Latest / CareTalk: Healthcare. Unfiltered. / A Brief History of AI in Healthcare w/ Lekan Wang, Partner, JSL Health Capital

Transcript

A Brief History of AI in Healthcare w/ Lekan Wang, Partner, JSL Health Capital · CareTalk: Healthcare. Unfiltered.

• 0:07
  John: Welcome to Care Talk America's home for incisive debate on healthcare, business, and politics.
• 0:12
  John: I'm John Driscoll, the chairman of the Yukon Health System, and today we have my old friend, Lekan Wang, partner at JSL an AI first.
• 0:24
  John: Healthcare investment fund that is joining with some of the cutting edge companies that are leveraging technology in advancing healthcare and hopefully creating better outcomes at lower costs.
• 0:36
  John: Lakan, welcome to the show.
• 0:38
  Lekan: Thanks, Ron for having me.
• 0:40
  John: Lakan, maybe we'll start with, since everybody claims to be an AI expert today, you actually have some experience in AI before this.
• 0:50
  John: AI moment.
• 0:51
  John: Maybe tell us a little bit about your own background and how you got interested and involved in technology in the first place.
• 0:59
  Lekan: Yeah, no, that's, that's a great question.
• 1:00
  Lekan: So, I, I was, uh, grew up in Minnesota, um, where, you know, actually a few blocks away from, from Optum headquarters, uh, and from a young age, you know, I was hearing about all these people in Silicon Valley and got super interested.
• 1:12
  Lekan: So, ended up at Stanford, um, went there for a couple of years.
• 1:16
  Lekan: And when I graduated, you know, a, a single genomic sequence was, I think it had just fallen under a million dollars a piece.
• 1:23
  Lekan: Um, and I was super interested in, in genomics.
• 1:25
  Lekan: Um, even thought about doing, doing a PhD for, for a little bit of time, but, um, ended up graduating into the recession and joined this little, uh, startup called Palantir Technologies.
• 1:34
  Lekan: Uh, at, at the time, um, when I joined, I was mostly doing some intel and, and, and defense work, but before I knew it, I, uh, had a kinda healthcare opportunity there and with a couple of people, uh, started delving into healthcare.
• 1:46
  Lekan: So
• 1:47
  John: maybe slow down there to describe.
• 1:49
  John: Who Palantir is, and they're, they're a, a, a mega tech company now.
• 1:54
  John: But how did it start and maybe tell the, what's the backstory about, uh, maybe Alex Karp and, and Palantir to kind of give folks a sense of, of of, of how novel it was at the time.
• 2:08
  Lekan: Yeah, I mean, the founders of Palantir can definitely give you the cleaner version of that story, but I think at the time, I mean the, the contrarian view of Palantir was that it could actually, uh, integrate lots of disparate pieces of data together.
• 2:20
  Lekan: You know, Palantir was built from.
• 2:23
  Lekan: Kind of, kind of the ashes of nine 11 when a lot of our intelligence and, and defense data was not well integrated.
• 2:29
  Lekan: And, and that partially allowed nine 11 to happen.
• 2:31
  Lekan: So the theory was that if we could, you know, integrate some of that data, but still really respect, you know, civil liberties, security rules and all of that, we could actually kind of get the both of, uh, the best of both worlds, which is that you, you could actually use data more effectively while still preserving, uh, uh, preserving privacy and civil liberties.
• 2:49
  Lekan: And, and I think in healthcare, when we first started, there were a couple of use cases.
• 2:52
  Lekan: We were.
• 2:53
  Lekan: Working on one of them was actually working with the Centers for Disease Control and Prevention, the CDC on foodborne epidemiology.
• 2:59
  Lekan: And, and one of the, I think, interesting things about foodborne epidemiology, especially with the CDC, is that a, a, a lot of, uh, you know, foodborne.
• 3:08
  Lekan: Investigations are actually done at the state and, and county levels.
• 3:11
  Lekan: Um, but the CDC kind of has a coordinating function.
• 3:15
  Lekan: So as you're investigating one of these large foodborne outbreaks, it's actually quite tough to, uh, you know, preserve the data integrity from each one of these counties and states integrate it together.
