WEBVTT - What's Your Problem? with Jacob Goldstein: Using AI to Help Doctors Save Lives 0:00:04.440 --> 0:00:12.399 Welcome to Tech Stuff, a production from iHeartRadio. Hey thereon 0:00:12.480 --> 0:00:15.440 Welcome to Tech Stuff. I'm your host, Jonathan Strickland. I'm 0:00:15.440 --> 0:00:18.640 an executive producer with iHeart Podcasts and how the Tech 0:00:18.680 --> 0:00:22.479 are you? Okay, So this isn't really tech stuff today, 0:00:22.640 --> 0:00:24.919 I thought it would do something a little different. So 0:00:25.000 --> 0:00:28.680 recently we had Jacob Goldstein on the show. And Jacob 0:00:29.000 --> 0:00:34.040 is a journalist. He's done tons of work for multiple 0:00:34.760 --> 0:00:38.919 prestigious news outlets, and he's also the host of a 0:00:39.000 --> 0:00:43.040 podcast called What's Your Problem with Jacob Goldstein? And on 0:00:43.080 --> 0:00:46.760 that podcast, Jacob talks with various smarty pants in the 0:00:46.840 --> 0:00:51.240 engineering field to talk about how technology can potentially help 0:00:51.320 --> 0:00:54.240 us solve some very difficult problems. And I thought it 0:00:54.240 --> 0:00:57.840 would be great to bring you an episode of his podcast, 0:00:58.040 --> 0:01:01.000 because I think if you dig text, you're also going 0:01:01.040 --> 0:01:03.880 to dig What's Your Problem. But I know it can 0:01:03.920 --> 0:01:07.360 be a hassle to go seek out another podcast, and 0:01:08.000 --> 0:01:11.480 a lot of y'all may never take that initiative. So 0:01:11.520 --> 0:01:14.679 I thought, well, I'll bring one episode in just for today, 0:01:15.000 --> 0:01:17.119 and we can listen to an episode of What's Your 0:01:17.160 --> 0:01:20.520 Problem and enjoy that, and then if you like it, 0:01:20.680 --> 0:01:23.119 you can go seek out that podcast and subscribe to it. 0:01:23.200 --> 0:01:26.280 And if otherwise you're like this isn't my bag, well 0:01:26.319 --> 0:01:28.520 don't worry. We'll have another tech stuff episode for you 0:01:28.560 --> 0:01:33.240 on Wednesday. So this episode is called using AI to 0:01:33.360 --> 0:01:36.760 help Doctors Save lives, and I think that's a cool 0:01:36.959 --> 0:01:39.720 topic to talk about. Often on this show, I'm talking 0:01:39.720 --> 0:01:44.560 about artificial intelligence in a rather skeptical way because I 0:01:44.640 --> 0:01:47.600 feel it's not a fault with AI necessarily. It's a 0:01:47.640 --> 0:01:50.800 fault in how lots of businesses are rushing to try 0:01:50.920 --> 0:01:56.320 and incorporate and adopt AI without fully baking in a 0:01:56.400 --> 0:02:00.280 business reason for it, and that kind of short sightedness 0:02:00.360 --> 0:02:04.360 can often have negative consequences. But I would never deny 0:02:04.400 --> 0:02:07.640 the fact that artificial intelligence does have its place and 0:02:07.680 --> 0:02:11.000 it can end up being a huge benefit to us 0:02:11.080 --> 0:02:14.720 if we design it properly and implement it properly. That's 0:02:14.760 --> 0:02:18.160 a big if, and I think in healthcare is one 0:02:18.200 --> 0:02:21.519 place where AI makes a lot of sense, again, assuming 0:02:21.560 --> 0:02:24.359 that we do take the care to design and implement 0:02:24.400 --> 0:02:28.919 it appropriately. Obviously, there are very high stakes when we're 0:02:28.960 --> 0:02:33.119 talking about healthcare. So let's listen in on this episode 0:02:33.200 --> 0:02:36.200 of What's Your Problem? And I hope you enjoy. 