Episode 53: Dr. Steve Labkoff and Dr. Leon Rozenblit Discuss Practical AI in Healthcare

Artificial intelligence is everywhere in healthcare, but how much of it is actually making a meaningful impact?

In Episode 53 of The Informonster Podcast, Charlie Harp sits down with Dr. Steve Labkoff and Dr. Leon Rozenblit, co-hosts of the Practical AI in Healthcare podcast, to discuss their backgrounds in informatics, the inspiration behind their podcast, and what they've learned from speaking with healthcare leaders at the forefront of AI adoption.

The conversation explores the difference between hype and real-world results, why AI literacy is becoming increasingly important, how healthcare organizations can identify practical use cases, and the critical role data quality plays in building trustworthy AI solutions.

Contact Clinical Architecture

• Tweet us at @ClinicalArch 
• Follow us on LinkedIn and Facebook
• Email us at informonster@clinicalarchitecture.com

Thanks for listening! 

Transcript

Charlie Harp (00:09):

I'm Charlie Harp and this is the Informonster Podcast. Today on the Infomonster podcast, I have Drs. Steve Labkoff and Leon Rozenblit. And we're going to talk about who they are, how they got into healthcare and their practical AI and healthcare podcast. Gentlemen, welcome to the podcast.

Dr. Steve Labkoff (00:31):

Thank you so much, Charlie. Nice to see you again.

Dr. Leon Rozenblit (00:34):

Thanks, Charlie. We're off for a good foot. You took away my doctorate. What the hell?

Dr. Steve Labkoff (00:39):

He's Dr. Rozenblit.

Charlie Harp (00:39):

I said Doctors.

Dr. Leon Rozenblit (00:41):

Okay. Okay. I'll let you go. I think I got a JD and a PhD. I know it doesn't add up to an MD, but sort of.

Charlie Harp (00:48):

Okay. Okay.

Dr. Leon Rozenblit (00:49):

I'm a respectable guy. I've been to school.

Dr. Steve Labkoff (00:53):

Boy, I knew that was going to go badly Charlie. I knew it.

Charlie Harp (00:57):

All right. Noted. There's sensitivity there.

Dr. Leon Rozenblit (01:00):

No, no. It makes for a good intro joke though.

Charlie Harp (01:05):

Well, hey, welcome today. And what I usually do on the podcast is I started off by having the guests kind of talk a little bit about who they are in case these folks haven't heard your podcasts or haven't interacted with you. And a little about your journey that what brought you into healthcare and you guys can decide who's going to go first.

Dr. Steve Labkoff (01:26):

You know what, Leon? I'm going to let Leon go first because I usually go first in this stuff. Leon, why don't you take it away?

Dr. Leon Rozenblit (01:31):

Yeah. I mean, my story is much more confusing, so why don't we get started by confusing everybody? There you go. So I went into informatics through the traditional old school indirect route. There used to be a period when you talk to an informaticist and they have some completely bizarre background where I started as a gardener and then I picked up a needle and I learned about computers. So mine is sort of like back then. Nowadays, you ask one of the younger folks is like, "I went to school for informatics and I got a PhD in informatics and I got an MD." It's like, "Oh, it's so boring." So I started out as a cognitive scientist. I actually had a legal background before that, but I was really interested in how communities build knowledge systems. So that translated into me building databases for large scientific projects, mostly things that were happening at Yale.

(02:22):

They were like large pediatric longitudinal studies that required multimodal data collection and organization. I turned that into a consulting practice and a business called Prometheus Research and we sort of became a niche provider of really complicated scientific database systems. Our customers gradually taught us to call them registries. I did not love that term at the time. I'm like, "I don't know what that is, but you listen to your customers and so eventually we would say we described ourselves as a registry vendor. We grew the company into supporting some of the larger national and international projects. We worked for some of the bigger medical specialty societies and patient advocacy groups and projects some of the audience may have heard of and eventually got annoying enough to like our bigger competitors where instead of like competing with us, just bought us. So we got bought out by IQVIA.

(03:23):

Actually very nice. It was a good relationship. They treated us well. I came in at IQVIA. I led an internal group that provided consulting services to the patient advocacy groups, medical specialty societies, and we tried to help them understand how to do a data strategy better and then transition them into to see if they wanted to build the registry, would help them build the registry as well. When my time was up, I decided I wanted to go back to academic work and kind of like shift into lower gear. So my wife describes me now as like the busiest semi-retired person that she's met and I'm working mostly in academic projects again in data governance. Specifically, I became really interested in AI and healthcare. So I've been working with Steve and others at the DCI network on AI in healthcare governance models and also with the DIA on AI governance models in the life sciences.

(04:21):

I'm generally interested in AI applications in healthcare. And of course, one of the more important and fun things that I've been doing is the working with Steve on the "Practical AI and Healthcare" podcast. So that's my short version of my story, how I got here and why my life makes absolutely no sense.

Charlie Harp (04:40):

That's great.

Dr. Steve Labkoff (04:43):

Yeah. So my background is a little more ... Well, I don't know if it's more traditional, but I started life as a clinician. I went and trained as an internal medicine doctor. I got board certified. I went and did a cardiology fellowship and was a cardiologist for a short bit, but I got bitten by the informatics bug and left cardiology and did a postdoc in clinical informatics at Brigham Women's Hospital at MIT. It was a joined project through Harvard Medical School and spent a total of four years up there and then got recruited out of that organization and into Pfizer and life sciences. And I spent well over a decade with Pfizer in a variety of roles. But because I was one of the only doctors in the organization that understood IT and understood medicine, I wasn't a typical pharma doc. I wasn't working in medical affairs.