• 3:26
  Lekan: Um, but you know, sometimes you can ask a certain question or you can't ask a certain question.
• 3:30
  Lekan: Sometimes you can see a certain field of data and you can't, that actually becomes a somewhat.
• 3:34
  Lekan: Challenging technical issue.
• 3:35
  Lekan: So that ended up becoming one of the first problems that, that we worked on.
• 3:39
  Lekan: Um, and then I think since then, I mean, I, I think Palantir these days, a lot of people, you know, really think of 'em as this, you know, defense intel company.
• 3:46
  Lekan: But, uh, I, I haven't been there in many years, but, but I believe the, the majority of its businesses outside of Intel and defense and, and in healthcare, the, the problem it's solving, it's a. Quite an unsexy one, but one I still really believe in, which is just how do you really integrate a lot of disparate data and really make it, make it useful for, for the people in the front lines who are using it.
• 4:07
  John: And, and, and obviously that is what a lot of the, the, the, the heavy metal in, in machine.
• 4:15
  John: Learning, uh, machine language, uh, work early on and deep database work.
• 4:20
  John: But then what, what's that transition into artificial intelligence?
• 4:24
  John: When did that sort of become a thing?
• 4:27
  John: Because initially in, in Palantir's defense work and I'll, and also the work with the CDC and also frankly during COVID, I.
• 4:35
  John: Pulling together different disparate forms of data that that, that, that, that needed to be tied together, but were different in their native state and, and needed to be kind of analyzed at some level in order to be normalized.
• 4:46
  John: And then draw insights from, when does that go?
• 4:49
  John: From just large databases and, and, and kind of, uh, you know, classic.
• 4:56
  John: Um, kind of regression analysis and, and being turned into these, these sort of automated loops that are now what we out sort of loop, bond together and call artificial intelligence.
• 5:08
  Lekan: Yeah.
• 5:08
  Lekan: You know, you're, you're totally right.
• 5:09
  Lekan: I think when, when Palantir was founded in kind of the, the early two thousands, it was definitely the era of, I, I would say kind of the era of AI before the current era.
• 5:18
  Lekan: Right.
• 5:18
  Lekan: So, so just to give you some, some, um, some, uh, dots in the timeline in 2024, uh, DARPA through the DARPA Grand Challenge, which was, you know, the first time they tried to get automated cars to essentially drive through a, a desert compared to what cars are doing today.
• 5:35
  Lekan: Yeah.
• 5:35
  John: People realize the Defense department.
• 5:36
  John: Was the first automated car investor.
• 5:40
  Lekan: Yeah.
• 5:41
  Lekan: The, the Defense Department has a very close tie with, uh, with, with the Valley and AI and everything that happens today.
• 5:47
  Lekan: Um, of course all people have conspiracy theories about that, but I mean, DARPA and, and the Valley and technology innovation has always had a very close relationship.
• 5:55
  Lekan: But it was, it was in 2004 that DARPA threw its first grand challenge and I think in that first grand challenge, it was like, no car made it past like 15 minutes or, or some super short distance.
• 6:03
  Lekan: Right.
• 6:04
  Lekan: The crazy, crazy thing.
• 6:05
  Lekan: A year and a half later in, uh, I believe fall of 2005, DARPA threw, uh, kind of a reprise of the DARPA Grand Challenge.
• 6:14
  Lekan: And at that time, a couple of cars made it, made it to, uh, made it to the end of, of this very long drive through the desert.
• 6:21
  Lekan: Um, what's even crazier is just two years later in the fall of 2007, DARPA through the DARPA Urban Challenge, which is essentially a simulation of an urban course and, and cars were, were able to do that.
• 6:31
  Lekan: Um, I, I do think between the DARPA challenges and in.
• 6:36
  Lekan: 2009 to 2012.
• 6:39
  Lekan: This is kind of the onset of convolutional neural networks.
• 6:42
  Lekan: This essentially is much more complex, kind of deep neural networks that are just far larger than before.
• 6:48
  Lekan: Um, trained on this dataset called ImageNet, that Fefe Lee from, from Stanford created, um, in order to to train these, these large neural networks.
