has a data breach or doesn't work they spent a lot of money on it, and now that hospital's liable. ⁓ and please feel free to the question or change the subject if you're not aware, but like what are some of the regulatory barriers with getting you know, like just a basic AI agent into an electronic health software. ⁓ what sort of for EHR there's a list a mile long of the, requirements in order to meet the Medicare standard for an EHR. Are there those same requirements for an AI agent aside from it having just to be HIPAA compliant?
Dan: Yeah, I'll just say my personal opinion again to sort of differentiate this and to, at the risk, right, of being a little bit speculative, ⁓ my answer is there aren't really any. which to say, like none these actual barriers, if you go to statute, if you go to compliance, which I've done in numerous entities, that I've been involved in, ⁓ you really can meet these barriers. Like it is really possible to do it. The problem is the change management, the problem is the culture and society around these entities. It's actually not necessarily the regulation. Now, ⁓ look, okay, building an EHR that's building Medicare, right? Like that's but that's like the last place I would ever want to do this. Like, why would I play that game? That game is not fun. Like, I think there's much more satisfying ways you can actually enact real change, touch millions of lives, and improve care for doc for patients and improve quality of life for doctors. and let that just be what it is. Like those organizations want to behave that way. That's okay. Like, I'm not trying to lead people kicking and screaming. I'm just want to show them that there's a better way. And when you're ready, here it is.
Anthony DiGiorgio: Love it. Totally agree. ⁓ Doctronic, what are your thoughts? So there's now an AI agent, AI app, whatever you want to call it, that has a license in Utah, right? Utah licensed this app to now prescribe medications. is the only thing standing between us ⁓ at least the some of the ⁓ thinking specialties from replaced by AI, is the only thing standing between that? licensure and regulation. I mean I heard somebody on Twitter was talking about how, AI is making more demand for doctors because people are gathering all their AI questions and going to the doctor and now they want their questions answered and then they want the prescription, right? The end point for a lot of the of physicians is a or an imaging study or some order. If AI is allowed to place these orders, is that gonna be a quick end to a lot of the the thinking specialists? Is Koka gonna have to go learn a new trade?
Dan: Yeah, I mean I think the the good news is, right, Or the bad news is, depending on, how you look at it, these things not going as quickly and as widely as you think they are. Like I don't think that there's a, ⁓ gold rush of people to be, the doctronics of the world. I think reimbursements are hard. I think value is in healthcare. and I think patients usually come second or third ⁓ best in a lot of these conversations. So I think what, the narrative may be San Francisco may not be the narrative in the Midwest or the Southeast where the majority, ⁓ a lot of American healthcare is actually delivered. ⁓ and you're gonna see real discrepancies and disparities in the use of technology in those kinds of places. So I was ⁓ kind of in love. love the fact that this happened in Utah, right? And, not in, the Bay Area. and you know full disclosure, I'm from the Bay Area. I'm a card carrying, member, so I'm trying not to, ⁓ be overly zealous about the other side. I think there's there's an So I completely disagree with the point that ⁓ people just go to the doctor to get their ChatGPT questions answered. They go to ChatGPT to get their ChatGPT questions answered. And ChatGPT will be increasingly satisfying in answering those questions, and that will be the end of that. I think if that's your business model as a cognitive specialist, it's a loser and and you should really rethink things. I think that there is a lot of organizational power and organized healthcare. I think there are a lot of people trying to break that down, but I don't see that regime changing outside of very specific contexts. Like I think one of the battles, of Doctronic is I, understand it, and again, I this is really not ⁓ not my world, but can we replicate this? Can we actually even sustain this? ⁓ Right? Will payers stand for this? Are going to, actually see this flourish? Or is this going to die in a bunch of legal battles and a bunch of refusals? And it'll be some sort of curiosity that is nice to have, but really doesn't end up moving the needle. I'm not sure that that's actually the way to ⁓ get, let's say, healthcare abundance if subscribe to that or want that narrative. I think there are better ways to do that, in my personal opinion, that will have a faster impact on human health in the United States the next 10 years.
Anish Koka MD: Dan, where ⁓ switching switching gears again, as we said, this would be wide ranging. Where did this interest come from? are extremely bright, of course. ⁓ but this interest in ⁓ science, machine learning any ⁓ place that arose from?
Dan: Yeah, I mean I think I've always been fascinated by data. I think I was, making spreadsheets on paper, ⁓ with paper and pen ⁓ when I was like six or seven, to track various sports statistics and various other things. So I think I've just always been fascinated by the idea that you can quantitatively understand the real world around you. I studied economics in college, I wasn't a med. and I think that, kind of thinking about ways modeling the world and you know, the late nineties and early 2000s was really fascinating to me. ⁓ suddenly, ⁓ 20 some odd years ⁓ now we all have tools at our fingertips that if you just know how to ask the right, questions will give you really fascinating answers, help you construct complicated systems that ⁓ would have taken many, many person of work that ⁓ really can extend your ability to understand complicated questions. And so I think it's just been a part of my life the entire journey. journey and I certainly ⁓ loved, being neurosurgeon and I think I love taking care of my patients and ⁓ I love a lot what I do. But ⁓ I think this part has really been there for my whole life and I think that's ⁓ probably core to identity, I guess I would say.
Anish Koka MD: Is your ⁓ dad is David Donoho? The the the statistician he's a statistician at Stanford, is that right? ⁓ Because he's no shrink in world of ⁓ data science, correct?
Dan: Mm-hmm. That's right. That's right. Yeah. Yeah. Yeah, yeah, I'd say he's ⁓ he's a fairly influential, thinker and effector. He's done a lot of phenomenal things and it's been a a privilege to, be able to learn and understand maybe a small fraction of what he does. but yeah, I think ⁓ seeing the change in, the rise of empiricism in the mathematical sciences more broadly, ⁓ some people would say, the death of theory. I don't know, not my field, but I've certainly heard that said. it certainly seems interesting when, every ⁓ one of my ⁓ friends, collaborators posted something on this. if you want if you know, the I think his take was, if you want any of your problems solved, just call them a math conjecture and somebody on Twitter will solve it in forty-five seconds using, chat GPT. So I think, ⁓ it's definitely times. It's been incredible to see the evolution, of that field from my decidedly ringside seat ⁓ and learn as much as can.
Anish Koka MD: your father was ⁓ I guess at some point disappointed where statistics and data science were going, in terms the real world application stuff. So I mean these last four or five years it'd be curious what your perspective or what your conversations have been in terms how seen things kinda progress. Is Yeah.
Dan: Yeah, I mean I'd let his, you know, his his published work, of which there are, you know, a few cited papers, ⁓ speak for themselves. but I think the perspective that I can certainly share is that it's in equal parts, maybe dispiriting and invigorating. I think there's a strong narrative that is being right? When you see things like ⁓ frontier math, when see people, really exemplary mathematicians sort of ⁓ lending their ⁓ credibility to the language model ⁓ attempts ⁓ sort of solve ⁓ for an entire field. I think that one of the narratives that gets pushed is like, similar to what we hear, right? Don't go into math, ⁓ right? Don't into medicine. ⁓ I think that was, an Elon thing six months ago. I showed that talk at a ⁓ neurosurgery conference that was speaking at. I was like, this is what Elon Musk is telling you to do this morning, right? Like, ⁓ don't do that, ⁓ right? Do not obey advance. think this is it's really for young people, for old people, for everybody, avoid this sort of delusional disillusionment. ⁓ we're working on a good term for this by the way. but saying that, ⁓ I just have to give up because AI has solved my field and so there's nothing meaningful that I can contribute. I think there's ⁓ a longer piece on this to come in the media outlet, but I think the the short answer is ⁓ there are a lot of forces that want people to believe that and are very important ⁓ to those particular ⁓ entities for people to believe ⁓ in that statement. and I think the truth is actually the opposite. It's never been a more exciting and invigorating time to be a neurosurgeon, to be a cardiologist, ⁓ maybe to be a I'm not quite sure. but certainly in our field, the ⁓ ability to create ⁓ and to explore and to with the assistance of a lot of tooling that's much easier to ⁓ is really incredible and and I love this time. I mean I think it's ⁓ my I I said I have five students in the operating room today I said it's never been a time to better time to be a doctor in my opinion. ⁓ because what it means to be a doctor ⁓ in their generation, right, they're one some of them high school kids, right? They're gonna if they go into medicine they'll be, they'll be in practice maybe in ten years, ⁓ right? It's gonna be so different for people recognize the differences while superficially probably seeming almost the exact same. Yeah.
Anish Koka MD: Yeah. it's fascinating to me, you talked about this earlier and I'll just kind of reiterate ⁓ the significant gap between where technology is and how it can be deployed, and then how it's actually deployed in as you called it, you know, these kind of failing systems, right? And there's no better use case for that than ⁓ COVID, where we've able to do telemedicine for long time. we've been able to, ⁓ zoom into conferences for a long time, and yet Yet the bureaucracy and administrators refused to allow that to happen just because they had control and they weren't allowed to happen. And then the moment they suddenly had these parameters these kind this straitjacket of well you can't have people in a room together because, you know, ⁓ because of COVID. ⁓ technology became suddenly everyone's and everyone's, ⁓ zooming into this and zooming into that. So it'll be Anthony, I talk about this all the time in terms of it does feel like in order to really advance ⁓ in a way that can utilize technology to it actually is, ⁓ it requires kind of creative it requires destruction. You know, ⁓ this Joseph Schumpeter you know talk a lot about creative destruction. and ⁓ do do that without kind of destroying what currently exists? So people get upset about talking about destroying these great institutions, these great health systems that we have that amazing stuff does happen in. But I don't know, it seems very hard to ⁓ imagine a where
Dan: Yeah.
Anish Koka MD: These folks are still controlling everything, because they it does seem to be a block on innovation in many ways.