0:02:37.480 --> 0:02:40.920 When you walk into a hospital, technology is everywhere. In 0:02:40.919 --> 0:02:43.800 one room, a surgeon is giving a patient a bionic knee. 0:02:44.040 --> 0:02:47.639 In another room, a CT scanner is creating this incredible 0:02:47.720 --> 0:02:50.640 three D picture of the inside of a person's body. 0:02:51.080 --> 0:02:55.240 But in other places the hospital feels less high tech. 0:02:56.160 --> 0:03:00.720 Doctors are still reading patients charts and making decision partly 0:03:00.800 --> 0:03:04.240 on evidence but largely on instinct. This part of the 0:03:04.240 --> 0:03:06.520 hospital is not so different from what it might have 0:03:06.560 --> 0:03:09.839 looked like, you know, fifty years ago, and bringing new 0:03:09.919 --> 0:03:13.560 technology to this part of medicine to care at the 0:03:13.560 --> 0:03:18.320 bedside is a really hard, really interesting problem, because you 0:03:18.400 --> 0:03:20.800 not only have to figure out how to use technology 0:03:20.840 --> 0:03:24.080 to deliver useful information to the doctor at the right time, 0:03:24.760 --> 0:03:27.520 you also have to figure out how to convince the 0:03:27.560 --> 0:03:36.920 doctor that the information is actually worth listening to. I'm 0:03:37.000 --> 0:03:39.480 Jacob Boldstein and this is What's Your Problem, the show 0:03:39.520 --> 0:03:41.760 where I talk to people who are trying to make 0:03:41.880 --> 0:03:46.360 technological progress. My guest today is Succi Sarya. She's the 0:03:46.440 --> 0:03:49.600 founder and CEO of a company called Baesian Health, and 0:03:49.680 --> 0:03:52.840 she's also a professor at Johns Hopkins, where she runs 0:03:52.840 --> 0:03:57.360 a lab focused on machine learning and healthcare. Succi's problem 0:03:57.480 --> 0:04:01.360 is this, how can you use artificially intelligence to detect 0:04:01.560 --> 0:04:05.760 when hospital patients are at risk of potentially deadly complications? 0:04:06.360 --> 0:04:08.920 And then once you've done that, how can you get 0:04:08.960 --> 0:04:12.520 doctors to believe that the AI's warning is worth paying 0:04:12.600 --> 0:04:15.800 attention to. She told me she first got interested in 0:04:15.840 --> 0:04:18.480 healthcare sort of by accident, when she was a grad 0:04:18.520 --> 0:04:21.440 student at Stanford studying AI and robots. 0:04:25.240 --> 0:04:28.080 You know, I grew up actually being fascinated by AI. 0:04:28.160 --> 0:04:30.919 I loved AI, and really most of my interest was 0:04:30.960 --> 0:04:33.200 in the algorithm front and like looking at robotics and 0:04:33.200 --> 0:04:36.360 building robots that were really smart, you know. And I 0:04:36.400 --> 0:04:41.000 got acquainted with medicine through a friend colleague who was 0:04:41.040 --> 0:04:44.880 a doctor taking care of babies. And what I learned 0:04:44.880 --> 0:04:47.640 through her was that this is all this data we're 0:04:47.800 --> 0:04:52.520 starting to collect, but literally nobody was doing designing any 0:04:52.520 --> 0:04:55.280 software to make sense of it. So it was just 0:04:55.360 --> 0:04:59.680 coming from a world where you know, I studied all 0:04:59.720 --> 0:05:05.240 kinds of data day in day out, with robots doing 0:05:05.279 --> 0:05:07.680 fun tasks like getting the robot to hold the ball 0:05:07.800 --> 0:05:11.480 or juggle the ball to then realizing, holy crap, there's 0:05:11.560 --> 0:05:14.160 like so many more useful things we could be doing. 0:05:14.200 --> 0:05:17.520 So that was really my first discovery of like how 0:05:17.520 --> 0:05:19.960 big a gap there was between people who thought about 0:05:20.000 --> 0:05:23.280 AI versus people versus the problems that needed to be solved, 0:05:23.279 --> 0:05:25.240 and how little we understood about these problems. 