Abridge’s Dr. Michal Nedoszytko on Physicians Who Build Software

The trip to San Francisco had nothing to do with hackathon. It was just like, you know, to learn. But then, like a few days before, I see this ad on Twitter, Anthropic organizes a hackathon. And I was using Claude Code for the past year, and I was absolutely obsessed about how it works. And funny thing, the finale actually coincided with the time that, that I will be in SF.

So, I said like, "Hell, why not? Never participated in such a thing. Let, let me try." So, the post-visit AI was a thing that I already had in the back of my head, so I coded it over the course of seven days, a little bit also between the patients, between continents. And it was really great because the Wi-Fi on the plane was patchy, and they just shipped this, you know, a team of agents.

I was able, like, to detach part of the coding tasks inside the Cloud while literally being inside the Cloud. And when I landed, you know, uh, the, it was, uh, it was really, really almost finished. I just had a massive jet lag, so I spent the following two nights polishing it, and earned me third place out of 13,000 applicants for this hackathon.

And yeah, I was there in, in SF, uh, in Shack 15 during the finale. I met all those guys from Anthropic and yeah, my story changed after this. Hi, and welcome to another episode of NEJM AI Grand Rounds. I'm your host, Raj Manrai. I'm here with my good friend and co-host, Andy Beam. And today, we are really excited to bring you our conversation with Dr.

Michał Nedoszytko. Uh, Michał is a clinician. He's a cardiologist, interventional cardiologist who has spent a lot of time in the cath lab, and he's also a coder and an engineer who builds and has been building software for many years. We had this absolutely fascinating conversation where he took us through his early days in building software, and getting the attention of the Polish Ministry of Health and having a lot of doctors in Poland adopt his early software. To then, uh, winning and coming in, I think, one of the leading positions in a big hackathon that was organized by Anthropic earlier this year, and to now his work as a clinician scientist at Abridge.

And Andy, I, I think this was just, it was absolutely fascinating to pick his brain about his skills as both an engineer and also as a practicing cardiologist, and hit on a recurring theme of the podcast that often the folks who are having the most impact in medical AI have really cultivated not one skill or the other, but really invested in both technical skills and also their deep domain clinical expertise.

Yeah. Raj, I mean, one of the best things about being a co-host of this podcast is, like, all of the awesome people that we get to talk to. And, like, Michał is just incredible. As an early, uh, doctor, as a doctor- in-training, I think actually he made, like, two or three separate EHRs during his career.

It's like every time he landed in some medical environment, there was not an EHR, and pre-vibe-coding days would just spin up an EHR to make his job easier. And he just has this sense of how to propose a technical solution to a health care problem. And that came through over and over again in this conversation, I think all culminating with this, like, incredible story about him vibe coding an app on the way to San Francisco and ended up winning this health care Anthropic Hackathon.

I think there were something like 13,000 entrants in this specific hackathon, and he was one of the top three. And again, just a, he's also, like, very articulate and very insightful in how he understands and analyzes not only, um, technical problems, but regulatory environments, the differences between Europe and the U.S.

And again, just a really wide-ranging conversation and lots of quotable lines in this one. I think there's gonna be lots of things that people can quote, um, in this episode. So, uh, I think one of the most fascinating guests we've had on the podcast, so really, really enjoyed the conversation. The NEJM AI Grand Rounds podcast is brought to you by Microsoft, Viz.ai, Lyric, and Elevance Health.

We thank them for their support. And with that, we bring you our conversation with Michał Nedoszytko on AI Grand Rounds. Well, Michał, welcome to the podcast. We're excited to have you on today. Well, thank you very much for having me. Great to be joining you. And thank you for joining us on the podcast.

Um, so this is a question that we always get started with. Could you tell us about the training procedure for your own neural network? How did you get interested in artificial intelligence? And what data and experiences led you to where you are today? Well, I think it, uh, it leads back to when I was 14 or 15 years old, when I was absolutely hooked on computers, and I was already creating websites and coding back then.

So, I believe that my neural network was just, like, repetitively trained over the past 20-ish years on coding, and, uh, the greatest source of inspiration and frustration was medicine. Just take us through some of those early coding projects. What were you interested in? What did you naturally gravitate towards?

When I was still back at high school, I was basically coding, uh, websites. I was, uh, once almost thrown out of the high school because I hacked the mainframe, but I did it so badly that they caught me. And then, uh, when I decided I'm going to go to med school, back in Poland, we had those, uh, preliminary exams and, uh, I was actually creating virtual exams.

It was the early days of Internet back then. So, I started, uh, preparing for those by actually creating the system for doing those tests. I also shared that online, and it was like this first virtual nationwide exam. During my med school, there were multiple projects. For example, uh, I was involved in Student Scientific Society of Pediatrics, so I created this first social network where, um, the, uh, students could, like, share their profiles.

It was great times. We were two guys and 70 great, uh, girls. Uh, so I created this network where we could, uh, exchange broad ideas. And then when I joined cardiology, I started creating those administrative tools, so I absolutely hated writing lab demands, uh, papers, referrals and so on. So, first it started out as a PDF generator, and then it grew into the full-on EHR, uh, where, uh, all of the members of my department were competing against each other by a sort of gamification.

So, uh, if you wanted to encourage them to actually make their documents better in quality, they gained, like, experience points, and they leveled up and so on. So, it was always a little bit part fun, part trying to make my life easier. But you know, an IT guy is somebody that spends days to save, uh, two minutes from a task that usually takes seven, and then you go down to five.

