I saw this patient. She had a 10-year history of breast cancer. And she was coming to see me because she needed preoperative cardiac evaluation prior to starting doxorubicin, and she was super nervous and anxious. So, at the end of the encounter, I asked her why and if there was something I did or something I said to make her clearly feel so uncomfortable.
And she told me that since she was diagnosed with breast cancer, her husband had come down to every single visit with a new type of doctor, but he couldn't make this visit. Him coming to those visits just helped her feel more present and I said, why? What does he do that's not obvious? And she said that he's quiet. He just, he's in the corner. He just, he just takes notes. Him taking notes meant that she felt like she could make eye contact with me.
She could feel more present knowing that they could go home and rewrite all those notes. Google all the big words and then go to the next clinician and feel like the main characters of their story retelling it as opposed to someone looking in from the outside. And I think the revelation for me and for so many of us, it's all about all of us really wanting to build better relationships, and have better conversations and be, be better to each other.
And so, wouldn't it be paradoxically profound if we could somehow leverage technology to bring those people closer together? Welcome to another episode of NEJM AI Grand Rounds. I'm Raj Manrai. I'm here with my co-host Andy Beam, and today we are delighted to bring you our conversation with Dr. Shiv Rao. Shiv is the founder and CEO of Abridge. Abridge is a health care AI company that's creating software for clinical conversations.
And during our conversation, we really got to dig into why Shiv started Abridge, what they're trying to do, and what traction they have. And I'd also say, Andy, this is a pretty wide-ranging conversation, even more than normal. Shiv is a pretty cool guy, and he has done some electronic music, lots of other things, I think maybe we talked about skateboarding at one point. He has a lot of interests outside of medicine and health care and AI. So this was really fun. Yeah, totally agree.
And you know, we first heard about Abridge through a friend of the show, Morgan Cheatham. We were at the NEJM AI event in Puerto Rico, and he gave us a demo, and we were both kind of like blown away. So, Shiv is an amazing CEO. You can tell that he's thought a lot about his company's mission. He can tell it very succinctly and when you hear him talk about it, it's tough not to get excited.
I think one of the reasons I was so interested in this conversation is that the revolution of AI and health care has been happening for a long time now, but we have yet to see very durable and sustainable business models behind that. And Abridge is really at the forefront of having not only compelling technology that doctors frankly like, you know, this is not a paid advertisement. But it does parts of their job that they just don't enjoy and it does that part well.
But they also seem to have traction commercially. So, it's been very interesting to see this new wave of AI companies that both have interesting tech but also interesting business models. And so it's really great to dig into Shiv on both of those topics. 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 Shiv Rao.
Shiv, thanks so much for joining us on AI Grand Rounds. We're excited to have you here. Oh, I'm super stoked to be here, thank you. Shiv, great to have you on AI Grand Rounds. So, this is a question that we always get started with. Could you please tell us about the training procedure for your own neural network? What data and experiences led you to where you are today? Yeah, I love that framing. Maybe I'll go backwards and then go forwards again.
And I think there, there's just like a few stories that are imprinted inside me, in my brain, that I think continue to be really informative and influential. So, prior to starting this company called Abridge in 2018, I was an executive at a large health system called UPMC. I sort of violated Peter's principle 10 times in five years, but I got to be the check writer for their corporate venture capital portfolio.
We were deploying their money into startups, but also into Carnegie Mellon University. We started a machine learning and health program. And I got to learn osmotically from those professors and those Ph.D.'s, those postdocs that we were funding, and it was just an amazing privilege. A couple lifetimes ago as a college student, I went to Carnegie Mellon University, and I think I was like subconsciously rebelling against all the Indian doctors in the world.
You know, against my dad, against my sister, against everybody I knew and grew up with. So, I studied a lot of the other stuff. I was in the negative space, if you will, of what I'm doing now. I got a scholarship year to program virtual synthesizers in the school of Computer Science and School of Fine Arts. Took a lot of film theory and art theory. I think I have a tendency to get really, really obsessive about something and then get obsessive about a seemingly like disparate thing.
And when I was a junior in college, maybe this is like a first story that's informative, I met this professor, this architecture professor named William McDonough. He was visiting and he was giving a lecture about sustainable architecture.
But really, like he was talking about design thinking when I think the d.school at Stanford was still pretty young, and he told this story about this ophthalmologist in India who designed this revolving platform that he sits on, and he used to do cataracts, but the way he would do it, he was, he'd sit on the platform, they'd bring a patient in at 12, he'd do a cataract, he'd say, spin, then do another cataract at three, spin five minutes later, cataract at six, and then spin, spin, spin.
He's like spinning all day long doing these. Five to 10 minute procedures, and by the time I'd heard this lecture, he'd given eyesight to over a million people, and he taught the procedure to his daughter who'd given eyesight to over 400,000 people. And I remember thinking biblical-like impact. This is, this is like next-level impact that he had delivered to the world. I remember like leaving that lecture thinking, you know, maybe I wanted to be in health care. So, I pivoted pretty hard.
I spent about a year making music and playing music at, like, art museums. It was like pretty weird music and like kind of avant garde electronic stuff where you just sort of like, you don't dance, and you just like, close your eyes and, like, listen to it. It used to be called intelligent dance music, like IDM, or Autechre or Aphex Twin, if you're familiar with those sorts of folks.
I, I, I think we should, not to interrupt you, I'm really enjoying this shift, but yeah, I feel like we should try to splice some of that into this episode if we can at, at some point. So, great to, great to, great to hear, uh, here's some of your work. I finally like figured out how to, in that scholarship year that I got at Carnegie Mellon, I got to do pre-med requirements and so like, kind of pivoted pretty late, went to med school, went to residency.