0:05:25.760 --> 0:05:28.800 So so you decide that this is going to be 0:05:28.880 --> 0:05:31.080 your thing, right, this is your life's work now. 0:05:31.240 --> 0:05:34.120 I mean in the beginning, I wasn't convinced. In the beginning, 0:05:34.240 --> 0:05:37.320 it was just about spending a few years helping out 0:05:37.520 --> 0:05:39.760 and making sure we are able to make you know, 0:05:40.080 --> 0:05:42.240 in the beginning, it was about my next three years. 0:05:42.960 --> 0:05:45.640 Like I was afraid of investine. I was afraid of 0:05:45.680 --> 0:05:48.839 the complexity of medicine. Like it wasn't an easy field. 0:05:48.880 --> 0:05:52.160 It's not one where they welcome you, right, just as 0:05:52.160 --> 0:05:54.680 an engineer, you don't come in and like at least 0:05:54.720 --> 0:05:56.960 twelve thirteen years ago, that wasn't the culture that. 0:05:57.120 --> 0:06:00.920 Like right, Like like an an MD a hospital does 0:06:00.960 --> 0:06:04.440 not want to hear from some AI researcher. They're busy, 0:06:04.440 --> 0:06:05.400 Oh no. 0:06:05.680 --> 0:06:08.120 For sure, and they're like, we're busy, we have real 0:06:08.160 --> 0:06:08.640 work to do. 0:06:08.880 --> 0:06:10.400 Yeah, what is this? 0:06:10.839 --> 0:06:13.200 Like this all sounds an esoteric mumbo jumbo. 0:06:13.440 --> 0:06:16.640 Yeah, And so you say, you know, we're collecting all 0:06:16.680 --> 0:06:19.200 this data in healthcare and we're not doing anything with it. 0:06:20.040 --> 0:06:24.039 That is not intuitive, Like, that's not you know, I 0:06:24.040 --> 0:06:26.200 think most people sort of prior And this is at 0:06:26.240 --> 0:06:28.839 an academic hospital, right, Your friend is at Stanford Hospital, 0:06:28.880 --> 0:06:31.719 a very prestigious academic hospital. I think Stanford Hospital, I 0:06:31.760 --> 0:06:34.320 think data. I think these are people doing research. So 0:06:34.400 --> 0:06:36.120 what do you mean when you say we're collecting all 0:06:36.120 --> 0:06:37.560 this data and not doing anything with it. 0:06:37.760 --> 0:06:40.880 Yeah, So twelve thirteen, fourteen years ago, this field was 0:06:41.000 --> 0:06:44.480 very new and at the time even collecting and storing 0:06:44.520 --> 0:06:47.960 this data, natural question was can be afforded? It costs 0:06:48.000 --> 0:06:50.200 dollars to store this data? Why would we do that? 0:06:50.520 --> 0:06:52.120 And when you say, what what kind of data are 0:06:52.160 --> 0:06:54.120 you talking about here? When you say collect and store 0:06:54.120 --> 0:06:54.920 this data? 0:06:55.040 --> 0:06:58.560 So literally, this was at the time babies entering, you know, 0:06:58.600 --> 0:07:01.480 in the new natle ICU, these are premature babies are 0:07:01.520 --> 0:07:04.800 born in real time. Devices are collecting heart rate and 0:07:04.880 --> 0:07:08.760 vitals and oxygen saturation data and like, and so that 0:07:08.839 --> 0:07:11.680 kind of detailed data, which is much more bulky, was 0:07:11.880 --> 0:07:14.320 historically not stored. Instead, what they would do is they'd 0:07:14.360 --> 0:07:18.840 take like fifteen minute averages and capture that okay, And 0:07:19.280 --> 0:07:21.320 naturally the question came up, do we need to store it? 0:07:21.360 --> 0:07:23.760 This is really expensive data. Let's just throw it away 0:07:24.160 --> 0:07:26.600 after forty eight hours, we don't need it anymore. Let's 0:07:26.640 --> 0:07:27.880 just throw a quick summary of it. 0:07:28.160 --> 0:07:30.080 Huh. So you might do a study, you might track 