So, I think, I believe I was a bit like this. Then when I went into my residency, I started doing internal interventional cardiology. So, I created full-on EHR for interventional cardiology, operative notes, stock, some, uh, coding, and it grew into a full-on startup, which I led for about three years. Well, it's not easy to have a startup along, uh, training in, in cardiology.

After three years I needed to renounce it. But, uh, in 2011, uh, there was a quite interesting story. So, uh, a Polish Ministry of Health has introduced this special law where the doctors needed to write on the prescriptions what's the level of reimbursement of the drug, and the funny thing is the only thing that they gave us was Excel files with about 13,000 medications.

So, the Polish doctors did what they deemed, uh, well, the most suitable, so they went on strike, and they decided, "We won't be writing those prescriptions." I was annoyed myself as well, but I thought it's quite easy actually to have a solution. I just imported this Excel file inside the database, and then I created this website where you have, like, this face of Minister of Health.

His name was Bartosz, and there was, like, this comic bubble, and you ask the Minister how much is this drug reimbursed. It start—. I'm sure he loved that. I'm sure he loved that. Oh. He, he actually did, but you know— Yeah. —it started out as a joke. Uh, because, you know, when he didn't know the answer, he was, like, looking into the void.

There were just some funny pictures of him. Um, and you knew in his silence that you won't actually get the answer. So, I shared the link on Facebook because it was supposed to be, uh, for, for my buddies from the department, and, like, two hours later the website crashed. And I'm looking over and it's, like, just, you know, the requests were all over the place, and it turned out that half of Polish doctors started using it.

And then, like, TV came. Uh, it, it went quite popular, and four days later I got a call from Ministry of Health, and it's like, uh, "Sir, the Minister would like to speak with you." And I'm like, "Oh my God, I'm going to jail for lese majesty, and my cardiologist career is over even before it started." So, you know, I go to the Ministry of Health in Warsaw, all dressed up and so on.

The Minister opens the door, and he, like, pulls up a pack of smokes. The Minister of Health warns smoking is deadly, and it's like, "Do you smoke, sir?" And it's like, "No, I, I, I'm a cardiologist. I cannot." "Because you kind of solved the problem. Uh, they, they stopped protesting and they are writing the prescriptions.

Would you like a job?" So, I was—. Amazing. Amazing. Yeah. So, I was working, uh, for about two years at the Ministry and, and this IT section. I tried to change, uh, a lot of things, but, uh, it turned out more political, as you might imagine. But it's just, like, to give an idea that sometimes even though you code, sophisticated solutions for a very long time.

Small things that just takes one afternoon and comes out of, you know, idea that comes into your head can actually change a lot, and this website is still used. It's like 15 years now, and I still have, like, 100,000 doctors using it every month, and it's, it's a funny story. I, I think another— So, yeah.

—another message for me for that takeaway is that not only there's a non-obvious and, of course, not always equal, relationship between effort in and sort of utility out, but I think also you're sort of cutting through the obscurity and complexity and making something simple for people. And I think that's, it's a pretty generally attractive thing, too, something simple and useful for folks to use who are very busy and practicing and otherwise having to navigate some overly complicated Excel sheet that just has enough activation energy that they won't do it to begin with, or they'll struggle to do it to actually get what they need to get done.

Yeah, I see. Do you think that's fair? I think that when you live the problem, when you're deeply in it, then this is what motivates you to, to find solutions. For me, it was always the administrative burden. It was always the inefficiencies of, of health care system, whether it was in Poland, whether it was in Belgium, where, where I used to live for the past nine years.

And I think that the doctors are the ones that truly understand the reason why we have this problem, and they also understand the solution. And I would actually like to argue that my 20 years of medical coding, of my programming career, is right now a little bit less relevant than ever. I think that we live in times which, uh, allow the physicians to prototype something so much faster and get those solution out much, much easier.

Yeah. No, I, I totally agree, and I think maybe that's actually a good transition to one of the topics that we wanna dig into. So, I think beyond getting half of Polish doctors to start using your platform and working for the Ministry of Health and all this exciting. By the way, all of this is happening while you're either in training or practicing, right?

Yeah. You're, you're doing your, your day job at the cath lab as a cardiologist while doing all of this as well. Is that right? Yeah, absolutely. Yeah. Yeah. I mean, always. At night or between the patients. So, after this period, I think you've been working a lot and you've been thinking a lot about developing software and developing programs that are useful for helping you, of course, leveraging your experience and your work as a physician, help your patients, and that's before the visit, that's after the visit, that's outside of just the encounter itself.

And so, uh, you've built Previsit and Postvisit. Can you take us through what those are? And of course, you know, how we, we connected was we saw that one of your tools had done very well at a big hackathon that happened, I think, earlier this year. It's, it feels like a, a very long time ago, but it wasn't that long ago.

It was just earlier this year itself. So, maybe just take us through, first, what Previsit and Postvisit are, and then tell us a little bit about that, that hackathon experience as well. Yeah. So, in order to, uh, to tell the story, I need to go back to when I moved to Belgium. So, about nine years ago, I moved from Poland to Belgium.

I was Head of Department in the French-speaking part. It was in the city of Mons. And I'm an interventional cardiologist in Poland. I spent much more time in the cath lab, uh, itself, but in Belgium it was, uh, a little bit inverse. So, I had a lot of administrative tasks, and I also had two private offices, and I, I needed to spend a lot of time at the outpatient clinic.

