Latest / What's Your Problem? / Using AI to Help Doctors Save Lives

Transcript

Using AI to Help Doctors Save Lives · What's Your Problem?

0:15
Speaker 1: Pushkin. When you walk into a hospital, technology is everywhere.
0:23
Speaker 1: In one room, a surgeon is giving a patient a
0:25
Speaker 1: bionic knee. In another room, a CT scanner is creating
0:29
Speaker 1: this incredible three D picture of the inside of a
0:32
Speaker 1: person's body. But in other places the hospital feels less
0:37
Speaker 1: high tech. Doctors are still reading patients charts and making
0:42
Speaker 1: decisions partly on evidence but largely on instinct. This part
0:46
Speaker 1: of the hospital is not so different from what it
0:49
Speaker 1: might have looked like fifty years ago, and bringing new
0:52
Speaker 1: technology to this part of medicine to care at the
0:56
Speaker 1: bedside is a really hard, really interesting problem, because you
1:01
Speaker 1: not only have to figure out how to use technology
1:03
Speaker 1: to deliver useful information to the doctor at the right time,
1:07
Speaker 1: you also have to figure out how to convince the
1:10
Speaker 1: doctor that the information is actually worth listening to. I'm
1:19
Speaker 1: Jacob Bolgstein and this is What's Your Problem, the show
1:22
Speaker 1: where I talk to people who are trying to make
1:24
Speaker 1: technological progress. My guest today is Succi Sariya. She's the
1:29
Speaker 1: founder and CEO of a company called Baesian Health, and
1:32
Speaker 1: she's also a professor at Johns Hopkins, where she runs
1:35
Speaker 1: a lab focused on machine learning and healthcare Succi's problem
1:40
Speaker 1: is this, how can you use artificial intelligence to detect
1:44
Speaker 1: when hospital patients are at risk of potentially deadly complications?
1:49
Speaker 1: And then once you've done that, how can you get
1:51
Speaker 1: doctors to believe that the AI's warning is worth paying
1:55
Speaker 1: attention to. She told me she first got interested in
1:58
Speaker 1: healthcare sort of by accident, when she was a grad
2:01
Speaker 1: student at Stanford studying AI and robots.
2:08
Speaker 2: You know, I crew actually being fascinated by AI. I
2:11
Speaker 2: love DAI, and really most of my interest was in
2:13
Speaker 2: the algorithm front and like looking at robotics and building
2:16
Speaker 2: robots that were really smart, you know. And I got
2:19
Speaker 2: acquainted with medicine through a friend colleague who was a
2:24
Speaker 2: doctor taking care of babies, And what I learned through
2:28
Speaker 2: her was that this is all this data we're starting
2:31
Speaker 2: to collect, but literally nobody was doing designing any software
2:35
Speaker 2: to make sense of it. So it was just coming
2:38
Speaker 2: from a world where you know, I studied all kinds
2:42
Speaker 2: of data day in day out, with robots doing fun
2:48
Speaker 2: tasks like getting the robot to hold the ball or
2:50
Speaker 2: juggle the ball to then realizing, holy crap, there's like
2:54
Speaker 2: so many more useful things we could be doing. So
2:57
Speaker 2: that was really my first discovery of like how big
3:00
Speaker 2: a gap there was between people who thought about AI
3:03
Speaker 2: versus people versus the problems that needed to be solved,
3:06
Speaker 2: and how little we understood about these problems.
3:08
Speaker 1: So so you decide that this is going to be
3:11
Speaker 1: your thing, right, this is your life's work now.
3:14
Speaker 2: I mean in the beginning, I wasn't convinced. In the beginning,
3:17
Speaker 2: it was just about spending a few years helping out
3:20
Speaker 2: and making sure we are able to make you know,
3:22
Speaker 2: in the beginning, it was about my next three years.
3:25
Speaker 2: Like I was afraid of investine. I was afraid of
3:28
Speaker 2: the complexity of medicine. Like it wasn't an easy field.
3:31
Speaker 2: It's not one where they welcome you, right, just as
3:35
Speaker 2: an engineer, you don't come in and like at least
3:37
Speaker 2: twelve thirteen years ago, that wasn't the culture.
3:39
Speaker 1: That like right, Like like an MD an MD at
3:42
Speaker 1: a hospital does not want to hear from some AI researcher.
3:46
Speaker 2: They're busy, Oh no, for sure, And they're like, we're busy,
3:50
Speaker 2: we have real work to do.
3:51
Speaker 1: Yeah, what is this?
3:53
Speaker 2: Like this all sounds an esoteric mumbo jumbo.
3:56
Speaker 1: Yeah, And so you say, you know, we're collecting all
3:59
Speaker 1: this data and healthcare and we're not doing anything with it.
4:02
Speaker 1: That is not intuitive, Like, that's not you know, I
4:06
Speaker 1: think most people sort of prior and this is that
4:09
Speaker 1: an academic hospital, right, Your friend is at Stanford Hospital,
4:11
Speaker 1: a very prestigious academic hospitel. I think Stanford Hospital, I
4:14
Speaker 1: think data. I think these are people doing research. So
4:17
Speaker 1: what do you mean when you say we're collecting all
4:19
Speaker 1: this data and not doing anything with it.
4:20
Speaker 2: Yeah, So twelve thirteen, fourteen years ago, this field was
4:23
Speaker 2: very new and at the time even collecting and storing
4:27
Speaker 2: this data, natural question was can be afforded? It costs
4:30
Speaker 2: dollars to store this data? Why would we do that?
4:33
Speaker 1: And when you say, what what kind of data are
4:35
Speaker 1: you talking about here? When you say collect and store
4:37
Speaker 1: this data?
4:37
Speaker 2: So literally, this was at the time babies entering, you know,
4:41
Speaker 2: in the new natle ICU, these are premature babies are
4:44
Speaker 2: born in real time. Devices are collecting heart rate and
4:47
Speaker 2: vitals and oxygen saturation data and like, and so that
4:51
Speaker 2: kind of detailed data, which is much more bulky, was
4:54
Speaker 2: historically not stored. Instead, what they would do is they'd
4:57
Speaker 2: take like fifteen minute averages and capture that okay, And
5:02
Speaker 2: naturally the question came up, do we need to store it?
5:04
Speaker 2: This is really expensive data. Let's just throw it away
5:07
Speaker 2: after forty eight hours, we don't need it anymore, or
5:09
Speaker 2: let's just throw a quick summary of it.