• 6:58
  Lekan: Um, the results of those two essentially swept in kind of the next era of AI where people were saying, Hey, it actually looks like, um.
• 7:06
  Lekan: We're actually able to do much more kind of real life scenarios, uh, with AI using very, very large, uh, uh, amounts of data.
• 7:14
  Lekan: Before that, um, we were using ai, but it was much more, I think you're right, it was much more statistical.
• 7:19
  Lekan: That was the era of statistical AI using, you know, what we today would probably call classical machine learning.
• 7:24
  Lekan: Um, using things like support vector machines and other types of methods to essentially, you know, classify one thing.
• 7:29
  Lekan: Versus another or use reinforcement learning, but using essentially more complex linear models.
• 7:34
  John: I think what's interesting Lekan, when you look back at ai, I mean, AI was a term that was coined in 1957, you know, with the, the, the, the, the, the, the Norbert Wiener kinda like.
• 7:45
  John: Don't even know.
• 7:47
  John: It was like the darmouth workshop, right?
• 7:49
  John: The Dartmouth workshop.
• 7:50
  John: Yeah.
• 7:51
  John: And, and, and the promise of that had been out for so long that people didn't really believe it.
• 7:55
  John: That when Fei Fei Lee and, and the ImageNet and you could identify a cat or a dog, the, the output of it seemed sort of, uh, pedestrian to outsiders.
• 8:06
  John: But the insiders realized that the machines were starting to work and starting to actually.
• 8:11
  John: Um, create novel insights and insights that hadn't been there before off of much larger data states.
• 8:17
  John: And the learning component, I think is the thing that was sort of triggered.
• 8:21
  John: Okay, so that's, that, that, so AI becomes relevant, you know, 10, 10 ish years ago.
• 8:27
  John: And also there's, which we don't need to get into, but the hardware really radically changes.
• 8:32
  John: And, and, and your ability to actually pull in more data to actually run the models.
• 8:37
  John: What's the relevance of tho that, of those models and those machines to healthcare?
• 8:43
  John: Like why is that?
• 8:45
  John: How do, how do we think about sort of stepping back why AI and all of these machines and all this data matters in healthcare?
• 8:53
  John: Why is it relevant?
• 8:55
  Lekan: Yeah.
• 8:55
  Lekan: I think, uh, so, so first, going back to one thing that you said, um, how, you know, even back in the day we talked about AI and, and I think one of the interesting things is.
• 9:04
  Lekan: You know, how, how do we even define AI and what is AI and what's good enough?
• 9:08
  Lekan: And I think this is one of the things that, that it keeps changing, right?
• 9:11
  Lekan: Uh, you, you can read papers back in the day that essentially says, you know, we have these amazing washing machines now that can essentially run a load all by itself.
• 9:19
  Lekan: And people, people honestly, of.
• 9:21
  Lekan: Call, call, call that artificial intelligence.
• 9:22
  Lekan: Right?
• 9:23
  Lekan: And, and if you think about, you know, what we did before then that was, that did actually seem pretty intelligent and, and pretty, pretty amazing.
• 9:29
  Lekan: Um, in, in the 1950s.
• 9:31
  Lekan: We also had, you know, the system, Eliza, which is essentially, today you would call it probably like a very simple rules-based therapist, chatbot, but literally it just, you know.
• 9:41
  Lekan: Essentially asks you Socratic questions back and people consider that ai, so the definition of AI really just keeps changing year over year.
• 9:48
  Lekan: Um, and especially today.
• 9:50
  Lekan: Um, I think from a, from a, you know, monthly basis probably, it seems like the definition of what we consider AI actually changes.
• 9:56
  Lekan: But to get back to your question, I, I think the current era of.
• 10:01
  Lekan: AI was really driven by three main things.
• 10:04
  Lekan: Okay, so number one, and I think this is the biggest thing by far, is that we have a much greater availability of, of data, of all kinds.
• 10:13
  Lekan: Um, today's models are trained on an absolutely, uh, almost unimaginable amount of data.
• 10:19
  Lekan: It's essentially a very large percentage of the entire internet together with, uh, a lot of other synthetic data together with, uh, armies of humans, essentially.
• 10:28
  Lekan: Also giving it, uh, essentially labels to train on.