(05:38):

I wasn't working as an MSL. I was working in planning and business development and helping to do some of the most cutting edge stuff that the organization undertook. So as an example, I helped Pfizer to investigate, eventually purchase and then deploy an EHR system back in 1998 through 2001. It became known as Amicor and was a JV between Pfizer and IBM and Microsoft. I got a chance to actually build a hospital in Africa between '01 and '05 where I was going back and forth to Kapala Uganda for building this hospital called the Infectious Disease Institute. And very recently, Leon and I got a call about a guy wanted to be on our podcast who guess what was working at the IDI in Uganda and I got to hear a little bit about the fact that the place is still up and running and is still a viable concern 20 something years later, which was really one of the nicest things I'd heard in a long time.

(06:29):

Fast forwarding after Pfizer, I went to AZ, AstraZeneca where I ran a few departments, one on real-world evidence, one on biomarkers and one on clinical trial interpretation and clinical operations later from there to intelligent medical objects where I helped work on interface terminologies and bring them into focus for that organization and then went back and became the Chief Data Officer at the Multiple Myeloma Research Foundation where that's where Leon and I crossed paths for a second time, but we actually started working together very closely on this multimodal registry for myeloma care called the Cure Cloud. It was one of the largest, if not the largest multimodal data registry that was stood up at that point between 2018 and 2021. And we built it, we stood it up, we've got it launched. And by the time I left at the middle of 21, in the middle of COVID, we had recruited over 750 patients to collect their medical records data, their proteomics, their genomics, nursing notes, patient donated data and a few other random data types and aligned them all by patients so we understood what the journey was looking like from multiple perspectives over time for the patients.

(07:46):

It was highly, highly complicated project and such. From there onto a bioinformatics consultancy known as Quantori and then most recently was the VP of Analytics for Clinical and Medical Affairs at Bristol-Myers Squibb. And that all came to an end and late last year where I then started my own consultancy called Luminant Consulting and Leon and I got together and said, Hey, you know what? It'd be really cool to focus in on this AI and healthcare work. I brought him into the work at Beth Israel. I've been there since 21 and we've been off to the races with that ever since.

Charlie Harp (08:28):

Now I have a few softball questions for you guys. The first one is, Steve, you have well more than one, but you have on particular hobby that I think is pretty cool. You want to talk about it?

Dr. Steve Labkoff (08:42):

Sure. So I got into astrophotography just before the pandemic and it turned out that's taking pictures of nebulas and galaxies and comics and sometimes the moon aligning around like the Statue of Liberty and things like that. And it became a bit of an obsession during COVID. I eventually got myself a master's degree in photography during COVID just because I had too much time on my hands. Yeah, so I don't usually put those credentials after my name because nobody actually knows what an M.photog means, but you actually have a master's degree in photography. That's awesome. People are interested. They could see my work at LumintPix, L-U-M-I-N-A-N-T-P-I-X, luminantpics.com.

Charlie Harp (09:33):

And they're breathtaking. I've seen a number of them. They're very cool.

Dr. Steve Labkoff (09:35):

Yeah. Thank you.

Charlie Harp (09:37):

All right. What fascinating hobbies do you have, Leon?

Dr. Leon Rozenblit (09:40):

Oh, man.

(09:43):

I sublimate all of my hobbies into work. I mean, honestly, in my free time, I'm a movie buff and I love a literature and science fiction. So I just read as a hobby and I do some writing recreationally. I've been working on a book and a background, but it's going so badly I'm embarrassed to even list it because I said I was going to write a book three years ago and people keep asking me, "How's the book going? " I'm like, "Just don't ask." I've learned a lot of humility trying to write a book about how little I know about writing books, but it's still technically in progress. But really, I mean, I'd say my biggest hobby in the last year has been getting, because I was able to be flexible with my time, was getting my hands dirty with AI and just getting into the more technical aspects of harness engineering and playing with CloudCode at a slightly deeper level to build skills, to create hooks, to think about how context Windows can be manufactured and supplemented by file systems.

(10:48):

So that's been so much fun. It's sort of like a drug, you're mainlining AI. You're like, "I can do this now. This is amazing. I just generated this little application in 30 minutes. What the hell?" So it's been a little bit addictive and I have to be honest, that's been my heroin. Oh,

Dr. Steve Labkoff (11:07):

And I'll give him a plug. He's amazing at it. We're going to get into talking about our podcast and we're a group of two with a young lady who helps us along the way for some of our social media work, but Leon has done just the work of like 12 people on the back end of the podcast when we do our digestion of information and organization and all of the data processing because there's a lot of data processing on the back end of a podcast, which I didn't appreciate when we dove into start. We kind of got into this just on a whim and we can get into that in a sec, I guess.

Charlie Harp (11:39):

Yeah.

Dr. Leon Rozenblit (11:40):

Well, thanks, Steve. I mean, Steve's our front man. I just drive the getaway car. So he lines up all the guests and may does all the cold calling and I go, "How do I automate everything?"That's sort of my role in the world. But it's been my role in the world for the past 30 years, Charlie, a litle bit like yours. You look at this problem and you're like, "This got to be a way to automate this. " And the tooling that's coming out is just so exciting. It's hard not to get into

Charlie Harp (12:07):

It. The real question is, how do we know this is really you?