0:07:30.200 --> 0:07:32.760 certain data points, but the idea that you're going to 0:07:32.840 --> 0:07:36.120 just as a matter of course, be storing all of 0:07:36.120 --> 0:07:40.080 this data that is now being generated and saved because 0:07:40.440 --> 0:07:44.360 electronic medical records are just being adopted. Nobody was doing that. 0:07:44.440 --> 0:07:46.000 Nobody had really thought to do it. It was an 0:07:46.000 --> 0:07:48.920 expensive prospect. It didn't seem like there would be a 0:07:48.920 --> 0:07:50.880 good reason to do it exactly. 0:07:51.320 --> 0:07:56.160 And coming from AI, where we looked at you know, 0:07:56.240 --> 0:08:00.680 fingerprint data on the internet in retail or finance, then 0:08:00.920 --> 0:08:04.200 the you know, we it was so natural to think 0:08:04.240 --> 0:08:07.880 about how this data teaches you things that it felt 0:08:08.120 --> 0:08:11.520 crazy to me that like me one similarly, all sorts 0:08:11.560 --> 0:08:14.520 of amazing things about these babies or human body or 0:08:14.560 --> 0:08:16.760 how we're involved, or like what are the signs and 0:08:16.800 --> 0:08:19.360 fingerprints of disease? How did they show up? 0:08:19.600 --> 0:08:22.400 When you say fingerprint data, that's a that's a metaphor, right, 0:08:22.560 --> 0:08:25.360 what does fingerprint data mean? In the context of sort 0:08:25.400 --> 0:08:27.440 of e commerce and online finance. 0:08:28.160 --> 0:08:30.400 Well, like they went to this site and then they 0:08:30.440 --> 0:08:33.400 came to this site, or like they saw and add 0:08:33.559 --> 0:08:36.480 somewhere else about this, and now you know they're searching 0:08:37.360 --> 0:08:39.920 for something, and it shows you intense It's. 0:08:39.720 --> 0:08:43.640 This moment ten years ago when like the when people 0:08:43.679 --> 0:08:46.480 are using data to know like everything about what I 0:08:46.559 --> 0:08:49.680 do when I'm shopping for new shoes. But you you're 0:08:49.800 --> 0:08:52.800 but they're not collecting data on like sick newborn babies 0:08:53.120 --> 0:08:53.960 exactly right. 0:08:54.200 --> 0:08:57.600 Does that mind blowing to you? Because it was crazy 0:08:57.679 --> 0:08:58.560 mind blowing to me. 0:08:59.000 --> 0:09:02.199 Okay, yes, my mind is blown. So what do you do? 0:09:03.880 --> 0:09:07.240 Well? I mean it seemed like such a pressing problem. 0:09:07.320 --> 0:09:09.960 It also helped that we were funded as a moonshot 0:09:09.960 --> 0:09:14.640 project by the Google founders, that it was a high 0:09:14.640 --> 0:09:17.400 profile investment, and it sort of naturally led way for 0:09:17.880 --> 0:09:21.520 United place like Stanford curiosity, and we had some amazing 0:09:21.559 --> 0:09:26.560 collaborators who were equally curious, who said, well, let's dive 0:09:26.600 --> 0:09:30.360 in and see what we'll understand. And that was the 0:09:30.360 --> 0:09:34.160 start of it. I literally got hold of this massive, 0:09:34.160 --> 0:09:38.360 twelve hundred page, like this huge thig book to learn 0:09:38.400 --> 0:09:40.880 about babies and what conditions they experience and what does 0:09:40.920 --> 0:09:42.920 it all mean, and then starting to understand how does 0:09:42.960 --> 0:09:45.800 it show up in the data, and you know, spent 0:09:46.040 --> 0:09:49.400 evenings and weekends, and actually I remember sitting in the 0:09:49.440 --> 0:09:55.560 basement of Stanford Hospital at over Christmas trying to work 0:09:55.600 --> 0:09:57.679 on trying to get data out of the health record 0:09:57.720 --> 0:09:59.640 in the first place. And we were trying to experiment 0:09:59.640 --> 0:10:02.360 with all of techniques for pulling the data out, which 0:10:02.640 --> 0:10:04.720 you know now is a whole lot easier than it 0:10:04.840 --> 0:10:06.439 was twelve years ago because. 