5:10
Speaker 1: Huh. So you might do a study, you might track
5:13
Speaker 1: certain data points, but the idea that you're going to
5:15
Speaker 1: just as a matter of course, be storing all of
5:19
Speaker 1: this data that is now being generated and saved because
5:23
Speaker 1: electronic medical records are just being adopted. Nobody was doing that.
5:27
Speaker 1: Nobody had really thought to do it. It was an
5:28
Speaker 1: expensive prospect, It didn't seem like there would be a
5:31
Speaker 1: good reason to do it exactly.
5:34
Speaker 2: And coming from AI, where we looked at you know,
5:39
Speaker 2: fingerprint data on the internet in retail or finance, then
5:43
Speaker 2: the you know, we it was so natural to think
5:47
Speaker 2: about how this data teaches you things that it felt
5:51
Speaker 2: crazy to me that like, we one similarly than all
5:54
Speaker 2: sorts of amazing things about these babies or human body
5:57
Speaker 2: or how we're involved, or like what are the signs
5:59
Speaker 2: and fingerprints of disease? How did they show up?
6:02
Speaker 1: When you say fingerprint data, that's a that's a metaphor, right,
6:05
Speaker 1: what does fingerprint data mean in the context of sort
6:08
Speaker 1: of eCOM online finance.
6:11
Speaker 2: Well, like they went to this site and then they
6:13
Speaker 2: came to this site, or like they saw and add
6:16
Speaker 2: somewhere else about this, and now you know they're searching
6:20
Speaker 2: for something, and it shows you intense.
6:22
Speaker 1: It's this moment ten years ago when like the when
6:26
Speaker 1: people are using data to know like everything about what
6:29
Speaker 1: I do when I'm shopping for new shoes. But you
6:32
Speaker 1: you're but they're not collecting data on like sick newborn
6:35
Speaker 1: babies exactly right.
6:37
Speaker 2: Is that mind blowing to you? Because it was a
6:40
Speaker 2: crazy mind blowing to me.
6:41
Speaker 1: Okay, yes, my mind is blown. So what do you
6:44
Speaker 1: do well?
6:47
Speaker 2: I mean, it seemed like such a pressing problem. It
6:50
Speaker 2: also helped that we were funded as a moonshot project
6:53
Speaker 2: by the Google founders, that it was a high profile investment,
6:58
Speaker 2: and it sort of naturally led away for United place
7:01
Speaker 2: like Stanford curiosity, and we had some amazing collaborators who
7:05
Speaker 2: were equally curious, who said, well, let's dive in and
7:09
Speaker 2: see what we'll understand. And that was the start of it.
7:14
Speaker 2: I literally, you know, got hold of this massive, twelve
7:17
Speaker 2: hundred page like this huge, big book to learn about
7:21
Speaker 2: babies and what conditions they experience and what does it
7:23
Speaker 2: all mean, and then starting to understand how does it
7:25
Speaker 2: show up in the data, and you know, spent evenings
7:29
Speaker 2: and weekends, and actually I remember sitting in the basement
7:32
Speaker 2: of Stanford Hospital at over Christmas trying to work on
7:38
Speaker 2: trying to get data out of the health record in
7:40
Speaker 2: the first place. And we were trying to experiment with
7:42
Speaker 2: all sorts of techniques for pulling the data out, which
7:45
Speaker 2: you know now is a whole lot easier than it
7:47
Speaker 2: was twelve years ago because.
7:49
Speaker 1: It's not built for that, right, It's basically built somewhat
7:52
Speaker 1: to track the patient and to a significant degree to
7:54
Speaker 1: like build insurance. Right, that's traditionally what electronic medical records
7:58
Speaker 1: were for.
7:59
Speaker 2: That's exactly right.
8:00
Speaker 1: Kind of amazing and kind of weird. I mean, I
8:02
Speaker 1: want to talk more about the bigger idea of data
8:06
Speaker 1: and healthcare, but just to kind of land this moment
8:11
Speaker 1: early in your career at Stanford, like, is there some
8:15
Speaker 1: project you do, Like what is the end of your
8:17
Speaker 1: work at Stanford.
8:18
Speaker 2: So the project was, you know, we're monitoring these premature
8:22
Speaker 2: babies right anywhere between twenty four week old babies which
8:26
Speaker 2: are very very tiny, like very twenty.
8:29
Speaker 1: Four weeks of gestation exactly.
8:31
Speaker 2: To like twenty eight thirty thirty two. And the idea was,
8:36
Speaker 2: these babies, you know, are like they're at risk for significant,
8:40
Speaker 2: like an array of complications. And the idea is, the
8:43
Speaker 2: sooner you know, the earlier you can do something about it,
8:47
Speaker 2: the greater the chance that you're going to actually resuscitate them.
8:50
Speaker 2: So our job was, like, could we look at this
8:52
Speaker 2: data from the second they're born and collect this data
8:55
Speaker 2: to start analyzing and modeling which babies at risk for
8:59
Speaker 2: which of these complications? And if you could, then you
9:02
Speaker 2: could start to put more of these preventative prophylactic type
9:06
Speaker 2: pathways or approaches in place for caring.
9:08
Speaker 1: Basically identify problems more quickly leading to better outcomes. That's
9:12
Speaker 1: the basic desire exactly.
9:14
Speaker 2: And in the process I discovered, like, you know, a
9:17
Speaker 2: long time ago, there was a physician named Virginia Apcar,
9:21
Speaker 2: and what she figured out is like, just by measuring
9:25
Speaker 2: five different things from when the baby's born, she can
9:29
Speaker 2: compute a very simple score that tells you how the
9:31
Speaker 2: baby's doing. And so so naturally, the question we asked is, okay,
9:35
Speaker 2: so now that we are seeing all these ways in
9:37
Speaker 2: which the machine learning and AI is discovering novel signs
9:41
Speaker 2: and patterns that are predictive. Could we just simply combine
9:45
Speaker 2: this to come up with a simple score. Huh that says,
9:48
Speaker 2: you know, can I predict complications? And what we found
9:51
Speaker 2: was this new simple score that uses data that no
9:54
Speaker 2: special thing you have to do, it's already being collected.
9:57
Speaker 2: We just analyze it and we aught to compute the
9:59
Speaker 2: score turns out to be much more predictive than the
10:01
Speaker 2: APCAR at predicting complications.