Dr. Leon Rozenblit (12:10):

I'm not sure nowadays. I mean, I tell people who ask me, "Did you do this using AI?" It's like, "Man, we're cognitive cyborgs at this point." Of course, I'm not sure this is really me and I couldn't prove it to you that I am. I mean, I could show you my cat. I

Charlie Harp (12:26):

Was going to say for the listeners, I can see that there's a cat and I don't think AI would have slipped that in.

Dr. Leon Rozenblit (12:30):

No, that would be weird. Yeah.

Charlie Harp (12:32):

So how did you guys meet?

Dr. Steve Labkoff (12:34):

Oh, we met through AMIA. Leon's company had a booth at AMIA, which these days is having a company have a booth at AMIA is not as common anymore as it used to be, but his company had a booth there. I was looking for companies that specifically had connections in Connecticut because I live in Connecticut and Leon's company was there and had a big Connecticut logo at the end of it or something and I linked on and we started talking and a few years later when I needed to bring in a partner to help us build the Cure Cloud, I called Leon and said, "You interested in bidding on this project?" And he was and he did and they won the bid, Fair and Square. I didn't put my thumb on the scale or anything and they did a fantastic job. They really, really did. They did a lift on this piece of work that I, even to this day, I still got nightmares about it because it was such a complicated thing.

Dr. Leon Rozenblit (13:26):

It's funny that you ended informatics project with a customer vendor relationship and you're still friends afterwards. That's a very good song.That's say something. But yeah, I think we were back then, Steve, we may have been even then one of the few companies and we were like AMIA sponsors for a while and were part of that. I feel a little sad that AME let that go. If we have any pull, Charlie, the three of us should really try to get back to their organization and explain to them that they do have non-academic members

Charlie Harp (14:02):

Who- Well, I'm the current chair of the IPC and I have to remind them that occasionally.

Dr. Steve Labkoff (14:08):

Just to put my own plug in there, you know where the IPC started, right? I started the IPC 15 or 15 years ago or so under the auspices when I think it was Don Demer was president back in the day and we built that in order to try to do precisely what Leon was saying was to bring in the vendor community and the business community into AMIA because again, it had been, and it may always be, but it had been very academically focused and I was working in Pfizer at the time when it was all started and I really felt that I was made to feel like a second class citizen at AMIA. And I worked very, very hard to get corporate members to be seen as partners and to be seen as resources for intelligent conversation and interesting work. And what I find very interesting is that since the administration turned over a couple of years ago, it's interesting how many of the AMIA folks have started looking back at industry for help and we can editorialize on why that is, but there's been a sea change in their attitudes in terms of now the grants are all drying up, where do they get their funding from?

Charlie Harp (15:24):

That's absolutely true. And this is my second stint as the IPC chair and we've got a decent membership and what I've tried to impress upon the folks at AMIA is that we're not in it just to sell. We're part of the community and it's a symbiosis. There are a lot of people that want to go from academia into industry and there are people that are in industry who want to participate in some of the things that happens in academia. And I think we started a webinar series last year where different IPT members talk about what it's like to work in their segment of the industry. So we're trying, Leon. We have our moments. They are in their roots and academic

Dr. Leon Rozenblit (16:08):

Academics. Yeah. And first, Charlie, thank you for your service. You're real genuinely, I know we say that in a joking way, and we've been joshing, but it's a hard job and I appreciate you stepping in and working on it. I'd say I've been learning a lot about the innovation ecosystem by serving as an entrepreneur in residence at Yale and there is a method to the madness, right? A hundred years ago, if you said, "I want to create a bunch of entrepreneurial companies," you wouldn't know what to do. You'd throw money at things, you'd know that a university is somehow involved. We actually know how to create crossovers. And one of the signs of a really healthy innovation ecosystems is where there are class of people who are very comfortable moving jobs from industry back in academia that happened in genetics and you know like, oh my God, biotech has really taken off when you can see professors going, "You know what?

(17:02):

I'm going to take off and work at this genomics company for a while." In informatics, it's surprisingly not as common. You just don't see people like you, me or Steve going into important academic roles and vice versa. You don't see folks in academics going like, "I'm going to go and work for this company for a while." I'd love to see what we and other folks like us, and there are not that many of us could do to make that more of a reality. I just think it'll be a much healthier innovation ecosystem in the field.

Charlie Harp (17:32):

I agree.

Dr. Steve Labkoff (17:35):

Atul Butte was one of those folks who actually did dance on both sides of the island.

Dr. Leon Rozenblit (17:40):

He really was.

Dr. Steve Labkoff (17:40):

He really was. I did really well at that. And I've been in both sides of that camp as well. I mean, like I alluded to earlier in 2021 just before, well, just after Charlie Saffron retired from the division of clinical informatics at Beth Israel, we started to resurrect some of the work that he and I had done while I was at Pfizer and later at AstraZeneca in terms of building out this network of industry and academic and government leaders to come together to solve complicated healthcare problems with informatics. And that eventually started into something that we call now the DCI Network, the Division of Clinical Formatics Network, which Yuri Kintana and I co-founded back in 21 and we're now into it five years. I brought Leon into it about three or four years ago and now we've been running conferences, we've published a bunch of academic papers, Leon's first author on a bunch, I'm first author on a bunch and it's been kind of refreshing to be able to walk back into the academic setting a little bit.

(18:42):

I got one foot in both camps now and it's nice to be able to do that. I think Leon's absolutely right. It'd be nice to see that happen more regularly across the board, but our industry is our industry and they don't necessarily make it so easy to do that yet.