0:10:06.160 --> 0:10:09.480 It's not built for that, right, It's basically built somewhat 0:10:09.480 --> 0:10:11.640 to track the patient and to a significant degree to 0:10:11.720 --> 0:10:15.400 like bill insurance. Right, that's traditionally what electronic medical records 0:10:15.400 --> 0:10:15.760 were for. 0:10:16.120 --> 0:10:17.240 That's exactly right. 0:10:17.280 --> 0:10:19.640 Kind of amazing and kind of weird. I mean, I 0:10:19.679 --> 0:10:23.120 want to talk more about the bigger idea of data 0:10:23.320 --> 0:10:28.559 and healthcare, but just to kind of land this moment 0:10:29.120 --> 0:10:31.960 early in your career at Stanford, like, is there some 0:10:32.160 --> 0:10:34.160 project you do, Like what is the end of your 0:10:34.200 --> 0:10:35.079 work at Stanford. 0:10:35.480 --> 0:10:39.559 So the project was, you know, we're monitoring these premature 0:10:39.600 --> 0:10:43.640 babies right anywhere between twenty four week old babies which 0:10:43.679 --> 0:10:46.479 are very very tiny, like very twenty. 0:10:46.200 --> 0:10:47.839 Four weeks of gestation. 0:10:47.600 --> 0:10:52.040 To be exactly to like twenty eight thirty thirty two. 0:10:52.880 --> 0:10:55.720 And the idea was, these babies, you know, are like 0:10:55.760 --> 0:10:59.800 they're at risk for significant, like an array of complications. Yeah, 0:11:00.040 --> 0:11:03.120 and the idea is the sooner you know, the earlier 0:11:03.200 --> 0:11:05.240 you can do something about it, the greater the chance 0:11:05.320 --> 0:11:08.680 that you're going to actually resuscitate them. So our job was, like, 0:11:08.800 --> 0:11:10.839 could we look at this data from the second they're 0:11:10.880 --> 0:11:15.160 born and collect this data to start analyzing and modeling 0:11:15.520 --> 0:11:17.960 which babies at risk for which of these complications? And 0:11:18.000 --> 0:11:20.959 if you could, then you could start to put more 0:11:21.000 --> 0:11:25.320 of these preventative prophylactic type pathways or approaches in place 0:11:25.360 --> 0:11:26.000 for carrying. 0:11:25.760 --> 0:11:29.640 Basically identify problems more quickly leading to better outcomes. That's 0:11:29.679 --> 0:11:31.920 the basic desire exactly. 0:11:31.920 --> 0:11:34.960 And in the process I discovered, like, you know, a 0:11:35.040 --> 0:11:38.320 long time ago, there was a physician named Virginia Apgar, 0:11:38.880 --> 0:11:42.120 and what she figured out is like, just by measuring 0:11:42.880 --> 0:11:46.040 five different things from when the baby is born, she 0:11:46.160 --> 0:11:48.600 can compute a very simple score that tells you how 0:11:48.640 --> 0:11:52.160 the baby's doing. And so so naturally, the question we 0:11:52.200 --> 0:11:54.319 asked is, Okay, so now that we are seeing all 0:11:54.320 --> 0:11:57.200 these ways in which the machine learning and AI is 0:11:57.240 --> 0:12:01.319 discovering novel signs and patterns are predictive. Could we just 0:12:01.360 --> 0:12:03.960 simply combine this to come up with a simple score 0:12:04.840 --> 0:12:08.160 that says, you know, can I predict complications? And what 0:12:08.240 --> 0:12:10.800 we found was this new simple score that uses data 0:12:10.840 --> 0:12:13.079 that no special thing you have to do, it's already 0:12:13.120 --> 0:12:15.719 being collected. We just analyze it and we ought to 0:12:15.760 --> 0:12:18.600 compute the score turns out to be much more predictive 0:12:18.600 --> 0:12:21.559 than the ABGAR at predicting complications. 