10:04
Speaker 1: And so so it worked. I mean, did do people
10:07
Speaker 1: use it? Is it standard of care? Now? What happened
10:10
Speaker 1: with that research?
10:12
Speaker 2: So at that point I was like, oh, this is
10:13
Speaker 2: so cool. And literally we got all these journalists who
10:16
Speaker 2: wanted to write about it, and it was on the
10:18
Speaker 2: fundraising you know, it was like Stanford's fundraising highlight for
10:22
Speaker 2: like the next five years, et cetera. But what was
10:25
Speaker 2: the saddest thing about it is that there was no
10:27
Speaker 2: natural mechanism for implementing it in practice, And it had
10:31
Speaker 2: to do with so many different pieces to it, Like
10:33
Speaker 2: we didn't have the infrastructure, we didn't have the like
10:37
Speaker 2: know how of like how do you get physicians to
10:39
Speaker 2: trust something like this? How do you build this in
10:42
Speaker 2: a way that is trustworthy and reliable? How do you
10:44
Speaker 2: do this so that it's not just like a pet
10:47
Speaker 2: project in one hospital, but it's like a system that
10:50
Speaker 2: is scalable nationally. And you know, what is the incentive structure?
10:54
Speaker 2: Who pays for it and why would they pay for it?
10:56
Speaker 2: And all of that is literally what sort of got me,
10:59
Speaker 2: like got me super interested in the field where I
11:02
Speaker 2: started to feel, wow, we're at the start of what
11:05
Speaker 2: feels like is a massive movement, has many component to
11:09
Speaker 2: be figured out, but we need to figure this out. Interestingly,
11:13
Speaker 2: at the time on Sandhill Road, you know why, virtually
11:17
Speaker 2: being in Palo Alto.
11:18
Speaker 1: Yeah, Santel Road, where all the venture capitalists are exactly.
11:21
Speaker 2: People were like, this is fantastic, here's money. Why don't
11:24
Speaker 2: you start a company on this topic? And I spent
11:28
Speaker 2: six months investigating, you know, talking to lots of peers
11:33
Speaker 2: health systems, hospitals and realizing were just too early. There's
11:38
Speaker 2: a lot of work that needs to go in place
11:40
Speaker 2: for this to become something that will stale nationally. Now
11:43
Speaker 2: fast forward ten years.
11:45
Speaker 1: Later, I want to fast forward, but give me just
11:46
Speaker 1: another moment when you say it's too early, Like in
11:50
Speaker 1: what ways was it too early? Like specifically what was
11:53
Speaker 1: not ready in the world to start a company at
11:56
Speaker 1: that time.
11:56
Speaker 2: So the first thing we needed is for hospitals to
11:58
Speaker 2: be ready to implement a system like that. For that
12:00
Speaker 2: to happen, they needed to have implemented the electronic health record,
12:05
Speaker 2: be stable users of the HR so that they'd be
12:07
Speaker 2: willing to plug in third party systems on top of it.
12:09
Speaker 1: And it's kind of amazing that ten years ago, you know,
12:13
Speaker 1: twenty whatever, twenty teens, still hospitals were not sort of
12:19
Speaker 1: ubiquitous users of electronic medical records, right, like doctors were
12:23
Speaker 1: still writing on paper.
12:25
Speaker 2: Honestly, coming from computer science where I did you know,
12:28
Speaker 2: where I was involved in other areas of AI and
12:30
Speaker 2: computer science, like this was like the biggest like shift
12:34
Speaker 2: in mindset I felt every time I came back into
12:38
Speaker 2: the healthcare side of the equation, it felt like I
12:40
Speaker 2: was going at least twenty thirty years back.
12:42
Speaker 1: Right like at a time machine going into the past
12:45
Speaker 1: when you walk into the hospital, which is particularly I
12:48
Speaker 1: don't know, ironic surprising, given how in some ways healthcare
12:53
Speaker 1: feels very cutting edge, right, Like A central interesting thing
12:56
Speaker 1: to me about the work that you do is the
13:00
Speaker 1: way in which healthcare is. You know, you go get
13:02
Speaker 1: a whatever, a CT scan. It's this incredible machine and
13:06
Speaker 1: it uploads to a computer and a whatever AI radiologists.
13:10
Speaker 1: Can you read the scan blah blah blah. And yet
13:14
Speaker 1: on the kind of data side, on the complicated patient
13:17
Speaker 1: at the bedside side, it's still very kind of old
13:20
Speaker 1: fashioned and almost artisanal.
13:23
Speaker 2: I mean, you raised like a fantastic point, which is
13:26
Speaker 2: I think when it comes to introducing and designing new medicines, Yeah,
13:31
Speaker 2: we've become really really good, But in terms of once
13:35
Speaker 2: the medicine is produced, in terms of actually accelerating the adoption,
13:39
Speaker 2: optimizing the update, designing who gets it and what does
13:43
Speaker 2: and when detecting early who would benefit from it, that's
13:48
Speaker 2: what I call the healthcare delivery side of the equation.
13:50
Speaker 2: I feel like there's a very very vast gap of
13:54
Speaker 2: what needs to happen to get better.
13:56
Speaker 1: So, okay, so you do this project. You see that
14:00
Speaker 1: it's too early to start a company because the world
14:04
Speaker 1: isn't ready yet, because hospitals aren't even widely using electronic
14:08
Speaker 1: medical record yet, much less being ready to sort of
14:11
Speaker 1: expert the data and listen to the data, et cetera,
14:15
Speaker 1: and you take a job as a professor at Johns Hopkins. Right,
14:19
Speaker 1: is that the next step?
14:20
Speaker 2: That's right? And part of the move to Hopkins was
14:24
Speaker 2: realizing there's so much depth and breadth of medicine, not
14:28
Speaker 2: just around on the actual devices or the engineering or
14:31
Speaker 2: the chemical or the drug development, but also on the
14:34
Speaker 2: delivery side, like how what does it take to scale
14:37
Speaker 2: ideas nationally? How do you design policy around it? There
14:41
Speaker 2: was sort of a whole institute dediclicated to scaling ideas nationally.
14:46
Speaker 2: So to me that was extremely exciting to learn about
14:50
Speaker 2: what would it take to really build the foundations of
14:54
Speaker 2: a field like this. And moving to Baltimore was a
14:56
Speaker 2: big move, but I was just excited by the idea
15:00
Speaker 2: of learning it all and learning it especially as an
15:02
Speaker 2: engineer as ERNII research as an outsider coming into healthcare.