And that was the time when the large language models, uh, started to emerge, and it was a ChatGPT moment. Before that, I was training neural networks in coronarography. I was toying a little bit with, uh, image recognition, but then I quickly realized the potential of LLMs. And one annoying element of a cardiology visit was repetitive questions and lack of, uh, certain quality data during the visit.

So, whenever you ask the patient, "What medications are you taking?" I would bet that more than half of the patients, they don't remember the names. They say like, "Hey, this drug for the heart, you know, the red pill." And you wanna just write down, is it, you know, Atorvastatin forty or it's a Bisoprolol five.

"My wife knows," and so on. So, I thought, like, why won't I use LLMs to do the pre-anamnesis with the patients? So, uh, my patients received a text message before the visit. They talked with the LLM that asked some pertinent questions, and then I received a summary. The goal was not to replace the anamnesis, but actually to carry the information from the house or, "My wife knows," so from the wife's neural network into the electronic health record.

And then when the patient arrived, he was already prepared. He already knew what I will be asking him, and the conversation had a completely different quality. So, it was not only about saving time, but also about improving the overall quality of the visit itself. So, it grew. The system was integrated in six countries.

We did over a hundred fifty thousand Previsits, and I was also, like, looking for a way, maybe it's time to launch my new startup. I'm not in training anymore, so, uh, maybe, maybe I could do something. So, that was the reason I actually came to San Francisco because I got connected with somebody who was creating a health care startup.

So, I had my trip planned to SF to understand what's happening in here, why this place is so special when it comes to, to building, when it comes to innovation, and so on. And I also had two other projects in the back of my head that I didn't dare to do in Europe because of regulation. Uh, I was involved in, in a scribe, uh, which was done by another company, and I also wanted to make, uh, an LLM that helps the patient after the visit.

So, the trip to San Francisco had nothing to do with hackathon. It was just like, you know, to learn. But then like a few days before, I see this ad on Twitter, Anthropic organizes a hackathon, and I was using Claude code for the past year, and I was absolutely obsessed about how it works. And funny thing, the finale actually coincided with the time that, that I will be in SF.

So, I said like, "Hell, why not? Never participated in such a thing. Let, let me try." So, the Postvisit AI was a thing that I already had in the back of my head. So, I coded it over the course of seven days, a little bit also between the patients, between conferences. I heard part of this was on the plane ride, too.

Yeah. Yeah. Yeah, it was crazy. Is, is that true? I mean, uh, yeah, yeah, absolutely. So, you know, I was so locked in during that time that, um—. Locked in. I love it. Yeah. Yeah, totally. I mean, uh, and, and it was really great because the Wi-Fi on the plane was patchy, and they just shipped this, you know, uh, team of agents.

I was able, like, to detach part of the coding tasks inside the Cloud while literally being inside the Cloud. The Cloud, yeah. And when I landed, you know, uh, the, it was, uh, it was really, really almost finished. I just had a massive jet lag, so I spent the following two nights polishing it. So yeah, and that's, uh, earned me third place out of 13,000 applicants for this hackathon.

And yeah, I was there in, in SF, uh, in Shack 15 during the finale. I met all those guys from Anthropic and yeah, my story changed after this because the hackathon precipitated…. Well, there were many big companies that heard about The Coding Doctor because the video that I made for the judges. We were supposed to make a pitch video for the judges.

It was supposed to be three minutes and, you know, describe the topic, and I just wanted to tell them that I'm gonna be in SF, so maybe we're gonna meet. So, I created this introduction where I'm coding. I heard about the hackathon. I flew to SF, and then I described the, uh, the project, and the video leaked somehow, and it went viral all over the Internet.

I think it was over 3.5 million views on Twitter. So, I got into some international media, um, China, Turkey, uh, and, and Korea, and then, well, in SF, it, it made some waves. So, big companies, they wanted to make babies with me. And, uh, they called out, like, "Hey, uh, would you like to, uh, to, to work with us?"

And, uh, I was like, "Oh, no, I'm, I'm on call next week in Belgium. I have to come back." And they were like, "Maybe you shouldn't." So, um, in all this noise, and it was really crazy time because, like, my LinkedIn literally exploded. I needed to vibe code like a parser to understand who's, who's texting me.

There was a message from Shiv, Shiv Rao, who's also a cardiologist. He's founder of Abridge. I also never heard about Abridge before. And then we meet, and like 15 minutes into the conversation, we're like, "Man, we do the same thing. Uh, our mission is aligned. We need to work together." And you know, in 15 minutes we made that happen.

The, the vibes were good. Yeah. Incredible. I mean, literally, what Abridge was doing for the past eight years and the mission was perfectly aligned with what I was doing in Europe. It just, I was doing it on a much smaller scale and mostly solo, while in here it's a successful company that's basically oriented towards the needs of physicians, and I had absolutely the same goal.

So, when I had the opportunity to join them and, and do what I do just on a much grander scale, safer, better, and having access to, like, millions of patients and physicians, then I said, "Yeah, why not hang my, uh, doctor scrubs on the wall for a while?" Although honestly, I will just have to do my USMLE because my papers need to be recognized in here.

So, uh, it's another challenge that I will have. Awesome. Uh, I, I wanna dig into what life is like at Abridge now. Um, but before we do that, you said something interesting that I wanted to follow up on. You made a passing comment like, "I could never do this in Europe," but I was interested to see why SF was so builder-friendly, especially for health tech.