And then kind of found my passion and I think like that passion is absolutely at this intersection of like health care and technology where I perceive the most impact. I think it's like back to that ophthalmologist story of, like, how do you create that kind of impact at that kind of scale? And it's not that seeing a patient on a weekly basis isn't any more or less profound than what you can do with technology. But technology for the people who want that flavor of impact, that impact at scale.
There's really nothing else like it. So, as I suck all the air out of this podcast, I'll maybe pause for a bit. No, no. So that's great. I think typically, at this point I would ask the guest to backtrack and say, how did you get into medicine? But I, I'm struggling to even come up with the right framing given the, like, journey you went on from intelligent dance music, but was there some inkling of a future physician in you as a kid? I know that you had family connections.
Were you just completely dispassionate about medicine and then you got interested in it again after this, or just like, was there something as part of your origin story that's relevant to this conversation? You know, I think that the passion has, has been for a specific type of very human impact. And how I was gonna get there, I don't think I ever knew for sure. But once it became very clear that this is a path, there was a path for me, that I could kind of go after.
And that there was creativity in health care as well, or like the variant of creativity that I've always been maybe like seeking. I think that's where, like, that resonant frequency got unlocked and I started to really go after it. But there were like three patients as a doctor, you know, when I was in med school life. I hated it. It felt like torture, rote memorization. Just felt, I never went to class. I spent most of my time undergrad and even much of med school, just like skateboarding.
I had, like, diluted myself into thinking like, that was, it was gonna, that was the right approach to academics. But I remember when I went to residency, things started to really come together. I went to Michigan and my first patient that I saw in Michigan, is so important to the moment that we're in right now with generative AI in care delivery. So, University of Michigan, Ann Arbor, so many of your patients are like academics. They're professors at the university, like highly educated people.
And I remember walking in the room with this newly M.D.'d swagger from med school thinking like, I could come in and like tell this patient what's up and here's the differential and this is what we're gonna do. And as soon as I walked in the room, she started to quiz me, this patient, on whether or not she had multiple endocrine neoplasia type 2A or 2B. That's like a, that's like an MCAT question, you know what I mean? Like that's not a question that like anybody has off the top of their head.
But it was like the most incredible, humbling experience. Like, the best way to like actually be thrown into the trenches is to be humbled. And thankfully, I think I pivoted pretty quickly and realized there was no way out. I just needed to own my ignorance. And so, on the spot I was like, you know what? I don't know. I think we should look this up together and maybe I can help us navigate the literature.
So, I remember turning on the computer in the corner of the room, going through some of the papers very quickly, kind of figuring this thing out and sharing it with her. And I remember, like, she appreciated that. I think we built some type of rapport, some type of trust in that moment. And that moment I feel is so much of every moment anymore in care delivery, like, especially with patients being equipped with all of this knowledge, it takes two hands to clap.
And I think both parts of that conversation, being able to explore symptoms, medications, diagnoses, procedures, differentials, like is, is really like the future of health care.
Yeah, I am just struck by the maturity of that response in your first year of residency, because I know so many doctors who have the, your Google search isn't equivalent to my M.D. and I understand the sentiment behind, behind that, given I've married into the profession, but I think that like, it's like such a, such a unique reaction to that type of challenge from a patient.
And I think most of us are getting a sense now of like the intellectual breadth that you've had, I guess like all of the things that you've done had, like, well-prepared you for that moment. Hopefully, I think so. Or maybe I, maybe I was caught off guard and just like somehow reflexively stumbled into the right reaction. But like humility is everything, right? I think in health care you're gonna be humbled on a daily, weekly basis by some patient case, some scenario.
There's like maybe two outta three patient stories that have been, like, super important. The second one was probably as a third-year resident. So, I'm a senior resident and my favorite place in the hospital is the ICU. It's like the high acuity stuff, the stuff where like it's life or death and you gotta react very quickly, and that it's something to do with maybe moving on to become a cardiologist. And I was in the ICU, we saw this patient.
He was in his 40s, he was transferred in the hospital. He was a former pro-athlete. And at that a pitcher, and he was coming in with septic shock from a rural health system. And sepsis, you guys are, are probably familiar with how the guidelines have shifted over time. But, like, I think one tenet of septic shock care is you wanna flood the patient with fluids because their arteries, their blood vessels can constrict.
And if they constrict, then you're not delivering enough blood to the rest of the body and your organs can start to die. And so there used to be this refrain before people realize that that Levophed leaves them dead. Meaning like if you give them a medication like Levophed that constricts your blood vessels without putting in anything inside those blood vessels, you're constricting nothing. You're not improving blood pressure. You're just cutting off circulation.
So that's kind of what happened in this patient before he came in. He was given pressors without fluids and, over the course of weeks, we got to the point where we were able to get him outta the ICU when he was able to walk out. But unfortunately, and this is a really sad story, but he needed, he got amputations of, like, his hands.
And it's like, especially tragic, obviously given his backstory, as a pro-athlete turned, like, physical therapist and so, I remember thinking like, there's gotta be a better way. There's gotta be a better way to be able to get, like, cutting edge literature, cutting-edge information, not just into doctor's brains, but into the point of care. And certainly, there's any number of ways that we've all probably explored, the industry has explored, trying to get this information to the point of care.
But there's no question that, especially with the new types of technology, like generative AI, we're gonna discover even better ways, like even, less friction-full ways to be able to bend the trajectory of outcomes and costs. So, Shiv, I just wanna ask one follow on to that. Yeah. So, I think you, referenced that after medicine residency you went into cardiology, right? So, you did cardiology fellowship and you're a cardiologist, and as I understand it, you're still practicing, right?