Charlie Harp (18:59):

Well, and I have a couple of my folks on my team are adjunct professors at the University of Utah and here at the Luddy School of Informatics. So we try. I think what we can do is try to strengthen those bonds as much as possible and see where we can go. But hey, I also met Steve at AMIA. So else AMIA brought us together here today. I'll give him credit for that.

Dr. Steve Labkoff (19:22):

Absolutely.

Dr. Leon Rozenblit (19:22):

That's right. They deserve it. That's right.

Charlie Harp (19:25):

Of course, he met with me to yell at me the first time we met. I

Dr. Steve Labkoff (19:27):

Didn't yell at you. Come on.

Dr. Leon Rozenblit (19:31):

I'm sure you deserved it too, just like Amya.

Dr. Steve Labkoff (19:34):

I did.

Charlie Harp (19:34):

I did. No, I'm just giving you a hard time. All right. Unless we're going to make this a two-parter, we should probably talk about your podcast.

Dr. Steve Labkoff (19:41):

Sure,

Charlie Harp (19:42):

Sure. So you guys, you host a podcast, Practical AI in Healthcare, what inspired you to start the podcast?

Dr. Steve Labkoff (19:52):

Do you mind if I start it, Leon?

Dr. Leon Rozenblit (19:53):

No, it was your baby. Go for it.

Dr. Steve Labkoff (19:55):

So in the course of the work we were doing at Beth Israel at the DCI Network, we ran a conference a couple years ago on AI and healthcare. I went to Yuri and said, in 2023, about three months into the new world of AI and healthcare, I said, "This is probably the biggest single thing that ever is going to happen in our lifetime in terms of informatics and we need to get in front of it. " So very quickly we mobilized, we hosted a couple of meetings at Harvard and they resulted in four academic papers and a bunch of additional work. And then last year in September we hosted another set of meetings and this one was called Signal Through the Noise. And the thing that was the generative piece of it was as I was going through my work at BMS and in the industry world, everything I was seeing was all hype.

(20:53):

Everything I could imagine or talk about, everything I saw was left, right and center, massive hype. And what I wasn't seeing at the time were where were the shining points? Where were the use cases that were making a significant difference? What were the things that were truly, as we say, moving the needle in a positive way? So we talked to Yuri about it. We said, "We're not sure we want to do this through Harvard with the Harvard brand because we were not clear if the lawyers would give us a hard time. We wanted to do this separately." Yuri said, "Fine. He's actually been a guest on the podcast." We spun it out and we've been focused in on seeking out new lives and new civilizations. Oh, I'm sorry.

Charlie Harp (21:32):

And

Dr. Steve Labkoff (21:32):

Boldly

Charlie Harp (21:33):

Going?

Dr. Steve Labkoff (21:33):

Boldly going where no inthemetician has gone before. We've been seeking out the use cases and from leaders of companies in academia and various aspects of healthcare delivery to find out what actually is moving the needle, what are the places where we're seeing shining spots of true examples of what does good look like? And we go to great pains. And I was coupling Leon on how much he does in the background for making the podcast a reality from doing the work of 12 with an AI agent or three or 12, but we seek out these business leaders. We've had a number of really strong leaders on the podcast, surprisingly, like we're punching over our weight class, I think. It's just literally the two of us and again, one young lady helping us, but we've interviewed folks like Bob Wachter. We've had Ted Shortlive. We're about to have Zach Kahani on the podcast.

(22:28):

We've had Matt Truppo, who's SVP at Sanofi and a number of CIOs of lesser known companies-

Charlie Harp (22:35):

You had Charlie Harp on there. Most importantly, we had Charlie Harp.

Dr. Steve Labkoff (22:39):

That's right. And the point of it was to ... And by the way, I want to talk about why we had you on there because the work you've done on the PIQI framework is actually insane. But

Dr. Leon Rozenblit (22:51):

Steve, before we get distracted

Dr. Steve Labkoff (22:53):

And

Dr. Leon Rozenblit (22:53):

Compliment, Charlie, I want to highlight one more really important point of why we felt we needed to get started. It was how hard it was to find anything that looked practical, right? There was conversations, there were pilots, people were doing stuff that was super exciting and you're like, "But yeah, but where's the signal, man?" And this was back when the first signal through the noise conference was 2024? 23.

(23:20):

23. And it was so hard. And then the next one, it was still hard. I mean, it wasn't as hard. We started pulling in a few people, but it was that difficulty of finding actual applications that really told us we needed to started bringing them to the surface because how else do we learn? If we only look at the failures and if we only talk about things that could happen in the future and are really exciting, it's fantasy land, right? It's like political utopias. This would happen if only human beings were different. It's like, no, what's happening now, right? Somebody's doing something good. And we were so lucky to be able to bring together, thanks to Steve's network and this lesser per my network, folks who are deep in the field and companies that are actually got traction on some problem. And I just think it's really fun to get those stories out there.

(24:11):

So I think that I've gotten really passionate about it. For me, it was a cool idea. It was Steve's baby. He's like, "Let's do a podcast." I'm like, "I love it. Let's do it. " And so it was like a crazy idea. Neither of us have knew ... I mean, Steve actually did a podcast before, but he had at least something under his belt, but it sounded like such a useful, interesting thing to do. And I guess as a middle-aged white guy, you're kind of required to have a podcast at a certain point. You're like, "I'm 56. I hear you get a podcast, right? So I enjoy." But I just thought that getting those stories and getting some of the theoretical background on it has been really cool. But back to Charlie, I mean, you're a great example, right? You're somebody who's an industry leader, leads a really important company in the field and has developed a really fascinating approach to data quality.