Could you elaborate on that a little bit? Because your solution doesn't use any medical data. It's all patient-submitted data, if I understand correctly. What's the difference in those ecosystems and, like, why…? Yeah. I, I have some, some high level, but it'd be great to elaborate on that for the listeners.

Yeah. So, I don't wanna divulge very much into politics, but the regulation in Europe and how it impacts the startup scene and innovation is somewhat different than in the U.S. So, long story short, Previsit is a system that's supposed to be an administrative tool. So yes, it does contain some medical data, but what matters whether these are identifiable data, so whether you can actually identify the patient.

So, I created Previsit in that way that possibility to identify the patient was completely separated. We didn't store the IDs. It was integrated with the EHRs. But Postvisit, for example, is a completely different type of tool which has full-on autonomy. It could suggest, uh, to the patient some decisions that he should undertake, and that was also the time when ChatGPT for Health, uh, launched with exactly the same functionality.

But in Europe, you have AI Act, you have GDPR, and a ton of other rules that unfortunately classify the health care-related system at the highest level of risk and also scrutiny, which implies that you have a lot of legislative work, that when you found a startup, you need to buy a lot of lawyer hours, and also you're very much thwarted when it comes to creativity.

To give you an example, Previsit was the first system of its kind that was launched in a hospital in Poland and in Belgium as well. So, when, uh, we went to the media in Belgium that we're, like, the first one with innovative AI, uh, three days later, we got the GDPR inspector asking some random questions, and the hospital, like, already under, you know, heels, like, "Oh my God, the fines for breaching the GDPR are so huge."

All was perfect. Like, it was really, so, right now there's, there's no issue, uh, because they did the full-on check, and we were fine. But, the whole, I would say vibe, if you could say, of building in Europe is that you have a lot of anxiety. There's a, a lot of difference between the mindset of people that come into SF, and it doesn't matter whether they're from Poland, Belgium, from Europe, from China.

It's just that in here you have not only a high concentration of capital talent, but also a certain trust which comes from well-made legislation and, and how the business is organized. In Europe, in order to launch a new technology, you need to have to first go through some checklists and guards, and then when you do that, then you can roll it out, and it takes a lot of time.

In the U.S. it's a little bit different. You get a full self-driving car, you launch it on the road, and if you don't cause any crashes, then it's fine, and if you will, then you will have an investigation, and you risk your capital reputation. But, you know, the balance is fine. I think you saw that about, like, yesterday literally, there was a call from the major frontier AI labs to regulate AI, so they're most likely afraid of, of something that they saw.

I believe that we need to find balance. So, Europe in principle has good intentions, but unfortunately the slowness of this legislation and lack of transparency and lack of unification, because it's 27 countries that need to be unified, that unfortunately impacts the speed of innovation. Yeah, thanks.

It's interesting because within the U.S., health care is often perceived to be over-regulated. Marc Andreessen has this relatively well-known blog post about how things that are not regulated change and the, the cost of it goes to zero. You can get a 55-inch TV for, like, 200 bucks now, but the cost of health care has gone up.

Do you think even within the U.S. its regulation, especially with, like, health tech and AI, is still too high, or do you think that actually the U.S. has gotten that balance right, especially, like, compared to Europe? I'm curious how you, how you think about that. I think when it comes to health tech, it's optimal.

So, my experience working at Abridge is that the private data and how you treat it, PHI is, uh, is treated with highest level of caution. So, like, even right now, four months into the job, in order to get into production, I don't even have access to this data. So, we are using, uh, special environment sandboxes to train the models.

So, I believe that it is done right, but at the same time, I don't see any inhibition. If there is a problem, usually two conversations away, you will find a solution. And moreover, if you look at the startup scene, everybody is building something. What it kind of means is that failure, in here, is just a part of process of learning.

While the cost of failure in Europe, if it won't succeed, is much higher, and then the iteration just takes so much longer. No, honestly, the moment that I landed in SF and I connected with people in the small health tech community, I just felt that, you know, everything is possible. And at the same time, they're treating the regulation with utmost respect.

I think FDA is currently trying to, uh, approach the regulation in somewhat a different way than U.S. is. Europe was the first one to introduce a regulation framework, and it was still at the time when we didn't know what AI actually is, where it will evolve. We had some idea, but right now we see that the agents can have an incredible amount of autonomy.

They have their kind of personalities, so to say. They have their roles, and there are some elements that EU required that, in my opinion, would be impossible, which would be explainability or getting deeper into the neural networks, which is just like cognitively impossible to understand of some of the parts of the neural networks are actually doing.

Currently, I think that U.S. does it a little bit better. On the other hand, I just came back from Europe. I was participating in this very special congress. Every two years, Polish doctors abroad come to a city in Poland. This time it was Wrocław. It was over one thousand doctors, and they're mostly in the U.S., and we just meet, uh, and discuss their scientific sessions.

Two years ago, when I was there in, in the city of Krakow, I participated in this small panel about AI and it was just like, you know, 15 minutes discussion, but it sparked like, you know, talks that lasted for two days. I did even like a guerrilla, uh, lecture somewhere in the halls of the Congress Center.