Yep. And we're gonna get into Abridge in a moment, but I'm wondering if you could reflect on what drew you to cardiology? When was the sort of moment where you knew that you wanted to pursue a fellowship in cardiology, practice cardiology? Can you trace it back to either an experience in residency or even before that? Yeah. Yeah. I think it has to do with cardiology being what I believe to be a really elegant mix of primary care and also surgery.
And for folks who have really short attention spans like me, for people who love data and certainly cardiology has seen a lot of funding. And so there've been a lot of clinical trials that are, there've been a lot of RCTs. You can always find some data, so that you're not walking in the room completely kind of naked, from that perspective.
And also, if you're someone who wants and sees that there's, like, an incredible privilege in building a relationship, like potentially a lifelong relationship with a patient, as you often do in cardiology, it's just one of those really, really magical fields. And there's other fields like this. Maybe gastroenterology and urology are like this as well. But I think cardiology especially resonated with me for all those reasons.
You know, you think about the heart, and I use this metaphor with patients, it's very much like a house. It's like a four-bedroom house. Two rooms at the top, two rooms at the bottom. And there's plumbing, just like plumbing in a house, there's plumbing in a heart, and that leads to a heart attack. Just like a house. There's electricity. The electricity in the heart leads to a whole field called electrophysiology. And that's where, like, many of the geeks in our profession end up gravitating.
But there's, there's also rooms there's all the physics related to valvular heart disease. There's also the pump, and it's not a water pump in the basement. It's the blood pump, and then you have problems there, you get heart failure.
And so, it's just like such a rich field from that perspective, especially for someone maybe who loves the physics of physiology and pathophysiology that you can geek out, but you can diagnose the problem, you can talk to the patient, build the relationship, and then you can treat it. You can go right to the cath lab and you can be the one who fixed it. And so that end-to-end, that ownership of the entire stack is something, pretty rare and pretty unique.
I think you just made a really compelling case for going into cardiology for some of the residents who are listening to this. Shiv, that was a great overview of the training procedure for your neural net. Currently, you're Chief Executive Officer at Abridge. I think we'd like to transition now and hear more. I think we got like a little foreshadowing, on that. But could you tell us about the origin story of Abridge?
Our machine learning colleague, I think Zack Lipton, features heavily in this origin story. If I have my facts right, could you tell us how this all came together and like what you're doing now? Yeah, absolutely. So, we started the company, we started Abridge in March of 2018, and a couple years before that, I was in that role at UPMC I told you about. And I was also overseeing some partnerships. So, we had a partnership that I was overseeing between UPMC and Microsoft Research for B as well.
And we were going after AI and health care challenges together. And that was certainly exciting, but sometime at the end of 2017, it became clear to me that I really wanted to be in the weeds and have access to that whole stack again. But on the technology side of things and wanted to be building with people who really inspired me on a problem that I thought would take forever to really figure out. And that gave us like infinite surface area, seemingly.
And I think the really, the most important epiphany that like underpins everything that we're doing at Abridge is that we think health care is about people. That's not gonna change. We don't think doctors and nurses are gonna be fully automated in the next decade. And if they're not fully automated, then you invoke first principles and you think what do they do? What's one of the original signals in care delivery? It's this: it's a conversation.
Those dialogues, those conversations drive so many workflows in health care. They're upstream of clerical work, like clinical documentation that crushes my soul every time I see patients on the weekend. Obviously, it's a different story with Abridge, but like, like it's a real challenge, and those conversations are also upstream of revenue cycle. You know, in this country, we're not compensated as doctors for the care that we deliver.
We're compensated for the care that we documented that we deliver. Those conversations are also upstream of clinical trial, recruitment of care management, of clinical decision support, ultimately experiences and outcomes. And so, there's that surface area, you know, there's that potential.
And so, maybe like the last story that I'll share with you that had everything to do with starting Abridge, was like in March of 2018, and this was when I still had a weekly clinic, a weekly cardiology clinic. At this point, I just see patients like one week in the month, and I saw this patient 50 years old.
She had a 10-year history of breast cancer. And she was coming to see me because she needed preoperative cardiac evaluation prior to starting doxorubicin, chemotherapy that could affect her heart muscle. And she was super nervous and anxious, like crawling out of her skin. So, at the end of the encounter, I asked her why and if there was something I did or something I said to make her clearly feel so uncomfortable.
And she told me that for the last 10 years since she was diagnosed with breast cancer, her husband had come to every single visit with a new type of doctor, but he couldn't make this visit for some reason. And she's an English professor at the University of Pittsburgh, super eloquent. She told me that him coming to those visits just helped her feel more present. And I said, why? What does he do that's not obvious? And she said that he's quiet. He just, he's in the corner. He just takes notes.
And she told me that him taking notes meant that she felt like she could make eye contact with me. She could feel more present knowing that they could go home and sort of rewrite all those notes. Google all the big words and then go to the next clinician and feel like the main characters of their story retelling it as opposed to someone looking in from the outside.
And I think the revelation for me and for so many of us in those early days of a bridge is that that story that she had as a patient is so similar to the story that doctors have, that nurses have as well. It's really about agency, it's all about all of us really wanting to build better relationships. And have better conversations and be better to each other.
And so wouldn't it be paradoxically profound if we could somehow leverage technology to bring those people closer together, to help them make more eye contact, to do what not only the doctor has to do for hours on end in their pajamas after the kids are in bed and the dinner's been eaten. But could we also do what her husband did, in the corner of the room?
And so threading that needle through the most important people in health care patients, obviously first and foremost, is like the ultimate ambition for everything that we're building at Abridge. I mean, that was amazing. Shiv, I love how you tied, all those different pieces together. When you were thinking about this from a technology perspective— Mm-hmm. —how did you start to formulate the core technological approach?