Dan: Just depends where you want to innovate, right? I think in what layer and what way. I think I mean I I'll just ⁓ my, particular personal take and experiences. ⁓ It's never a better time, in my opinion, to do innovation in healthcare. ⁓ I mean, much ⁓ much more exciting, more vibrant than any other time that that I've been involved, right? I mean, ⁓ you back maybe 15, 20 years ago, innovation basically meant like consulting for a big med tech company or strategic to help them, market their latest device. I think today innovation means building your own product, going out into the market and finding willing adopters, you know, hopefully, of said product ⁓ and seeing it adopted in real clinics impacting real patients' lives and doing so at a time scale of weeks, ⁓ days. ⁓ and really be builders. we're the people who understand these problems because we live in these systems ⁓ and, are confronted with I think it's for me invigorating because I think lot of medicine is a learned helplessness model, ⁓ right? You're sort of standing on this, and you get shocked unpredictably and then you go back and you're just cowering in the corner and someone opens the door, you don't even leave. I think this is a huge antidote to that mindset. you know, through I think, folks like yourselves and many others who are showing, ⁓ a lot of to be creative. And is there gonna be some destruction in that? I don't know, like ⁓ let Godzilla fight, ⁓ fight Mothra. I don't know, let's let it happen. Like, let's just see. but I know that, for me, ⁓ I want to be out there building and advancing the state of play across a variety of different domains, variety of different concepts, and ⁓ just sort of
Anish Koka MD: where do see regulation? ⁓ We talked touched on this with doctronix of but where do you regulation fitting in this? Like for instance ⁓ say a commercial product that videos surgeons, gives feedback, that something that should be by the FDA, or is it best not to be regulated by some ⁓ FDA type?
Dan: Yeah, I mean I think it's a it's a perilous question for ⁓ somebody who may be facing the agency not not terribly far down the road. but I think ⁓ I think ⁓ ⁓ we had heard from ⁓ a ⁓ I guess former head of FDA you know at one of our national meetings a few months ago, who I can now blame him for this, right? Who shared his opinion that such products actually ⁓ because the surgeon is actually making the decision, that such products ⁓ may be outside of the purview of the FDA. and I'd say that was a very invigorating opinion ⁓ to hear that. what does this actually mean? I think seeing and knowing a little bit about some products that are coming out in the market, some of which I'm again investing in or some of which I'm involved in, ⁓ I think that there really are going to be slew of such products that ⁓ the regulatory barrier ⁓ again, it may not be as high as people think. Like it's really important to push this stuff through and see what is actually required. And if you Do this in an intelligent way, even in the current Ancien regime, right? Like, can you find a way that you can do this with just human factors in a 510K? Right? Like, how can you actually do this? I think, I think it's more feasible than people think. And think there may be tailwinds. and there's some, I think perhaps not for the interoperative, performance analysis and performance improvement and guidance, and taking over the surgeon and being the co-pilot or co-brain of the surgeon, maybe that will be a high regulatory bar. But for a lot of other things that we've talked about. I think there are ways to implement these products that can be done intelligently, that can conform even to very antiquated regulatory standards. And I don't think we need to wait for some magic moment when the floodgates will open and now we can innovate in healthcare. And now we can have AI agents in healthcare and now we can have, all of these amazing technologies that are available to everyone else on earth. I think we can have them now. I just think we have to do them intelligently. We have to do them with the patient in mind ⁓ and to do them understanding the business some medicine if you can do those three things like I think it's a product you can sell tomorrow.
Anish Koka MD: Without the FTA.
Anthony DiGiorgio: Are we
Dan: Again, I think that's where it comes back to doing it intelligently, right? I think that's where it comes back to, doing this ⁓ with an understanding of what the current regulatory structures really do and don't say, not what they're imagined to say. ⁓ some of which in conversation with the agency, some of which perhaps ⁓ I don't want to say in opposition to, but maybe in more of a sort of firm stance. And there's a company out there ⁓ that, I'm not invested in, but but love a lot. company and Andromeda Surgical, ⁓ and their mission is to provide a platform intelligence technology for autonomous surgery. Right. they got Health Canada clearance, right? So they're ⁓ they're out and about, right? They'll in FDA soon. ⁓ and ⁓ I'm assuming I don't know. But this is coming. how are we going to we talked a little bit about humanoids in the operating room. ⁓ so to go back to this for a second, there was a cool paper by ⁓ Mike Yip, who's at ⁓ UCSD, where they basically chained ⁓ a robotic controller, like you can think about the control station for Vinci Robotic Surgical System, up to the ⁓ arms of a humanoid robot operating conventional laparoscopic instruments. And so when the surgeon Would move their hands in the robotic control system that currently exists today. The humanoid robot would mimic those movements in the patient. Of course, it's an animal model. And there were, real problems, with some latency, real problems with some accuracy, but they were able to accomplish the surgical procedures ⁓ in these, in vivo models, even despite some of limitations from very ⁓ early, early studies. So how do we that? Nobody knows, but I'll tell you, you can buy a humanoid for $30,000, you can buy a Da Vinci for question mark, but maybe 30 times that, maybe 50 times that, right? I'm somebody's figure it out, right? Because that margin, right, to the the prior turn of phrase is a huge opportunity for some enterprising young person out there who's gonna solve for this.
Anthony DiGiorgio: Well not only can you you buy a humanoid robot, you said what, thirty thousand dollars? It's a lot cheaper than than a than a neurosurgeon, right? So I don't know.
Dan: Something like that. Absolutely. Right. Yeah. So where where does this converge? I think I think we don't we don't know. ⁓ but I think it's a really exciting time. I think like all of us, right, we wanna, we wanna create the future. and I think that's it's again, it's never been more feasible to do that. So I think it's that's if there's one overriding message, Is this overly optimistic? Is this Pollyanish? I think we're putting things out in the real world across a variety of things that I work on. and are not theories, right? These are, ⁓ real things that that we really have conviction in. And I think I guess ⁓ we'll the receipts and we'll what happens.
Anthony DiGiorgio: What about the Luddites? the anti data center movement. are data centers poisoning our water? Are they driving up energy costs? Do should we be banning them or are the Luddites gonna stand in the way of our progress?
Dan: Yeah, so I showed this fun this fun slide at ⁓ at CSNS. ⁓ actually the first sabotage, right? So Sabot, right, throwing the shoes into the thing. The thing that was being thrown the shoes thrown into ⁓ was actually, an autonomous system back in, you know, hundreds of ago. So I think this tradition of trying to, slow down progress and I think is ⁓ ages old. I think it's dangerous time. ⁓ Right. I think it is a violent time. ⁓ and I think we are going to see, I think quite regrettably, ⁓ potentially significant violence around ⁓ some of these ⁓ based on that I think are largely incorrect and are being pushed. ⁓ but overall I think ⁓ the need for here, the need to improve human conditions, the need to create greater human opportunity ⁓ is just too and ⁓ I don't ⁓ it's reasonably likely that those kinds of movements will will really have any meaningful success. They may reshuffle the deck of where these things happen, potentially, sadly for the economic advantage of of certain areas. But ⁓ I don't that, these movements are to be in the large part successful in ⁓ an increasingly, global world where there's global need and global
Anthony DiGiorgio: are there other countries that are ⁓ that are gonna compete with us on AI or is this a US dominated thing?
Dan: Yeah, there's no question that there's a global competition right now. there's no question that this includes what we might benignly call countries of concern or simply say China. but I that so I'll I'll say this as somebody who downloaded the Kimmy K3 weights the other day. very excited, gotta figure out some things of how to run a three trillion Prams model. And, you know, there's a lot of a lot of learnings happen around this. but I think ⁓ there's lot of strategy and games playing that's occurring right now ⁓ around technological advances, a lot of, interesting moments and movements. But ⁓ ultimately, I think that this kind global competition ⁓ really is for the better. I think will encourage folks to ⁓ continue to put their best work out there, the best products out there because they know they're gonna be challenged. from a global standpoint or national security standpoint, say that's probably outside of my area of expertise. But we're watching those things and I think it's ⁓ it's fascinating to think about, ⁓ deep seek markets basically as a hedge fund, ⁓ creating a large language model. ⁓ there's a lot of strange and fascinating things that are occurring for people who are paying attention.
Anthony DiGiorgio: ⁓ you used a couple of terms there that went over my head, so and you gave this great talk that C S N S that also largely went over my head, but our audience is a lot smarter than I am. So yeah, you explained what you meant by what was it, prams and and chimney three something?
Dan: Yeah, so like the news of like last week in technology, which is already now boring and we're on to the next thing, was there was a Chinese lab called Moonshot AI and they released a very large ⁓ ⁓ generalist foundation model ⁓ that they Kimi and the version is K three. and this is ⁓ an open weights model, which means they give you ⁓ all of the internal sort of structures the model that allow you, if you have the computational abilities. So for a model of this size, require, depending, there's some you know inside baseball of how you do this, you can do this potentially on some smaller devices with some tips and tricks. But the vanilla, benign, not really ⁓ thoughtful way doing it, not there's anything wrong with it, is just you get $100,000 and buy a very significant amount of computation to be able to run these models efficiently. And ⁓ a parameter count is a way of ⁓ thinking about the size, the number of sort of different you can. can think about it as nodes within the or neurons within the neural network. It's not not quite accurate. ⁓ one way of thinking about it. And so this model was sort of order of magnitude 3 trillion. And ⁓ ⁓ we don't know exactly if you think about the latest claw models or open AI models, but they're probably ⁓ around that order of magnitude. ⁓ a of the models that you might try to run on your laptop at home ⁓ would significantly smaller in the few billions. So we think that these larger models are more capable, and so far they've given us some very interesting results in early testing.
Anthony DiGiorgio: So you don't is that how you measure I guess the advancedness of the AI models, these nodes or prams, or are there other tests that they can to see how good these models are when they stack up against each other? I seem to remember some graph that you showed during your talk.