And this year, two, two years later, we had like five sessions about AI. Everybody was talking about AI, AI in this specialty. I brought a friend from Abridge who's dealing with cybersecurity, and right now everybody just wants to build something. And what's even more interesting is that the new generation of people that just graduate right now, I believe that almost all of them, they're having some sort of a projects on the side and they either wanna pivot a little bit into business side or they want to create some scientific, uh, project using AI.

So, it changed in a matter of really two years. Awesome. Thanks. I think that's a good transition to your role at Abridge. So, before we dive in, could you tell us what you do at Abridge, what your role is, and sort of what day-to-day life looks like for you there now that you've had this rapid transition into the startup scene.

Yeah. So, my official role is clinician scientist, which kind of translates to a member of technical staff doctor, I would say. So, basically, it is a doctor that codes or has some certain technical skills. So, what I do, it's really difficult actually, you know, to, to, to like say every week looks somewhat different, you know?

Like, literally last week, uh, we had this event with our partners in Chicago, and we're like, "Hey, guys, do you have any pressing problem that you need solving?" And they were like, "Yeah, we have problems with triage on emergency department, and we have care gaps that we don't understand why our patients are not coming for colonoscopy."

So, we created a prototype for them within, you know, really two weeks, and they were so happy that it was pitched on stage, and I think that it will be carried forward. I'm, for example, currently responsible for clinical trial matching. So, I built a system that matches the patient context with clinical trials and a lot of other functionalities that stem from it.

Uh, I'm also building a prototype of operative notes, so there's a lot of stuff. I meet with community a lot, so for example, in July, we organized our own hackathon, but it was a hackathon basically for doctors, and the event was just crazy success. So, we had eight hundred people that applied. They flew in from seventeen countries.

So, we teamed up engineers from the best companies in the industry. It was, like, even NASA was there, and all the major health care systems were there as well. And we spent the whole day building, and everybody just wants more. So, we're gonna soon announce the second one. I meet with a lot of people.

Um, I try to also learn a lot about the U.S. health care system, which still is a really crazy place for me. I think it's… I, I won't stop learning about it. I thought that Belgium was, uh, was a crazy system, but this is on another level. So, yeah, I think that I divide my time between coding and spanning like, you know, numerous agents and community and events, yeah.

Yeah. I mean, that's awesome. I'm curious how many people like you there are at Abridge and similar places. I think, like, your background is extraordinary. You have this, like, engineer's mentality, but you're also an interventional cardiologist. That's, uh, uh, I think a recurrent theme of this podcast is, like, two c— two capabilities living in one brain turns out to be, like, a very powerful thing.

Are there more people like you, or are you sort of the coordinator of the human agents where you have both the domain expertise and the coding expertise, and you can kind of direct traffic? Or are you, like, essentially the new 10X engineer because you know sort of both sides of the equation? So, right now we grow the team, so there are a few people like me.

Everybody has a, a bit of a different skill set. We recently opened up a few positions for clinician scientists, and we hired some extraordinary people. Usually, it's somebody that has a lot of clinical experience, so 10 plus years, and lived the pain, and at the same time he was either coding or deeply involved in tech.

Right now, I believe we have a team of 10 or 11 clinician scientists, not only doctors, but there's also nurses, NPAs. Since this is a company that's led by a doctor and by basically doctor values, so we really, really heavily value the physician's perspective, and, uh, we also encourage all the physicians to actually continue practicing because you have to use the product and find the ideas to do that.

So, that's why I'm still a little bit struggling to cope with my putting scrubs on the shelf, but, um, I'm quite active, uh, and, and, and follow my fellow doctors. There are, uh, some great people that we have recently hired also, and, uh, before there were four clinician scientists, I believe, in the team.

Honestly, I've been a little bit around, and I saw the companies that are in this industry, and I think that the density of the talent that we have right now is unmatched. So, just as a quick follow-on, and then I think we're gonna jump into the lightning round. Just to echo what Andy said, I think this has absolutely been a theme of the podcast.

Some of the folks who are having the most impact in this space are not wielding one skill, but they really are not only conversant, they're deeply experienced in both. And so, maybe if I can press you a little bit more to just reflect on your own background. So, you are a practicing interventional cardiologist who's gone through full medical training and until very recently was seeing patients every day and sounds like maybe will start seeing patients again at some point in the near future, hopefully.

And you're also, you're not a random or an average physician. You have, I think, a deep engineering background, and you've been building, and you've been learning through building for many years. And so, I, I don't think it's an accident. I think both of those skills are essential because I think if you take a good engineer and a good doctor and neither of them have background in the other, other's discipline, it's often difficult for them to talk to one another, and it's difficult for them to rapidly iterate towards what are good ideas unless those sort of skills exist in the same mind, and there's low latency to, to rapidly, rapidly iterate.

And so, we've seen this, you know, people who come to mind, who've come on the podcast, Ziad Obermeyer, right? Has sort of deep quantitative economic training and also is a practicing physician. Uh, Lily Peng, you know, many others who've come on the podcast I think have this sort of deep respect for both disciplines, and they've spent many years of their lives cultivating that skill set to bring it together and then to do what they do.

And so, the question that I have is, you know, as you think about how training should evolve, and I actually, I think I have two questions here. Uh, the first is, you know, as you think about how medical training should evolve and for folks who are, and I realize we didn't really ask you too much about this, for folks who know even before they go through medical training that they want to go into medicine, but they also wanna be you.