You know, I think it even was not obvious to me and lots of others, like, the doctor-patient interaction as being the place to insert technology. So, I guess, like, how did, how did the, this, the technical roadmap come together? Yeah, totally. And there were venture capitalists back then who categorically avoided the point of care, and they would tell us like, we're avoiding that place. We want nothing to do with it. So, Zack would say, so Zack was a founding advisor.
It's funny, like I remember calling him and asking him if he would co-found the company with me, and he was just starting as a professor at Carnegie Mellon University, it wasn't exactly the right time. But here we are, and he's our Chief Technology and Science officer. So, it all worked out in the end. But as a founding advisor, he obviously has his fingerprints all over our original approaches to this challenge. And he would scoff at the idea of attention is all you need.
Potentially changing the trajectory of our early research, because certainly I think to folks like you, to the researchers out there, probably like everyone was already experimenting with transformers and trying to figure out their place and their place in workflows like this in health care. But that was certainly the moment where we started. We started with pre-trained models like Burr, bioBert, long-form Pegasus. Like we experimented with them all.
In August ’21, Zack and his lab published a paper and presented a paper at ACL that really kind of like started to make clear that we could do this. And this is pre-LLM. And in that paper, he described this two step, two stage modular summarization pipeline, basically to create doctor notes, also known as soap notes. And the first step was like utterance extraction.
And again, we were using a BERT-based model, like it was fine-tuned, but that model would classify every utterance from a dialogue, from a transcript for quote unquote note worthiness. And then there was an abstractive summarization piece or step where we would use a fine-tuned like T-five transformer model to basically write the note in the style of a doctor's note. And we were leveraging this huge corpus of data that we had aggregated and paid for to annotate over many, many years.
And that's really, so much of our early days as a company was about. Kind of like doing two things at once, walking and chewing gum, if you will. The walking piece was around this long game around R&D that we were embarking upon. But then the other aspect was we didn't hold our breath either to, like, learn from users. So, we put a consumer app out, believe it or not.
So, like recognizing that it takes two to create that dialogue and that the patient story and being able to serve patients over time was gonna be such a core kind of aspect of our mission. We put a free consumer app out in July 2019.
And so, while that was going, we were also asking those users for permission to their data so that we could then add, annotate it, aggregate it, create these new notes that were annotated in exactly the right ways with all the right timestamps to be able to go after this problem, like, with those types of models. So should that, that's actually the question that I wanted to ask, which is, so you guys started in 2018, right?
And this is now hard to believe, but this is like maybe three, four years before ChatGPT large language models and the current moment that we're in technologically, right? Where now I'd say we have, we went past the ChatGPT moment, and now we have this rich ecosystem of competitive LLMs at the foundation layer, both proprietary and open source.
And so, what I'm really curious about is, how you have navigated that sort of technological shift from 2018 BERT-based models, very different I imagine ways that you were approaching or even thinking about problems then versus the sort of LLM-era that we're in now. How have you adapted the technology, what you're doing, the models, how you even approach R&D or even maybe even how you think about R&D for the company? Yeah, it's a great question.
It was 2022 when we started to pull LLMs into the stack, and where we obviously saw all, like, these stepwise improvements in terms of the output. And I'd say the way we were initially integrating LLMs in was fairly basic compared to what that stack looks like today. Today, we have this whole reasoning engine, and it's any number of different models, maybe like 20, 25 models that are all orchestrated together in order to create that clinically useful, but also billable node and all.
And where there's also models that are looking for hallucinations, like guardrail models or models that are doing information extractions so that we can template orders inside the medical record. And so, the level of complexity, you know, now compared to 2022, it's like off the charts. But what our prepared mind and all that potential energy that we had stored up from all those experiences framing these challenges, albeit leveraging tools like BERT and
T-five. What it meant is that in 2023, as we were inserting and starting to leverage LLMs in our stack, we just, all that potential energy turned kinetic immediately. And it's because of multiple things happening at the same time, like multiple stars aligning. The most important star for any company is the market. You know, you can have the best product, the best technology. If you don't have like a real market problem, you're not gonna really create that much impact.
But in this case, the market needed this. It still needs this. Two outta five doctors don't want to be doctors in the next two to three years. 27% of nurses per JAMA don't want to be nurses in the next 12 months. There was one survey that suggested that like 60% of medical students don't wanna be full-time clinicians, you know, when they graduate.
So, we have this public health emergency where patients are having to drive sometimes five, six hours from rural settings because their hospitals have shut down, and that's the only way they can see a rheumatologist in the inner city hospital who can prescribe the biologic that can save their lives. So, we have to do something to address this challenge.
And so I think what that led to in 2023 was C-suite executives across the country understanding, recognizing that we need tools for the first time ever, maybe. We need automation in a way that we didn't before. We need to augment clinicians somehow by hook or by crook. So what I didn't know was that all those other years, those years leading up to 2023, I was pre-selling. So, when we were doing these demos.
And we were leveraging that stack that I was telling you about, that Zack published on that, that his team built. We were really pre-selling because that product, by the way, was very compelling. It wasn't quite as magical as where we are now, obviously, but it could do a job for some clinicians out there. And so, all those C-suite executives we had demoed for who didn't have time before, for whatever reason, they called us back.
And one of the big learnings, like, I think anyone building in health care AI or health care tech in general needs to understand is, like, health care is not homogenous. And on one end of the market there are the large integrated delivery networks, the academic medical centers, the payer providers, the complex systems. That's where 70% of doctors are. On the other end of the market though, there's the direct primary care doctor down the street.