Dan: yeah. So I definitely don't think that model size tells us much about ⁓ performance, right? Like ⁓ I could give three trillion parameters right now of ones and zeros and it wouldn't do very much. but ⁓ I think speaks to some extent to the complexity and perhaps capability of these models. The best way to measure these models is actually by having meaningful tasks that a community agrees on and is new are neutrally adjudicated. ⁓ those kind of competitive frameworks ⁓ benchmarks are really how I think most of us would look at a model and say, is it better than the prior generation? And ultimately it for each person, ⁓ right? For each use ⁓ right? There's, these things are ⁓ unusual and they're Behavior and they can get better and worse at the same time at different tasks. so ⁓ goes back to sort of the what is surgery's unicorn, drawing moment or an autotyping on an airplane, ⁓ right? What is your personal evaluative set for model? ⁓ I strongly encourage people to have one, to a perspective, to say this is a kind of task that really matters to me. ⁓ And when I want to know, ⁓ here's I'm gonna measure this. And it could be common community task, right? Like you could find a benchmark that someone's developed and says, Yes, I really care about like how many times. it can say like, is the gallbladder pictured in this field of surgery? And that's real, you that's what I want to measure. Like, that's cool, right? Or maybe it's your EKG analysis task of like, how accurate is this or can it predict structural heart disease or can it predict a future cardiac event, in two years, right? We can each have those perspectives, have our own data sets and, run these and see what happens and keep the model makers on us. I think that approach is really important. ⁓
Anthony DiGiorgio: So how much of that relies on, the I guess the these prams and how advanced the model is versus just how it was trained, right? I think a lot of the at least I'm again pretty much a novice of this, but you know, if you ask Claude and ChatGPT to do different things, you get probably two different answers. But that's more the training of the model rather than how advanced the model it is, right?
Dan: Yeah, I mean I think there's tremendous advances that are occurring under the hood ⁓ of these models. It's definitely not just the size, let's say, of the neural network. There's ⁓ advances that in the process of how these models are trained. There's advances that occur in what's, in mid-training and post-training and all of these different sort of, ⁓ stages when you can the output for the individual user. and advances in how you interact with the models. ⁓ Certain kinds of interaction techniques or prompts that you deliver to a large language model ⁓ may dramatically change its. performance. ⁓ simple example, if you ask a model to talk to you like a three-year-old, it you know, ⁓ try generate plausible text that looks like a three-year-old would have made it. If you ask it to talk to you like a Nobel laureate, it'll try to generate plausible text that looks like a Nobel laureate made it. And there's a whole process for how it sort of accomplishes that. but yeah, think there's a lot be written on this story and it's a tremendously exciting time.
Anthony DiGiorgio: what are some the tasks that the community has agreed on to to judge these models? And then what do you use for for your models? Is it, did I did I do the perfect pituitary surgery and the model only passes if it says yes, you are the perfect surgeon?
Dan: That's exactly right. That's that that's what every neurosurgeon should do is just, you know, only accept a model when it tells you that you're the best. I would definitely recommend that. ⁓ I ⁓ Yes. Yes. Yeah, mean think it's an interesting question. think there are
Anthony DiGiorgio: I mean, that's like all the all the models are sycophants, right? So that's basically any model.
Anish Koka MD: Yeah.
Dan: The things that the surgical community and there's a paper that's in review that I think is gonna speak to some of this. the tasks that the community currently has to measure model performance are ⁓ probably challenging. They're probably not ⁓ quite the ones. ⁓ And a part that is something we talked about earlier, which is they ⁓ never go back to the patient. and they're tasks that, do particularly care Dr. DiGiorgio or Dr. Koka knows that's a gallbladder in a in a picture? Like if we're asking question, we're in big trouble, right? Either there's been a apocalypse and we're all taking out gallbladders because there's, no one else to do them, or we're way out of our strike zone. so I think these are a lot of these benchmarks I think are are phenomenal community and they've been extremely hard and all credit to the myriad of groups and and and institutions and people who've done these Over the years. But I think there's another leap to be made in of getting these models to actually perform meaningful tasks for clinicians. And so we look at things around surgical judgment, ⁓ tasks around surgical Those are things that we're ⁓ really interested in measuring.
Anthony DiGiorgio: you mentioned ⁓ obviously you don't I can't give away anything proprietary, but what are some of the most exciting things you've seen out there that you can talk about?
Dan: Yeah, sure. I think for me personally, there's ⁓ sort two sets of things that are that are really exciting. I think the first is, something we've already already talked about or sort of touched on or maybe skirted around, which is that ⁓ we can deploy, ⁓ sort of frontier style products in healthcare and make a real difference for patients and for physicians. and that is really happening ⁓ today. it just won't happen in the places that are the hardest to move. and when those systems collapse. think that's going to be really ⁓ interesting development. I think ⁓ far as you know other advances, we're running ⁓ one the first clinical trials of surgical AI right now in Tanzania. ⁓ So is a place where nobody ⁓ would say is where you should start with AI trials, but I think we know that's where the need is. We're really passionate about it. We have phenomenal partners there. And, there is a particular neurosurgical procedure that needs to get performed ⁓ a hundred times more often than is currently performed. And we want to see if we can make a dent in that process. I think we're seeing national level, ⁓ health systems taking an interest in some of this work. so I think there's real opportunity to actually get of these deployed. They just might not happen in the places where you would have thought, or, maybe where innovation has maybe traditionally happened before. And that kind of innovation path is going to look, I think, very different ⁓ to the conventional device narrative. So ⁓ you know, for me it's about the real world. For me, it's about seeing real impact on the lives of patients and surgeons. and, I we're we're on our way.
Anthony DiGiorgio: ⁓ what scares you about AI? Are you does any of the doomerism ever creep in? Is there anything that scares Dan donoho about the robots taking over the world?
Dan: no, no. For for the record, right? when the robots look back on this transcript, three to five years from now, I said that I was a fan and was a fan from the beginning and please spare my family. I think that And ⁓ the short answer is there are a lot of things that scare me. there are a lot of things that I don't understand, right? So I'm a dad and my son is ⁓ about to turn six. he's growing up in a world where AI will always be smarter than him ⁓ in almost any way you can it, at almost any task that would be meaningful to measure. And I think that collision we really don't understand. I'll say I don't understand it. we're very guarded about how we use, technology in our household, right? And this may be kind of funny for somebody who's, quite, AI pilled, I guess the kids would say, to see that like we are, very anti-screens and very, you know, just not a part of our lives when when we're at home. and I think the way that we ⁓ grow symbiotically, and co-evolve and co-adapt with artificial intelligence is really not known. and I think we're muddling around doing the best our we can and we'll have evidence in ten years or twenty years, long after it's too late for the people I care about. I think what also scares me is that things don't move fast enough, right? we're, not seeing in my opinion, an Enough advances in AI drug discovery, although there's a lot of activity in that area. We're not seeing as much of the digital transformation of healthcare as as we should. And we're seeing a lot of preventable mistakes happen that perhaps some of these intelligent systems could have helped with. So I think, the thing that scares me is not the doom. The thing that scares me is the slowness and the of getting these systems really deployed in the real world. I think there's plenty of doom ⁓ about, and I every generation has grappled. This, whether we're talking about, the Cuban Missile Crisis or ⁓ the Cold War more or real world wars that were really fought, approximately every 20 or 30 years. I think there are real existential threats out there. ⁓ and then there are things that people, kind of like to talk about on in perpetually online communities, like no offense. but ⁓ I think there's enough real-world suffering out there that we can alleviate, and that's what I focus on.
Anthony DiGiorgio: So is this is AI ultimately going to democratize and decentralize, or is it going to concentrate? is it whoever owns the big models and can deploy them, are they gonna have all the power, or is it gonna be something that can be democratized and everyone's supercomputer in their pocket and their own access to all the world's knowledge and potentially their own doctor, their own drone army, who knows what?
Dan: Yeah, I I think it's a profoundly democratic force. I think there certainly are greater for things like state control. you know, who has a ton of information and is highly motivated to act on it. those those things definitely do exist. but I also think that the net effect of all of these systems existing all around the world is a really profound opportunity both to accelerate human progress and increase human freedom. and increase. human autonomy and I think that's that's the world I'm really excited about.
Anthony DiGiorgio: All right. Anish closing questions?
Anish Koka MD: No, this has been really great. It's been ⁓ it's great to see ⁓ you pushing kind of at the frontier of AI and And so ⁓ I think the future is super bright with ⁓ guys like you doing that. So ⁓ so ⁓ it ⁓ awesome stuff. What what you have what's kind of like the short term runway? Like what's next for you?
Dan: Yeah, so I mean I think most ⁓ most immediately top of mind were ⁓ Going back actually to to Ethiopia ⁓ in a few days. we're ⁓ advancing our health system partnership with the Ministry of Health there again to get to real world AI deployment ⁓ in surgery in their health system. ⁓ so that's one of the really ⁓ I think exciting opportunities. and ⁓ broadly other than that, I think I think there's are a lot of exciting things that are coming the corner, for me. some of Which ⁓ hopefully will be ⁓ kind of rolled out and maybe more apparent over the next ⁓ the next three months. But ⁓ I guess I would say ⁓ broadly stay tuned. ⁓ I think there's gonna be a lot of exciting developments from SDSC. for folks who haven't heard about us or want to learn more, you can always come check us out at surgicalvideo.io and learn more about our offerings what the community looks like. And we're you know very excited to collaborate with folks, work with surgeons in more than 60 countries. really welcome that sort of global spirit. what's exciting for me for the back half of the year, ⁓ hopefully at least Anthony, you're coming to Washington, D.C. in a in a few months for the CNS conference. So ⁓ that's gonna be really cool. We got the American College of Surgeons coming to town a couple months before that. So ⁓ it's it's a packed agenda. and I think we'll see, even more disruptive change at each of those meetings and look forward to sharing more thoughts about it.
Anthony DiGiorgio: Excellent. And you remember to plug your website. Thank you. Perfect guest. All right. Dr. Dan donoho, You can follow him on Twitter ⁓ D Donaho, D D O N O H O ⁓ his website, ⁓ Surgical Data Science Collective. Say the website again, just in case, and we'll link to it.