You know, they're listening to this podcast and they want to be a physician, but they wanna be a physician who's influencing AI and helping lead AI. What is the sort of skill set that you would encourage them to cultivate maybe even before medical school? Is it what you did, which is essentially going and building things and putting it out there in the world, getting feedback, iterating, seeing what works, what doesn't work, sharing things openly, uh, sometimes irreverently, right?

You know, and challenging, I'd say, some of the, the norms, which I think is very sort of, uh, open source and wonderful in, in many ways. Is it some type of training? What is that sort of quantitative or computational skill set for future doctors that maybe they should start cultivating before medicine to be in a position to succeed like you're succeeding now?

Yeah, that's a great question, and I believe that currently we have much better answer than we had twenty years ago. Before I would say, yeah, you have to take some time to learn some programming language, some library, and reiterate and so on. But right now, coding became so easy and also so much fun that I believe that most doctors already have all those virtues and the skills that are necessary, which is curiosity and persistence.

I know that it might sound cliché, but you know, most of us, they live some sort of a problem. And for many of us, the outlet of creativity is usually creating papers. It's about publishing. I have my PhD in, in cardiovascular genetics, but I wasn't always fertile when it comes to, uh, creating papers, but I was very fertile in creating code.

And I believe that right now, those two things actually converge very much. So, whenever I go around and I talk a little bit about coding and I try to encourage physicians to, to start creating software, I ask them, "What's the problem?" So, for example, I went to UCSF the other day to meet with interventional cardiologists, and they tell me, "We have problem with patient consent."

And they start, uh, telling me what's the problem, and I just bring microphone to the mouth of the person that is talking, and it starts, you know, relaying the context inside the coding environment, which is Codex or Claude Code. And I, I believe that if you are able to verbalize and, and catalyze the problem inside the working document or inside a conversation session, then this is the way that you will be able to create a prototype.

And when you see the prototype, it actually hits you, "Oh, we could reiterate, we could change the interface. We could, like, try it with somebody." And then it takes much less time than it did before, you know. Before, prototyping took months. Right now, it can literally take hours. So, I believe that if you're curious, if you're able to verbalize the problem and the solution to it, and then you're persistent, then you're able to reach this level when the software will be functioning.

However, what is important is that we are dealing with private data in here, and also, it's a regulated environment, so before we go forward and we launch it on so-called production, which means in the hospital environment, or we actually start using patients' data, then this is the moment that we team up with engineers.

So, I think that to do something for your local use, it's, it's quite easy right now, and it can actually change the way you think and you, and you function. But I just wanna stress that although the barrier to building has dropped, the barrier to building safety has not. And right now, the engineers are actually required for, not coding, but to help deliver it.

That's why I'm happy at Abridge, because there's, like, a ton of people that will make sure that what I freely create will be safe. Right. I think that, that's very well put. I, I think also, you know, I'm totally with you. Uh, and you can, the barrier to vibe code something and to prototype something and to feel it and see how it works, that has never been lower.

But I also think therefore the barrier to sort of vibe code something that is actually broken underneath, even though it on the surface level looks very good and is doing something potentially dangerous or passing data around in a way that you don't want it to pass data around, that's also been in a way never easier to put something out there that gets you in trouble.

And so, as I just think about it, I think I love the energy I think that you're gonna inspire and that you've already inspired a lot of physicians taking on more of a leadership role in AI. Uh, but as I think about all the wonderful physicians that I'm privileged to work with, I think some of them are sending me amazing vibe-coded prototypes, and then others are also sending me interesting things that they wanna put into production that I think are missing some of the engineering substance.

So, I, not to press on this again, I think you have a very unique background, and I know you have a, you have something interesting coming out in NEJM AI in the near future, too. But I think your background is not just vibe coding. I think your background was putting things into production, having them break, having websites be, you know, overrun with interest, and then patching that and iterating.

And that's, I think, a real set of engineering skills, which again, I think is potentially also responsible for your sort of outlook beyond just your ability to verbalize and, you know, get your domain expertise into Claude Code or into, uh, into Codex. So, I think that engineering mindset, I, I think is, is really, really key, at least in my mind, to cultivate alongside the deep clinical expertise that you get as a physician.

Yeah, go ahead. Go ahead. I mean, well, thank you for this compliment. I would maybe just conclude it as, you know, maybe my neural network was actually more wired into mathematics, but I have considered that the human organism is the most sophisticated of the machines— Love it. I love it. —and the most noble.

I love it. And then one last question before we jump into the lightning round. You did very well at the hackathon. You got all this attention, and your LinkedIn was exploding, and you needed to write a script, right, I think to just parse through the responses. Mm-hmm. So, you probably had more than one offer at the time and a lot of interest from other folks in the Bay Area and, you know, other tech companies that are interested in bringing on the doctor who codes.

How should physicians think about… And I, I think I, I love that Abridge, for example, is physician-led and, you know, Shiv is great, and we've had him on the podcast, and I can see the appeal. But how, I guess, how should physicians who are in a similar position as you think about, uh, or maybe the simpler question, how did you think about which company you would wanna go to?

Was it important for it to be physician-led? Was it important for it to be sort of working on a certain class of problems? You know, why, why Abridge versus the other, other potential companies you could have gone to? Well, absolutely. So, um, first of all, there are some physicians that are kind of considering whether actually they want to pursue residency after the med school.

I have some questions, and I would absolutely encourage them to do so. I still believe that the doctor's job and the way we learn, and especially when we graduate from med school, is absolutely unique. It is impossible to learn any neural network on the experience that you gather, on the level of responsibility, and on the complexity of it.