There's the independent PCP, there's people taking cash payment out of pocket, off the insurance grid, so to serve the individual doctor. Or the small provider group is a totally different undertaking than serving the large complex system. And so that's also probably why in any given space in health care, like any given use case, you might see a ton of startups competing.
But then there's probably only two or three who are trying to actually, or who have this stuff to actually try to serve those large systems where you have to come correct off the bat. You have to serve all the, in our case, all the specialties, all the different settings, outpatient, inpatient, urgent care ERs, and in our case, all those different spoken languages. We're in Boston, for example, today.
And there's, where you guys are and there's probably gonna be tens of thousands of conversations spoken on Abridge in Haitian Creole, and Brazilian Portuguese. So yeah, it's a different kind of complexity. So, you know, you mentioned some of the technical challenges that you address by having this mixture of models and this complex system that's underneath the actual product itself. And one of the ones that caught my attention was hallucinations, right?
So, this is something that we've been talking about with LLMs and with ChatGPT that I think has now entered mainstream, right? With the general LLMs from the tech companies, we all know that they hallucinate and I think they still do, right? I think this is not a solved problem at that sort of base layer, but what's interesting is now there are companies that are serving kind of medical needs like yours, like a few others that are trying to address this problem, right?
This problem with the base layer, with the LLMs that are involved in your stack. And I'm curious, you know what, whatever you can tell us about how you solve that and how you sort of validate, 'cause it's actually, it's a very hard problem, right? To even know how much your system is hallucinating to have well adjudicated labels of what a hallucination is. To take expensive sort of human time to verify whatever are the automated parts of the evaluation itself.
Like how do you really detect and reduce those hallucinations? How do you approach that problem? Maybe just to dovetail into that, one thing I get asked a lot is how do you know these products are safe? And I think that that's like a distillation of what Raj is asking. That's, that's a shorter and better question than what I just asked, so let's go with that one.
Yeah. And safety is a really interesting topic because in many ways you could argue that health care, like, the health care industry has adopted AI faster than any other industry out there. Over these last few years, like health care has taken up AI and scaled AI. Unlike any other industry, it's wild, it's historic, it's never happened before. We're live in well over a hundred health systems. We're at scale. We're probably touching millions of clinicians and millions of patients every week.
And that happened relatively quickly over a few years, and that's never happened before. When I was a corporate VC, if we had invested in a startup that was provider facing and the founder said, you know, that they, they, they closed two pilots a quarter, it'd be champagne bottles all around the room. Like, going to health systems for oftentimes for the right reasons means you, you need to be prepared to eat glass. It's gonna take a long time to inflect.
It's gonna take a long time to build trust, but I think because of all these, like, sort of, challenges in health care because of like burnout and burden and all the other supply-demand mismatches. That certainly created a moment. But what happened, what we saw happen was that the health care industry sort of intuited a kind of like risk profile around a company. So, what's really taken off over these last few years are companies focused on use cases that are high frequency and low stakes.
So, documentation when the clinician is in the loop can to trust and verify. And when they also have tools, like, we have a feature called linked evidence where you can highlight a word, sentence fragment, paragraph, and we show you where the evidence came from, all the way down to the actual associated snippet of audio. You can trust and verify at that level. You can, it's all auditable.
And so with that human in the loop, that clinician in the loop, we certainly de-risk, I think, many of the issues associated with AI at the point of care. But then you think about the other end of this matrix, high risk. So, high stakes and high frequency. I think clinical decision support, for example, in the ICU, I think about all the sepsis work and like what it takes to actually deploy sepsis predictors to the point of care.
Those companies have to verify, they have to validate, they have to publish. It's a different type of gauntlet. Yeah, I think that's very well said. And I think what is fascinating here is that we have evidence, about the accuracy, the diagnosis rate. We're almost sick of these studies now, all of us, right around—should be careful with what I say. But we're, you know, all of us are very used to these studies showing multiple choice question dominance of LLMs.
But I think what you're saying is that the use cases are very different from one another in terms of the risks attached to different errors. And I think you're right. Also, clinical decision support feels different than documentation. Although I think documentation, my guess is you would say it's also very important not to hallucinate there. Mm-hmm. As well.
And it's important that it reflects what actually happened and that it reflects the patient's story going back to the way you so eloquently put it at the beginning. And so, I think this is still a general challenge that the field is facing not only for the sort of the foundation, the base layer, but I think for the companies that are providing medical information, whether it's through, documentation, the serving medicine, right through documentation,
like your company. Or I think some of the others that are trying to live in the space of providing useful information to physicians at the point of care? I think of something like open evidence, right? This is a hard problem that they're trying to solve around hallucinations, around the accuracy and the fidelity of the information that's provided and the references. And I think it is an open both academic and technical problem and very interesting to sort of hear how you think about it.
It's actually pretty related to that line of questions. You mentioned something about trust, and I think this is a big part of what you have done very well. You have doctors trusting your product and I'm wondering if you can tell us, how did you go about that? How did you actually get doctors to trust Abridge? Yeah, trust ends up being the only currency that matters
in health care. Clinicians or executives inside of a health system only want to know so much about your technology and your stack and all your guardrail models. They wanna understand that it works, that it's reliable. They wanna understand if you are transparent and if your technology is somewhat transparent and audible. They wanna understand if you're credible people as well. And so, I think for us, transparency, reliability, credibility are the most important.
Those are the dimensions of like trust that we focus on, and we just try to unpack what that could mean at every single layer of our company. Whether that's on the go-to-market and sales side, whether that's on the technology side. I mentioned that feature linked evidence. We published white papers like we did one some months ago where we sort of broke down a lot of our kinds of evals, and our benchmarks, and how we hold ourselves accountable, and what our word error rates look like today.