Dan: It's ⁓ yeah, surgical video dot io.
Anthony DiGiorgio: Great. Thank you. And for everyone on listening, don't for don't forget to like and subscribe on Apple, Spotify, and YouTube. Thank you.
Anish Koka MD: Thanks so much.
Dan: Thanks guys.
Dan: Thanks, Tony and anish it's a lot of fun to be here with you guys. I think ⁓ longtime listener, first time caller. So ⁓ very excited to see you. I know the conversations can be wide ranging, so we're open to it all. Let's do it.
Anthony DiGiorgio: Right. All right. So tell me about the surgical data science initiative. What are you doing with AI? are you just focusing on surgery, all of medicine, all of healthcare, world domination? Where are we going with it?
Dan: definitely ⁓ we're going to Mars, but first we got a couple things to fix on this planet. ⁓ I think ⁓ lot the excitement around this stems from our real experiences as doctors and as surgeons, particularly. as an example, ⁓ I did a surgery today, you know, a couple cases today, did a pituitary case today. and during this case, I have, my mentors ⁓ continual running sort of ⁓ talk in my head as I'm operating. You know, ⁓ do do that, do This, don't do that, etc. ⁓ But the is at the end of that case, I get no feedback whatsoever. ⁓ I today wild. I had ⁓ a film crew, ⁓ five students, ⁓ the operating room staff, know, fellow, all this stuff. And I do this every case, I kind of look around and I realize, like There's nobody in this operating room who's seeing the same thing. There's nobody in the operating room who has the whole picture, certainly not even the surgeon. how we improve in that environment? How do do the best for our patients and maybe most importantly for our specialty and our field? ⁓ How do actually save the knowledge that we're generating during that surgery? Because I definitely learned I grew as a surgeon, as I try to do in every single case. How do we those learnings, but maybe even more importantly, how do we transmit them? ⁓ that key insight there's a broken piece in surgery. ⁓ Where the routine day-to-day learning and teaching doesn't make it out of the operating room. And think at the Surgical Data Science Collective, and basically in all of my work, is around this one core problem ⁓ of do we share and transmit and learn ⁓ as as we can from these incredible experiences that we're privileged to have as surgeons. ⁓ And do we get the insights from our mentors to live on and hopefully to benefit many surgeons who they've never met? We've tried a lot of things in the that maybe haven't worked and I'm really excited about the promise of the level of understanding we can get from tools like artificial intelligence from helping, use those tools to actually connect people around extremely meaningful work.
Anthony DiGiorgio: what does the AI do? How is AI going to? I have the same thing, right? I'm finishing a case and my mentors are in my mind. A lot of times I think I act a little bit too much like my program director in the OR. His voice even comes out. ⁓ so how AI gonna that? Is it gonna watch the videos? Is it gonna give us continual feedback? Is it gonna be like a life coach when we're done? What what are your thoughts?
Dan: Yeah, I think it's an iterative process, right? And so we have to start from where we can get the most benefit and balance that against risk and potential harm because I think as we've seen from a lot of these technologies, they are powerful. And that means they do have the potential to have missteps or, out and out ⁓ rampage. and ⁓ I think when we think about really important high-risk domains like surgery and technical crafts, ⁓ we have to be measured in our approach. ⁓ that being said, there are other areas where we can crap let things fly a little bit more. ⁓ But As far as surgery, I think we're really focused on understanding really important operative decisions ⁓ helping surgeons optimize their performance. So we start in the post-operative and pre-operative settings because those are the lowest risk areas where the surgeon ⁓ an intermediary between the information from the system and the patient. And start simply by ⁓ cataloging understanding patterns. And it doesn't sound very exciting, know, it's extremely profound once you can actually do that. You understand: hey, this is what, you know, ⁓ we have thousands of cases right in this catalog. And so what have thousands of surgeons done in this similar scenario is a core question, right? So we can break down these sort of game tapes, create, ⁓ an understanding of what the actual actions were of surgeons, and then feed them back to our, you know, our surgeons and our research team to understand ⁓ exactly mattered during that surgery, what drove that outcome. What were the other possibilities or ways you could have played that particular game? ⁓ And then it into the preoperative world of, you're Going into this case, what should you know about the last five cases that you did, the last 500, or the best five cases that anyone's ever done for this kind of operation, for this kind of patient in this scenario? ⁓ And I that's really those are the most important, you know, terrains to start in. ⁓ Meanwhile, ⁓ does escape us that the exact same system capable of doing these things ⁓ could provide profound insight during surgery. but I will say that the way to provide that information ⁓ is not settled. It is not exactly. Exactly the best way to mix human and artificial intelligence during a very time-sensitive performance. the common thing if you're, coaching elite performers, right, ⁓ is ⁓ you'll tell a of stuff potentially before surgery, before an before a game. But during the game, I think really good coaches are really effective and minimalistic at what they say. They're only communicating what players need to hear. How do we get an AI to do I think ⁓ is not 100% subtled. we're very excited to be at the forefront of efforts that are trying to figure that out and trying to figure out how to bring that into training, into early practice, and to really, get the best out of everyone who's trying to master their craft.
Anthony DiGiorgio: So AI is essentially watching the video, I assume you're plugging in pre-operative imaging ⁓ ⁓ a history. ⁓ there's a lot in surgery that I don't think AI can capture. Like it can't really capture like the the tactile feel, right, that the tissue gives us. at least not yet. and I so much of that is just, the instinct of of how does the tissue feel. I don't know, is the video capture technology three-dimensional enough where AI can pick up depth? Like it seems, AI coaching me through a surgery, yeah, maybe saying, you've made one, two extra movements here versus there versus AI, like truly getting that same instinct that I as a surgeon as I feel the tissue planes, So I'm putting a pedicle screw in, there's a there's a certain feel to it that tells me that it's good or as I'm dissecting tissue off certain tissue planes have a very tactile feel. Is AI gonna get to that point too, or are we really just looking at, yes, you could have made a different cut here or took one too many extra steps there.
Dan: Yeah, I think humans operating do rely on a variety of senses. But as a thought experiment, imagine what would happen if you told your patient you're gonna go into the operating room with the blindfold on, right?
Anthony DiGiorgio: Wait, you don't tell your patients that?
Dan: I I don't routinely tell my patients that. But we all wear one or two pairs of gloves. We diminish tactile sensation really, incredibly. so I don't want to say it's unimportant. I think it's really important, but I wonder, particularly for MIS and and endoscopic cases, ⁓ how much of that is sort ⁓ able to be parsed from video. we're very excited to, ⁓ systems that have some physical intelligence, are still, ⁓ I think, a little bit of a ways away, there are some exciting new models ⁓ that have come out in the last few months. That I think show the promise of where this technology can go. if it can understand and create physical relationships and physically intelligent AI, ⁓ think the promise of a technology that understands physical properties in the world ⁓ is coming. I don't want to discount that there's a lot that goes into an operation. ⁓ but I think All of these problems, ⁓ are whittled away. And, what would have said, was absolutely impossible. There's no way we could get this kind of performance out of a thing ⁓ four years ago, five years ago, eight years ago, ten years ago. ⁓ Each of those barriers is, being in certain ways. I don't want to be overly optimistic because I do that there's a lot, I see this every day, there's a lot of difficulty in bringing the promises of this technology. To bear on the real world, is extremely must ⁓ messy and sort of human-centered in very important ways. But I think the raw technological barriers, sort of the things that are solvable with compute, ⁓ the things that ⁓ could potentially understood, I think we're going to see more and more of those falling away and things that we thought, ⁓ there's no that the system could understand this, or there's no way that this property could be measured. I think, the answer from, a lot of people is like, okay, hold my beer, right? We're gonna go figure this out and we'll report back. So that's where I'd leave that particular question.
Anish Koka MD: clarify for our audience, where we are ⁓ based on some of the trials you've done. ⁓ there's trial ⁓ you actually did a randomized trial with ⁓ Canadian medical students, is that correct? Can you ⁓
Dan: Yeah, yeah. So there's there's a lot of trials that have been done, particularly in education, that tell us that, particularly when it comes to teaching and learning, ⁓ where, we would have said that there's, there's no way if an AI can't even feel how pedicle screw goes in, there's no way he can teach a good pedicle screw, right? Or can't feel a brain tumor section. I think that in ⁓ particular scenarios, which were, in simulation, to be clear, we're not, turning medical students loose on the ⁓ unsuspecting populace, or at least yet. we found that we can't know human performance until we measure it. So this case, measuring the performance of humans as teachers. And once we measure that, you can be surprised. ⁓
Anish Koka MD: So what so what exactly s so ⁓ but tell me what exactly was the trial? How what was the randomized control trial? What was the design of it?
Dan: Yeah, so the the trial there were sever actually there's a number of these efforts that were done over the years. But essentially in this in this crossover design, the students were ⁓ randomized to either receive feedback from ⁓ a human right, or to receive several of different degrees of feedback over the years, but sort of culminating in ⁓ a full that could give them real meaningful feedback about their surgical performance in a brain tumor resection task. So you I mean, you guys know this, you do this all the time, but for our audience, ⁓ removing a tumor is one of the
Anish Koka MD: Okay.