15:08
Speaker 1: In a minute, Succi and her colleagues figure out how
15:10
Speaker 1: to use AI to detect when certain patients are at
15:13
Speaker 1: risk for complications and also how to get doctors to listen.
15:28
Speaker 1: So Sucia is at Johns Hopkins in Baltimore and she
15:31
Speaker 1: has this big idea using AI to help doctors treat
15:35
Speaker 1: hospital patients, but she has to figure out exactly what
15:38
Speaker 1: to focus on.
15:39
Speaker 2: One of the big areas was this idea of like
15:42
Speaker 2: early detection of patients at risk for complications and diagnostic
15:48
Speaker 2: errors being the third leading cause of death. Like, huh,
15:50
Speaker 2: that's nuts. Like, so today, you know there are critical
15:54
Speaker 2: moments that are missed. We get patients the wrong diagnosis
15:58
Speaker 2: or that they're developing something subtly and slowly. That's like
16:01
Speaker 2: a whole branch of diagnostic errors where you know, complication
16:04
Speaker 2: or a condition develops, but they don't get noticed in
16:07
Speaker 2: a timely fashion. And so these seemed perfect for AI
16:12
Speaker 2: to come in with the kind of data that exists
16:15
Speaker 2: to be able to flag patients that are high risk
16:18
Speaker 2: and make it easy to provide a second pair of VICE.
16:20
Speaker 1: Because it's basically pattern matching, right, I mean, differential diagnosis
16:24
Speaker 1: is taking lots of different variables from the patient and
16:29
Speaker 1: trying to put those variables together to match the patient
16:33
Speaker 1: to you know, thousands of other patients and say, oh,
16:36
Speaker 1: all of these, all of these variables, all of these
16:40
Speaker 1: health indicators suggest that the patient has disease X. Like
16:43
Speaker 1: that's fundamentally what a differential diagnosis is, and like machine
16:47
Speaker 1: learning should be very good at that exactly.
16:50
Speaker 2: And previously people have attempted a differential diagnosis with very
16:54
Speaker 2: coarse symptoms like high level description of like you have
16:57
Speaker 2: cop your fever. What was different this time around is
17:00
Speaker 2: because of the HR, we had very detailed data.
17:03
Speaker 1: The EHR, the electronic health record right.
17:05
Speaker 2: Exactly, and so it provided this brand new opportunity to
17:10
Speaker 2: do this. And then you know, naturally when you go
17:12
Speaker 2: down the list and start looking at problem areas, sepsis
17:16
Speaker 2: is a model disease. We chose to demonstrate the idea.
17:20
Speaker 1: So let's just talk about sepsis for a minute. What
17:23
Speaker 1: is sepsis?
17:24
Speaker 2: So let's say a patient gets infected. Your immune system
17:27
Speaker 2: is now going to do respond in order to protect
17:30
Speaker 2: your body. But in sepsis, it overreacts and starts attacking
17:34
Speaker 2: your organ systems, leading to organ failure and your depth.
17:41
Speaker 2: And so the idea of its sepsis treatment is very
17:43
Speaker 2: much the earlier you can detect it, the better you
17:46
Speaker 2: are at like tackling it.
17:48
Speaker 1: Right, Okay, so I buy it. It seems seems like
17:54
Speaker 1: a big problem and it seems like one that might
17:55
Speaker 1: be solved or at least, you know, made less bad
17:59
Speaker 1: by with the application of machine learning. So how do
18:03
Speaker 1: you how do you actually do it? What do you
18:06
Speaker 1: have to do to build a model and see if
18:08
Speaker 1: it were and get people to use it.
18:10
Speaker 2: Yeah, so this is almost like what you're about to
18:13
Speaker 2: describe in two minutes what was almost a five year journey.
18:16
Speaker 2: So first, it's collecting a huge amount of data where
18:18
Speaker 2: you can identify both patients of subtic versus non septic
18:22
Speaker 2: and when they had it, and what other conditions did
18:24
Speaker 2: they have, and what else was happening in their life,
18:27
Speaker 2: and you know, all the data leading up to that
18:29
Speaker 2: episode and what was done after the fact. So you
18:32
Speaker 2: get the data. Then the next part is, you know,
18:34
Speaker 2: you have to actually understand the biological process or the
18:36
Speaker 2: clinical process that's happening and layer that on top of
18:39
Speaker 2: the data to make sure you're going from like just
18:41
Speaker 2: bits and bytes to data that makes sense, okay, And
18:44
Speaker 2: then you implement lots of different learning algorithms to be
18:48
Speaker 2: able to experiment, you know, the thing that we first
18:51
Speaker 2: did versus the thing we do now. There's like lots
18:54
Speaker 2: of generations of improvements in order to get to a
18:56
Speaker 2: place where you're going from, like you know, not very
19:00
Speaker 2: good signal to very good signal.
19:03
Speaker 1: So you're you're building a model through trial and error,
19:06
Speaker 1: basically trying to get an AI model that that has
19:09
Speaker 1: a high sensitivity and specificity that's good at issuing an
19:13
Speaker 1: alert when a patient has sepsis and doesn't issue too
19:17
Speaker 1: many alerts when the patient doesn't have sepsis.
19:19
Speaker 2: Basically exactly, and also does it in a way that
19:21
Speaker 2: you know, when it says somebody has sepsis, it's able
19:24
Speaker 2: to explain why. It's able to provide enough information so
19:27
Speaker 2: that the clinician can act on it. And it's not
19:30
Speaker 2: doing it sorely that there's not enough to work on,
19:33
Speaker 2: and it's not doing it so late that it's useless.
19:36
Speaker 1: Like often people talk about AI models machine learning models
19:40
Speaker 1: as black boxes, right, like very good at pattern matching,
19:44
Speaker 1: very good at predicting the next word, but we don't
19:46
Speaker 1: know why, And so you're saying, in this instance, you
19:49
Speaker 1: sort of need to know why.
19:51
Speaker 2: My very key evolution of a scientist working in this
19:54
Speaker 2: area was, in the beginning, I saw it all as
19:56
Speaker 2: data and math, and then as I started working more
20:00
Speaker 2: and more in interfacing and actually deploying systems like this.
20:03
Speaker 2: What I started realizing it's actually not math and data.
20:06
Speaker 2: It's about trust, huh, Because ultimately to get adoption and
20:10
Speaker 2: to get outcomes, I need to get trust from these
20:14
Speaker 2: highly trained clinicians who studied this year and year out,
20:19
Speaker 2: and they have a process and a system for working,
20:22
Speaker 2: and you have to fit within this system.
20:25
Speaker 1: And they're very busy, and it's very high stakes, and
20:27
Speaker 1: they kind of think they know everything, and it's so
20:31
Speaker 1: presumably very hard to get them to trust you in
20:36
Speaker 1: making their clinical judgments exactly.
20:38
Speaker 2: But moreover, I've also been on the other side of
20:41
Speaker 2: like tons of engineers making all sorts of claims about
20:44
Speaker 2: their system knows better. But when you actually go and
20:47
Speaker 2: make sense of what the evaluations they've done, they literally
20:50
Speaker 2: have very little understanding of medicine and the practice of healthcare.
20:54
Speaker 2: So like their claims are mostly not good. So a
20:58
Speaker 2: huge part of it is like developing respect and humility
21:01
Speaker 2: for the system, the complexity, so that when you're bringing
21:04
Speaker 2: in this new thing, it really truly fits. It's easy
21:08
Speaker 2: to use sense it creates value. Without all that, you're
21:13
Speaker 2: not going to get to the benefit.
21:16
Speaker 1: So now you say it creates value and suddenly you
21:19
Speaker 1: sound like a founder, an entrepreneur, and not like an academic.
21:24
Speaker 1: Where where in this arc do you start a company?
21:28
Speaker 2: You know, it was somewhere in twenty eighteen. I remember
21:31
Speaker 2: twenty eighteen was a transformative video for me for a
21:33
Speaker 2: number of reasons. I'll start with the very simple thing
21:37
Speaker 2: of like, when we first built this system and deployed it,
21:42
Speaker 2: only like two or three clinicians used it, and it
21:46
Speaker 2: was the two to three clinicians who were involved in
21:48
Speaker 2: working on the project with us. What I realized was
21:51
Speaker 2: we knew from looking at large amounts of data that
21:53
Speaker 2: the system was working, it was working correctly, and we
21:57
Speaker 2: could identify these cases. We could identify them early, and
21:59
Speaker 2: even from interacting the clinicians, we knew you could do
22:02
Speaker 2: something differently about it. So it's one thing for system
22:05
Speaker 2: to detect. You know, clinicians will say so what, so
22:08
Speaker 2: what am I supposed to do well about it? And
22:10
Speaker 2: in this scenario, we've even done studies to know that
22:13
Speaker 2: actually they could be acting, you know, they could use
22:15
Speaker 2: this output to meaningfully change the patient's care. So then
22:19
Speaker 2: to me, the question was, Okay, if we know this
22:23
Speaker 2: thing works, why the heck are we not succeeding. And
22:25
Speaker 2: that's kind of where it went from the puzzle of
22:28
Speaker 2: math and data to trust. You know, how do we
22:30
Speaker 2: develop and deploy it in a way that's transparent. How
22:33
Speaker 2: do we understand like what are the top of mind
22:36
Speaker 2: issues from a practicing clinician's point of view, and how
22:38
Speaker 2: do we address it? Where are we creating value? How
22:41
Speaker 2: do we start quantifying value?
22:44
Speaker 1: Now? Are there any moments where you're like, you know,
22:46
Speaker 1: you have this thing that can be helpful, and yet
22:49
Speaker 1: someone a doctor, a hospital administrator, whatever is telling you
22:53
Speaker 1: why they're not going to use it?
22:55
Speaker 2: Basically, I mean so many moments I can't even like
23:00
Speaker 2: begin so I think I remember this time when they
23:03
Speaker 2: basically were like, Okay, this thing is flagged the system.
23:07
Speaker 2: What do I do with it? And I was like,
23:09
Speaker 2: you should look if the patient has a scepting and
23:11
Speaker 2: they were like, are you kidding me? How many flags?
23:13
Speaker 2: Do you know? How many alerting systems exist? If I
23:16
Speaker 2: were to take every single alerting system and start to
23:19
Speaker 2: use that to start informing when I'm doing a diagnostic
23:23
Speaker 2: workup and what am I doing, I basically would not
23:26
Speaker 2: get my day to day work done right.
23:28
Speaker 1: It's like if you're it's like when you're if you're
23:29
Speaker 1: ever in an emergency room, like everything is beeping all
23:32
Speaker 1: the time, and your system is just one more beep
23:36
Speaker 1: in a sea of beeps that everybody ignores, and.
23:38
Speaker 2: You feel passionately about it.
23:41
Speaker 1: Your reasons, who care about this beap? But nobody else
23:44
Speaker 1: cares about this beap?
23:45
Speaker 2: Nobody gives a damn And it was just like so
23:49
Speaker 2: it was difficult, right like you come. I was sort
23:52
Speaker 2: of like, you know, I felt defeated. I sat that
23:54
Speaker 2: I was like, this is so unbelievable. This is like
23:57
Speaker 2: so powerful. Why aren't they believing me? And so there
24:00
Speaker 2: was an information gap right like then it was like understanding, oh,
24:04
Speaker 2: you know the system in which they live, Okay, I
24:07
Speaker 2: understand that all these different alert exist. How are these
24:10
Speaker 2: alerts created? How are we different? How can we demonstrate
24:13
Speaker 2: we're different? Why should we be trusted? And so that
24:17
Speaker 2: was as an example starting point. Like another one was
24:20
Speaker 2: like we deployed it, and we deployed it in a
24:22
Speaker 2: way where it was you know, within the electronic health record,
24:26
Speaker 2: but it was done in a way that was really cumbersome,
24:28
Speaker 2: like every time they needed to respond, it was like
24:31
Speaker 2: a few you know, it was like a minute and
24:34
Speaker 2: a half of work, and you know, honestly they're so busy.
24:38
Speaker 2: A minute and a half extra to do something that
24:40
Speaker 2: they don't already have total conviction in is like a
24:43
Speaker 2: lot to ask. So then we spend a bunch of
24:47
Speaker 2: time optimizing, well, how do we go it from me
24:49
Speaker 2: take it from a minute and a half to like
24:50
Speaker 2: three seconds? How do we optimize it so that it's instantaneous.
24:55
Speaker 2: It's easy, it's just there.
24:58
Speaker 1: So this isn't about the data at all. This is
25:01
Speaker 1: just user experience basically.
25:04
Speaker 2: Hugely human factors, like human factors and human factors here
25:08
Speaker 2: is very different and complicated because you're trying to optimize
25:11
Speaker 2: human factors within a chassis that is very complicated. Right,
25:14
Speaker 2: Like you're not like standalone software. This is like you're
25:18
Speaker 2: within an electronic health record, and like, how do you
25:21
Speaker 2: do this in a way that the electronic health record
25:23
Speaker 2: providers will.
25:24
Speaker 1: Allow you information software?
25:28
Speaker 2: It's not your software? And how can you do it
25:30
Speaker 2: in a way that like is smooth and seamless and
25:33
Speaker 2: they actually like it? And then you can do this
25:36
Speaker 2: in a way where it's not just custom built for
25:38
Speaker 2: a Johns Hopkins, but it's something that you can send
25:41
Speaker 2: to take to a rural hospital.
25:44
Speaker 1: So so you're doing all this, at what point in
25:46
Speaker 1: this arc do you start the company?
25:48
Speaker 2: So another like personal thing happened, which is I lost
25:53
Speaker 2: my nephew to sepsis. And you know, it was the craziest,
25:59
Speaker 2: like saddest, like you know, most insane feeling to be
26:04
Speaker 2: able to like, you know, as like a researcher, as
26:07
Speaker 2: a scientist, I'm like net deep in these research areas.
26:10
Speaker 2: And then it's one thing to go and talk about it,
26:14
Speaker 2: to say, well, here's how you do it, and here's
26:16
Speaker 2: how it works, and here's why it will work, and
26:19
Speaker 2: here's why this is a great idea. And it's another
26:21
Speaker 2: to then come to that moment of realization where like,
26:24
Speaker 2: well I haven't actually done anything to make a difference.
26:27
Speaker 1: So you're already working on sepsis. Yes, and your nephew,
26:31
Speaker 1: you say, nephew meaning younger than you?
26:33
Speaker 2: Is this a young much younger than me?
26:35
Speaker 1: Wow?
26:36
Speaker 2: And realizing like I was doing, like it all sounded
26:41
Speaker 2: like an excellent like it all sounded great on paper.
26:43
Speaker 2: You know. It was like you know, I'd go to
26:46
Speaker 2: meetings and lots of people would listen and they'd say, Yay,
26:49
Speaker 2: great idea, et cetera. But then at the end of
26:51
Speaker 2: the day, for me, it was like I'd gotten too
26:55
Speaker 2: used to you know, it's easy. It's easy to like
26:57
Speaker 2: talk about something smart and then people say it's a
27:00
Speaker 2: great idea, and then you leave the room and you
27:01
Speaker 2: feel good about it, and then you go back and
27:04
Speaker 2: you work on it some more. And I think it
27:07
Speaker 2: was hard, like hard for me to sort of realize
27:12
Speaker 2: like I had gotten to carry it away, and I'd
27:15
Speaker 2: gotten to carry it away like not thinking about what
27:19
Speaker 2: is it actually going to take to make it real?
27:21
Speaker 2: And the making it real is what's like just so
27:24
Speaker 2: much harder than I thought. But part of it is
27:27
Speaker 2: I also felt like this isn't just a sad This
27:31
Speaker 2: isn't just like a you know, for an idea for sebsis.
27:34
Speaker 2: This is really like crazy to me that this isn't
27:37
Speaker 2: how we operate the like, I think the time has
27:39
Speaker 2: come and what is exciting to me is in the
27:42
Speaker 2: last year or two, I'm starting to see the world
27:45
Speaker 2: has shifted. There's been a very meaningful change in the
27:49
Speaker 2: last few years. I think losing my like losing my nephew,
27:53
Speaker 2: made it very real. It went from this idea to
27:57
Speaker 2: feeling like this was an opportunity where it's very real.
28:01
Speaker 2: Now we can make a difference. The pieces exist, and
28:04
Speaker 2: I need to make it happen. I can't hide anymore.
28:07
Speaker 2: And in twenty eighteen I went from like started to
28:11
Speaker 2: realize like most systems that finished implementing the health record
28:16
Speaker 2: electronic health record policies were starting to change. The AI
28:22
Speaker 2: was mature enough that it was really clear we could
28:24
Speaker 2: do a lot with it. And it was my very
28:27
Speaker 2: little part I could do to you know, address my
28:33
Speaker 2: my you know, my part of grief related to my nephew.
28:36
Speaker 2: Like it was the very little role I could play.
28:38
Speaker 2: So so in twenty eighteen I started to, you know,
28:42
Speaker 2: think go after it with the idea that we're going
28:44
Speaker 2: to actually start a company. We're actually going to turn
28:46
Speaker 2: this into something that scales nationally. And that's where it
28:49
Speaker 2: all began.
28:50
Speaker 1: So you start the company, and you do build this
28:55
Speaker 1: AI model to detect sepsis in hospitalized patients, and you
29:02
Speaker 1: do this study and you wind up publishing the outcome
29:07
Speaker 1: in the journal Nature Medicine, which seems like a big,
29:10
Speaker 1: big moment in your work, in the life of your company.
29:13
Speaker 1: So tell me about that study.
29:17
Speaker 2: Yeah, So in twenty two, in July twenty two, we
29:20
Speaker 2: had three studies. They were featured on the cover of
29:22
Speaker 2: Nature Medicine. These were very big studies for the field.
29:25
Speaker 2: Then the studies that came out in twenty two were
29:29
Speaker 2: basically showing how we implemented the system by five different sites,
29:35
Speaker 2: like both in the emergency department, the floor, the hospital
29:38
Speaker 2: flow is the ICUs across academic and community hospital, So
29:42
Speaker 2: five different hospitals in totally different geographic region right in
29:46
Speaker 2: Maryland in DC, rich communities, poor communities. And what we
29:52
Speaker 2: were able to show was the system both like you know,
29:56
Speaker 2: almost three quarter of a million patients in the study
29:59
Speaker 2: forty four hundred physicians and nurses who were part of
30:02
Speaker 2: the study that you could detect sepsist significantly earlier than
30:07
Speaker 2: they were currently detecting an acting on. So that was
30:10
Speaker 2: one second we showed that. In fact, when we then
30:15
Speaker 2: implemented the system, we show saw meaningful reduction in treatment timing,
30:20
Speaker 2: like patients were getting treatment in a more timely fashion
30:24
Speaker 2: when providers were seeing the alert and acting off of it.
30:27
Speaker 2: And then the third we know early detection is possible
30:30
Speaker 2: now and we know treatment timing is moved and we've
30:33
Speaker 2: known in sepsis that early treatment is the key to
30:36
Speaker 2: better outcomes, So the questions do we see that in
30:38
Speaker 2: our population as well? And we saw that in patients
30:40
Speaker 2: who actually got you know, early alerts. On who got
30:45
Speaker 2: the alerts and providers acted on it, we actually saw
30:48
Speaker 2: much better outcomes in terms of reductions in mortality, morbidity,
30:53
Speaker 2: length of state, fewer complications, secondary complications that arise out
30:57
Speaker 2: of sepsis. So it was just extremely exciting to see
31:02
Speaker 2: that we could go from you know, a technical idea
31:06
Speaker 2: to actual outcomes. And then one of the most interesting
31:08
Speaker 2: things we've studied here was adoption. Will clinicians adopt? It
31:13
Speaker 2: was a very real world study to show, like, can
31:16
Speaker 2: a system like this actually work? And you showed ninety
31:20
Speaker 2: percent physician adoption. So that was extremely exciting to see.
31:23
Speaker 2: And that's what I call that's what you know was
31:26
Speaker 2: about closing the trust gap.
31:28
Speaker 1: So, okay, so you published this paper whatever a year
31:32
Speaker 1: and a half ago, where are you now? What's your
31:34
Speaker 1: company doing?
31:35
Speaker 2: One thing that's very also that I didn't cover earlier
31:38
Speaker 2: is that we expanded the system dramatically from not just
31:41
Speaker 2: working on sepsis but a variety of other conditions like sepsis,
31:47
Speaker 2: where there is very significant both clinical benefit but also
31:51
Speaker 2: financial benefit for the health system. The reason the financial
31:53
Speaker 2: piece matters is, you know, ultimately health systems are working
31:56
Speaker 2: on one two percent margin. For them to be able
31:59
Speaker 2: to implement systems that actually improve care, they still need
32:02
Speaker 2: to be able to financially justify that this can be done,
32:05
Speaker 2: and that was crucial.
32:07
Speaker 1: So what are some of the other things you're working
32:10
Speaker 1: on besides sepsis? Now?
32:12
Speaker 2: Like another example area is fresh ulcers? Okay, huge area
32:16
Speaker 2: where like bed so like a bed source exactly like
32:20
Speaker 2: it's an area where again huge patient impact in terms
32:24
Speaker 2: of like you know, if you do end up getting
32:26
Speaker 2: a serious beds or how detrimental it is for the patient,
32:29
Speaker 2: sometimes leading to death, sometimes leading the need for amputation,
32:33
Speaker 2: but even more interestingly, huge burden on the caregivers themselves,
32:37
Speaker 2: like nurses today have to do a huge amount of
32:40
Speaker 2: work to take care of these patients. Like today, there
32:43
Speaker 2: are lots of scenarios where these patients are missed, and
32:45
Speaker 2: there's an opportunity where you can actually use this data
32:48
Speaker 2: to identify this higher school and start again implementing these
32:51
Speaker 2: new ways in which you can do targeted you know,
32:55
Speaker 2: preventative measures.
32:56
Speaker 1: What has to happen for you to you know, for
33:00
Speaker 1: your software to get adopted at hospitals all around the country.
33:03
Speaker 1: Like I buy that it's helpful. How do you get
33:06
Speaker 1: from it being a kind of researchy thing to being
33:08
Speaker 1: a thing that everybody uses?
33:10
Speaker 2: So the hurdles we needed to cross was one. We
33:13
Speaker 2: needed to figure out a way to get approvals from
33:14
Speaker 2: the electronic health records to be able to integrate it.
33:16
Speaker 2: We did. That took a couple of years.
33:18
Speaker 1: From like the just the big software makers, Epic, whatever,
33:21
Speaker 1: the companies that make the electronic health records. They have
33:23
Speaker 1: to say yes, okay, so that's done. Check. Great. What
33:26
Speaker 1: has to happened next? Yeah?
33:28
Speaker 2: Next, you need a system that is able to you know,
33:30
Speaker 2: when you go from one side to the next, to
33:31
Speaker 2: the next to the next. You need the ability to
33:33
Speaker 2: be able to measure and generalize as you core, cross
33:35
Speaker 2: site and reliably perform.
33:37
Speaker 1: So it has to work in lots of different kinds
33:39
Speaker 1: of hospitals that collect different kinds of data in different settings.
33:43
Speaker 2: And in our partnerships we've shown that data.
33:45
Speaker 1: Okay, Third check.
33:47
Speaker 2: Like I said, we have to show that basically people
33:49
Speaker 2: will adopt in these different environments. So we have data
33:51
Speaker 2: to show that.
33:52
Speaker 1: Okay.
33:53
Speaker 2: For in some of these areas you need fd approval okay,
33:56
Speaker 2: and in the areas we need f the approval. We're
33:58
Speaker 2: working with the FDA to get those approvals.
34:00
Speaker 1: Okay. So that's kind of the next step, correct.
34:03
Speaker 2: And then once that's done, you can now start to
34:06
Speaker 2: you know, it's it's available, it can be market it,
34:09
Speaker 2: you can scale it nationally. All very exciting things.
34:13
Speaker 1: So so if things go well for you, what will
34:19
Speaker 1: the world look like in say five years.
34:23
Speaker 2: Oh my god, so exciting. I think we will actually
34:26
Speaker 2: be implemented at sixty seventy eighty percent of the market,
34:30
Speaker 2: I hope in the US. What's interesting now is like,
34:35
Speaker 2: you know, healthcare is a market which is a leader
34:37
Speaker 2: follow up market. And once you show things that work,
34:40
Speaker 2: it makes logical sense. You have the proof points, you've
34:43
Speaker 2: tackled most of the common issues that people struggle with.
34:46
Speaker 2: Then this is an area where you can scale. And
34:48
Speaker 2: when it comes to like the areas we're working in,
34:50
Speaker 2: which is clinical, unlike some of the other areas like
34:53
Speaker 2: billing and messaging and back office, you know, the years
34:57
Speaker 2: of development required to build. What we build is very long,
34:59
Speaker 2: Like it's taken us eight to nine years to do
35:01
Speaker 2: all the pieces necessary to get to where we are,
35:03
Speaker 2: So there aren't as a lot of like other competitors
35:06
Speaker 2: in the market.