And, like, our models are trying to continually improve we try to teach our clinicians and our health systems about preference tuning and DPO and try to tell 'em about like reinforcement learning and our LHF and what that means when we go live. And I think that transparency, like, they appreciate, too. But I think ultimately for us, the challenge in health care, like the exciting part of that end of the spectrum, like the large end of the spectrum, is that we have to thread
these needles. And number one, this has to work for the end user, and it's not just any end user, it's like all the different specialties, et cetera. Number two, it needs to work for executives, and one is the CMIO. The CMIO wants to know roadmap. Where are you going? Are you credible people? Can you get there? Do you have science at the center of your company? Who are those people? Do they have the stuff, do you have proprietary data sets?
Like they might ask those questions to understand that like these slides that you put in front of them, that you're gonna be able to execute against them. There's the CIO. And the CIO has any number of concerns. But some of those concerns relate to can you integrate, can you be a credible part of a very complicated stack of investments that they've made? Are you gonna be around for the decades to come? Are you funded as such? How much are you investing into R&D?
How do you think about enterprise? Do you have all the certifications? Can you check off all the compliance boxes? But then there's the CFO. And that's like the third most important, or maybe they're all the same. But like that, that's the third stakeholder that we encountered over these last few years. And now they're a regular part of any partnership. And the CFO wants to know the bottom line. They wanna understand the metrics and any more of the metrics. Are not just around burden or burnout.
They're around, revenue cycle. Because again, like these notes are bills essentially. And so, like, how do we quantify the impact that we're creating and helping clinicians capture and get credit for the complete care that they delivered? Fantastic Shiv. Actually, maybe one final question then, I think we're gonna jump to the lightning round.
Could you give us like a quick summary of what scale you're operating at now, and to whatever extent you can, a preview of the roadmap for what's next, where you're looking to expand within the United States, outside of the United States. What's on the horizon for you?
Yeah, so far I'm happy to share that we're live in over a hundred health systems across the country, and they tend to be large academics and integrated delivery networks, and they tend to also be scale deployments across specialties and settings. That means we're touching millions of lives a week. So, we're impacting millions of patients and doctors every week, which is obviously incredibly fulfilling and exciting. The roadmap is equally exciting.
So, we've raised a lot of capital over the years, and 80% of the capital that we've raised is really all about research and R&D. It's all about R&D, like 80% of it needs to be invested into R&D. We believe that will translate into just, like, better product. That's what it means to play a long game. Like at the end of the day, it's just about creating impact. So, what that impact looks like is just changing so quickly.
So as an example, about a month ago we announced what we call a contextual reasoning engine. That reasoning engine is pulling in disparate data from multiple sources. Including the medical record, but also including payer systems, including rev cycle systems, including textbooks, including coding best practices, including potential rules that are a health system specific. I remember I used to be on a clinical documentation improvement team at UPMC and we would go from department to department.
We do these lunch and learns, pizza and PowerPoints, try to teach doctors how to write notes that are billable. How to reach PCPs. What risk adjustment is. What an HCC is. How an HCC maps to an ICD. How to think about me criteria. Make sure to include what you discuss. Monitoring, evaluating, assessing, or treating, 'cause that's gonna make a huge difference on how much you get back, how much credit you get, and how you're compensated by Medicare.
Every single doctor would have a thousand yard stare. Nobody, no doctor wants to be told how to be a biller or a coder or an auditor. No one went to medical school or nursing school to memorize any of that stuff. They just wanna go back to see their patients. And so what happens when AI, when like this age agentic system in the background that's not even in your face, is actually taking care of all that work that nobody wanted to do in the first place.
That's really what that reasoning engine's all about. And I think that's where we'll be continually investing, but also, like scaling across more settings. So, in the coming weeks we'll be scaling across inpatient. We've mostly been in the outpatient space, urgent cares, ERs. Now we'll start to get into the inpatient world as well. Wow. Fantastic. Andy, are we ready for the lightning round? Shiv, are you ready for the lightning round? Yeah, I think so. Awesome. So, let's hop into it.
We rewrote some of these on the side based on your intro. Uh, 'cause I think your background, it'll be fun to get your take on them. So, do you think things created by AI can be considered art? Yeah, absolutely. Could you give us a little, give us a little, I mean, so, so brevity is the soul of wit, and you nailed it by that metric. Uh, but give us a little bit more there. I, I, I think art is as much about who made it as the impact that it can create. And how it can affect you as a human.
And so, if something generated by AI can impact me in a very human way, then you know, I would consider that art personally. Do you think that art is, I'm gonna get way? Oh geez. I don't know that I'm like the right guy for all these. No, no, no. That's right. There's do you, most people are reacting that there's a recognition of a shared experience or something in art sometimes. So, so maybe, I don't know if, if AI, like, do you think.
Is there something about an aggregate shared experience that a generative model is conveying that's still recognizable as a shared experience? Or some people would say that it's just vacuous. Again, way outta my lane here. Yeah, I don't know if there's anything to that. I think there is something to that. There, there are like some philosophers out there that talk about an aura around art that's been created by a human. And I don't know what that means.
What I do know though is that so many of my favorite artists are like increasingly leveraging technology or they've always leveraged technology. We had a company conference last year called The Conversation where we were explicitly just trying to have the weirdest, most inspiring kind of firesides we could imagine. Like just put people together that maybe shouldn't be talking. And I got to have a fireside with Rick Rubin. I don't know if you know who Rick Rubin is, and it was awesome.