Dan: Riskiest things, that we can do in all of surgery. And ⁓ I think learning how do that is often left until sometimes the later years, even of neurosurgical training, of ⁓ being, the who's actually doing the difficult part of the case. ⁓ And ⁓ so in scenario, were able to take a group of medical students measure their force, their bleeding risk, and many, many other parameters in terms of how effectively, ⁓ how riskily, and how efficiently they were these simulated tumors. ⁓ And either try to coach them and improve their performance with humans or with AI. And surprisingly, we not only found that AI led to more ⁓ performance gains in the students in this actual task, but that when we switched students from getting AI to getting human coaching, they got worse. And this is ⁓ unexpected and maybe controversial. But ⁓ and similarly, when we switch students from human to AI feedback, they again improved much more quickly, ⁓ as though received perhaps little performance improvement at all. So I think is one example of a nice piece of work that shows that one of the domains that we thought ⁓ really would be as you said, what can AI teach me, right? How can it do all of this stuff? ⁓ I think there's a real that we have to measure the benchmark ⁓ of human Performance before we assume this. So, you know, I think very, very recently, Ethan Go had this nice letter with this sort of laundry list of all-star co-authors in Nature Medicine ⁓ talking about the need for human benchmarking of human performance tasks. And secondly, something that I completely agree with of measuring meaningful outcomes, not just looking at individual care processes when we're adjudicating whether AI works or not, but actually looking on whether it improves the ultimate outcomes that we really care about ⁓ terms of patient care. I think the more and more of those two things that we do, the better off we'll.
Anish Koka MD: You it's always hard to get at what the gold standard is ⁓ in many things that that we're doing, especially in when you're trying to push the field forward. So what is the gold standard in ⁓ grading medical students who are attempting to do a resection in a ⁓ cadaver or simulation? that's a simulation ⁓ virtual simulation. Okay, So sorry, so this simulation ⁓
Dan: In in a simulation, a virtual simulation in this case. But we've I've done cadaver work in other studies too.
Anish Koka MD: what's the gold standard? Because you have AI kind of giving a score to the students, correct? And then you have nurse surgery, ⁓ I assume veteran experts like yourself kind of looking at it and giving also giving a score or ⁓ what's the yeah.
Dan: Mm-hmm, mm-hmm. mean I think there there is a scoring component to it, but part of it is actually objectively measuring the task. So this is sort of analogous to stuff we've done in cadaver work where you can actually measure blood loss or whether someone succeeded or failed in, a binary task. in this particular model, they can measure the force applied to tissue, the amount of injury that was caused, the amount of ⁓ hemorrhage that occurred, ⁓ right? Because this is a fully paradigm. But even in the physical world, right, in let's say cadaveric simulators, right, we have those objective measures that we can create. And once we have those benchmarks, right, once we have those real-world data points, we can then model them and predict them using ⁓ that maybe we wouldn't have thought were possible.
Anish Koka MD: fascinating. there were human experts in this I'm sorry, I keep going back to this one trial. I know you did a bunch of them, but it was so interesting to see how how you did this. I would never have have thought I just didn't realize that ⁓ AI was at a point that w it could actually do this. So it it's fascinating to me. I I'm curious if right, okay.
Dan: Yeah, and this was old AI. This this was, you know, really most of this work was done now, years ago. ⁓ So when think about model generations occurring ⁓ over weeks ⁓ or ⁓ right? again, I don't want to be the straight line go up kind of guy, right? I think the world is a little bit more complicated than that, ⁓ but
Anish Koka MD: Right.
Dan: it just shows like something that we do all the time is we, talked about this at a talk I gave at the Society for Robotic Surgery this past weekend, ⁓ it's really important for ⁓ I people in a particular craft, medicine, whatever it is, to have your own internal test for AI, right? Your own reality and sanity checks that mean something to you. so in this one example, there's a mathematician ⁓ who at that time was working at Microsoft and working with a lot of open AI guys, ⁓ and his thing was. I'm gonna ask this particular task of an AI, it's gonna draw a unicorn using a particular kind of software library and a particular kind ⁓ sort of manuscript preparation software. And we're gonna see how good it is, right? Other folks have different tests that they ⁓ And once have one of these things that you can throw at the AI, over and over again from these different systems, you can chart their progress. So in our world, we look at you know its ability to understand surgery, to decompose a surgery into its consistent. Gestures, steps, and phases to predict its outcomes. And I just say over the last three or four months we've been continually surprised to see how much progress has been made. And I think these kinds of benchmarking efforts where we keep the AI honest using things that we really know and understand are really critical. ⁓
Anish Koka MD: Dan, who's who's validating the AI? we've had some type of AI at EKGs, and putting out these automatic interpretations. most of my consults from neurosurgery are because, ⁓ the computerized algorithm says abnormal EKG. And then it's like, well, call Koka. But but ⁓ but so who it is, it is it it it no, no, no, I I like I it. ⁓ I I love it.
Dan: Hmm. Yeah, well it sounds like it's keeping you in business, so why are you biting the hand that feeds?
Anish Koka MD: who's validating these AI in terms of its interpretation of whether or not the trainee or whoever is doing a good job? Like do the experts agree with most of the time, whether it be the model, that initial model, or maybe is it improving now with the current models? and the problem there is that
Dan: Yeah.
Anish Koka MD: I've come to know plenty of neurosurgeons in my short career so far. two ⁓ skull based surgeons, one at ⁓ X place, another trained at Y place. If look at the same video, are both gonna be like, Whoa, what the heck is this guy doing? Like he should be doing this. I mean, what's the gold standard?
Dan: Yeah, yeah. I mean I think to be honest, ultimately the patient is the judge, not the surgeon. Right. And I think that's really the critical distinction. We can look at all of these operational things of do surgeons agree. And we've done these studies of radar agreement on various judgments that surgeons can make. Those are all great, but they're instrumental, right? Where we really need to go to your point about EKGs, right? Yeah, it's great that we flag abnormal EKGs. Maybe I don't know. I mean, I haven't, I forget what the letters are even of an EKG, right? but what ultimately is does it save your life? ⁓ Right. And so there's lot of really interesting. So Ziad Obermeyer had a ⁓ recent paper around like population level EKG analysis showing that when you look at large scale amounts of of EKGs, you can predict critical events, right? Even long after those EKGs occur. You can diagnose other phenomena that aren't commonly thought of as present in an EKG itself. things ⁓ like upcoming cardiac events and even potentially other things as well. ⁓ And I think that's you know it's one direction that we have to go in. I think to your point about I don't want to lose your point though, which I think is really important, which is a lot of the technology we use today in medicine sucks, right? was made ⁓ decades ⁓ right? doesn't even incorporate the things that we've learned, ⁓ that basic, ⁓ first-year computer science student in could, vibe code up in a night, could probably beat any of those automated EKG algorithms that come pre-installed in your EKG machine. ⁓ Right. We could talk about, the broken of medical innovation, and challenges in regulatory and in all of those things, right? That's your answer for why your EKG sucks. And my personal opinion, it's don't blame the technology. Blame the barriers that are preventing actual real technology from getting to you and getting to your patients at the speed that technology can improve.
Anthony DiGiorgio: I think that's a really good point. And the futurist in me says, Yeah, but probably at some point in the future there's gonna be humanoid robots that can do a craniotomy in the middle of the night better than I can and don't get tired. But I think that's probably a long way off, right? We can we still don't even have driving cars that self-driving cars that work on, most roads in the US without some sort of human input. ⁓ right, 'cause all the Waymo's have a human behind them at some point in that chain. ⁓
Dan: Yeah, one to forty, I think, is the the rumored number.
Anthony DiGiorgio: Yeah. So, yes, maybe we'll get there, but there's AI applications right now that we're really falling short on. Right. So we in a lot of clinics, they will pay a human to go through the charts for the clinic for the neurosurgeon the day before clinic ⁓ and collect all the in a way that can be summarized for that neurosurgeon before they go into the exam room. Right. That is like the the right now AI can do that probably better than that human can. Yet we're still paying a human an extraordinarily large amount of money. typically it's like an advanced practice provider. So someone making, well into six figures gets paid to do this that, the current models of Chat GPT could do right now better than that human can. So why aren't we? I mean, this is like low hanging AI fruit. Why isn't this being done right now?
Dan: I think it is. I think it's just not evenly distributed. And why are our broken businesses in medicine not applying best practices? I mean, I think that's a tale, a tale as old as time. I think the nice thing about, you know, innovation at the pace of software is really are, real-world applications that ⁓ you can, and apply, and they're HIPAA compliant and cybersecurity certified and all of these things ⁓ that do this. But think that ⁓ the broader lesson is deployment of technology into healthcare requires Sustained focus and iteration, working hand in hand with a customer that is a really arduous ⁓ intense process. ⁓ it's a lot easier to sell a consumer SaaS tool, ⁓ You want to be your first vibe coded, one employee, a billion dollar startup, ⁓ right? You're probably not do that in healthcare, ⁓ right? You're probably gonna that in I don't know, some you know absurd consumer application, right? SMB application. but I think there are some reasons. Really positive signs. are entities that are really interested in funding disruptive technology companies in healthcare because ⁓ these exist. And guess I would say, like, ⁓ where do I think that innovation going to be distributed? ⁓ I wouldn't go to an insurance hospital and expect anything like the best in class technology. I actually think you'll find small agile, operators of practices of ASCs that are willing to innovate, potentially smaller, ⁓ health systems that. understand this and sort of imbibe this from the beginning and say, hey, we want to disrupt this. And I think we see real world examples of how that can drive, ⁓ tremendous that it can improve quality of life for healthcare providers. And yeah, happy to happy to go into more details about those things, but I think ⁓ the answer is like in healthcare is really hard, particularly in calcified, insurance pay And ⁓ innovation is going to occur elsewhere. Those hospitals, personal opinion, they all want to be second. But there are places that want to be first, ⁓ we should look to those as our examples.
Anthony DiGiorgio: See, Koka I knew he'd belong. He wants more competition to drive innovation, just like us. I mean, I think that's a great point, right? So ⁓ think the large hospital takes a big risk by deploying something, it has a data breach or doesn't work they spent a lot of money on it, and now that hospital's liable. ⁓ and please feel free to the question or change the subject if you're not aware, but like what are some of the regulatory barriers with getting you know, like just a basic AI agent into an electronic health software. ⁓ what sort of for EHR there's a list a mile long of the, requirements in order to meet the Medicare standard for an EHR. Are there those same requirements for an AI agent aside from it having just to be HIPAA compliant?