I mean, you need to live the health care system in order to understand it, because nothing in theory translates into that work. When you choose a company, that a little bit depends on what you would like to do. So, I chose Abridge because I absolutely had no doubt that we're doing the same thing. We were talking the same language with Shiv directly five minutes into the conversation, and it was just like, you know, stars aligned.

I was waiting for this for years. And I'm still in my honeymoon period, and I hope it's gonna last for a very long time. But everybody is a little bit different, so maybe somebody is a little bit more into science, maybe somebody is a little bit more into surgery, um, dermatology, visual. So, I, and, and maybe somebody wants to sacrifice less time.

He wants to continue his medical career. I would definitely say that, yeah, go with the heart. That's, uh, that's one thing. But another is also, for me what was important is that I wanted to evolve. All those years that I spent in Europe, I felt that I am evolving and I progressed. However, the weeks that I spent in San Francisco being exposed to all this talent and all those unique solutions and, and, and basically the sheer speed, uh, that motivates me very much.

I think that it also should be fun. So, the work should be something that gives you a dopamine hit and encourages you more. It's really difficult to say. So, like, to conclude, absolutely continue the medical career if it is your passion, if it is something that you sacrifice years of, uh, of your life to finish the med school, then definitely pursue with this, gain enough experience, and then build something on the side, I, I believe, to see whether it will continue.

I think that's a good point to transition to the lightning round. Um, are you ready? Yeah, absolutely. What's the lightning round? Yeah, good question. We're gonna ask you a bunch of unrelated questions, and the goal is to give a short answer. You can take them seriously, you can take them un-seriously.

That choice is up to you. Okay. Let's do it. Cool, so the first one, and I find this to be super revealing, we like to lead with it, um, what was your first job? I was distributing pamphlets as a 15-year-old guy somewhere on the streets just to earn a little bit more, and then I created websites. What kind of pamphlets?

What kind of pamphlets? Jesus. It was, I think, some quick cash loan or something like this. So not, like, political material. It was, like, same-day loans. No, no. We, we were a full-on grown democracy back then, so. Cool. Awesome. Alright. Should medical AI be explainable? Yes. Yes. Uh, it should be explainable, but I think the term explainability is not mathematic.

It's a little bit about understanding which part of the output relates to which part of evidence, and it already is. So, this kind of explainability. Okay. Got it. What person has had the biggest influence on your life? Oh my God. I cannot name those. Shiv Rao. No, I'm just kidding. But, uh yeah. Yeah, absolutely.

Um, I think I remember the, uh… And, um, so it's, it's just a colleague of mine who said a quote that just resonates with me, uh, for a very, uh, resonated with me for a very long time. "Sometimes I have to think for a very long time before I don't say nothing." It's just like the thought process, uh— Yeah. —really, really needs to, needs to transpire before you actually verbalize something.

I love that as a variation on, "If I'd had more time, I would've written you a shorter letter." Um, there's something about, like, thinking, uh, to be brief, um, in, in this case to abstain, so. Yeah, exactly. Alright, here's our next one. Will a doctor's job look dramatically different in five years because of AI?

Depends from the point of view. I believe that yes, because you will have to cognitively detach and some of the parts of, um, of the job that we do right now, even analytical, will be separated, and you will have to kind of coordinate the agents. But in terms of, of the care and of the feel, I believe that we will be more grounded in the pure principles of medicine.

We will be closer to the patient, and we will be actually, uh, in the moment with him instead of being in the moment with our PCs and, and documentation. Cool. Um, this is one that we ask a lot, but I feel like you have a very unique perspective on it. Will AI in medicine be driven more by computer scientists or clinicians?

Clinicians. Definitely clinicians. Right now, I believe that the computer scientists have created such an incredible machine that I, I, I cannot imagine how much better it can actually become. But what we are lacking is actually, you know, translation of the pure clinical knowledge and domain expertise inside work.

But it might be sinusoids, you know? It might be just, you know, eons or eras of, uh, discovery and scientific advancements that can be driven by both. I, I don't know. Hmm. Alright, here's our last one, and, uh, this one's admittedly a little bit unfair, but I think given all the, uh, the recent news stories and the obsession that I think this has, this has been as a story, we, we had to ask this.

Um, and also, you're in San Francisco, so I think, uh, this has probably been a, a daily conversation as well. What is your P-doom, your probability of doom? Where I'll define that as a shorthand for the probability that you personally put on something with AI going catastrophically wrong. And so, you can define exactly what catastrophically wrong means, but some people have talked about things as extreme as human extinction.

Others have talked about a loss of control, economic devastation, but the probability that AI does something very, very harmful, catastrophically bad, uh, for humanity in the next five or 10 years. I believe that the risk that the humanity will do something more catastrophic is much higher than, than what, uh, what actually AI might do.

So, we all have the fear of this automated, you know, improvement, auto-improvement right now. I believe that if you look at the side of the world where I come from, then we have some more persistent issues right now that could lead to, uh, some sort of doom. I am somewhat of an optimist because when you look at the history of, of, of humanity, uh, there was always fine balance between the good and the evil.

And if the good wouldn't be more than the evil, then we wouldn't be here and, and, and we wouldn't talk. So, I honestly am an optimist, like most of those researchers, that AI can lead to, uh, to improvement and maybe, I don't know, maybe the concept of countries one day will be just, you know, thrown away and we will be happy as John Lennon, uh, sang in, in "Imagine."