(25:08):

I mean, when we heard about the PIQI framework, and I'm a litle embarrassed to say I wasn't following the work, right? You show up on a podcast and I'm like, "This is amazing. This is the work that people should hear about. " And we need to, because it's a cool approach that it's very, very timely and it's certainly a foundation. The fact that you guys are doing this open source, I think is really atractive. So let's talk about that. That's exactly the kind of thing we want to elevate.

(25:32):

So thank you for being a guest.

Dr. Steve Labkoff (25:36):

The nature of what you're doing with that, I mean, I assume that most of the folks who listen to your podcast know about what you've done, but in case they don't, I mean, you're building this framework of looking at data quality where data quality is actually probably, if it's not the single most important aspect of building models, it's probably in the top three. As we move forward in this space, especially in healthcare where quality of decision making matters,

(26:04):

The decision quality is going to be predicated on what's the quality of the information that's used to build the models. And the work that you've done with the PIQI framework is it's like a Robinhood arrow hitting an arrow. It's like so bullseye, it's just incredible. And it's funny because I know the clinical architecture is a well oiled machine and a going concern. I look at the PIQI architecture and think, I don't know why that's not your leader here, why that's not the thing that is actually putting you guys front row, center, and everybody should be adopting this.

Dr. Leon Rozenblit (26:38):

More podcasts. More podcast.

Charlie Harp (26:43):

I think that is, it was one of those things where I think I'm going to look back on this crazy idea, which by the way, PIQI was started by a presentation I gave at AMIA on data quality and taxonomy and understanding data quality and it just kind of sparked something in me. And then I got to the point where I'm like, this could be a product, but if it's a product, it's going to take forever to get traction.

(27:06):

And the other thing too is you see something like AI looming on the horizon and the fact that people ... I just gave a talk at the Amplify conference where it was all about why the data is good enough argument is no longer good enough with where we are in healthcare and that we're at this watershed moment where we have a very short runway to correct some of these things before the option to fix it gets taken away from us by the proliferation of bad data generated by AI and why this is such a critical moment. But yeah, I appreciate the kind words and I think that there's been a lot of amazing feedback from the HL7 community from the federal community around what we've been doing with PIQI. And so it is kind of something that's become very important to Clinical Architecture and what we're doing going forward.

(28:03):

It's probably the most significant thing we can do over the next five to 10 years. But going back to the podcast, one of the things I love about your podcast is it's in the name, it's practical AI because you're right, there is so much out there where, oh, this is going to replace doctors and, oh, there's not going to be no need for terminologies in the future. And somebody came to me, a guy who runs another software company in healthcare, I won't name names, but he's like, "Charlie, is it exciting? Terminologies are going away." And I said, "Dude, we still use fax machines." I don't think terminologies are going anywhere anytime soon. I think that we do need to dial up the resolution of what we're doing and I think there's so many ways that AI can help us with that. We just got to make sure that we don't screw it

Dr. Leon Rozenblit (28:57):

Up.This is where it helps to have cranky old guys who've been around the block a couple of times. Honestly, when I was younger, I'm still pretty excitable and I love news things. You show me a cool new approach, like PIQI. I'm like, "This is cool." I mean, I can really appreciate it, but you also develop a nose for bullshit and you develop a sense of like, okay, we've seen this before. That's not how things change. Change itself is a process and it's more gradual than people think and there's a lot more inertia in existing systems. In a way, this comes back to Chesterson's gate principle where what Chesterton wrote to his son is that look, if you're walking down the country road and you see a closed gate, don't open it and unlock it until you know why somebody put it there and closed it.

(29:52):

It's there for a reason. Somebody had a reason. Don't assume that you know more than the people. So a lot of the infrastructure that we're annoyed by is there to serve a purpose and we really need to understand it well before we say off with their heads, let's get rid of ontologies. They're too difficult and too boring.

Charlie Harp (30:10):

Yeah, no, that's absolutely true. So you've talked to a lot of folks in the industry in the time that you've spent doing the podcast. What are some trends and themes and what are some of the things that got you the most excited about some of the folks you've had on the podcast?

Dr. Steve Labkoff (30:30):

So when we first started, one of the early ahas that we came out with and it came out in talking to multiple folks in multiple different venues was how important mundane use cases were. This was something that came out at the Signal Through Noise Conference and we weren't really clear going in that that was really an important concept, but all of a sudden we realized, oh my God, amongst all this hype, it's the really boring, really mundane use cases that were actually moving mountains and making hard stuff easier, faster and more efficient. And that's where we started. And if you listen to our podcast, every five or six episodes or seven episodes, we do what's called a reflections episode where Leon and I try to synthesize what we're seeing and what we're trying to suss out, what does it all mean in the context of what we've heard?

(31:28):

So the first block of podcasts really drove us into the domain of that mundane stuff matters probably a whole lot more than people were giving a credit for. We went from there into different types of toolings that are out there. There's a foundational piece of the tooling. The tooling has to be done properly and we ran into folks that were doing the tooling. We then ran into another piece, which was that literacy around AI actually matters. That's another lesson that came from the second signal through the noise conference where the more we started digging, the more we kept coming up with there might be a tiny fraction of folks who really get it deeply and the masses don't. And if you're going to put these models into clinical practices, an example, if you're not literate with the tools, it's like giving an intern a colonoscope and saying, "Here, go do a colonoscope on this guy.