Dan: Yeah, I'll just say my personal opinion again to sort of differentiate this and to, at the risk, right, of being a little bit speculative, ⁓ my answer is there aren't really any. which to say, like none these actual barriers, if you go to statute, if you go to compliance, which I've done in numerous entities, that I've been involved in, ⁓ you really can meet these barriers. Like it is really possible to do it. The problem is the change management, the problem is the culture and society around these entities. It's actually not necessarily the regulation. Now, ⁓ look, okay, building an EHR that's building Medicare, right? Like that's but that's like the last place I would ever want to do this. Like, why would I play that game? That game is not fun. Like, I think there's much more satisfying ways you can actually enact real change, touch millions of lives, and improve care for doc for patients and improve quality of life for doctors. and let that just be what it is. Like those organizations want to behave that way. That's okay. Like, I'm not trying to lead people kicking and screaming. I'm just want to show them that there's a better way. And when you're ready, here it is.
Anthony DiGiorgio: Love it. Totally agree. ⁓ Doctronic, what are your thoughts? So there's now an AI agent, AI app, whatever you want to call it, that has a license in Utah, right? Utah licensed this app to now prescribe medications. is the only thing standing between us ⁓ at least the some of the ⁓ thinking specialties from replaced by AI, is the only thing standing between that? licensure and regulation. I mean I heard somebody on Twitter was talking about how, AI is making more demand for doctors because people are gathering all their AI questions and going to the doctor and now they want their questions answered and then they want the prescription, right? The end point for a lot of the of physicians is a or an imaging study or some order. If AI is allowed to place these orders, is that gonna be a quick end to a lot of the the thinking specialists? Is Koka gonna have to go learn a new trade?
Dan: Yeah, I mean I think the the good news is, right, Or the bad news is, depending on, how you look at it, these things not going as quickly and as widely as you think they are. Like I don't think that there's a, ⁓ gold rush of people to be, the doctronics of the world. I think reimbursements are hard. I think value is in healthcare. and I think patients usually come second or third ⁓ best in a lot of these conversations. So I think what, the narrative may be San Francisco may not be the narrative in the Midwest or the Southeast where the majority, ⁓ a lot of American healthcare is actually delivered. ⁓ and you're gonna see real discrepancies and disparities in the use of technology in those kinds of places. So I was ⁓ kind of in love. love the fact that this happened in Utah, right? And, not in, the Bay Area. and you know full disclosure, I'm from the Bay Area. I'm a card carrying, member, so I'm trying not to, ⁓ be overly zealous about the other side. I think there's there's an So I completely disagree with the point that ⁓ people just go to the doctor to get their ChatGPT questions answered. They go to ChatGPT to get their ChatGPT questions answered. And ChatGPT will be increasingly satisfying in answering those questions, and that will be the end of that. I think if that's your business model as a cognitive specialist, it's a loser and and you should really rethink things. I think that there is a lot of organizational power and organized healthcare. I think there are a lot of people trying to break that down, but I don't see that regime changing outside of very specific contexts. Like I think one of the battles, of Doctronic is I, understand it, and again, I this is really not ⁓ not my world, but can we replicate this? Can we actually even sustain this? ⁓ Right? Will payers stand for this? Are going to, actually see this flourish? Or is this going to die in a bunch of legal battles and a bunch of refusals? And it'll be some sort of curiosity that is nice to have, but really doesn't end up moving the needle. I'm not sure that that's actually the way to ⁓ get, let's say, healthcare abundance if subscribe to that or want that narrative. I think there are better ways to do that, in my personal opinion, that will have a faster impact on human health in the United States the next 10 years.
Anish Koka MD: Dan, where ⁓ switching switching gears again, as we said, this would be wide ranging. Where did this interest come from? are extremely bright, of course. ⁓ but this interest in ⁓ science, machine learning any ⁓ place that arose from?
Dan: Yeah, I mean I think I've always been fascinated by data. I think I was, making spreadsheets on paper, ⁓ with paper and pen ⁓ when I was like six or seven, to track various sports statistics and various other things. So I think I've just always been fascinated by the idea that you can quantitatively understand the real world around you. I studied economics in college, I wasn't a med. and I think that, kind of thinking about ways modeling the world and you know, the late nineties and early 2000s was really fascinating to me. ⁓ suddenly, ⁓ 20 some odd years ⁓ now we all have tools at our fingertips that if you just know how to ask the right, questions will give you really fascinating answers, help you construct complicated systems that ⁓ would have taken many, many person of work that ⁓ really can extend your ability to understand complicated questions. And so I think it's just been a part of my life the entire journey. journey and I certainly ⁓ loved, being neurosurgeon and I think I love taking care of my patients and ⁓ I love a lot what I do. But ⁓ I think this part has really been there for my whole life and I think that's ⁓ probably core to identity, I guess I would say.
Anish Koka MD: Is your ⁓ dad is David Donoho? The the the statistician he's a statistician at Stanford, is that right? ⁓ Because he's no shrink in world of ⁓ data science, correct?
Dan: Mm-hmm. That's right. That's right. Yeah. Yeah. Yeah, yeah, I'd say he's ⁓ he's a fairly influential, thinker and effector. He's done a lot of phenomenal things and it's been a a privilege to, be able to learn and understand maybe a small fraction of what he does. but yeah, I think ⁓ seeing the change in, the rise of empiricism in the mathematical sciences more broadly, ⁓ some people would say, the death of theory. I don't know, not my field, but I've certainly heard that said. it certainly seems interesting when, every ⁓ one of my ⁓ friends, collaborators posted something on this. if you want if you know, the I think his take was, if you want any of your problems solved, just call them a math conjecture and somebody on Twitter will solve it in forty-five seconds using, chat GPT. So I think, ⁓ it's definitely times. It's been incredible to see the evolution, of that field from my decidedly ringside seat ⁓ and learn as much as can.
Anish Koka MD: your father was ⁓ I guess at some point disappointed where statistics and data science were going, in terms the real world application stuff. So I mean these last four or five years it'd be curious what your perspective or what your conversations have been in terms how seen things kinda progress. Is Yeah.
Dan: Yeah, I mean I'd let his, you know, his his published work, of which there are, you know, a few cited papers, ⁓ speak for themselves. but I think the perspective that I can certainly share is that it's in equal parts, maybe dispiriting and invigorating. I think there's a strong narrative that is being right? When you see things like ⁓ frontier math, when see people, really exemplary mathematicians sort of ⁓ lending their ⁓ credibility to the language model ⁓ attempts ⁓ sort of solve ⁓ for an entire field. I think that one of the narratives that gets pushed is like, similar to what we hear, right? Don't go into math, ⁓ right? Don't into medicine. ⁓ I think that was, an Elon thing six months ago. I showed that talk at a ⁓ neurosurgery conference that was speaking at. I was like, this is what Elon Musk is telling you to do this morning, right? Like, ⁓ don't do that, ⁓ right? Do not obey advance. think this is it's really for young people, for old people, for everybody, avoid this sort of delusional disillusionment. ⁓ we're working on a good term for this by the way. but saying that, ⁓ I just have to give up because AI has solved my field and so there's nothing meaningful that I can contribute. I think there's ⁓ a longer piece on this to come in the media outlet, but I think the the short answer is ⁓ there are a lot of forces that want people to believe that and are very important ⁓ to those particular ⁓ entities for people to believe ⁓ in that statement. and I think the truth is actually the opposite. It's never been a more exciting and invigorating time to be a neurosurgeon, to be a cardiologist, ⁓ maybe to be a I'm not quite sure. but certainly in our field, the ⁓ ability to create ⁓ and to explore and to with the assistance of a lot of tooling that's much easier to ⁓ is really incredible and and I love this time. I mean I think it's ⁓ my I I said I have five students in the operating room today I said it's never been a time to better time to be a doctor in my opinion. ⁓ because what it means to be a doctor ⁓ in their generation, right, they're one some of them high school kids, right? They're gonna if they go into medicine they'll be, they'll be in practice maybe in ten years, ⁓ right? It's gonna be so different for people recognize the differences while superficially probably seeming almost the exact same. Yeah.
Anish Koka MD: Yeah. it's fascinating to me, you talked about this earlier and I'll just kind of reiterate ⁓ the significant gap between where technology is and how it can be deployed, and then how it's actually deployed in as you called it, you know, these kind of failing systems, right? And there's no better use case for that than ⁓ COVID, where we've able to do telemedicine for long time. we've been able to, ⁓ zoom into conferences for a long time, and yet Yet the bureaucracy and administrators refused to allow that to happen just because they had control and they weren't allowed to happen. And then the moment they suddenly had these parameters these kind this straitjacket of well you can't have people in a room together because, you know, ⁓ because of COVID. ⁓ technology became suddenly everyone's and everyone's, ⁓ zooming into this and zooming into that. So it'll be Anthony, I talk about this all the time in terms of it does feel like in order to really advance ⁓ in a way that can utilize technology to it actually is, ⁓ it requires kind of creative it requires destruction. You know, ⁓ this Joseph Schumpeter you know talk a lot about creative destruction. and ⁓ do do that without kind of destroying what currently exists? So people get upset about talking about destroying these great institutions, these great health systems that we have that amazing stuff does happen in. But I don't know, it seems very hard to ⁓ imagine a where
Dan: Yeah.
Anish Koka MD: These folks are still controlling everything, because they it does seem to be a block on innovation in many ways.
Dan: Just depends where you want to innovate, right? I think in what layer and what way. I think I mean I I'll just ⁓ my, particular personal take and experiences. ⁓ It's never a better time, in my opinion, to do innovation in healthcare. ⁓ I mean, much ⁓ much more exciting, more vibrant than any other time that that I've been involved, right? I mean, ⁓ you back maybe 15, 20 years ago, innovation basically meant like consulting for a big med tech company or strategic to help them, market their latest device. I think today innovation means building your own product, going out into the market and finding willing adopters, you know, hopefully, of said product ⁓ and seeing it adopted in real clinics impacting real patients' lives and doing so at a time scale of weeks, ⁓ days. ⁓ and really be builders. we're the people who understand these problems because we live in these systems ⁓ and, are confronted with I think it's for me invigorating because I think lot of medicine is a learned helplessness model, ⁓ right? You're sort of standing on this, and you get shocked unpredictably and then you go back and you're just cowering in the corner and someone opens the door, you don't even leave. I think this is a huge antidote to that mindset. you know, through I think, folks like yourselves and many others who are showing, ⁓ a lot of to be creative. And is there gonna be some destruction in that? I don't know, like ⁓ let Godzilla fight, ⁓ fight Mothra. I don't know, let's let it happen. Like, let's just see. but I know that, for me, ⁓ I want to be out there building and advancing the state of play across a variety of different domains, variety of different concepts, and ⁓ just sort of
Anish Koka MD: where do see regulation? ⁓ We talked touched on this with doctronix of but where do you regulation fitting in this? Like for instance ⁓ say a commercial product that videos surgeons, gives feedback, that something that should be by the FDA, or is it best not to be regulated by some ⁓ FDA type?
Dan: Yeah, I mean I think it's a it's a perilous question for ⁓ somebody who may be facing the agency not not terribly far down the road. but I think ⁓ I think ⁓ ⁓ we had heard from ⁓ a ⁓ I guess former head of FDA you know at one of our national meetings a few months ago, who I can now blame him for this, right? Who shared his opinion that such products actually ⁓ because the surgeon is actually making the decision, that such products ⁓ may be outside of the purview of the FDA. and I'd say that was a very invigorating opinion ⁓ to hear that. what does this actually mean? I think seeing and knowing a little bit about some products that are coming out in the market, some of which I'm again investing in or some of which I'm involved in, ⁓ I think that there really are going to be slew of such products that ⁓ the regulatory barrier ⁓ again, it may not be as high as people think. Like it's really important to push this stuff through and see what is actually required. And if you Do this in an intelligent way, even in the current Ancien regime, right? Like, can you find a way that you can do this with just human factors in a 510K? Right? Like, how can you actually do this? I think, I think it's more feasible than people think. And think there may be tailwinds. and there's some, I think perhaps not for the interoperative, performance analysis and performance improvement and guidance, and taking over the surgeon and being the co-pilot or co-brain of the surgeon, maybe that will be a high regulatory bar. But for a lot of other things that we've talked about. I think there are ways to implement these products that can be done intelligently, that can conform even to very antiquated regulatory standards. And I don't think we need to wait for some magic moment when the floodgates will open and now we can innovate in healthcare. And now we can have AI agents in healthcare and now we can have, all of these amazing technologies that are available to everyone else on earth. I think we can have them now. I just think we have to do them intelligently. We have to do them with the patient in mind ⁓ and to do them understanding the business some medicine if you can do those three things like I think it's a product you can sell tomorrow.
Anish Koka MD: Without the FTA.
Anthony DiGiorgio: Are we
Dan: Again, I think that's where it comes back to doing it intelligently, right? I think that's where it comes back to, doing this ⁓ with an understanding of what the current regulatory structures really do and don't say, not what they're imagined to say. ⁓ some of which in conversation with the agency, some of which perhaps ⁓ I don't want to say in opposition to, but maybe in more of a sort of firm stance. And there's a company out there ⁓ that, I'm not invested in, but but love a lot. company and Andromeda Surgical, ⁓ and their mission is to provide a platform intelligence technology for autonomous surgery. Right. they got Health Canada clearance, right? So they're ⁓ they're out and about, right? They'll in FDA soon. ⁓ and ⁓ I'm assuming I don't know. But this is coming. how are we going to we talked a little bit about humanoids in the operating room. ⁓ so to go back to this for a second, there was a cool paper by ⁓ Mike Yip, who's at ⁓ UCSD, where they basically chained ⁓ a robotic controller, like you can think about the control station for Vinci Robotic Surgical System, up to the ⁓ arms of a humanoid robot operating conventional laparoscopic instruments. And so when the surgeon Would move their hands in the robotic control system that currently exists today. The humanoid robot would mimic those movements in the patient. Of course, it's an animal model. And there were, real problems, with some latency, real problems with some accuracy, but they were able to accomplish the surgical procedures ⁓ in these, in vivo models, even despite some of limitations from very ⁓ early, early studies. So how do we that? Nobody knows, but I'll tell you, you can buy a humanoid for $30,000, you can buy a Da Vinci for question mark, but maybe 30 times that, maybe 50 times that, right? I'm somebody's figure it out, right? Because that margin, right, to the the prior turn of phrase is a huge opportunity for some enterprising young person out there who's gonna solve for this.
Anthony DiGiorgio: Well not only can you you buy a humanoid robot, you said what, thirty thousand dollars? It's a lot cheaper than than a than a neurosurgeon, right? So I don't know.
Dan: Something like that. Absolutely. Right. Yeah. So where where does this converge? I think I think we don't we don't know. ⁓ but I think it's a really exciting time. I think like all of us, right, we wanna, we wanna create the future. and I think that's it's again, it's never been more feasible to do that. So I think it's that's if there's one overriding message, Is this overly optimistic? Is this Pollyanish? I think we're putting things out in the real world across a variety of things that I work on. and are not theories, right? These are, ⁓ real things that that we really have conviction in. And I think I guess ⁓ we'll the receipts and we'll what happens.
Anthony DiGiorgio: What about the Luddites? the anti data center movement. are data centers poisoning our water? Are they driving up energy costs? Do should we be banning them or are the Luddites gonna stand in the way of our progress?
Dan: Yeah, so I showed this fun this fun slide at ⁓ at CSNS. ⁓ actually the first sabotage, right? So Sabot, right, throwing the shoes into the thing. The thing that was being thrown the shoes thrown into ⁓ was actually, an autonomous system back in, you know, hundreds of ago. So I think this tradition of trying to, slow down progress and I think is ⁓ ages old. I think it's dangerous time. ⁓ Right. I think it is a violent time. ⁓ and I think we are going to see, I think quite regrettably, ⁓ potentially significant violence around ⁓ some of these ⁓ based on that I think are largely incorrect and are being pushed. ⁓ but overall I think ⁓ the need for here, the need to improve human conditions, the need to create greater human opportunity ⁓ is just too and ⁓ I don't ⁓ it's reasonably likely that those kinds of movements will will really have any meaningful success. They may reshuffle the deck of where these things happen, potentially, sadly for the economic advantage of of certain areas. But ⁓ I don't that, these movements are to be in the large part successful in ⁓ an increasingly, global world where there's global need and global
Anthony DiGiorgio: are there other countries that are ⁓ that are gonna compete with us on AI or is this a US dominated thing?
Dan: Yeah, there's no question that there's a global competition right now. there's no question that this includes what we might benignly call countries of concern or simply say China. but I that so I'll I'll say this as somebody who downloaded the Kimmy K3 weights the other day. very excited, gotta figure out some things of how to run a three trillion Prams model. And, you know, there's a lot of a lot of learnings happen around this. but I think ⁓ there's lot of strategy and games playing that's occurring right now ⁓ around technological advances, a lot of, interesting moments and movements. But ⁓ ultimately, I think that this kind global competition ⁓ really is for the better. I think will encourage folks to ⁓ continue to put their best work out there, the best products out there because they know they're gonna be challenged. from a global standpoint or national security standpoint, say that's probably outside of my area of expertise. But we're watching those things and I think it's ⁓ it's fascinating to think about, ⁓ deep seek markets basically as a hedge fund, ⁓ creating a large language model. ⁓ there's a lot of strange and fascinating things that are occurring for people who are paying attention.
Anthony DiGiorgio: ⁓ you used a couple of terms there that went over my head, so and you gave this great talk that C S N S that also largely went over my head, but our audience is a lot smarter than I am. So yeah, you explained what you meant by what was it, prams and and chimney three something?
Dan: Yeah, so like the news of like last week in technology, which is already now boring and we're on to the next thing, was there was a Chinese lab called Moonshot AI and they released a very large ⁓ ⁓ generalist foundation model ⁓ that they Kimi and the version is K three. and this is ⁓ an open weights model, which means they give you ⁓ all of the internal sort of structures the model that allow you, if you have the computational abilities. So for a model of this size, require, depending, there's some you know inside baseball of how you do this, you can do this potentially on some smaller devices with some tips and tricks. But the vanilla, benign, not really ⁓ thoughtful way doing it, not there's anything wrong with it, is just you get $100,000 and buy a very significant amount of computation to be able to run these models efficiently. And ⁓ a parameter count is a way of ⁓ thinking about the size, the number of sort of different you can. can think about it as nodes within the or neurons within the neural network. It's not not quite accurate. ⁓ one way of thinking about it. And so this model was sort of order of magnitude 3 trillion. And ⁓ ⁓ we don't know exactly if you think about the latest claw models or open AI models, but they're probably ⁓ around that order of magnitude. ⁓ a of the models that you might try to run on your laptop at home ⁓ would significantly smaller in the few billions. So we think that these larger models are more capable, and so far they've given us some very interesting results in early testing.
Anthony DiGiorgio: So you don't is that how you measure I guess the advancedness of the AI models, these nodes or prams, or are there other tests that they can to see how good these models are when they stack up against each other? I seem to remember some graph that you showed during your talk.