But, uh, aside of romanticisms, uh, romanticism, I believe that, uh, mm. No. I'm not a fan of, the doom predicament. I believe that this will lead us to good things. Okay. Alright. Well, I think that was, uh, it was a great lightning round, and so we, we actually just have a few last questions here just to conclude, some kinda big picture closing thoughts.

Um, and I think we've touched upon this, but maybe we can drill more into it. So much of what I think is critical to what you're building and your role at Abridge is your clinical expertise and your medical training, right? The perspective— Mm-hmm. —that you bring as a physician who's on the ground practicing, that you're infusing into what you're building and how you're influencing what the engineers are also building, uh, at Abridge.

And so, um, this is, uh, I think a question that maybe the answer, it's totally fine if the answer is, like, it shouldn't, but I think, uh, we've asked this question to a lot of folks, and I've been fascinated that we've gotten very, very different answers. You know, so you're a practicing clinician.

Clearly, your medical training has helped you influence AI and the development of AI in a positive way. So how, if at all, do you think medical education should evolve in the era of AI that we're now living in? Yeah. So, I believe that the current AI tools are the kind of technology that's maybe not equivalent, but the successor of Word, Excel, and what led us from writing on paper into, uh, into using, uh, computers, uh, as our daily interface for, and, and Internet.

I believe that the whole curriculum should be adjusted to it. I believe that there should be classes, uh, and there already are. I was at Stanford the other day, and there's, like, this whole course about using generative AI and understanding the limitations and, and also the potential productivity benefits of it.

So, this is going to be, and this already is our reality. We are surrounded by AI, whether it is discreet, whether it is something straightforward, whether it is proactive, retroactive. This should be a large part of the curriculum, and it should be taught early on. How we learn, how we train, how we exercise.

This is already influenced very heavily by AI, so it should be a large part of the curriculum. Awesome. I think maybe one question to, like, take us home. We like to ask these in a specific order. So first, what are you most pessimistic about? I think that to Raj's point during the lightning round, like, pessimism is kind of peaking right now in terms of notable AI leaders asking to pause research or to pace the frontier.

I think that if you look at public polls, sentiment around AI and data centers has never been lower. It actually is the most bipartisan issue in the United States right now, that data centers stink. Um, so I'm curious, like, what are you pessimistic about? And we'll ask the inverse of this question next, but, uh, first, like, what gives you the most concern right now?

Well, having a transatlantic perspective, I'm mostly pessimistic about this compartmentalization and isolationism that we are currently divided between U.S., China, and Europe that tries to struggle. You know, they say that America innovates, uh, China replicates, and Europe regulates. So, that's something that worries me a little bit, that instead of joining forces, we're just, you know, divulging into trade wars, technology wars, and so on.

And this trend is going to be something that might derive the use of AI to some more nefarious, um, usages. I hope not. So, this trend worries me, and I truly believe that somehow if we could inverse that, then we could, instead of using those tokens for, you know, intelligence or analyzing satellite imagery and, and, and warfare, we could actually, uh, cure diseases which were incurable before, and we could make life multi-planetary by quoting somebody.

Sure. Uh, but, um, yeah, this is, I, I, I believe that this is the trend that worries me the most because— Hmm. —deeply we, we have talent. Like, when I went to Poland, and there, there were so many ideas that could just flourish. But somehow because of the fact that we are separated, it's, uh, the, the, the speed of improvement is, is not as it should be.

Hmm. Yeah, I think, I think I share all those concerns. So, to avoid ending on a pessimistic note, what are you most excited about? What are you most optimistic about, and what gets you out of bed every day excited to, to do the things that you do? I, I, I think it's the variety, you know, of things that, that just arrive, you know.

Like, uh, every day I find some crazy solutions that, that I've never thought that are possible before. So, like, the, the Postvisit thing, that was crazy that I was able to do it, like, within really, it was three days because the rest was creating a video. Uh, but, um, right now, um, the speed with which the frontier models are coming out is just like on this hockey stick curve.

And, uh, right now we have Astra that, that went out, which is just a killer of a model. It's, it's incredible. And I just recently realized how well you can actually control your computer with those models. Like, for example, when I'm doing a presentation right now, I'm talking to Codex that controls the computer, and it controls PowerPoint, which has Claude inside, and Claude talks to the slides.

So there's, like, agents talking with agents. There's just, like, multiple layers. I'm, I think I'm excited about, you know, the next week, the next two weeks to see, you know, what's, what's gonna launch, what incredible ways to interact with voice, with thoughts, with uh, with just another level of efficiency of, of transmitting your intent, you're going to be able to do, like, hell, maybe one day we're gonna have Neuralink and we will just be able to close our eyes and, and work there.

Like, it wouldn't be work there. But, uh, yeah. We live in sci-fi a little bit. We do. I think that's a, that's an amazing note to end on. Yeah. Thank you so much. Thanks, uh, thanks for coming on the podcast today. Oh, thank you very much for having me. It was really, really great discussion. Awesome.

This copyrighted podcast from the Massachusetts Medical Society may not be reproduced, distributed, or used for commercial purposes without prior written permission of the Massachusetts Medical Society. For information on reusing NEJM Group podcasts, please visit the permissions and licensing page at the NEJM website.

Abridge’s Dr. Michal Nedoszytko on Physicians Who Build Software

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin