Latest / Vanguards of Healthcare by Bloomberg Intelligence / Abridge Moving AI From Scribing to Clinical Workflows

Transcript

Abridge Moving AI From Scribing to Clinical Workflows · Vanguards of Healthcare by Bloomberg Intelligence

0:19
Speaker 1: Welcome to another episode of Bloomberg Intelligence's Vanguards of Healthcare podcast,
0:23
Speaker 1: where we speak with the leaders at the forefront of
0:25
Speaker 1: change in the healthcare industry. My name is Jonathan Palmer,
0:28
Speaker 1: and I'm a healthcare analyst at Boomberg Intelligence, the in
0:31
Speaker 1: house research arm of Bloomberg. I'm thrilled to welcome Zach
0:34
Speaker 1: Lipton for today's episode. Zach is the co founder and
0:37
Speaker 1: chief technology officer of a Bridge. He brings a really
0:40
Speaker 1: unique perspective to the table as both a builder and
0:43
Speaker 1: a scholar.
0:44
Speaker 2: He's currently an.
0:44
Speaker 1: Associate professor at Carnegie Mellon University, where his research focuses
0:48
Speaker 1: on the foundations and impacts of machine learning. Since its founding,
0:52
Speaker 1: a Bridge has become a breakout leader in the generative
0:54
Speaker 1: AI space and they did a very sizable series E
0:57
Speaker 1: round last year and our partnering with some of the
1:00
Speaker 1: largest health systems in the country. Zax, thank you so
1:03
Speaker 1: much for joining us today. Maybe just to kick things off,
1:05
Speaker 1: how would you describe a Bridge as mission? And I
1:08
Speaker 1: don't know, A few sentences.
1:09
Speaker 3: Save time, save money, save lives. I think, yeah, we
1:14
Speaker 3: we see maybe like there there's there's I think a
1:18
Speaker 3: very near term mission, which is about removing sort of
1:20
Speaker 3: like all sucking all the friction out of out of
1:22
Speaker 3: the practice of healthcare, getting doctors above water and out
1:26
Speaker 3: of you know.
1:26
Speaker 2: Sort of like paperwork trap.
1:28
Speaker 3: That's you know, they spend two hours doing clerical work
1:32
Speaker 3: for every one hour of direct patient care. So I think, like,
1:34
Speaker 3: that's that's sort of the been the like the beachhead
1:38
Speaker 3: is like the operational gains. And at the same time,
1:41
Speaker 3: like my background came from, you know, a place of
1:44
Speaker 3: dealing with clinical medical data, working in the methodology of
1:47
Speaker 3: like statistical inference, causal identification, and ultimately, you know, I
1:52
Speaker 3: think the big dream for machine learning and healthcare has
1:54
Speaker 3: always been not just to not just to mitigate paperwork,
1:57
Speaker 3: but also to provide truly no clinical insights. And I
2:01
Speaker 3: think the kinds of partnerships and scientific relationships that we
2:05
Speaker 3: are making with our our partners put us in a
2:08
Speaker 3: position to sort of lead on both sides. And said,
2:12
Speaker 3: you know, so, so I think if we were to
2:13
Speaker 3: stage it sequentially save lives.
2:17
Speaker 1: Well said, that was scinct. Maybe just rewinding a little bit,
2:22
Speaker 1: where did your interest in the field in machine learning
2:24
Speaker 1: and natural language processing really kick off?
2:28
Speaker 3: Yeah, I got a taste in undergrad So I was
2:31
Speaker 3: a musician first, that was my big love. I was
2:34
Speaker 3: a saxophone player. I got to being in the New
2:37
Speaker 3: York area. God is like a young kid to grow up.
2:40
Speaker 3: But I was also had a little bit of like
2:42
Speaker 3: a math science brain, and I think I was always
2:44
Speaker 3: caught between these two worlds. And my way of threatening
2:46
Speaker 3: the needle at the time was to when when you
2:51
Speaker 3: hit a fork in the road, take it. So I
2:54
Speaker 3: I My way of like punching kicking the can down
2:57
Speaker 3: the road was to go to school in New York.
2:59
Speaker 2: So I into.
3:00
Speaker 3: Colombia and I got more of a secular education while
3:04
Speaker 3: getting to be on the music scene. And so so
3:06
Speaker 3: I was kind of like floating in these worlds. I
3:08
Speaker 3: wasn't a computer scientist and I hadn't really programmed much
3:10
Speaker 3: by then, but I sometime, I think sophomore year, was
3:15
Speaker 3: dating an engineering student and I ended up just reading
3:19
Speaker 3: all her textbooks on the side for fun and as
3:21
Speaker 3: it does, a little bit, Yeah, I got a little
3:23
Speaker 3: bit of a taste something felt right about There was
3:27
Speaker 3: something very powerful about about a way of thinking, of
3:32
Speaker 3: a structured way of thinking about problems that came from
3:35
Speaker 3: trying to sort of map them onto a computational model
3:39
Speaker 3: such that you could manipulated algorithmically or build software around it.
3:43
Speaker 3: So that was like my little taste, and I was like,
3:44
Speaker 3: you know, being like, I think that combination like machine learning,
3:47
Speaker 3: really sits at that intersection of you know, math, especially
3:51
Speaker 3: on the side of like probabilistic reasoning and statistics and
3:54
Speaker 3: computer science engineering, and so I had like those two
3:57
Speaker 3: ingredients and I got a little bit of a tape
4:00
Speaker 3: is done.
4:00
Speaker 2: But music was.
4:01
Speaker 3: Still was still my first love, and I spent my
4:04
Speaker 3: early twenties primarily playing the saxophone, and it was actually
4:09
Speaker 3: it was only my life got kind of sidelined for
4:12
Speaker 3: a bit by a health situation, and for me, that
4:17
Speaker 3: experience of going through sort of seeing, you know, health
4:22
Speaker 3: care is so weird. I think people want to think
4:24
Speaker 3: it's like healthcare is great or healthcare is terrible, but
4:26
Speaker 3: it's really a mix. There's things that we understand really well,
4:28
Speaker 3: and there's places where of like extreme competence, and there's
4:33
Speaker 3: also there's there's areas where we're guessing it's like kicking.
4:36
Speaker 2: The vending machine.
4:37
Speaker 3: I think autoimmunity is that kind of situation where we
4:40
Speaker 3: don't really have a very strong mechanistic understanding. I think
4:43
Speaker 3: it's improving a little bit in recent years. But I
4:45
Speaker 3: think as a patient navigating a system, realizing that like
4:49
Speaker 3: you're in this like nebulous cohort of like hundreds of
4:51
Speaker 3: thousands of students that sit somewhere between the diagnostic lines,
4:55
Speaker 3: and you're trying to triangulate, like what is the state
4:58
Speaker 3: of medicine based on supports of two patients here, ten
5:01
Speaker 3: patients there. And there's a little bit feeling of like,
5:04
Speaker 3: we have to be doing something smarter and and I
5:07
Speaker 3: tried to put my finger on like what what is
5:09
Speaker 3: it that we could be doing more intelligently? And I
5:12
Speaker 3: think there was this mix of Part of the problem
5:14
Speaker 3: was like just people didn't even have access to the
5:17
Speaker 3: right data. Part of the problem is that the data,
5:20
Speaker 3: the right data wasn't even always being captured. Part of
5:23
Speaker 3: the problem was that even if you had all the
5:25
Speaker 3: data from all the systems in the world, there's still
5:27
Speaker 3: like some fundamental mathematical problems of how do you go
5:30
Speaker 3: from longitudinal data to actionable you know, uh sort of
5:36
Speaker 3: recommendations that that would actually improve outcomes for people. I mean,
5:40
Speaker 3: that's that's a fundamentally hard problem that we call it
5:43
Speaker 3: causal inference and academia. But you know, to be able
5:48
Speaker 3: to produce action guiding insights without you know, being completely
5:54
Speaker 3: bobblenacked only on you know, isolated like randomized control trials.
5:58
Speaker 3: To be able to actually harness the entirety of longitudinal
6:00
Speaker 3: data to produce those kinds of insights requires a set
6:04
Speaker 3: of strategies that you know, people in statistics and public
6:08
Speaker 3: health and econometrics have been studying for years. So for me,
6:12
Speaker 3: I think I came together with this like wild idea
6:15
Speaker 3: in twenty twelve, which was, hey, like, if I want
6:19
Speaker 3: to have an impact on healthcare, like the route to
6:22
Speaker 3: do it would be to do a PhD in machine learning,
6:24
Speaker 3: which you know, maybe in twenty twenty six seems like
6:27
Speaker 3: extremely obvious, but at the time it was it was
6:30
Speaker 3: a quaint community. There was there was a little workshop
6:33
Speaker 3: that I actually ended up I'm still deep in that community.
6:36
Speaker 3: But there was a workshop that was called Meaningful Use
6:40
Speaker 3: of Complex Medical Data, which actually gave you the lovely
6:43
Speaker 3: acronym MUCKMED. And this is like it wasn't even a
6:46
Speaker 3: publishing conference, it was it was a group of like
6:49
Speaker 3: I don't know, seventy people that would get together like
6:51
Speaker 3: once a year at Children's Hospital Los Angeles and have
6:54
Speaker 3: like poster presentations and like that was the community. But
6:58
Speaker 3: that was that was kind of my idea, was like
7:00
Speaker 3: machine learning was my way in to have an impact
7:02
Speaker 3: in healthcare, and I ended up backing up very far
7:05
Speaker 3: from like the kind of impacts I wanted to have
7:08
Speaker 3: into the theory of machine learning and neural network methodology.
7:13
Speaker 3: Deep learning was exploding at the time. I got to
7:15
Speaker 3: be like kind of a core member of the community
7:17
Speaker 3: and build a whole career as an academic in machine learning.
7:20
Speaker 3: But I always kept that threat in healthcare, and it
7:23
Speaker 3: was always kind of like working backwards from that personal
7:26
Speaker 3: experience as being a patient wanting the wanting system to
7:30
Speaker 3: be better.
7:31
Speaker 1: Well, that's great. I mean, if we work back from
7:33
Speaker 1: that on that arc from twenty to twelve to now,
7:36
Speaker 1: what have you seen in your career that had really
7:39
Speaker 1: been the defining moments in your journey, you know, as
7:42
Speaker 1: an academic, as a business leader, and really maybe wind
7:46
Speaker 1: up with how you got to a bridge. Sure, so
7:49
Speaker 1: I was like sitting on two sides. I had one
7:52
Speaker 1: office in the medical school and one in the School
7:53
Speaker 1: of Computer Science. I was funded by a grant from
7:56
Speaker 1: the NIAH that came through the medical school. My advisor
7:59
Speaker 1: was a user of computer science and machine learning, and I.
8:03
Speaker 2: Spent a lot of that early.
8:05
Speaker 3: You know, this like big opacal moment was twenty twelve,
8:09
Speaker 3: there was a paper published out of by Jeff Hinchin's
8:12
Speaker 3: lab in University of Toronto that was a breakthrough result
8:14
Speaker 3: using neural networks for image classification. Very quickly folks started
8:19
Speaker 3: catching on and other folks were able to figure out
8:21
Speaker 3: how to train neural networks for various image problems, and
8:25
Speaker 3: they started taking off, and there was this there was
8:27
Speaker 3: still this general feeling in the field that like neural
8:29
Speaker 3: networks were a backwater and kind of you know, a
8:34
Speaker 3: not so rigorous area that was like more appealing because
8:38
Speaker 3: of the storytelling then because of practicability, And suddenly there
8:42
Speaker 3: was there was this area and inside computer vision, suddenly
8:44
Speaker 3: neural networks were really working, and I think it really
8:47
Speaker 3: helped that. You know, I'd been a musician, so I
8:48
Speaker 3: actually came into PhD with almost no prior commitments about
8:52
Speaker 3: methodology and also no skills that I had already invested
8:55
Speaker 3: in such that like I was.
8:57
Speaker 2: I wasn't like resistant.
8:59
Speaker 3: I think a lot of peopleeople who hadn't spend a
9:01
Speaker 3: lot of time becoming like learning a lot of mathematics
9:03
Speaker 3: in order to reason about like Hilbert's spaces, because they
9:07
Speaker 3: were working on kernel machines, where like, oh, this neural
9:09
Speaker 3: network stuff is pretty pretty ad hoc, you know it,
9:14
Speaker 3: it's pretty scruffy. I think like I saw that wave
9:19
Speaker 3: coming clear eyed because I wasn't deep in any rabbit hole.
9:23
Speaker 3: And it hit image recognition, then it hit speech recognition,
9:27
Speaker 3: then it started hitting natural language processing. And I think
9:30
Speaker 3: a big moment for me personally was around twenty fourteen
9:33
Speaker 3: twenty fifteen recognizing, hey, the source of problems that this was.
9:38
Speaker 3: These were neural sequence models basically the forerunners or the
9:44
Speaker 3: ancestor of what became like the transformer is now the
9:46
Speaker 3: dominant frequence architecture. At the time, it was the recurrent
9:49
Speaker 3: neural network that was the dominant workhorse. And I saw
9:53
Speaker 3: that there was there was The kinds of problems we
9:55
Speaker 3: had in language were very similar to the kind of
9:57
Speaker 3: problems we had with medical record data. You can't just
10:01
Speaker 3: it's not like an image where you could just crop
10:02
Speaker 3: every image to be the same size and just feed
10:04
Speaker 3: it into the model. You sort of have like a
10:06
Speaker 3: long sequence, and some documents are one sentence and some
10:09
Speaker 3: are one paragraph, and some are two pages, and the
10:11
Speaker 3: order in which the words occur kind of matters. You've
10:14
Speaker 3: got this like big sequence of events, and I saw
10:18
Speaker 3: the like sequence of events whether it was like lab
10:21
Speaker 3: readings or medications administered or diagnoses applied to a medical
10:26
Speaker 3: record kind of has a similar structure. Of the order matters.
10:30
Speaker 3: So some could be long, some could be short. There
10:31
Speaker 3: were some additional problems, like you might go long periods
10:34
Speaker 3: of time without an observation. But for me, I got
10:37
Speaker 3: to basically be like the first paper that applied like
10:41
Speaker 3: modern neural sequence models to electronic health record data. And
10:44
Speaker 3: I got that to be like my first paper in PhD.
10:48
Speaker 3: That was like a huge moment, and I think I
10:50
Speaker 3: was just very lucky and to sort of point in
10:54
Speaker 3: the right direction, sitting at the intersection of the right
10:56
Speaker 3: two fields at the right time, and just to be
10:58
Speaker 3: in the field at a moment of great change.
11:00
Speaker 2: You know. I think when the field is possified.
11:02
Speaker 3: And plateaued at that point, it's really run by the
11:06
Speaker 3: people with the deepest kind of like tribal knowledge and
11:09
Speaker 3: the moment when like a new lane opens up, it's
11:12
Speaker 3: actually the scrappiest people who don't know anything and don't
11:15
Speaker 3: have any commitment that get in. And so I got
11:18
Speaker 3: to get in very early in my reck my PhD,
11:20
Speaker 3: and that kind of opened a path to sort of
11:24
Speaker 3: have a voice and some kind of like leadership in
11:26
Speaker 3: this community and other opportunities to collaborate. Because if you
11:29
Speaker 3: were looking around in twenty fifteen saying, hey, I want
11:31
Speaker 3: to do deep learning and healthcare, there were like three
11:33
Speaker 3: people you could talk to, and so that was that
11:35
Speaker 3: was a great place for me personally. Maybe I'll fast
11:39
Speaker 3: forward like all the way to come into a bridge. Sure,
11:41
Speaker 3: let's do it, you know. So Okay, So I told
11:43
Speaker 3: you about that muckmed which had now graduated and become
11:46
Speaker 3: like a publishing conference machine learning for health care. So
11:48
Speaker 3: I was there, and I was flying to Pittsburgh to
11:54
Speaker 3: I just I was about three and a half years
11:55
Speaker 3: into PhD and I had just been offered a faculty
11:58
Speaker 3: position to be a professor at cry Melon. So I'm
12:01
Speaker 3: like flying there to like, I guess after they offer
12:04
Speaker 3: you a position, they fly you out to go on
12:06
Speaker 3: like a house hunting trip.
12:07
Speaker 2: It's like to figure out where you're going to live.
12:09
Speaker 3: And I happened to be sitting on this flight from
12:13
Speaker 3: Boston to Pittsburgh, and I guess, like I was younger
12:16
Speaker 3: and a little bit more you know, I don't know,
12:20
Speaker 3: less grizzled, like maybe now on a planet, just put
12:22
Speaker 3: my headphones on and mind my own business. But I
12:24
Speaker 3: used to make a little bit of a sport of
12:25
Speaker 3: like chatting up the person next to me. And I
12:28
Speaker 3: got to know this guy sitting next to me, and
12:30
Speaker 3: it turned out and he was the chief of medicine
12:32
Speaker 3: at UPMC. So yeah, fortuitous, right, Like there were some
12:37
Speaker 3: moments in the history of this company and his career
12:39
Speaker 3: for me that I believe in free will, but I
12:41
Speaker 3: also I also believe in fate. And so he was
12:45
Speaker 3: this guy, and I was, you know, despite being like
12:48
Speaker 3: a very important person, what was just down to chat
12:51
Speaker 3: And you know, I think at that moment in twenty seventeen,
12:54
Speaker 3: now deep learning was really especially in computer vision, becoming mature,
12:58
Speaker 3: you know, nothing like it is now, but in the
13:00
Speaker 3: sense that they were already front page New York Times
13:03
Speaker 3: articles about what was happening in neural networks research. There
13:06
Speaker 3: was already a little bit of an arms race, and
13:08
Speaker 3: leading AI researchers were already saying things like, basically, this
13:11
Speaker 3: is the end of cardiology. I mean, of course, I
13:14
Speaker 3: of radiology, right because of that ability to sort of
13:18
Speaker 3: recognize findings and images and to classify them.
13:22
Speaker 2: I think I think the.
13:22
Speaker 3: AI community underestimated the complexity of practicing medicine. But the
13:26
Speaker 3: point is like it was coming as like a major
13:29
Speaker 3: title wave that was some way or another going to
13:32
Speaker 3: impact healthcare, and I think the hospital systems were caught
13:36
Speaker 3: flat footed. You know, they hadn't really been on the
13:39
Speaker 3: leading edge of the trend. They were like, oh wait, wait,
13:42
Speaker 3: you know, they're suddenly like caring from their board like
13:44
Speaker 3: what's your AI strategy?
13:45
Speaker 2: And they're like what so.
13:49
Speaker 3: So I had this moment of like meanwhile, like I
13:53
Speaker 3: was someone who's like my life's purpose, you know, or
13:56
Speaker 3: like my career mission was to have an impact in healthcare,
13:59
Speaker 3: and I had just accepted a faculty position at a
14:02
Speaker 3: university that has no medical school. Like I'm sitting there
14:05
Speaker 3: being like all right, like I need to find a
14:07
Speaker 3: way to turn this into a collaboration. And I think
14:10
Speaker 3: from his side, he was probably thinking like, oh wait,
14:13
Speaker 3: there's a there's a new professor at CMU who's excited
14:15
Speaker 3: about healthcare, you know, in the machine learning department, like this,
14:18
Speaker 3: this could be an opportunity to do something interesting in Pittsburgh.
14:21
Speaker 3: So Steve actually invited me out to a happy hour.
14:25
Speaker 3: At first, he drove me home to my Airbnb, and
14:28
Speaker 3: then he invited me out the next day to a
14:29
Speaker 3: happy hour and he kind of like rolled in thick.
14:31
Speaker 3: So I'm like this like third year PhD student, And
14:34
Speaker 3: he shows up for this happy hour in a strip
14:37
Speaker 3: mall in Pittsburgh with the CFO of UPMC, the CTO
14:42
Speaker 3: of UPMC, and with the executive vice president of UPMC Enterprises.
14:46
Speaker 3: And it was that person which is basically like this
14:49
Speaker 3: sort of like innovation arm of the hospital system. That
14:53
Speaker 3: was Shivrao, who's my co founder in a bridge. That's
14:56
Speaker 3: an amazing that's an amazing origin story. So then what
15:00
Speaker 3: did you guys decide? I mean, you make the introduction,
15:02
Speaker 3: you get you get to get the introduction to Shive.
15:04
Speaker 3: What did you guys first start working on? I mean,
15:06
Speaker 3: were you kicking around different ideas? Did you land on
15:10
Speaker 3: what is now a bridge pretty quickly.
15:13
Speaker 2: You know, we actually did. And this is a wild thing.
15:16
Speaker 3: I think, you know, it sounds arrogant or something, and
15:18
Speaker 3: you're expecting to hear like, no, we wandered in the
15:20
Speaker 3: desert for a while. We did wander in the desert
15:23
Speaker 3: for a while, but like first we saw clearly what
15:25
Speaker 3: would be a bridge, and then we wandered in the desert.
15:28
Speaker 3: So ship has this incredible ability despite like you know,
15:31
Speaker 3: he's he's not technical at all himself, but he has
15:34
Speaker 3: this remarkable ability to like see the forest and to
15:39
Speaker 3: connect a lot of dots across a lot of places.
15:42
Speaker 3: And so he he already had landed in a couple places.
15:45
Speaker 3: One was that like the burnout crisis and health care
15:49
Speaker 3: was ascendant, Like this was something that you could already
15:52
Speaker 3: see in two thousand January twenty eighteen, or when we
15:54
Speaker 3: were having this conversation. Yeah, maybe around then, like did
15:57
Speaker 3: this was maybe like had already risen to be like
16:00
Speaker 3: already number four, but it was going to be priority
16:02
Speaker 3: number one. Okay, Like this was only getting worse. Doctors
16:05
Speaker 3: were starting to leave the profession. Med students were starting
16:08
Speaker 3: to say that they didn't they didn't even plan to practice,
16:11
Speaker 3: and The other side was that I think shift and
16:15
Speaker 3: I had maybe from different angles, him as an investor
16:18
Speaker 3: on the sort of like venture side of the hospital system,
16:21
Speaker 3: me as like an academic watching some of my peers
16:24
Speaker 3: try to start companies. Was that we had watched a
16:26
Speaker 3: lot of companies try to get into the get into
16:29
Speaker 3: the workflow and fail. And I think they never really
16:31
Speaker 3: thought deeply about the workflow or about the economic structure
16:35
Speaker 3: of the health system. So I think they were just saying, like, hey,
16:38
Speaker 3: what if we could predict why from X, and that
16:40
Speaker 3: would be a cool model, right, and the NSR having
16:42
Speaker 3: a strategy for like and who's going to buy it?
16:44
Speaker 3: And when is the doctor going to see it? And
16:46
Speaker 3: in what surface area? And how do you how do
16:50
Speaker 3: you find a way into the doctor's hands, how do
16:51
Speaker 3: you find a way to to their to their eyes.
16:54
Speaker 3: And we set it quickly on this idea that like
16:56
Speaker 3: it's all about the conversation, Like the conversations the entry.
17:00
Speaker 3: This is the like great missed opportunity in healthcare. This
17:03
Speaker 3: is where the whole story happens. It's where all the
17:07
Speaker 3: details are shared by the patient, It's where all the
17:09
Speaker 3: medical reasoning takes place. Like this is the juiciest place
17:12
Speaker 3: to play, and right now, there's no like technology wasn't
17:17
Speaker 3: additive in the conversation. It was like, I was sorry,
17:19
Speaker 3: were subtractive. It was like degrading the quality of conversation.
17:23
Speaker 3: Doctors were sitting there with their back to the patient
17:25
Speaker 3: because they were pre writing their paperwork, not sitting there
17:29
Speaker 3: being aided by technology that was doing their paperwork for them,
17:33
Speaker 3: providing context, helping them to surface important details from from
17:38
Speaker 3: past records, from the medical literature, Like none of that
17:42
Speaker 3: was happening. So we kind of sat on this idea
17:44
Speaker 3: that like it was sort of intelligence at the point
17:47
Speaker 3: of care, like this was the place to be be
17:49
Speaker 3: inside the conversation. Only problem is that the technology in
17:52
Speaker 3: twenty eighteen was was nothing like what we have now.
17:55
Speaker 3: For reference, the largest language models at the time were like,
17:58
Speaker 3: you know, quaint by today's standards, like a thousand times
18:02
Speaker 3: smaller two thousand times smaller. They tended to have a
18:05
Speaker 3: look back of They could process about one hundred tokens
18:07
Speaker 3: of text, so it was like they were looking at
18:08
Speaker 3: about a paragraph, generously a small paragraph. So you had
18:12
Speaker 3: to figure out how am I going to summarize a
18:14
Speaker 3: four thousand word conversation or a ten thousand word conversation,
18:18
Speaker 3: let alone a ten thousand word conversation plus the entire
18:21
Speaker 3: medical record of like supporting context. How am I going
18:24
Speaker 3: to like transform that into medical documents using the technology
18:28
Speaker 3: of twenty eighteen? And so I kind of was like, well,
18:31
Speaker 3: sign me up for that problem, because that's that's the
18:34
Speaker 3: machine learning problem, that's the AI problem. And I was like,
18:37
Speaker 3: I'll work on that. And not only that, but like
18:39
Speaker 3: I kind of took a posture of instead of optimizing
18:43
Speaker 3: for like you know, you know, like sort of just
18:46
Speaker 3: like personal stake, I was sort of like, support my lab.
18:49
Speaker 3: I'm going to work on this is this is I
18:50
Speaker 3: don't know if it's going to take one year or
18:52
Speaker 3: if this is going to take twenty years, but probably
18:54
Speaker 3: not one year. Support me to work on this as
18:56
Speaker 3: an open ended research problem in the university, and support
18:59
Speaker 3: me to pay my PhD students. And I'm going to
19:02
Speaker 3: work on this like the R and D side of this.
19:05
Speaker 3: And in the meanwhile, the company experimented with all kinds
19:08
Speaker 3: of ways of getting to market. A bridge launched a
19:12
Speaker 3: patient facing application. But the thing is that your hands
19:15
Speaker 3: were tied behind your back about like it's much more
19:17
Speaker 3: what kind of AI features you could ship directly to
19:19
Speaker 3: the patient at the time without without a doctor in
19:21
Speaker 3: the loop to to you know, was the wherewithal to
19:25
Speaker 3: be the filter to verify the accuracy of the content.
19:29
Speaker 3: But that was like, you know, like from a pure
19:31
Speaker 3: distribution angles like, well, what value could we give to them,
19:33
Speaker 3: even if it was lighter weight On the AI side.
19:36
Speaker 3: During pandemic there was an attempt to build telemedicine platform.
19:40
Speaker 3: But that whole time on the on the R and
19:43
Speaker 3: D side, I was chugging along working on the end
19:45
Speaker 3: to end AI scribing side with my students, and we
19:49
Speaker 3: actually got it to work surprisingly well. And as this
19:54
Speaker 3: sort of knowledge in the commons and open source capabilities
19:58
Speaker 3: got much stronger, and as we worked with a bridge,
20:01
Speaker 3: you know, we're developing more expansive data sets, we were
20:04
Speaker 3: able to keep improving the quality to the point that
20:07
Speaker 3: in twenty twenty two it started to become obvious that hey,
20:11
Speaker 3: like twenty eighteen, this was a pipe train. In twenty
20:13
Speaker 3: twenty two is this is going to happen in the
20:16
Speaker 3: next year. And the question is like do we go
20:19
Speaker 3: all in or do we allow or do we like
20:23
Speaker 3: you know, like let someone else like kind of really
20:25
Speaker 3: get there. So I think we were a company that
20:27
Speaker 3: was almost founded too soon, but really engaged at the
20:33
Speaker 3: right time. So January twenty twenty three around there, I
20:36
Speaker 3: ended up stepping more fully into day to day operator role,
20:40
Speaker 3: took over as chief technology officer at the company. We
20:43
Speaker 3: had maybe like five technical staff inside the company at
20:46
Speaker 3: the time. Now we're a four hundred and fifty person company,
20:50
Speaker 3: and so over those over that first year we completely
20:56
Speaker 3: remade the AI brains and the sort of product experience.
21:00
Speaker 3: We got to market with the first enterprise product of Virginia,
21:03
Speaker 3: I product in the history of healthcareacter my knowledge, at least,
21:05
Speaker 3: the first successful one came out around April May with
21:08
Speaker 3: a closed contract, and a month later I was like
21:13
Speaker 3: we were raising a Series B. By the end of
21:16
Speaker 3: twenty twenty three, we had I think eleven enterprise partners
21:21
Speaker 3: and raised a Series C, including with a check from
21:25
Speaker 3: Nvidia and Jensen, which involved at a pretty pretty pitch,
21:29
Speaker 3: and did twenty twenty four with over one hundred hospital
21:33
Speaker 3: system partners and ended twenty twenty five with over two
21:36
Speaker 3: hundred hospital system partners that represent among them over forty
21:41
Speaker 3: percent of the doctors in the country.
21:43
Speaker 1: That's amazing. Congratulations. Please tell me that UPMC was the first.
21:47
Speaker 3: Or close to the first, close to the first, they
21:51
Speaker 3: were the I think they might have been the first
21:53
Speaker 3: for some of some stages along that journey.
21:57
Speaker 1: Well maybe for some of the people who work in
21:59
Speaker 1: healthcare or or don't work in healthcare. For the lay person,
22:02
Speaker 1: can you just talk about what a scribe does and
22:05
Speaker 1: you know, the AI scribe versus maybe traditional describing services
22:09
Speaker 1: that that acquinition would use. Sure, maybe start with the
22:13
Speaker 1: job to be done. Yeah, that's even better.
22:15
Speaker 3: Doctors have a ton of clerical work that they have
22:19
Speaker 3: to do to document the visit that they have. And it's
22:22
Speaker 3: not just about documenting. It's also like the actions that
22:24
Speaker 3: they take have to be So maybe this the big
22:28
Speaker 3: gorilla is notes. You know, I'm I'm a PhD advisor
22:32
Speaker 3: as a recently tenured professor, and you know when I
22:37
Speaker 3: when I work with a student, I work with them
22:40
Speaker 3: for six years. Every six months I have to write
22:43
Speaker 3: like about a page and a half of like a
22:44
Speaker 3: report of how they're doing. Doctors see a patient for
22:47
Speaker 3: fifteen minutes and they have to write like a two
22:49
Speaker 3: and a half page report about what is going on
22:51
Speaker 3: with a patient and I complain about my PaperWorks. It's
22:55
Speaker 3: a colossal burden. They have to document the patient's story.
22:59
Speaker 3: It's called like a history of present illness. They have
23:02
Speaker 3: to document in a very structured form, their assessment and
23:06
Speaker 3: plan for each problem that the patient has. They have
23:08
Speaker 3: to document a family history, the social history, the past,
23:12
Speaker 3: medical history, the family that I think I already said family.
23:15
Speaker 3: They have to document a complete inventory of all the
23:19
Speaker 3: medications that the patient is on and what if anything
23:22
Speaker 3: has changed. They have to write patient facing after visit
23:26
Speaker 3: summary that is duplicative in its contents, but reformulated specifically
23:30
Speaker 3: to be the subset of things that are like don't
23:33
Speaker 3: forget this, and you must take these action items. So
23:37
Speaker 3: those are part.
23:37
Speaker 2: Of the plan.
23:38
Speaker 3: But they're they're they're they're written a little bit differently
23:40
Speaker 3: in a slightly different maybe like tone, and assume the
23:43
Speaker 3: level of like reading level or understanding of medical jargon.
23:47
Speaker 3: So doctors they also have to enter the medication orders.
23:49
Speaker 3: They have to they have to fix billing codes to
23:54
Speaker 3: particular problems, and it have to be if they attach
23:57
Speaker 3: a billing code because they're billing for treatment of a problem,
23:59
Speaker 3: they better have documented details AGMC.
24:02
Speaker 2: For each one.
24:03
Speaker 3: So it's this colossal tangle of paperwork that the doctors
24:07
Speaker 3: are accountable for. The idea of a scribe, a human
24:10
Speaker 3: scribe in the first place, that you have someone in
24:11
Speaker 3: the office who is basically like your note taker, and
24:15
Speaker 3: while you're tending to the patient and thinking a little
24:17
Speaker 3: bit more first principles, they are busily writing out all
24:21
Speaker 3: their notes and then digitizing them for you, and very often,
24:24
Speaker 3: like in a teaching hospital, they might make you know,
24:27
Speaker 3: when possible, make some of the trainees do some of
24:29
Speaker 3: describing for the attending and the more senior doctors. Remote
24:34
Speaker 3: scribing services started as a sort of very very often
24:38
Speaker 3: like a kind of like record the audio with a
24:41
Speaker 3: conversation and there's a human listening in a call center
24:44
Speaker 3: and basically trying to do the same thing, often with
24:47
Speaker 3: a much longer turnaround for.
24:48
Speaker 2: The notes, and maybe with lower quality.
24:50
Speaker 3: They also have less signal because they have the audio
24:52
Speaker 3: signal but they didn't see what was going on. They
24:56
Speaker 3: these services, they just wouldn't scale doctors who could afford them.
25:00
Speaker 3: You know, you're you're very well paid, like you know,
25:02
Speaker 3: your kind of like highest compensation specialties. You know, to
25:06
Speaker 3: to to get that yoke of of clerical work off
25:10
Speaker 3: their nat It might have been worth it, but that
25:13
Speaker 3: the line share of doctors couldn't afford. We're talking like
25:15
Speaker 3: twenty thousand dollars per doctor per year. It was a
25:18
Speaker 3: kind of extreme expense for these sort of like remote
25:20
Speaker 3: human and the loop scribing services, so they weren't The
25:23
Speaker 3: vast majority of doctors never had the benefit of a scribe,
25:27
Speaker 3: even though like for the subset who did, it was
25:28
Speaker 3: kind of a big business. AI scribing is just that
25:31
Speaker 3: it is. It is the idea that an AI agent
25:35
Speaker 3: that is we could call we call the genre here
25:39
Speaker 3: ambient listening, but that that's sort of in the background
25:41
Speaker 3: listening because like the doctor has a device, you know,
25:44
Speaker 3: usually their phone that has a secure application on it,
25:47
Speaker 3: and they say they consent the patient, this is my
25:49
Speaker 3: AI note taker. Would you be okay if I recorded
25:52
Speaker 3: this conversation, so you know, the AI could could take
25:55
Speaker 3: the notes and I could focus on focus on you patient,
25:59
Speaker 3: you know, nine nine point you know something percent of
26:01
Speaker 3: the time says yes, they hit record.
26:03
Speaker 2: And it is on the basis.
26:06
Speaker 3: Of the raw audio together with the information that we
26:10
Speaker 3: are able to the relevant context that we could pull
26:13
Speaker 3: from the EHR, the AI agent is able to transcribe
26:16
Speaker 3: the audio structure that content, collate it with what we
26:20
Speaker 3: already know from the problem list, from prior notes, and
26:24
Speaker 3: synthesize it all into a kind of comprehensive first draft
26:27
Speaker 3: of the documentation, so that the doctor kind of gets
26:30
Speaker 3: to like lift up in the stack from being like
26:34
Speaker 3: you know, on on the front lines to writing copy
26:36
Speaker 3: to being more of an editor.
26:39
Speaker 1: Right, they can they can practice at the top of
26:41
Speaker 1: their license versus doing all that quirical work exactly. So
26:44
Speaker 1: what how I don't know how you define it, but
26:47
Speaker 1: how accurate is it when when you go through and
26:51
Speaker 1: you know, piece all these pieces of structured and nonstructured
26:55
Speaker 1: data together, is there a metric that you look at
26:57
Speaker 1: that says, I don't know, we're x percent accurate with
27:00
Speaker 1: what was actually discussed.
27:02
Speaker 2: In the room.
27:03
Speaker 3: Yeah, it's a great question. You know when you have
27:06
Speaker 3: a very straightforward classification task, like I have a test
27:10
Speaker 3: for a blood task for HIV, and I want to say,
27:12
Speaker 3: how accurate is it? Just straightforward definition of accuracy which
27:15
Speaker 3: is out of end cases, what fraction are classified correctly?
27:18
Speaker 2: Sure?
27:18
Speaker 3: And you know you have other majricyclet you say, what
27:21
Speaker 3: is the recall out of out of every case that exists?
27:23
Speaker 3: What fraction did I call positive? Or what is the
27:25
Speaker 3: precision out of everyone that I called positive? What fraction
27:27
Speaker 3: were actually positive? And these three numbers kind of tell
27:29
Speaker 3: you everything you need to know. They they identify something
27:32
Speaker 3: called as called the confusion matrix, and it sells you
27:35
Speaker 3: everything you need to know about the performance of the classifier.
27:38
Speaker 3: I should caveat that even then things could get more
27:41
Speaker 3: complicated and interesting ways, because you could say, oh, you
27:43
Speaker 3: tried it on this population, but do those numbers hold
27:46
Speaker 3: on this different you know, subset of the population. Are
27:49
Speaker 3: you sure that it works, you know, with the same
27:51
Speaker 3: accuracy over here? With something in the generative AI space,
27:57
Speaker 3: evaluation is a little bit more like evaluate the outputs
28:00
Speaker 3: of labor, okay. And so for example, if I were
28:04
Speaker 3: to ask you for like a particular like you know,
28:08
Speaker 3: young journalist, and from like how accurate are they? And
28:13
Speaker 3: you'd have a whole set of like conceptually fraught questions
28:16
Speaker 3: of like well, I could I could try to say, well,
28:19
Speaker 3: like how many out of all the claims they make,
28:23
Speaker 3: what fraction of those claims are are correct? But then
28:27
Speaker 3: even like is there an unambiguous way of mapping from
28:31
Speaker 3: an article to a collection of claims.
28:33
Speaker 2: So you're gonna have.
28:34
Speaker 3: To like like you have a like the gap between
28:37
Speaker 3: the construct and the measurement is much larger maybe than
28:40
Speaker 3: it would be for for you know, hot dog not
28:43
Speaker 3: hot dog, or for for a blood test. For us,
28:45
Speaker 3: it's very similar. So so this is the the the
28:49
Speaker 3: pursuit of an entire team within a bridge. And you
28:51
Speaker 3: also see within the big labs like Anthropic or then
28:54
Speaker 3: opening out, you have entire teams that are devoted to
28:56
Speaker 3: accuracy and or we don't say it's sort just accursus
28:59
Speaker 3: Exaccuracy has a literal meaning as like a simple metric
29:02
Speaker 3: for forstem, but more broadly to evaluation. And when we evaluate,
29:08
Speaker 3: what we have is not just a single metric. What
29:10
Speaker 3: we have is a broad sensorium. You know, the leader
29:14
Speaker 3: board doesn't have one column and has you know, fifty columns,
29:17
Speaker 3: and contributing to that are all kinds of measures. We
29:21
Speaker 3: care about of every kind of pertinent medical detail that
29:25
Speaker 3: is present in the transcript, what fraction are captured in
29:28
Speaker 3: the note. We also care of the pertinent details in
29:31
Speaker 3: the audio, which fraction are captured in the transcript. We
29:36
Speaker 3: also that that would get at some notion of like
29:38
Speaker 3: recall right is something and analogous like you don't want
29:42
Speaker 3: to have omissions? And we also everything in the note
29:46
Speaker 3: what is substantiated in the transcript? So we have a
29:48
Speaker 3: bunch of metrics that we create around factuality. And the
29:52
Speaker 3: question is, well, well, who's in a position to assess that?
29:54
Speaker 3: Because it's not as straightforward as like was positive here
29:57
Speaker 3: you said negative, therefore wrong? You know some you have
30:00
Speaker 3: to look at a natural language expression together with a
30:03
Speaker 3: natural language expression and say is this adequately substantiated by
30:07
Speaker 3: the evidence over here? So there there is there's a
30:11
Speaker 3: dance where basically what we construct is there's a system
30:15
Speaker 3: of sort of the AI. Well you you need you
30:18
Speaker 3: know AI all the way down. You need an AI
30:19
Speaker 3: system that could evaluate the AI system. This goes by
30:23
Speaker 3: many names in the field. You hear people talk about
30:25
Speaker 3: automated evaluation. You people if people are l l M
30:28
Speaker 3: as judge. And the challenge here is that you can't
30:33
Speaker 3: just simply say to the to an l M, give
30:35
Speaker 3: it a prompt say is it accurate or or like?
30:38
Speaker 3: You need to be extremely precise in how you structure
30:41
Speaker 3: the task and what are the how, how how the
30:44
Speaker 3: model is supposed to break it down? What exactly costanus
30:46
Speaker 3: a claim? What is it looking at and what is
30:47
Speaker 3: it associating with? And then you also need to engage
30:51
Speaker 3: humans too, in a kind of ongoing, never ending process
30:56
Speaker 3: of continually calibrating the automated men that you have to
31:00
Speaker 3: assure that their judgments continue to align with human experts.
31:07
Speaker 3: So you have a whole apparatus of metrics that are
31:09
Speaker 3: looking for all these very important things. Some of them
31:12
Speaker 3: are looking at appropriate use of it, like very in
31:16
Speaker 3: the weeds, aspects of style of medical language. Some are
31:20
Speaker 3: targeted to catching omissions. Some care about misattributions, right, Like
31:25
Speaker 3: is the when the patient says something, it should show
31:28
Speaker 3: up in a history of present illness, in the subjective
31:30
Speaker 3: section of conversation. When the doctor says something, it should
31:34
Speaker 3: show up in the assessment or in the objective section.
31:37
Speaker 2: It's like donning on me.
31:38
Speaker 3: Now that that might sound offensive to some patients, but
31:41
Speaker 3: you know, the point is there's a whole panel of
31:45
Speaker 3: sort of like attributes of interest. You care about accuracy, expelling,
31:49
Speaker 3: you care about style, you care about organization, you care
31:53
Speaker 3: about desertererata that relate to compliance with downstream billing processes.
31:59
Speaker 3: And for each of these you need to construct a judge,
32:03
Speaker 3: and you also need to make sure that you're checking
32:06
Speaker 3: in with an ultimate gold standard source of truth, which
32:08
Speaker 3: is both the judgments of human experts and also downstream outcomes.
32:13
Speaker 2: And this is a.
32:16
Speaker 3: Massive pursuit that we pour incredible amounts of energy into.
32:20
Speaker 3: And it is not something where you ever say no
32:23
Speaker 3: one neither at a bridge, nor an anthropic nor an
32:25
Speaker 3: opening I would ever tell you and now evaluate again.
32:28
Speaker 2: We have solved evaluation. It is.
32:31
Speaker 3: It is a task that requires persistence, and I don't
32:36
Speaker 3: like humility.
32:37
Speaker 2: No, that's great.
32:38
Speaker 1: I mean as a late person, I think I got
32:40
Speaker 1: all that. I guess one of the things that it
32:42
Speaker 1: kind of spurred in my mind was, you know, you
32:44
Speaker 1: have all this messy clinical conversation that's happening in this
32:49
Speaker 1: ambient environment. Is there one thing that's technically the biggest
32:53
Speaker 1: challenge to structure from a data perspective in that conversation.
32:57
Speaker 1: Maybe I'm not framing it in.
32:59
Speaker 2: A good way.
33:00
Speaker 1: So maybe you could take the question, however you think
33:03
Speaker 1: is it's interesting?
33:05
Speaker 2: I think it's like all like you know, it's it's
33:09
Speaker 2: all challenging.
33:10
Speaker 3: Really like I'll get a little bit meta that you know,
33:12
Speaker 3: I think when you're dealing with you know, a source
33:15
Speaker 3: of truth that is fundamentally like as you find it
33:18
Speaker 3: in the wild, so organic and unstructured, like everyone makes
33:21
Speaker 3: the question hard to answer is when you say, is
33:25
Speaker 3: there one thing that it's hard to deal with? And
33:29
Speaker 3: like the hardest thing is the meta thing, which is
33:31
Speaker 3: the the problem doesn't arrive to you with enumeration of
33:35
Speaker 3: these are the things, And so I would say, is
33:38
Speaker 3: at any given point in time, like those areas that
33:40
Speaker 3: we talk about, the ones that we have almost like
33:43
Speaker 3: made into a bit of a science, like we put
33:46
Speaker 3: our finger on it, we can measure them very cleanly,
33:50
Speaker 3: those are those are actually the easier ones, Like these
33:53
Speaker 3: are the easiest aspects of you know. Again I was
33:56
Speaker 3: giving like the metaphor of like imagine trying to assess
33:59
Speaker 3: your journalists and all number. I think a similar one
34:02
Speaker 3: would be a software developer. Okay, how good is your
34:04
Speaker 3: software developer? And what you find is like you say, well,
34:07
Speaker 3: how many lines of code are they producing? You know,
34:09
Speaker 3: how many commits are they making? When fraction of their
34:11
Speaker 3: commits have to be rolled back? What if I were
34:13
Speaker 3: to look at their teams, you know the linear tickets
34:16
Speaker 3: that were assigned to them, you know how many were
34:18
Speaker 3: closed and by what time, And you could construct a
34:20
Speaker 3: whole bunch of metrics and then you say, well, which
34:23
Speaker 3: one of those is the hardest to work on it?
34:24
Speaker 3: And in some ways it's like all the things that
34:26
Speaker 3: were easy to measure, that you knew how to measure,
34:28
Speaker 3: that you had to clean word for, that you had
34:30
Speaker 3: a great statistic for, those are actually the easy ones.
34:32
Speaker 3: And the hardest thing is when you make a change
34:36
Speaker 3: to a model and then say, so, we have a
34:38
Speaker 3: cast gade here, we have we have this very built
34:41
Speaker 3: out machinery of automated metrics. But we don't just say
34:44
Speaker 3: then okay, the machines like it, we're going to deploy it.
34:47
Speaker 3: We then have an internal review with a set of
34:50
Speaker 3: like actual trained clinicians, and then we do a blinded
34:53
Speaker 3: review with a set of external clinicians. And it's only
34:56
Speaker 3: then that we start this cascade. And the hardest thing
34:58
Speaker 3: of all is when you're in a situation where all
35:01
Speaker 3: the metrics look look great and some of us they
35:03
Speaker 3: don't like it, and you know it is the hardest
35:06
Speaker 3: thing is the one you don't have a word for.
35:08
Speaker 3: That you don't have your finger on. It is the
35:11
Speaker 3: very like open ended nature of generative AI that you
35:15
Speaker 3: are you're ultimately producing a document and however however comprehensive
35:21
Speaker 3: we make the rubric, you know, there will always be
35:25
Speaker 3: probably some some desiderata that isn't yet captured.
35:32
Speaker 2: Got it?
35:32
Speaker 1: That makes that makes a ton of sense. You know,
35:34
Speaker 1: we've been spending a lot of time on I guess
35:37
Speaker 1: the nuts and bolts of the technology. Maybe we're just
35:42
Speaker 1: switching gears a little bit when we when we think
35:44
Speaker 1: about the business side of it and implementation. It's very interesting.
35:49
Speaker 1: I was having a conversation just the other day with
35:50
Speaker 1: a very known, well known VC and she said something
35:53
Speaker 1: along the lines of, you never would have thought that
35:55
Speaker 1: hospital would in providers would be at the forefront of
35:58
Speaker 1: adopting AI if you would asked us maybe five ten
36:01
Speaker 1: years ago, but that's really the case today. So what
36:04
Speaker 1: does implementation of an abridged solution look like for a provider?
36:09
Speaker 1: And have you seen that sophistication rise kind of similar
36:12
Speaker 1: to that comment over the last couple of years.
36:14
Speaker 3: Yeah, I think there's there's no doubt that there's in
36:19
Speaker 3: you know, I think in a lot of people's minds,
36:20
Speaker 3: like a like a kind of dissonance between a perception
36:24
Speaker 3: of healthcare as this slower, low moving right dominated by
36:28
Speaker 3: legacy software. At the same time, maybe it's been one
36:33
Speaker 3: of the what if not the fastest, one of the
36:35
Speaker 3: one of the two or three fastest kind of adoption
36:37
Speaker 3: stories in any vertical. I think it's a couple things.
36:43
Speaker 3: One I think is just that you know, the like
36:45
Speaker 3: pressure can only build for so long in a system,
36:48
Speaker 3: or it as we really and I think that the
36:50
Speaker 3: untenable position that doctors have been in with the demands
36:54
Speaker 3: for through put, the demands for completing ever more kind
37:00
Speaker 3: of like digital paperwork, and like the limited hours in
37:03
Speaker 3: the day, like it was just like it really is
37:06
Speaker 3: just such a perfect use case. That's something something had
37:10
Speaker 3: to give. And so I think that moment that once
37:13
Speaker 3: people realized it worked, and then we continue to iterate
37:16
Speaker 3: and improve at this high VELOSSI I think things change.
37:19
Speaker 3: And I say from an implementation standpoint, you know, people
37:22
Speaker 3: are so used to software that changes like on an
37:24
Speaker 3: annual cadence or a quarterly cadence, that like for them
37:28
Speaker 3: to suddenly be in a world where first, first, like
37:30
Speaker 3: there's the adoption story, and that I think actually has
37:32
Speaker 3: been smooth. It's been kind of wild. I think once
37:34
Speaker 3: people realize what was happening, we just the shift in
37:39
Speaker 3: the ground was so sudden that we went in one
37:42
Speaker 3: or two years from the first system trying it to
37:46
Speaker 3: system of feel like if they don't adopt an ambient system,
37:48
Speaker 3: their doctors will leave. And so I think like at
37:51
Speaker 3: that point like getting in the door or like them
37:54
Speaker 3: being convinced they need to that this is the way
37:57
Speaker 3: things have not even are going the way things like
38:00
Speaker 3: have already gone. And this question of like I think
38:02
Speaker 3: adopting ambient right now is and is no longer being
38:05
Speaker 3: on the leading edge.
38:06
Speaker 2: It is catching up.
38:07
Speaker 3: And then I think where there's had to be like
38:10
Speaker 3: truly some I think like I know it's.
38:11
Speaker 2: The right word.
38:12
Speaker 3: I'm less of a corporate guy, more of a scientist,
38:14
Speaker 3: but some some kind of like culture change, change management.
38:19
Speaker 1: Everybody's resistant to change.
38:21
Speaker 3: I get that, some some kind of you know, insert it'sert,
38:24
Speaker 3: the it'sert the word the business school, the right business
38:26
Speaker 3: school word. I didn't spend enough time there, you know,
38:29
Speaker 3: to uh to retain. But I think the part that
38:31
Speaker 3: has been the hardest for folks is to like also
38:35
Speaker 3: recognize like they're living in a world where like, in
38:38
Speaker 3: order to the way that you use this technology effectively
38:41
Speaker 3: and adopt it is not to sort of say like, okay, now,
38:44
Speaker 3: don't change anything. Ever, It is actually that like to
38:47
Speaker 3: accept that, like, hey, we're going to have improvements, and
38:50
Speaker 3: they're going to be coming on some on the daily,
38:53
Speaker 3: some on the hourly, some of the weekly. I mean,
38:55
Speaker 3: I think big macro changes will be signaled with a
38:58
Speaker 3: lot of advance notice, but that hey, like our ability
39:03
Speaker 3: to roll out improvements for each niche workflow, each niche
39:07
Speaker 3: specialty features that improve the functionality the agent, like these
39:12
Speaker 3: are going to be happening on a on a somewhat
39:15
Speaker 3: consistent cadence and for the partners to enter not like
39:19
Speaker 3: kind of a state of like ossification, but a rhythm
39:23
Speaker 3: of and a genuine partnership where we are we are
39:26
Speaker 3: in direct communication, and I think it's helped so much
39:30
Speaker 3: that we have I think truly like the best partner
39:32
Speaker 3: experience team in the field. And I get zero percent
39:34
Speaker 3: of the credit for it, but I'm I'm very grateful
39:37
Speaker 3: for for everything that they do, and you know, they
39:41
Speaker 3: are they are very high touched there. They are meeting
39:45
Speaker 3: with these partners, whether it's weekly or bi weekly, they
39:48
Speaker 3: understand what's coming, They're walked through how to use new features.
39:52
Speaker 3: They understand they're also walked through like what for them
39:56
Speaker 3: is often, like I think, like a very new way
39:58
Speaker 3: of doing business when it comes so like procuring technology
40:03
Speaker 3: and what I say now, I mean, we spend so
40:05
Speaker 3: much time time about like the we're kind of going
40:07
Speaker 3: down memory lane talking about describing. But I actually think
40:09
Speaker 3: that as much as there's no limit to how deep
40:12
Speaker 3: describing rabbit hole goes, of just how far we can
40:15
Speaker 3: go into pursuit of note quality and rigor of evaluation
40:19
Speaker 3: and sort of bespoke describing workflows that nail all these
40:24
Speaker 3: nich Es cases in profound ways and give a degree
40:28
Speaker 3: of like billing compliance and concordance. There's so much more
40:31
Speaker 3: to the story now. Like now, the big adoption is
40:34
Speaker 3: I think people realizing that we're coming out of the
40:37
Speaker 3: era of points solutions and entering the kind of AI
40:40
Speaker 3: platform era, and.
40:42
Speaker 1: Well, that's that's great. I'd love to get your view
40:45
Speaker 1: on that, because I think if I heard before you know,
40:47
Speaker 1: we had the sea change with ambient and you know,
40:50
Speaker 1: and now it's somewhat.
40:51
Speaker 2: Table stakes, right.
40:52
Speaker 1: Yeah, as you think about either the industry's roadmap or
40:55
Speaker 1: your roadmap, what are you most excited about? I guess
40:57
Speaker 1: maybe everything's moving so quick, but maybe the next two
41:00
Speaker 1: or three years. I didn't mean to cut you off
41:02
Speaker 1: because it sounded like you were going in that direction,
41:03
Speaker 1: but I want maybe I wanted to frame it a
41:05
Speaker 1: little bit to think about, you know, where's a bridge
41:08
Speaker 1: going to go over the next couple of years, and
41:09
Speaker 1: where do you think the industry is going to go
41:10
Speaker 1: that's going to have profound change.
41:14
Speaker 3: I think twenty twenty six is the year of a fundamental,
41:20
Speaker 3: kind of agentic transformation of the workflow. You should not
41:23
Speaker 3: work with one vendor to summarize the health record and
41:28
Speaker 3: another one to ask a question about the medical literature,
41:31
Speaker 3: and then another one writes your note, and another one
41:34
Speaker 3: edits the note, and another one looks at the as
41:36
Speaker 3: a billing code, and then a different product looks at
41:39
Speaker 3: the note and the billing code and tells you whether
41:41
Speaker 3: or not you need to ping the doctor to update
41:43
Speaker 3: their note to better support the billing code. Like that's
41:46
Speaker 3: not how how any of this I think should work.
41:50
Speaker 3: And I think we're living in this moment now where
41:53
Speaker 3: we really have the ability. I think we have the
41:56
Speaker 3: ability now to actually weave together all of these kinds
42:01
Speaker 3: of skills. You know that that's the sort of like
42:04
Speaker 3: in that agent engineering is sort of the framing that
42:07
Speaker 3: we have. There's content sources, there's skills, there's there's patterns
42:11
Speaker 3: of interaction. And you know, for example, the system that
42:15
Speaker 3: prepares you for the visit and says, hey, you're seeing
42:18
Speaker 3: this patient. It's their annual visit. Don't forget that. You know,
42:21
Speaker 3: you have to check in with them about these three
42:23
Speaker 3: chronic conditions and more over, like you have to because
42:26
Speaker 3: they're a Medicare advantage patient and like in order to
42:29
Speaker 3: support them for them, like you need to document that
42:31
Speaker 3: you've done it. Then, like when you're during the visit,
42:34
Speaker 3: you know, the system should be able to dynamically when
42:36
Speaker 3: it hears you discuss each condition and sequence kind of
42:39
Speaker 3: cross it off your list. When you mentioned, hey, you know,
42:42
Speaker 3: I really think we should bring you in for an MRI.
42:45
Speaker 3: That knee the page that immediately the systems recognize, hey,
42:49
Speaker 3: that's that's a diagnostic procedure that requires prior authorization from insurance,
42:53
Speaker 3: and you need to check that they meet these certain
42:55
Speaker 3: eligibility criteria and if they do, you know, let you
42:59
Speaker 3: know it's cool. And if they don't, let you know
43:01
Speaker 3: it's not and give you an opportunity to change course.
43:04
Speaker 3: And and sometimes whether or not they qualify depends on,
43:09
Speaker 3: for example, information that you need to ask the patient about,
43:12
Speaker 3: and you need to be reminded at the moment of
43:14
Speaker 3: the visit, like, hey, don't forget to ask them. So
43:19
Speaker 3: at the end of the visit, for example, you should
43:21
Speaker 3: be able to say something like ask the agent, so, like,
43:24
Speaker 3: remind me what were the latest results from your thyroid
43:27
Speaker 3: function tests and were they abnormal? And you know, what
43:31
Speaker 3: would be the standard of care, like update to their dosage?
43:35
Speaker 3: And the models should be able the agent should be
43:37
Speaker 3: able to like compose these skills together, consult a record,
43:40
Speaker 3: retrieve the right information, check the latest article and up
43:43
Speaker 3: to date with whom we have a content relationship. See
43:46
Speaker 3: how this sits relative of a normal range, and depending
43:48
Speaker 3: on how far outside it is, like what would be
43:50
Speaker 3: the recommended increment in the dosage? And come back to
43:53
Speaker 3: you with like here's what the literature says, here's a citation,
43:56
Speaker 3: here's what was in the medical record, here's the reference
43:59
Speaker 3: and would you like me to update the plan and
44:02
Speaker 3: the note, you know, with.
44:03
Speaker 2: The following text.
44:04
Speaker 3: I think that that exit from a period of you know,
44:08
Speaker 3: little black boxes here and here and here and there
44:11
Speaker 3: to a complete end to end, integrated agent first experience,
44:16
Speaker 3: I think is where healthcare technology, at least that's the
44:19
Speaker 3: bedside I think needs to go, and that's the future
44:23
Speaker 3: we're really excited to be building towards right now.
44:25
Speaker 1: So one platform to kind of rule them all, to
44:27
Speaker 1: get a little nerdy there is the way you think
44:29
Speaker 1: about it. Do you think that's going to happen organically
44:32
Speaker 1: or do you think there's going to be some consolidation
44:35
Speaker 1: between the different players who are maybe you know, more
44:38
Speaker 1: focused on one of those you know black boxes and
44:41
Speaker 1: have you know, a pretty big deep expertise in it.