35:06
Speaker 1: You have a mode, and FDA approval is going to
35:09
Speaker 1: be even more of a.
35:09
Speaker 2: Mode among other things. Exactly, so we have a very
35:13
Speaker 2: very significant like moat and hurdles people have to cross
35:17
Speaker 2: to really get it to work, and we've invested in them.
35:20
Speaker 1: And so in your happy five year future, most of
35:24
Speaker 1: the hospitals in the country will be using your software,
35:27
Speaker 1: your models to detect sepsis, to detect bedsores earlier than
35:33
Speaker 1: in a variety.
35:33
Speaker 2: Of for the conditions. Like we've looked at our own
35:36
Speaker 2: financial models and show that like a you know, modest
35:40
Speaker 2: four to five hospital health system stands to gain like
35:44
Speaker 2: fifty two hundred million dollars from the implementation of our
35:47
Speaker 2: system in some you know, the condition areas we're tackling.
35:50
Speaker 1: And people will die less and be less sick as
35:52
Speaker 1: a benefit also, And.
35:54
Speaker 2: That is honestly the biggest maturity I've had in building
35:57
Speaker 2: this company. I started from like the cause of caring,
36:00
Speaker 2: and it was realizing like It's funny. In healthcare, they're
36:05
Speaker 2: so used to caring for patients who are dying every day.
36:07
Speaker 2: They've gotten the sensitive. You then come back to realizing
36:11
Speaker 2: you need the other things to follow, like the money.
36:14
Speaker 2: You need to figure out a way to make it
36:16
Speaker 2: easy for them to do the right thing, And when
36:19
Speaker 2: you do that, then they do actually care about doing
36:22
Speaker 2: the right thing, because that's why they were there in
36:24
Speaker 2: the first place.
36:27
Speaker 1: We'll be back in a minute with the lightning round. Okay,
36:40
Speaker 1: I'm going to keep you another two minutes or something
36:43
Speaker 1: to do a lightning round. You went to college at
36:46
Speaker 1: Mount Holyoke and all women's college. Yeah, and so I'm curious,
36:50
Speaker 1: what is one thing you would tell someone considering attending
36:54
Speaker 1: an all women's college.
36:55
Speaker 2: Oh, I loved Mount Holyoke. It was so much fun.
36:57
Speaker 2: It's where I got my confidence that I could do
37:00
Speaker 2: really really hard things and not be, you know, not
37:02
Speaker 2: feel defeated.
37:04
Speaker 1: If you weren't working in healthcare, where would you be
37:08
Speaker 1: trying to apply a.
37:10
Speaker 2: Oh my god, I've just been so obsessed with healthcare
37:13
Speaker 2: for the last decade. I haven't really lifted my head
37:15
Speaker 2: to think about other things. I mean, honestly, there are
37:17
Speaker 2: a million areas you could apply it, but I don't
37:21
Speaker 2: like thinking about it because it's just that the need
37:23
Speaker 2: is so dire in health care and it's so hard.
37:25
Speaker 2: It's so hard for an II research to focus in
37:27
Speaker 2: healthcare because they don't make it easy. You can make
37:31
Speaker 2: a lot more money doing the same kind of things
37:33
Speaker 2: in finance. You can get the data more easily, you
37:35
Speaker 2: can make money off of it more easily. Like it
37:38
Speaker 2: is annoying, It is really annoying.
37:41
Speaker 1: Is chet GPT overrated or underrated?
37:44
Speaker 2: Actually? I think it's underrated.
37:46
Speaker 1: Okay, go on, I think.
37:49
Speaker 2: You know, when we see the math, we're like, okay,
37:52
Speaker 2: that's the math. That's interesting to me. What was really
37:54
Speaker 2: informative was like the experience, the social experience. It was
37:58
Speaker 2: so exciting to see people who first interacted with it
38:02
Speaker 2: and you know, have the head mind be blown by
38:05
Speaker 2: the experience. And that's sort of then informing how important
38:09
Speaker 2: the user experience out of the houses, like you know,
38:11
Speaker 2: we had some of the chatbot technology before we had
38:14
Speaker 2: some of the interactive but it's sort of how opening
38:17
Speaker 2: I designed it in the use cases like storytelling, poems,
38:22
Speaker 2: like the use cases where they trained the system to
38:25
Speaker 2: be very good at conversant like was what made the
38:29
Speaker 2: experience so exciting because then people could start, you know,
38:32
Speaker 2: like experiencing it themselves and that sort of opened up
38:36
Speaker 2: their mind to what else could it do?
38:37
Speaker 1: Analogous to the lesson you were talking about in your
38:40
Speaker 1: own work, where getting the answer right figuring out if
38:44
Speaker 1: the person has sepsis is actually only part of what
38:47
Speaker 1: you have to.
38:48
Speaker 2: Do huge and that's I think where ais a field
38:52
Speaker 2: that a lot has a lot of growing up to
38:54
Speaker 2: do because historically the people who entered this field are
38:57
Speaker 2: you know, they gravitate towards the math, they gravitate towards
39:01
Speaker 2: the hired science. But what they don't realize is ultimately
39:05
Speaker 2: it is a people problem that you're solving. You have
39:08
Speaker 2: to get people to love it. You have to get
39:11
Speaker 2: people to incorporate it in their daily lives for this
39:13
Speaker 2: to be successful, and you have to operate in a
39:16
Speaker 2: world which is not very precise, Like people have their
39:20
Speaker 2: faults and their mistakes, and they work in a particular way,
39:22
Speaker 2: and you've got to get this thing to fit.
39:28
Speaker 1: Suchi Saraya is a professor at Johns Hopkins and the
39:31
Speaker 1: founder and CEO of Asian Health. Today's show was produced
39:37
Speaker 1: by Edith Russolo and Gabriel Hunter Chang. It was edited
39:41
Speaker 1: by Karen Chakerji and engineered by Sarah Bruguer. You can
39:45
Speaker 1: email us at a problem at pushkin dot FA. I'm
39:48
Speaker 1: Jacob Goldstein, and we'll be back next week with another
39:51
Speaker 1: episode of What's You're Talking