• 10:31
  Lekan: Um, this is.
• 10:32
  Lekan: Then driven by, by two things that, that make this possible.
• 10:36
  Lekan: Number one is we have a much larger amount of compute available.
• 10:42
  Lekan: So, uh, Moore's Law, um, has, uh, you know, you ask some people, it's, it's sort of trailed off.
• 10:47
  Lekan: But if you actually look at, um, GPUs or graphical processing units, and these days we have, you know.
• 10:52
  Lekan: You know, essentially tensor processing units or other types of specialty chips, specialty asics, that are designed specifically for training neural networks.
• 11:01
  Lekan: These are still getting, uh, you know, faster, uh, in an exponential rate.
• 11:05
  Lekan: Um, but a lot of this is also driven by this very specific model architecture called the transformer architecture.
• 11:13
  Lekan: Uh, this was a debuted in a paper that is very well known today in the industry called Attention is All You Need came out June of 2017.
• 11:21
  Lekan: Uh, and what this.
• 11:22
  Lekan: Uh, essentially transformer architecture did, is it took some of the older styles of, of neural networks that, uh, you know, honestly we've known since the fifties and sixties.
• 11:31
  Lekan: Uh, and it essentially augmented it in, in a few very key ways.
• 11:36
  Lekan: One of the key ways is that instead of having to look at all of your previous, uh, previous weights in each.
• 11:44
  Lekan: Uh, iteration of training.
• 11:45
  Lekan: It can essentially do it in kind of a much more scalable form by just multiplying matrices together.
• 11:50
  Lekan: Um, this is very simplified.
• 11:52
  Lekan: Sorry for everyone who's technical.
• 11:53
  Lekan: I'm super simplifying this, but it essentially allows this to be, to be, to be very scalable,
• 11:57
  John: but simplifying this.
• 11:58
  John: There is a, there's a, a, a, a, a reach of data and the internet.
• 12:03
  John: Plus the ability to build synthetic data or, or, or sort of create more data plus humans.
• 12:09
  John: The data piece is there, it's really a three part piece.
• 12:12
  John: The hardware is novel, much more powerful and much faster.
• 12:18
  John: Then you take those two components and you add a different way of architecting the learning, the the ability to drive.
• 12:27
  John: To, to answer questions, drive insights, and that, and then that loops because it, it, it, it, you're, you're going back through the, the, the armies of humans back through the data, running it through bigger and bigger hardware and Yep.
• 12:41
  John: That attention is all you need basically was a, a better model to actually.
• 12:47
  John: To, or better architecture to build faster and more insightful models.
• 12:51
  John: And I, I think, I think those, no, I think that's a, that's, that's a, I don't think you're, I think that's exactly that, that, that created the moment we're in right now.
• 13:00
  John: That's awesome.
• 13:02
  John: Why is it relevant to healthcare?
• 13:05
  Lekan: Yeah, I mean in healthcare, um, as you know, many, uh, may recall, like we, we've tried using, uh, you know, what we, we call AI in healthcare many times, right?
• 13:16
  Lekan: Um, anywhere from the expert systems to, you know, we use a lot of graphical models and bays nest in the statistical era.
• 13:22
  Lekan: Uh, and each time, you know, there's been aspects of AI that's been promising.
• 13:28
  Lekan: But really, really underdelivered.
• 13:30
  Lekan: Um, and we also know that healthcare is an area that, uh, traditionally has been relatively low to relatively slow to adopt technology.
• 13:40
  Lekan: Um, today I think we have a few big problems in healthcare.
• 13:43
  Lekan: We have, uh, a surge in documentation and administration, and every clinician will tell you that they spend so much time doing documentation and, and administration.
• 13:54
  Lekan: Um, there's what I would.
• 13:55
  Lekan: Think of as almost competitive documentation and administration between payers, providers, and other entities as well.
• 14:02
  Lekan: Uh, and also we have a huge access problem in healthcare too, right?
• 14:05
  Lekan: I mean, I, I, I think for, for both you and me, we, we are in healthcare and sometimes, you know, I even have trouble understanding how to get the care I need and how to navigate my own care.