44:44
Speaker 3: I think the growth in technology is always I think
44:50
Speaker 3: in a healthy state of affairs, some mixture of organic
44:53
Speaker 3: both the nonlinear moves, you know. I think that you know,
44:57
Speaker 3: tremendous respect and admiration for founders who build great things
45:02
Speaker 3: at all stages, and I think it's it's a beautiful
45:04
Speaker 3: thing about about this industry in this town that you know,
45:07
Speaker 3: when you look back at it, you know you don't.
45:09
Speaker 2: Have one winner.
45:10
Speaker 3: You have even when you look at big tech at
45:12
Speaker 3: the consolidation, it's how many thousands of companies that came
45:15
Speaker 3: together and became Google, Google, Google, the whole Google Documents
45:19
Speaker 3: suite was one product and Google YouTube was a different
45:22
Speaker 3: product exactly. And so I was deep mind and I'm
45:26
Speaker 3: I'm I'm definitely at once very excited about bringing in
45:30
Speaker 3: new talent and starting kind of organic zero to one efforts.
45:33
Speaker 3: In some ways that's my happy place, and I'm also
45:36
Speaker 3: very happy to sort of win together. And so I
45:39
Speaker 3: think there will be some some amount of consolidation for sure.
45:42
Speaker 1: Well said, you know, we're getting to the end of
45:44
Speaker 1: our time here, Zach and I like to wrap these
45:46
Speaker 1: conversations up by like focusing on a life lesson that
45:49
Speaker 1: the guest really holds dear to them and that maybe
45:51
Speaker 1: drives them in their.
45:52
Speaker 4: Day to day.
45:53
Speaker 1: Is there something for either your personal or professional life
45:56
Speaker 1: that you know you share that really kind of informs
45:58
Speaker 1: what you see as your mission?
46:00
Speaker 3: You know, I think if I were going for I
46:03
Speaker 3: already shared with you a bit of my personal health journey.
46:06
Speaker 3: So I think when it comes to mission. That that
46:09
Speaker 3: I think is the biggest north star for me. It's
46:11
Speaker 3: having been a patient, having been struggling, having had that
46:15
Speaker 3: frustration with a system, as it is with a state
46:17
Speaker 3: of medical knowledge, as it is with our ability to
46:20
Speaker 3: leverage data to produce fundamentally new clinical insights. So on
46:24
Speaker 3: that side, that's that's what propels me from from a
46:27
Speaker 3: pure mission standpoint. It's like if I if I feel
46:30
Speaker 3: like I'm getting disengaged with that, like if we were
46:32
Speaker 3: to go down a path that obviously you have to
46:34
Speaker 3: do so many different things to run a company, you
46:36
Speaker 3: have you have to be a recruiter, you have to be.
46:39
Speaker 2: More design guru.
46:41
Speaker 3: But if I felt like I couldn't ladder backwards to
46:44
Speaker 3: improving patient lives, then like if I if I couldn't
46:47
Speaker 3: see like myself in my twenties, like struggling, as like
46:50
Speaker 3: being like glad that I was doing this now, then
46:54
Speaker 3: then then I think I'd be mission on aligned. I
46:56
Speaker 3: might add, you know, on the standpoint of just like
46:58
Speaker 3: selling personal that gives you a little bit of gave
47:01
Speaker 3: me a little bit of grit, and that maybe is
47:04
Speaker 3: not about mission, but it's about ethic because you know,
47:08
Speaker 3: I'm speaking you from San Francisco, which is you know,
47:11
Speaker 3: there's probably like did dare I sound like a kind
47:14
Speaker 3: of like old grumpy guy, but like, you know, in
47:17
Speaker 3: a moment of like truly maximum just.
47:22
Speaker 2: Like I don't know it's right word.
47:23
Speaker 3: It's like some combination of like entitlement and decadence and
47:27
Speaker 3: a sense that like, well, I got a master's you know,
47:30
Speaker 3: I applied to YC. Why aren't you giving me twenty
47:32
Speaker 3: million dollars or something? And you know, I was a
47:35
Speaker 3: jazz musician in like two thousand, you know, early like
47:40
Speaker 3: mid offs through early twenty tens New York City, and
47:45
Speaker 3: like my is is at a different universe, in a
47:48
Speaker 3: different world in terms of the extent of like competition
47:54
Speaker 3: of just like economic reality, and you know, like like
47:59
Speaker 3: a state where where you can only be doing what
48:02
Speaker 3: you're doing because you absolutely love it and can't imagine
48:05
Speaker 3: doing anything else. And also even then you better be
48:10
Speaker 3: one of the like one hundred or so best players
48:13
Speaker 3: in the world to even like get a gig paying
48:15
Speaker 3: like forty dollars at the local bar, because it's only
48:18
Speaker 3: like that. There's all the best jazz musicians in the
48:20
Speaker 3: world descended on New York and there's maybe like any
48:22
Speaker 3: give a night to like twenty places that have little
48:24
Speaker 3: gigs around town. And I used to in like twenty
48:28
Speaker 3: I don't know, twenty ten or something I used to have.
48:32
Speaker 3: I was like ran the house band playing like the
48:34
Speaker 3: one Am set like, which was like we play a
48:37
Speaker 3: set and run the jam session at a at a
48:39
Speaker 3: dive on the Lower east Side called the Moldy Fig
48:43
Speaker 3: And on a good night we maybe got between like
48:46
Speaker 3: thirty and forty bucks a man, and we used to
48:49
Speaker 3: like it became like our unit of currency, like you know,
48:52
Speaker 3: you thought about like, oh, like my car got towed,
48:54
Speaker 3: like how many moldi figs do I have to you know,
48:57
Speaker 3: did that cost?
48:57
Speaker 2: And it was like it was like four nights so work.
49:00
Speaker 2: It's nights of work, maybe more.
49:02
Speaker 3: And so I just I come back to that from
49:06
Speaker 3: like that experience of like working in that environment where
49:09
Speaker 3: it was just like you were you're just grateful to
49:12
Speaker 3: have an opportunity to do the thing that you loved
49:15
Speaker 3: and where when I started as a PhD student, it
49:18
Speaker 3: was a race, like I was saving money hand over fist,
49:20
Speaker 3: making like you know, thirty thousand dollars a year and
49:23
Speaker 3: that was probably less than dirty that I was probably
49:25
Speaker 3: thirty thousand, including some like site income because I was
49:27
Speaker 3: still playing some jazz gigs, you know, and I just
49:31
Speaker 3: I just try to bring that back that like that
49:34
Speaker 3: level of rejection, that level of like economic austerity. To
49:39
Speaker 3: live in that and and find purpose and enjoy in it,
49:43
Speaker 3: I think is like a grounding like brings me back
49:45
Speaker 3: at anytime, anytime you think about like what's going on
49:47
Speaker 3: and the funding climate and who's raising what. And I
49:51
Speaker 3: just think that that is like the most you know,
49:53
Speaker 3: if I need to just like like wait.
49:55
Speaker 2: Around yourself a little bit where I and find my core.
49:59
Speaker 2: I go back to the multi fig nice. I love that.
50:02
Speaker 1: Thank you so much for sharing it. I love the
50:04
Speaker 1: juxtaposition of you know, jazz and jamming, you know, this
50:08
Speaker 1: very unstructured musical environment to now, you know, the structured
50:12
Speaker 1: data that you're dealing with in your professional career. It
50:16
Speaker 1: paints a nice picture. This has been a fascinating discussion
50:19
Speaker 1: on you know, all things at Bridge and I can't
50:21
Speaker 1: thank you enough for joining us today.
50:24
Speaker 2: Yeah, thank you, Jonathan. Really have been enjoy Thanks for
50:26
Speaker 2: having me. It's our pleasure.
50:28
Speaker 1: So that's Zach Clifton, co founder and CTO of a Bridge,
50:31
Speaker 1: Thank you so much for joining us on our latest episode,
50:33
Speaker 1: and please make sure to click the follow button on
50:35
Speaker 1: your favorite podcast app or site so you never miss
50:38
Speaker 1: a discussion with the leaders of the future innovation in
50:41
Speaker 1: healthcare and Jonathan Palmer. And you've been listening to the
50:43
Speaker 1: Vanguards of Healthcare podcast by Bloomberg Intelligence. Until next time, take.
50:47
Speaker 4: Care uses as basssssss bases