(32:21):

You're doing an intern picking up a colonoscope and doing a colonoscopy. You want somebody who's been trained and knows how to use the instrument properly and won't cause damage and won't hurt you. And AI is very much in the same vein. Leon, we have a couple more things.

Dr. Leon Rozenblit (32:37):

Yeah. I'll mention two classes of insights. I'd say one thing that was really exciting to see where people were succeeding was it was building the right frameworks and workflows around the model. There was so much interesting innovation about how to put them into production in the right way and to create supporting data sets, for example, to create the right interaction models with the staff. And that's what people were succeeding on. It wasn't necessarily like, oh my gosh, this is a better model. There was some interesting cases of cool pre-training and there was some phone specializations, but there was so much room for improvement just there, right? Just like in creating that right institutional context and creating the right technical context. The other one that I think about a lot now is that depth is going to remain the competitive mode.One issue with all of these companies is how do you build a company when the intelligence is a commodity?

(33:40):

What's the defensible IP? I can build whatever, a podcast creating automation. If it's all built on cloud code, what's my defensible IP? Anybody can recreate it and just look at it, "Oh, that's pretty clever. Let's do that in our own way." Where I'm seeing that competitive mode is in really deep understanding of workflows, relationships and really domain specific concepts. And the people who are digging deep and really understanding the customer's need, my sense is they're the ones who are going to survive even as the models improve non-linearly. But I don't know. I'm curious what you guys think, but that's sort of my impression right now. I may see 17 things next week that will change my mind, but that's my sense at the moment.

Dr. Steve Labkoff (34:30):

To amplify what you're saying, people have been sort of positing, oh, AI get rid of doctors or AI will get rid of nurses. I actually think that's exactly wrong. I don't think it's going to happen. What I think's going to happen is that the role of the physician's going to change as the role of an engineer on a train may have changed into something else when trains went to planes, but the role of a physician's going to change in a way where they have to be able to navigate the AI, understand what the AI is doing, be able to interpret the results in a way that is not just keeping them relevant, but keeping them in a position where they're able to, at least for the moment, they raise their own capability level to become more synthesis driven as opposed to mechanic. So an AI model is probably never going to run a colonoscope or it's never going to run a cardiac catheterization lab.

(35:28):

But if you're really talented at what you're doing with AI, you can automate tools or automate processes that will support those things and do it in a much cleaner, faster and intelligent way. And as the clinician, you're at the top of that food chain. Your deep expertise is going to help you to drive those successes in a much more, I don't know what the right wording is here, but in a more fascial way. The wins are going to be from those who are really deep and understand that exactly what Leon just said.

Dr. Leon Rozenblit (36:00):

So Charlie, do we have time for one more reflection? Can I bounce off of Steve? Yeah, Steve, I agree with the overall concept and thank you for enriching what I try to say. I want to refine that vision a little bit. I mean, I think there's a tendency today when we talk about all of us, when we talk about AI, we think of LLMs just because they're so visible. But I mean, there's enormous innovation happening in non LLM AI and Nvidia just came out with a new Cosmos three model, which is super, super interesting. And that movement to world models is going to represent physical intelligence and action-based intelligence in a much richer way. It's a litle bit hard to predict, but if those research efforts pan out, you're very likely to see an AI model that can control a surgical robot better than a human.

(37:02):

I think even in that, or the question is, in that world, what is the role of a human? I still would want a human overseeing it. We want human judgment who decide what do we trust the robot? Is the robot good enough? Does this case fit the robot's parameters? But I would probably draw that line a litle bit differently because I'm trying to think hard about where do the non- LLM models fit in and how are they going to change the space of the possible?

Dr. Steve Labkoff (37:31):

Again, to add another layer to this is in one of the podcasts, the reflections Leah and I were talking through the role of people may be changing and the use cases that are there today and this was something Leon brought up in one of, I don't remember which reflection it was, the very first set of use cases when new technology hits the scene are like automating or changing what you already know how to do. But the real innovation comes when all of a sudden somebody has this massive insight and all of a sudden something like Uber happens where all the pieces were there, taxis have been around for hundreds of years perhaps and then somebody says, "Hmm, there's a bunch of tech that we could throw together in a way that no one has thought of before and we can crowdsource the demand and we could funnel that demand away from the medallions right to the people who want to do this and all of a sudden you've got in a completely new industry." And it's those innovations which I don't think have happened yet.

(38:28):

I think they are yet to come and that'll be in healthcare, it's going to be all over the place, but there will be different ways of approaching medical problem solving that people have not yet figured out and that's what's on the horizon and that's what's the next, let's call it five to 10 years is what's going to happen. Those novel approaches that we haven't seen or understood yet will start to manifest from the young people who aren't as old and as grouchy as we are.

Charlie Harp (38:54):

I think one of the things that's interesting and you both touched on it is the mundane and the things that technology is really good at. Like for example, we'd be in a different world if providers didn't have to document like they do today. And I think AI can absolutely help with that. I think when I was at Zinx, one of the things that Scott Weingarten said at a conference is that there's 30,000 medical journal articles that are published every month in the United States alone. And if a newly minted med school graduate spends three hours a night their first year after they finish reading medical journal articles at the end of one year, they're only 12 years behind. That's the kind of thing where if we can use technology to bring cutting edge information to the point of care, if we can use technology and have a feedback mechanism where a human is still a human with a conscience and practical experience can ride shotgun on that process, I think technology can help, but that's kind of where I go back to the technology is relying on the data that we give it and if the data we give it is bad, it's going to end up ... You guys seen the movie Idiocracy, I imagine.