• 14:14
  Lekan: Uh, so it, it's, it's really, really challenging to get here at care.
• 14:17
  Lekan: And, and I think with.
• 14:18
  Lekan: The current wave of AI in healthcare for, for the first time.
• 14:23
  Lekan: Um, and, and you know, I've, I've always been the person who, who's been an AI who's been a little skeptical of, you know, actually AI applications in the real world.
• 14:30
  Lekan: I'm actually, for the first time, actually extremely optimistic about how AI can actually affect these huge problems in healthcare from, from, uh, decreasing administration to, uh, to increasing access, to actually increasing the number of, uh, cures and treatments in, in healthcare as well.
• 14:47
  Lekan: Uh, and this is actually being born in data too.
• 14:51
  Lekan: This is not just, you know, uh, VCs, which I am one investing in healthcare is hoping that this is gonna come in seven years.
• 14:57
  Lekan: There's actually real, you know, clinical data, third party validated data, showing that in many of these areas, AI is, uh, really.
• 15:05
  Lekan: Uh, really causing an improvement, um, of, of these things.
• 15:09
  Lekan: This is not to say that, you know, AI will cure everything.
• 15:12
  Lekan: It'll make everything better.
• 15:13
  Lekan: We actually have to be quite careful about how we approach this.
• 15:16
  Lekan: But, but overall, I think now is the time to be very optimistic in the use of AI in, in, in healthcare.
• 15:20
  John: So maybe, maybe, maybe give folks a as little bit of a sense of what, what's going on at JSL and kind of what you're doing.
• 15:29
  John: What, what brings you to the, the point in investing and what kinds of things you're investing in.
• 15:34
  Lekan: So JSL Health, uh, investment is a, a healthcare venture capital investment fund that started in around 20 21, 20 22.
• 15:42
  Lekan: We invest only in healthcare companies that serve the US market, uh, and only healthcare companies that have some kind of technology, AI software, data angle.
• 15:52
  Lekan: Um, and, and I think like when we think about that, there's a couple of different areas.
• 15:57
  Lekan: I mean, number one, we definitely inv invest in kind of AI first companies.
• 16:02
  Lekan: Um, and this is in both, you know, healthcare and life sciences.
• 16:05
  Lekan: Uh, I, I think maybe a good example of this would be a company called Symphony based in the Bay Area.
• 16:10
  Lekan: They do medicinal chemistry.
• 16:12
  Lekan: Uh, they, uh, essentially use AI to predict properties of certain chemicals.
• 16:17
  Lekan: They predict the synthesis path of how you create these small molecules.
• 16:21
  Lekan: And then they actually actually have robots that then you can, you know, kinda just double click a button and a real drug will be printed out of the other side in both, you know, tiny flow chemistry, nano molar quantities, all the way up to gram quantities that you can actually administer.
• 16:33
  Lekan: Uh, and this is of course a nice feedback loop.
• 16:36
  Lekan: So, so I, I think like a company like that is interesting from.
• 16:40
  Lekan: Perspective because, you know, like I said before, a lot of AI is you're feeding it in, feeding a lot of data into these models, it's learning and then it's spitting stuff back out.
• 16:49
  Lekan: Right?
• 16:50
  Lekan: Uh, a lot of people, this gets into a little bit of, you know, the technical side again, but one of the big questions in AI these days is how well these models can actually generalize, um, out of the training domain.
• 17:02
  Lekan: And this is a very complex way to say.
• 17:05
  Lekan: D is the AI just repeating what exactly what it's learned or connects the stochastic pair for itself?
• 17:11
  Lekan: The stochastic pair.
• 17:12
  Lekan: Yeah.
• 17:12
  Lekan: I, I, I, you know, I sort of dislike that term, but that, that is kind of the, probably the, the short way.
• 17:18
  John: Yeah.
• 17:18
  John: F folks who virtue the, the premise here, the, there, the, the pro ai is that it, there is learning and insight and novel things that are where a brilliant.
• 17:28
  John: Human would be, or better.
• 17:30
  John: The stochastic parrot is, it's just repeating back to us in a off of more data what humans already know and, and so, but presumably you're not investing in parrots.
• 17:43
  Lekan: That That's right.