The vibe master legendary. Legendary music producer, right? There we go. Here we go. Yeah. We have a company value. You have to taste good things, to have good taste. And I think like now more than ever, like taste is a really big part of how you build defensibility and moat in, in the AI world. It's like sensibilities mean so much. My twin 8-year-old boys are like vibe coding right now in Cursor. They just got through a hundred days of Python and Replit and they don't have good taste.
It's very clear like they're gonna have to create, figure that out over time. But Rick made this point around the wah-wah pedal, interestingly enough, and how that was like really radical technology at one point and very few people wanted to adopt it. But then this sort of like guitarist who was not well known at the time figured out that he could express himself in a, like a totally differentiated way. And then he changed his name to Jimi Hendrix and the rest is history.
And I think that more people will figure out that like, AI is a superpower that we can all leverage in our daily lives. I totally agree, and I also totally agree about the importance or the growing importance of taste. I think it's always been that very, very key factor. But I think, in this era where you can produce so much, so easily. Totally. Taste, and taste and selection is so much more important. Alright. I have no idea what you're gonna say to this next lightning round question.
I've been trying to simulate and trying to guess, but I truly have no idea. I feel like I just built it up too much. It's fine. But what is your favorite novel? Oh man. Um, that's, that's a, that's a hard one. I guess like different, different novels for different reasons. I don't, there's not one novel. The book that like maybe had the most profound impact on me growing up was Autobiography of a Yogi.
I don't know if you've ever read that book, but for anyone who has a tendency towards magical thinking in any way, shape, or form, it's a pretty profound book. I feel slightly contrived mentioning that book because apparently it was one of Steve Jobs's favorite books. And I don't mean to like bring up a book that like tech companies or Silicon Valley folks, might revere for whatever reason. But certainly growing up.
I was born in Pittsburgh, but I lived in India from age nine to age 15, and that was a book that I was gifted when I got there. And I remember it was, like, just wild. Like I just could not believe what I was reading about these yogis and the things that they could do. And it makes you wanna believe in this other plane, like, of existence and magic, frankly. In college, I'd say like the most important book to me was Franny and Zooey by Salinger.
I just remember reading about one of the character's steadfast dedication to trying to find meaning and she was like repeating this mantra like over and over and over, trying to find meaning and finding there to be sort of some, some profundity in that as well. There's like something meditative about getting obsessed with something that's worth it and she had something that was worth it.
And you know, maybe at this point in my life, I'm working like I did as an intern, and this will be the rest of my life. But this is how the last like eight, nine years have been certainly like 80 hours a week. And that's how so many of us are working at Abridge. And it's worth it. And I think we're doing it because we love it, like, we we're obsessed with it. And that's what that book means to me a little bit, too. Amazing.
Awesome. Thanks, Shiv. So, speaking of different planes of existence in an alternative universe, if you weren't a doctor, an entrepreneur, what would the alternative universe the alternative Shiv be doing for his professional life?
So, in our Slack last night, we were there was like this thread, where we were talking about, I think I had brought this up, but I had brought out how Shaquille O'Neal owns 150 Five Guys, 17 Auntie Annes, 150 car washes, 40 24-hour fitness centers, several Las Vegas nightclubs, and a movie theater. And my argument was that he is way better prepared for like AGI than like any of us. Like he's just like, he's totally there. He is gonna crush it.
But we were trying to like, I was trying to like debate or bring—. Sorry, why does owning a lot of franchises mean he's prepared for AGI? Uh, uh, you think about physical AI and I'm sure Jensen would say physical AI is like coming pretty soon, but like, I wanna believe that it'll take a little bit longer. And a lot of those, like physical experiences, a lot of those like types of jobs and experiences that like humans still want are gonna be around for a beat.
But like, maybe this somehow dovetails with the taste thing, but one craft that I'm excited about is Japanese indigo farming. And it's like this monastic almost lifestyle that these folks have. And it apparently can take hundreds of hours to make one garment. But they're YouTube videos that detail every single step of the process. And like a few of us were like unpacking some of those details in a Slack channel last night.
But if I could go have this sort of monastic, almost Walt Whitman sort of like lifestyle off the beaten path, like, you know, focused on indigo farming, that sounds pretty awesome off the AGI path. Nice. Alright, Shiv, our next question is, will AI and medicine be driven more by computer scientists or clinicians? It is neither, I think it's gonna take all of us. I think that's where the magic's gonna be.
If there's one thing that I, I really will die on the hill of is like domain transfer and, you know, having your foot in multiple spots and like learning from one domain and like pulling those learnings into another. And putting weird combinations together. Like that sort of alchemy is where like really creative stuff happens. And so like, we need that. This has been such a consistent theme of what guests have said on the podcast too.
And it's like the existence of both of those skills within the same mind is very different than sort of two experts coming together. A hundred percent. Who can't actually, who can't actually, there's, it's a latency, right? Yeah. They can't actually, they either can't talk to each other or they can't see the connections because they're, there's no sort of shared vocabulary and ability to like rapidly iterate on what a good idea is from one field to the other.
Yeah. So, it's great to me 'cause it's been like such a consistent theme. Yeah. Of the, of the, yeah. Zack has—Zack, our CTO, he's a professor Carnegie Mellon, but he, he's like obviously fully focused on Abridge right now. But he has a team that he calls the mutants and the mutants are medical professionals who are actually like very serious computer scientists as well, or programmers.
And so those people just have like these multidisciplinary meetings in their own mind, you know, like they can skip so many different steps and oftentimes they also have like taste, too. And so, you're getting these intangibles. Totally agree. Totally agree. Cool. Awesome. Last question. If you could have dinner with one person, dead or alive, who would it be? So, my reflexive take is Yohji Yamamoto. You know that guy, like, like Japanese designer, but he's still alive.