Dr. Steve Labkoff (40:14):

Oh yeah. That

Charlie Harp (40:15):

Hospital scene and idiocracy, I'm so terrified that that's where we're going to go if we let AI run down the wrong path. But yeah, I think that it's scary, but it's also incredibly exciting what's happening right now.

Dr. Steve Labkoff (40:31):

You know what you're describing in some ways? I saw an article the other day about the Voyager probe. So the Voyager Probe has been in space for like 50 years literally and it's still chugging away heading out of the solar system and it's like 12 light minutes away from the sun and Alpha Centauri is like four light years away from the sun. So the question is who's going to come up with the medical equivalent of warp drive to cut that distance down into something that we can all manage. I don't know what that looks like yet, but I think it's going to happen sooner rather than later and someone's going to come up with a medical equivalent of warp drive that's going to like all the interpretation of DNA as an example and how you rate how DNA, the genomes and the transcriptome and the metabolome and all the, how they all interrelate in a way that no one has come up with yet.

(41:27):

And what's the expression from the X files? The truth is out there we just need to figure out how to see it. Sorry for all the science fiction references, but it's going

Dr. Leon Rozenblit (41:38):

To- Yeah, it just made me realize like I'm less optimistic than Steve in some ways. I would settle for like-

Charlie Harp (41:45):

Now Leon, come on. Yeah.

Dr. Leon Rozenblit (41:46):

I'm like, I'm going to

Charlie Harp (41:47):

Depress it.

Dr. Leon Rozenblit (41:49):

So yeah, I would settle for just like a nuclear rocket. It's like we're using like old fashioned chemical rockets in healthcare and I'm like, there's a lot of room for improvement before you have to violate fundamental laws of physics. So let's just tackle a few of those. And I know what you're saying, Steve. I mean, there are rooms for transformative innovations that none of us seeing. I totally agree with that. But Charlie, what your question made me think about how do we condense the incredible production of knowledge into something that's useful to the individual? There's opportunities that automation creates that simply weren't possible like even three months ago. So let me give you like a ridiculous use case that I just tried. I can now take Claude code. They just released this new functionality called Workflows. I can take CloudCode and apply the functionality to every conversation I've ever had with it.

(42:46):

And I've got a year's worth of actual sessions where I'm trying to struggling to get to do stuff. And I could say, analyze every single thing we did and tell me which of the things I did are like a bad idea and tell me which things I could do better given this new model. And also a lot of the stuff I did work fine when you were like at version four, but you're at 4.8. What am I missing? It can actually do that, right? You could actually say, "Well, now that you're version 4.8 and your thinking was still a version 4.6, you can stop doing this stupid thing. It's a lot easier to just do it this way."That's very focused learning because otherwise I have to read the documentation and like yes, in theory I can catch it if I read all the docs, but I don't have the time.

(43:30):

They're releasing too many versions and I'm too slow. So that's an interesting example. Imagine that with a medical student, the question we ask not give me new stuff that's new is given everything you know about what I've learned and what I've done, like here's every procedure I've ever done, here's every article I've ever read because why don't ... Of course you know this, right? If you're my personal AI and you're watching this, tell me what I'm missing. Here's a new case. Tell me what I'm missing. What else do I need to know in order to be a best doctor possible for this case? That is possible today. That's something we could do right now that we could not do three months ago.

Dr. Steve Labkoff (44:06):

Yeah. I did a similar, probably a less insightful version of that. I recently went to, I have a year plus history with ChatGPT and I said, "Tell me some things about what you know about me that I don't know about myself What are the kinds of things that you see given my conversations with you, given my questions that I've asked you, what do you see about me that I don't see myself?" And God, I could probably have saved a fortune with psychiatrists over my lifetime because all joking aside, it identified characteristics of how I think and how I synthesize in ways that literally I hadn't really realized and it brought to mind a different ... And with that information, I can start potentially processing my own thinking differently. So that's an interesting little thought exercise everybody ought to give a try to and see what it does.

(45:06):

If you're like Leon and me and have been working with these models like hours in a day quite literally for the past year, two years, three years, you're going to find that it has ... And because the models have improved lately, it will find and synthesize the answers to those questions in ways that you really ... And sometimes it could be a little scary, I think. But getting back to that medical student, what would make that medical student advance in their career or advance in their thinking or help them to connect the dots in ways that they at that point haven't been able to do?

Charlie Harp (45:37):

Well, and the thing that concerns me is we human beings tend to have this thing we do with technology where once we get comfortable with it, we kind of go on this cognitive autopilot with it. And my biggest concern about AI, and I see this in some kids, some relatives where ... And we already live in a world where people just believe whatever shows up on their social media feed or whatever and it kind of makes their life choices for them. It's take the medical student, for example, it's one thing to kind of synthesize somebody and say, "Hey, watch out. This thing has been uncovered. Here's a better treatment modality. Here's what you could do. Here's what I've observed about their genome and what we know about whether they're a fast metabolizer or pharmacogenomics, but at what point do we run the risk of where we're no longer processing in our own heads anymore, we're just asking questions and giving ourselves over to what the machine tells us.