• 17:44
  Lekan: Yeah.
• 17:44
  Lekan: I, I mean, well, I, I,
• 17:46
  John: so I think the, by think, just, just, just to fair disclosure, I sit on the JSL board where Lekan and his partner sometimes listen to my, my, my questions.
• 17:56
  John: But, but I, but pre presumably we're a little stubborn
• 17:58
  Lekan: sometimes,
• 17:59
  John: presumably, um, uh, that, but that is one of the interesting questions among folks who are, who are sort of on the theoretical and the practical side, honestly, because.
• 18:10
  John: Whether it's learn, whether it's just grabbing more information and getting to the answer faster, or actually creating novel insights or both is, is, is, is, is quite relevant to the application in healthcare where we're dealing with, you know, human lives that are, that are fragile.
• 18:27
  Lekan: Yeah, that's exactly right.
• 18:28
  Lekan: And, and you know, in, in some ways I think the prob this, you know, open question of whether these models are, uh, you know, generalizing beyond, beyond its initial domain.
• 18:39
  Lekan: Uh, it, it, it's sort of, you know, not, not that helpful of a question in other ways it's very important, but, but I think the way that we think about it is that, uh, no matter whether it can generalize beyond it having more training data.
• 18:52
  Lekan: Within the domain that you're working in, uh, and using the same, almost like s. Semantics that you are using is very helpful to improve the model, no matter wh whether or not it can actually generalize.
• 19:05
  Lekan: And, and this is actually why we are very bullish on companies that have some kind of proprietary data or even better generating interesting novel proprietary data or data that's.
• 19:17
  Lekan: Better capturing, uh, kind of the, the full operations of clinical care better?
• 19:23
  Lekan: Um, I think, you know, if you go to a hospital, uh, especially in certain settings, um, I think everyone knows that, you know, parts of it are captured, um, in administrative data, in clinical records.
• 19:33
  Lekan: But I think everyone knows that a lot of both clinical data, clinical discussion, as well as operational data is actually not that well captured in the documentation itself.
• 19:43
  Lekan: Right.
• 19:44
  Lekan: Um, and that actually I think does lead to a gap in some of these systems that, that's trained on AI that isn't able to use some of that data in, in the understanding of, of patients in clinical care.
• 19:56
  Lekan: Um, so I mean, we, we have some other companies too.
• 19:58
  Lekan: So another company called, uh, general Medicine.
• 20:01
  Lekan: Um, you know, founded by former founders of PillPack, they're really thinking about using, uh, AI for, uh, creating much more accessible care to a lot of Americans by making costs transparent, by having clinical navigation, uh, be very, very easy by having, um, you know, very experienced physician consults be very easy.
• 20:22
  Lekan: Um, and I think that's kind of a company.
• 20:25
  Lekan: Where it's still very physician first.
• 20:27
  Lekan: Um, at every stage there's a physician there, but it's really using AI to really think about, um, how do we use that better for, uh, for patient triage?
• 20:38
  Lekan: How do we use that better for scheduling, for making the experience for the patient, uh, much better and and much clearer while using all the, all the data that you currently have.
• 20:47
  John: I mean, you have, you have some pretty clear problems in healthcare.
• 20:50
  John: It takes more than 15 years between the time that, uh, an, an area of medicine actually changes a clinical, uh, criteria or pathway.
• 21:00
  John: Yep.
• 21:00
  John: And the time it becomes generalized.
• 21:01
  John: There are an estimates of 500 to 750,000 medical errors every year that are avoidable.
• 21:09
  John: That happen in Yep.
• 21:10
  John: The administration, I think it's, that's the 20% of all autos, autopsies pick up medical things that might have been, uh, uh, uh, that were missed in diagnosis that might have had an impact on the, on the, on the, on patients patient, whether the patients would've lived longer.
• 21:27
  John: Um, so there's, this is a target rich environment for novel intervention interventions, but historically.
• 21:34
  John: The promise of technology and data, um, has run well ahead of the success of using that data and changing either the administrative load to your point, like it's a pain in the butt to make an appointment, see a doctor, even get through the documentation between meetings in a hospital, to real clinical insight in terms of what's, what's new about.