And somehow, I don't know how, but I somehow I wanna meet him while he is still around. But just like a pretty, so he just, to give you a sense like, he's played one note his entire career as a fashion designer, but it's been the right note, and he's stuck to it. Like some ideas you need to hold very, very, very tightly to, you know, and like for a company that ends up being your mission. And if you can be blown with the wind, then who are you anyway?
And like Yohji from a very young age, had this very ascetic aesthetic and he's held so strongly to it that he's like, you know, produced entire, uh, entire fields of people who are just like following in his footsteps. This just continuing with our theme of taste and the aesthetic and the sense of—. Yeah, I'm like the worst person. All important, all important. Talking about all this. It's good. It's good. No, it's fantastic.
I think, I think it's both interesting, but I really don't, I also think it's not accidental in your success as well. So actually, this is the, this is just, so, first of all, you did great with the lightning round. We're done with that. Awesome. And we just have two, sort of final questions for you, Shiv. And there's sort of big picture concluding questions parting thoughts that we'd love your take on.
And I think we, the first one, we've actually talked about this, I think a fair amount, but maybe you can try to distill this. There's a lot of medical students, residents, interns, fellows, young physicians, physicians in-training who listen to the podcast. And I think a lot of them are gonna love this episode because they're gonna love, they're gonna imagine themselves in your shoes and how they could sort of get there, how they could be leading an AI company as a physician.
How they can bring their clinical perspective to AI and how they can get involved. And so maybe you can just distill down for the physician audience, how can physicians like yourself, become leaders of AI companies. I think that what clinicians sometimes might lose track of is, now more than ever before, you can just do things, you can just learn things. You can just go deep.
You can just go build. Like the distance between idea and execution right now, especially like on a prototype for example is smaller, is narrower than ever before. And so, the ability to validate something that could be useful in the world, is easier than ever before. Now scaling and all the other things that like need to come later is a different matter, but like you can get your hands dirty in a way.
Like the barrier to entry on being able to like just go build and pursue what you're really passionate and obsessed about is not there. And so, people should recognize that and go after it. I think even for clinicians like older, like me I think what we sometimes forget is that so much of our training has equipped us for the tech world, for the business world. The example I use a lot is like when I'm in ICU and you're seeing
a patient who's dying. With respect, obviously I'm like abstracting, lessons. But a startup is like a dying patient. You are forced to figure out to react. Very, very quickly. You build a prototype, you deploy it in the ICU, you start pressors. You start up inotropes, you measure are things improving or not, and then that helps you decide whether or not you persevere.
Or you pivot and you just go through these build, measure, learn cycles as a doctor in those acute settings, the same way you go through build, measure, learn cycles as a startup, especially in those early stages and sort of getting outta the ICU is like getting to product market fit. Getting out of the hospital is maybe like free cash flow or profitability, but like there's a lot of like lessons that you can take from your training as a clinician.
All the people you managed, those residents, those trainees, those pharmacists on your team, like how you coordinated care, how you insured people, were also like growing in their roles. Like all of that, I think I've been able to like leverage in some way, shape, or form. But the most important thing is just surround yourself by really, really amazing people and that's where I've been the luckiest, I think the mission has been pure enough.
That we've been able to attract people like Zack and Julia, our COO and Brian, our Chief Commercial, and, Saga our CFO. Like the list just goes on and on and on. And those are the people, honestly, that I'm learning way more from, than vice versa. Cool. Last question, Shiv. So you often hear folks like Eric Topol say that AI will restore the doctor-patient compact. And that presupposes a sort of platonic ideal of what the relationship between patients and doctors was at one time.
But I, I guess I'd like to rephrase it and say like, how will AI change the doctor-patient relationship? You know, I think you, you said you get more eye contact now. Mm-hmm. But what do you think it looks like going forward? I think some of us doctors have nostalgia for something we've never really known.
You know, we have this like idea, this romantic picture of a mid-century doctor who visited you at home, put your, put their hand on your back when you were suffering and knew your grandparents and also your children, and helped you create the Google Maps to whatever your outcome, your chosen outcome, your personalized outcome was, or you wanted it to be. And I think that we're seeing a path to that sort of romantic story, but that's not terminal. Like, that's not the end.
We're clearly getting to a world where it's more like that first patient I saw at Michigan where roles I think will get increasingly blurred lines, and we're both gonna be there for each other. But maybe one thing I can kind of leave you with is just like, why we should revel in the moment that we're in right now. Inside of our company, like HIPAA compliant, you know, sort of communication channel.
We have a channel called Love Stories, and so every day we get positive feedback from users, and I think most of these users are sort of like realizing like maybe that that nostalgia that they have for something that they own, maybe it's like it's coming true right now. So, this comes from a rural health PCP at Tanner Health. And she writes to us, this is some time ago. This is one of my favorites. I was sitting at dinner last week and my son asked me, mommy, why aren't you working right now?
I literally took my phone out and explained to him that Abridge is a new tool that lets mommy come home early and eat dinner with her family. I started to tear up and looked over at my husband, who then said. Mommy's gonna be able to eat dinner with us every night now. That's the moment, right now that, and that's the moment that we've gotta scale as fast as we possibly can. Everybody deserves that and the patient experience is improving as well.
And then I think like, we'll all gradually and probably pretty quickly actually relatively speaking, like move into that new world of blurred lines and where like knowledge is just basically completely democratized. And it's more about, maybe to our other conversation, it's more about taste and sensibilities and like empathy and like humanity. But we have a lot of work to do like in the right now with technology. Cool.
I think that's an excellent place to end, Shiv. So, thanks for joining us today on AI Grand Rounds. Thanks so much, Shiv. That was great. Thank you. It's been a privilege, 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.