(46:47):

And what does that mean in the long term for us as a species

Dr. Steve Labkoff (46:52):

That ... So now you're entering the philosophical domain mentioned my other podcast because the other podcast that I did that I still do before I got into doing practical healthcare was this podcast with my rabbi and that one's called Judaism of 21st Century. And the premise of that is, and it's a much tinier audience. It's a very niche kind of thing, but the premise of it is how do you look at the world we have today and are there ... I'm asuming that the lessons that we are doing today, the things we're doing in our world, we're not necessarily the smartest version of who we are. And my presumption is that there may well have been smarter people before us and the folks who wrote the Bible, the Toro, the Talmud especially, may have had insights. They may not have a technology, but they may have had insights that may apply today that we haven't seen.

(47:45):

So the premise of that podcast is how do you connect those lessons to today? Now I bring that up in this context because the question of bioethics and how we are turning over ... To your point, Charlie, about are we going to just turn things over or are we going to become better versions of ourselves? Most human beings are pretty lazy. The human brain is actually very, very lazy. It wants to do the least amount of work that it can and get by. My concern like yours actually is that we will stop thinking. We will stop thinking critically. And this is where the AI literacy thing really comes into sharp focus and contrast. Those that are able to take the initiative and really up their game in terms of thinking and doing it critically are going to be the ones that are going to actually excel.

(48:35):

People are afraid they're going to lose their jobs. Well, if you can do that and use your brain differently, you're going to excel with this new world. If you can't, you're going to become one of the masses that just listens to whatever ChatGPT tells you or what Claude tells you and you're just going to go on like a lemming.

Dr. Leon Rozenblit (48:51):

So can I add one a different perspective? I mean, I like the philosophical angle, but I actually want to share that I think you can also see it as an engineering problem.

(49:05):

So yes, automation bias is a real thing, right? If you have to hit a button and every time you hit a button, you get a good result, you can just keep hitting that button. This is all perfectly true. We see it. Cognitive science tells us that's the way human brains work, but I think part of our job as technologists and as designers is to think of countervailing measures. If we want humans to think, build systems that encourage thinking. If you want humans to only be able to pick out trustworthy information and high quality information, build systems that send high signals about what high quality means, like the PIQI framework, for example, in healthcare, right? I think I resist a little bit treating it as a philosophical problem. There's certainly a philosophical dimension to it when we start talking about questions like what does this mean for human beings and what does it mean for us spiritually, but I want to solve it and I want to encourage other people to think about as an engineering problems because they're solvable.

Charlie Harp (50:05):

Which means we have to be practical, which takes us back to your podcast. And I think we can probably wrap up today, but I think I have to have you guys on again. I think- We'd love to. ... a moral imperative, but

Dr. Steve Labkoff (50:20):

We spend a lot of time on our podcast trying to be really focused on our guests and trying to ask intelligent questions, but we don't get a chance to riff like this as often as I'd like. And I find this one refreshing because we're able to really do a little on the fly synthesis of our thinking here and it's nice. It's nice to be able to do that. And I really enjoy having you ask these questions and letting us kind of go on here and it's been a nice experience. So thank you for that. Oh, it's

Charlie Harp (50:45):

My pleasure.

Dr. Leon Rozenblit (50:45):

I think Steve is saying, Charlie, you're a fantastic host and we're so glad you were able to have us on. Thank you

Charlie Harp (50:50):

Wow, please say more. No,

Dr. Leon Rozenblit (50:54):

I- What's the Mark Twain line? Compliments make me uncomfortable. It doesn't matter how much I want to said that there's what you left with the feeling they haven't quite said enough.

Charlie Harp (51:03):

It's like in the old Bugs Bunny cartoons, tell me more about my eyes. All right, gentlemen, so for the audience, tell them where they can find the podcast.

Dr. Steve Labkoff (51:14):

So we're available on both Apple and Spotify and lately we just opened up a new channel on YouTube and you can find us at our website. It's www.practicalaiinhealthcare, no spaces or dots, just practicalaianhealthcare.com. And Spotify and Apple are the core places to find it. Our website has a library of every podcast we've ever published. At this point, we're approaching episode number 40. We try to do on a week and we do take some time off for good behavior around 4th of July and Christmas. So we will stop. We're going to have a break in a few weeks. I think Leon's going to be shouting for joy for that break because we've been a little overloaded lately, but yeah, that's where to find us. And if you want to email us, steve@practicalaiinhealthcare and Leon@practicalaiinhealthcare. And we're happy to get recommendations of guests and recommendations for themes and anything else you want to chat with us about.

Charlie Harp (52:09):

And for the listeners, I have a face for radio, so my podcasts are audio, but these guys also, they do it on video as well.

Dr. Steve Labkoff (52:18):

Which is brand new. We just opened that a month ago. That's

Charlie Harp (52:20):

Right. So check it out. I have a feeling we're

Dr. Leon Rozenblit (52:21):

Going to lose a lot of people when they see what they look like.

Charlie Harp (52:24):

You just got to put down the filters. You just got to do the full-

Dr. Leon Rozenblit (52:26):

Yeah, I think that's right.

Charlie Harp (52:28):

All right, gentlemen, it's been a pleasure. Thank you so much for being on today.

Dr. Steve Labkoff (52:32):

Thank you, Charlie. Thank you so much.

Dr. Leon Rozenblit (52:34):

Thanks, Charlie. It was great.

Charlie Harp (52:35):

Great. And to the listeners, I'm Charlie Harp, and this has been another exciting episode of the Informonster Podcast. Thanks for tuning in. I'll talk to you next time.