• 21:58
  John: Personalized medicine.
• 21:59
  John: Again, promise 20 years ago to you started in this 20 years ago, we're still kind of scratching at whether we can truly personalize drugs and, and genomics proteomics and come up with any insights that it'll help people live longer.
• 22:14
  John: What's different this time?
• 22:16
  John: And with every, there's a lot of hype about this.
• 22:20
  John: Uh, there are a lot of problems to be solved.
• 22:22
  John: What's different now that gives you optimism that your investments, um, and, and the broader investments in AI and data are gonna actually improve care or lower costs.
• 22:34
  Lekan: Yeah, I mean, I think there's a couple of things I'll I'll say about that.
• 22:37
  Lekan: Um, the first thing is just that the, like the, the ROI is is definitely there already, right there, there's been papers that have shown that, um, and I know this is sort of a cartoon problem, but it, but it is real problem where, uh, if you just look at differential diagnosis, right?
• 22:53
  Lekan: Something that physicians do every day.
• 22:55
  Lekan: I know it's not the whole job by far, but if, even if you look at differential diagnosis, there's a bunch of papers out.
• 23:01
  Lekan: Now that essentially show that, uh, AI physicians are able to perform at a very, very high level, you know, depending on which paper you read anywhere between, you know, 70th and 90th per percentile, right?
• 23:12
  Lekan: Um, of course, again, this is, I'm not saying this is what physicians do all day.
• 23:17
  Lekan: But this is getting to the point where, you know, it, it's i'll, I'll even go ahead and say it.
• 23:21
  Lekan: It's starting to get to the point where it's almost going to become, you know, unethical to not at least consult an AI in certain types of diagnoses.
• 23:30
  Lekan: Right.
• 23:31
  Lekan: Um,
• 23:31
  John: that's a big deal.
• 23:32
  John: Yeah.
• 23:33
  John: That's saying if you don't use the machine, you're practicing at a, at, at a, at a, at a suboptimal level and you're dealing with human life.
• 23:41
  Lekan: Yep.
• 23:41
  Lekan: I actually agree with that.
• 23:42
  Lekan: Yes.
• 23:43
  Lekan: Um, and, and this is, again, you need to look at it on a case by case basis for, for certain specialties, doesn't make sense for other specialties.
• 23:49
  Lekan: A hundred percent make sense, right.
• 23:51
  Lekan: Um, some other results that are interesting and, and, you know, because of just how our industry works today, they've had, you know, we've had studies having patients rank, uh, an AI chatbot versus a physician on a chat about how empathetic they seem.
• 24:05
  Lekan: Um, and AI seems to win.
• 24:07
  Lekan: The, the vast majority of the time too.
• 24:09
  Lekan: And of course this may be because due to physician burnout, due to just, you know, how much time you can actually take to, to put pen to paper to explain certain things.
• 24:17
  Lekan: But, but I find that really interesting as well.
• 24:18
  Lekan: But, but now of course, in, in a real world scenario, um, we all know that care is much more complex than that.
• 24:25
  Lekan: You often times have a much longer conversation with a physician.
• 24:29
  Lekan: You may not want to share everything.
• 24:31
  Lekan: Um, with a physician, you may even, you know.
• 24:33
  Lekan: Everyone describes their symptoms incorrectly.
• 24:35
  Lekan: Sometimes there's a, you know, misunderstanding.
• 24:37
  Lekan: Oftentimes there's coordination of care that has to happen.
• 24:40
  Lekan: Oftentimes, uh, clinical care goes more than just beyond clinical, the actual clinical care.
• 24:45
  Lekan: Right.
• 24:45
  Lekan: And in those scenarios, I mean, there's also been studies that show that in a complex care scenario, ais.
• 24:51
  Lekan: Perform, you know, potentially only slightly better than a physician or in other cases may even under perform a physician.
• 24:57
  Lekan: And, and I often chalk that up to the fact that a lot, again, a lot of clinical operations are just not digitized right now.
• 25:07
  Lekan: Um, so, so, you know, one of my,
• 25:09
  John: or maybe there's still a role for, maybe there's still a role for some humans.
• 25:13
  Lekan: Oh, that's actually exactly what I mean, right?