Dan: yeah. So I definitely don't think that model size tells us much about ⁓ performance, right? Like ⁓ I could give three trillion parameters right now of ones and zeros and it wouldn't do very much. but ⁓ I think speaks to some extent to the complexity and perhaps capability of these models. The best way to measure these models is actually by having meaningful tasks that a community agrees on and is new are neutrally adjudicated. ⁓ those kind of competitive frameworks ⁓ benchmarks are really how I think most of us would look at a model and say, is it better than the prior generation? And ultimately it for each person, ⁓ right? For each use ⁓ right? There's, these things are ⁓ unusual and they're Behavior and they can get better and worse at the same time at different tasks. so ⁓ goes back to sort of the what is surgery's unicorn, drawing moment or an autotyping on an airplane, ⁓ right? What is your personal evaluative set for model? ⁓ I strongly encourage people to have one, to a perspective, to say this is a kind of task that really matters to me. ⁓ And when I want to know, ⁓ here's I'm gonna measure this. And it could be common community task, right? Like you could find a benchmark that someone's developed and says, Yes, I really care about like how many times. it can say like, is the gallbladder pictured in this field of surgery? And that's real, you that's what I want to measure. Like, that's cool, right? Or maybe it's your EKG analysis task of like, how accurate is this or can it predict structural heart disease or can it predict a future cardiac event, in two years, right? We can each have those perspectives, have our own data sets and, run these and see what happens and keep the model makers on us. I think that approach is really important. ⁓
Anthony DiGiorgio: So how much of that relies on, the I guess the these prams and how advanced the model is versus just how it was trained, right? I think a lot of the at least I'm again pretty much a novice of this, but you know, if you ask Claude and ChatGPT to do different things, you get probably two different answers. But that's more the training of the model rather than how advanced the model it is, right?
Dan: Yeah, I mean I think there's tremendous advances that are occurring under the hood ⁓ of these models. It's definitely not just the size, let's say, of the neural network. There's ⁓ advances that in the process of how these models are trained. There's advances that occur in what's, in mid-training and post-training and all of these different sort of, ⁓ stages when you can the output for the individual user. and advances in how you interact with the models. ⁓ Certain kinds of interaction techniques or prompts that you deliver to a large language model ⁓ may dramatically change its. performance. ⁓ simple example, if you ask a model to talk to you like a three-year-old, it you know, ⁓ try generate plausible text that looks like a three-year-old would have made it. If you ask it to talk to you like a Nobel laureate, it'll try to generate plausible text that looks like a Nobel laureate made it. And there's a whole process for how it sort of accomplishes that. but yeah, think there's a lot be written on this story and it's a tremendously exciting time.
Anthony DiGiorgio: what are some the tasks that the community has agreed on to to judge these models? And then what do you use for for your models? Is it, did I did I do the perfect pituitary surgery and the model only passes if it says yes, you are the perfect surgeon?
Dan: That's exactly right. That's that that's what every neurosurgeon should do is just, you know, only accept a model when it tells you that you're the best. I would definitely recommend that. ⁓ I ⁓ Yes. Yes. Yeah, mean think it's an interesting question. think there are
Anthony DiGiorgio: I mean, that's like all the all the models are sycophants, right? So that's basically any model.
Anish Koka MD: Yeah.
Dan: The things that the surgical community and there's a paper that's in review that I think is gonna speak to some of this. the tasks that the community currently has to measure model performance are ⁓ probably challenging. They're probably not ⁓ quite the ones. ⁓ And a part that is something we talked about earlier, which is they ⁓ never go back to the patient. and they're tasks that, do particularly care Dr. DiGiorgio or Dr. Koka knows that's a gallbladder in a in a picture? Like if we're asking question, we're in big trouble, right? Either there's been a apocalypse and we're all taking out gallbladders because there's, no one else to do them, or we're way out of our strike zone. so I think these are a lot of these benchmarks I think are are phenomenal community and they've been extremely hard and all credit to the myriad of groups and and and institutions and people who've done these Over the years. But I think there's another leap to be made in of getting these models to actually perform meaningful tasks for clinicians. And so we look at things around surgical judgment, ⁓ tasks around surgical Those are things that we're ⁓ really interested in measuring.
Anthony DiGiorgio: you mentioned ⁓ obviously you don't I can't give away anything proprietary, but what are some of the most exciting things you've seen out there that you can talk about?
Dan: Yeah, sure. I think for me personally, there's ⁓ sort two sets of things that are that are really exciting. I think the first is, something we've already already talked about or sort of touched on or maybe skirted around, which is that ⁓ we can deploy, ⁓ sort of frontier style products in healthcare and make a real difference for patients and for physicians. and that is really happening ⁓ today. it just won't happen in the places that are the hardest to move. and when those systems collapse. think that's going to be really ⁓ interesting development. I think ⁓ far as you know other advances, we're running ⁓ one the first clinical trials of surgical AI right now in Tanzania. ⁓ So is a place where nobody ⁓ would say is where you should start with AI trials, but I think we know that's where the need is. We're really passionate about it. We have phenomenal partners there. And, there is a particular neurosurgical procedure that needs to get performed ⁓ a hundred times more often than is currently performed. And we want to see if we can make a dent in that process. I think we're seeing national level, ⁓ health systems taking an interest in some of this work. so I think there's real opportunity to actually get of these deployed. They just might not happen in the places where you would have thought, or, maybe where innovation has maybe traditionally happened before. And that kind of innovation path is going to look, I think, very different ⁓ to the conventional device narrative. So ⁓ you know, for me it's about the real world. For me, it's about seeing real impact on the lives of patients and surgeons. and, I we're we're on our way.
Anthony DiGiorgio: ⁓ what scares you about AI? Are you does any of the doomerism ever creep in? Is there anything that scares Dan donoho about the robots taking over the world?
Dan: no, no. For for the record, right? when the robots look back on this transcript, three to five years from now, I said that I was a fan and was a fan from the beginning and please spare my family. I think that And ⁓ the short answer is there are a lot of things that scare me. there are a lot of things that I don't understand, right? So I'm a dad and my son is ⁓ about to turn six. he's growing up in a world where AI will always be smarter than him ⁓ in almost any way you can it, at almost any task that would be meaningful to measure. And I think that collision we really don't understand. I'll say I don't understand it. we're very guarded about how we use, technology in our household, right? And this may be kind of funny for somebody who's, quite, AI pilled, I guess the kids would say, to see that like we are, very anti-screens and very, you know, just not a part of our lives when when we're at home. and I think the way that we ⁓ grow symbiotically, and co-evolve and co-adapt with artificial intelligence is really not known. and I think we're muddling around doing the best our we can and we'll have evidence in ten years or twenty years, long after it's too late for the people I care about. I think what also scares me is that things don't move fast enough, right? we're, not seeing in my opinion, an Enough advances in AI drug discovery, although there's a lot of activity in that area. We're not seeing as much of the digital transformation of healthcare as as we should. And we're seeing a lot of preventable mistakes happen that perhaps some of these intelligent systems could have helped with. So I think, the thing that scares me is not the doom. The thing that scares me is the slowness and the of getting these systems really deployed in the real world. I think there's plenty of doom ⁓ about, and I every generation has grappled. This, whether we're talking about, the Cuban Missile Crisis or ⁓ the Cold War more or real world wars that were really fought, approximately every 20 or 30 years. I think there are real existential threats out there. ⁓ and then there are things that people, kind of like to talk about on in perpetually online communities, like no offense. but ⁓ I think there's enough real-world suffering out there that we can alleviate, and that's what I focus on.
Anthony DiGiorgio: So is this is AI ultimately going to democratize and decentralize, or is it going to concentrate? is it whoever owns the big models and can deploy them, are they gonna have all the power, or is it gonna be something that can be democratized and everyone's supercomputer in their pocket and their own access to all the world's knowledge and potentially their own doctor, their own drone army, who knows what?
Dan: Yeah, I I think it's a profoundly democratic force. I think there certainly are greater for things like state control. you know, who has a ton of information and is highly motivated to act on it. those those things definitely do exist. but I also think that the net effect of all of these systems existing all around the world is a really profound opportunity both to accelerate human progress and increase human freedom. and increase. human autonomy and I think that's that's the world I'm really excited about.
Anthony DiGiorgio: All right. Anish closing questions?
Anish Koka MD: No, this has been really great. It's been ⁓ it's great to see ⁓ you pushing kind of at the frontier of AI and And so ⁓ I think the future is super bright with ⁓ guys like you doing that. So ⁓ so ⁓ it ⁓ awesome stuff. What what you have what's kind of like the short term runway? Like what's next for you?
Dan: Yeah, so I mean I think most ⁓ most immediately top of mind were ⁓ Going back actually to to Ethiopia ⁓ in a few days. we're ⁓ advancing our health system partnership with the Ministry of Health there again to get to real world AI deployment ⁓ in surgery in their health system. ⁓ so that's one of the really ⁓ I think exciting opportunities. and ⁓ broadly other than that, I think I think there's are a lot of exciting things that are coming the corner, for me. some of Which ⁓ hopefully will be ⁓ kind of rolled out and maybe more apparent over the next ⁓ the next three months. But ⁓ I guess I would say ⁓ broadly stay tuned. ⁓ I think there's gonna be a lot of exciting developments from SDSC. for folks who haven't heard about us or want to learn more, you can always come check us out at surgicalvideo.io and learn more about our offerings what the community looks like. And we're you know very excited to collaborate with folks, work with surgeons in more than 60 countries. really welcome that sort of global spirit. what's exciting for me for the back half of the year, ⁓ hopefully at least Anthony, you're coming to Washington, D.C. in a in a few months for the CNS conference. So ⁓ that's gonna be really cool. We got the American College of Surgeons coming to town a couple months before that. So ⁓ it's it's a packed agenda. and I think we'll see, even more disruptive change at each of those meetings and look forward to sharing more thoughts about it.
Anthony DiGiorgio: Excellent. And you remember to plug your website. Thank you. Perfect guest. All right. Dr. Dan donoho, You can follow him on Twitter ⁓ D Donaho, D D O N O H O ⁓ his website, ⁓ Surgical Data Science Collective. Say the website again, just in case, and we'll link to it.
Dan: It's ⁓ yeah, surgical video dot io.
Anthony DiGiorgio: Great. Thank you. And for everyone on listening, don't for don't forget to like and subscribe on Apple, Spotify, and YouTube. Thank you.
Anish Koka MD: Thanks so much.
Dan: Thanks guys.