• 25:15
  Lekan: I think there is, there is a role for humans, but I think that the key we need to think about is how do we.
• 25:20
  Lekan: Uh, capture these operations digitally.
• 25:23
  Lekan: Yeah.
• 25:23
  Lekan: So that we can insert either insert humans or insert the best ai what, what, whatever the, the scenario calls for.
• 25:30
  Lekan: Right?
• 25:30
  Lekan: Yeah.
• 25:31
  Lekan: Um,
• 25:32
  John: and figure out.
• 25:32
  John: So I'm figure, but figuring out that dance between.
• 25:35
  John: The machine and, and the doctor is gonna be interesting.
• 25:38
  John: Maybe Lekan, we could, we could do a lightning round to, to sort of wrap up.
• 25:44
  John: You've done a great job at grounding us in kind of where we're going and what the opportunities are, but I'd like to ask you sort of from a predictive perspective, some, some, some, uh, uh, I want your, your predictions on some of the things that people have talked about.
• 25:58
  Lekan: Let's do this.
• 25:59
  John: How soon, how soon before we have an AI doc replacing the regular doc.
• 26:06
  Lekan: I think for many Americans now already.
• 26:08
  Lekan: Uh, and I think if I can just elaborate on this, uh, many people just cannot get access to care, period.
• 26:14
  Lekan: And compared to that, um, pulling up, you know, chat GPT or you know, you should use our company General Medicine.
• 26:21
  Lekan: Uh.
• 26:22
  Lekan: I, I think that is already better than not getting care.
• 26:25
  John: How important is it for patients in terms of self-care to use the AI versus trying to find a way to talk to the doctor about how they're gonna manage their life?
• 26:37
  Lekan: I think these days, I, you know, this is what I'm actually personally doing.
• 26:40
  Lekan: It's about 50 50.
• 26:41
  Lekan: I still trust a real world physician.
• 26:44
  Lekan: I know there are biases in some of these models and they miss certain things.
• 26:47
  Lekan: They have lack of stability sometimes.
• 26:49
  Lekan: So I, I would do 50 50, but, um, so I would use AI to pull up all your records, make sure that you kind of understand your own condition, but then have the final conversation with a physician.
• 27:01
  John: Great drug discovery, the promise of technology, machine learning, AI has been out there that we're gonna be leveraging the machines to finally discover drugs faster, better, smarter.
• 27:14
  John: It costs over a billion dollars to, to, to bring a drug to market.
• 27:17
  John: It goes through multiple ma, many of the most novel, the biggest problems, cancer, heart disease, whatever, and the, and, and some of those smaller categories where there's a lot of.
• 27:29
  John: Very specific data available.
• 27:31
  John: Most drugs fail.
• 27:33
  John: How long before we can leverage AI to actually materially improve our ability to produce drugs for some of the, the hardest categories of illness we have.
• 27:45
  Lekan: Yeah, this is a, a tricky area, I think in the very early stages of just drug discovery for, you know, target selection, molecule discovery.
• 27:52
  Lekan: That's already happening today, without a doubt.
• 27:54
  Lekan: Um, I also do think in later stage, uh, you know, doing feasibility studies on clinical populations and running clinical trials, that's also, uh, starting to, uh, get better.
• 28:06
  Lekan: But I, I think it's a, there's a lot of, a lot of the middle stages due to a combination of, you know, existing processes that are hard.
• 28:13
  Lekan: As well as regulation.
• 28:14
  Lekan: Um, I think in probably five to 10 years we're going to start having a lot more drugs go through this type of AI first drug pipeline.
• 28:23
  John: Awesome.
• 28:23
  John: I think we'll leave it there, Lakan.
• 28:25
  John: And this has been a wonderful introduction to ai, uh, a tour, tour zone of how it's being used right now.
• 28:31
  John: And I really appreciate you, you joining us.
• 28:34
  John: Uh, if you, but like what you heard or you didn't, we'd love you to subscribe on your favorite service to Care talk.
• 28:40
  John: I'm John Driscoll, the chairman of the Yukon Health System and Lekan, thank you for joining.
• 28:46
  Lekan: Thanks for having me, John.