0:12:22.080 --> 0:12:24.959 And so so it worked. I mean, did do people 0:12:25.040 --> 0:12:27.320 use it? Is it standard of care? Now? What happened 0:12:27.360 --> 0:12:28.640 with that? With that research? 0:12:29.160 --> 0:12:30.840 So at that point I was like, oh, this is 0:12:30.880 --> 0:12:33.679 so cool. And literally we got all these journalists who 0:12:33.679 --> 0:12:35.360 wanted to write about it, and it was on the 0:12:35.360 --> 0:12:39.400 fundraising you know, it was like Stanford's fundraising highlight for 0:12:39.480 --> 0:12:42.360 like the next five years, et cetera. But what was 0:12:42.400 --> 0:12:44.120 the saddest thing about it is that there was no 0:12:44.280 --> 0:12:48.439 natural mechanism for implementing it in practice. And it had 0:12:48.480 --> 0:12:50.400 to do with so many different pieces to it, Like 0:12:50.760 --> 0:12:54.040 we didn't have the infrastructure, we didn't have the like 0:12:54.320 --> 0:12:56.800 know how of like how do you get physicians to 0:12:56.840 --> 0:12:59.240 trust something like this. How do you build this in 0:12:59.280 --> 0:13:01.640 a way that is true, us worthy and reliable. How 0:13:01.679 --> 0:13:03.160 do you do this so that it's not just like 0:13:03.240 --> 0:13:06.199 a pet project in one hospital, but it's like a 0:13:06.600 --> 0:13:10.000 system that is scalable nationally. And you know, what is 0:13:10.040 --> 0:13:12.520 the incentive structure? Who pays for it and why would 0:13:12.520 --> 0:13:15.240 they pay for it? And all of that is literally 0:13:15.280 --> 0:13:18.600 what sort of got me, like got me super interested 0:13:18.640 --> 0:13:20.920 in the field where I started to feel, Wow, we're 0:13:20.920 --> 0:13:24.480 at the start of what feels like is a massive movement, 0:13:25.440 --> 0:13:27.920 has many components to be figured out, but we need 0:13:27.960 --> 0:13:32.920 to figure this out. Interestingly, at the time on sand 0:13:32.960 --> 0:13:35.400 Hill Road, you know why, virtually being in pal Aalto. 0:13:35.240 --> 0:13:38.880 Yes Santel Road where all the venture capitalists are exactly. 0:13:38.440 --> 0:13:41.920 People were like, this is fantastic, here's money. Why don't 0:13:41.920 --> 0:13:45.320 you start a company on this topic? And I spent 0:13:45.440 --> 0:13:48.840 six months investigating, you know, talking to lots of peers 0:13:50.440 --> 0:13:55.680 health systems, hospitals and realizing we're just too early. There's 0:13:55.720 --> 0:13:57.680 a lot of work that needs to go in place 0:13:57.920 --> 0:14:00.840 for this to become something that will scale nation. Now, 0:14:00.880 --> 0:14:02.360 fast forward ten years. 0:14:02.200 --> 0:14:03.920 Later, I want to fast forward, but give me just 0:14:03.960 --> 0:14:07.360 another moment when you say it's too early, Like in 0:14:07.440 --> 0:14:10.440 what ways was it too early? Like specifically, what was 0:14:10.520 --> 0:14:13.200 not not ready in the world to start a company 0:14:13.240 --> 0:14:13.680 at that time? 0:14:13.800 --> 0:14:15.720 So the first thing we needed is for hospitals to 0:14:15.760 --> 0:14:18.000 be ready to implement a system like that. For that 0:14:18.080 --> 0:14:21.800 to happen, they needed to have implemented the Electronic Health record, huh, 0:14:22.120 --> 0:14:24.560 be stable users of the HR so that they'd be 0:14:24.600 --> 0:14:26.840 willing to plug in third party systems on top of it. 0:14:27.080 --> 0:14:30.080 And it's kind of amazing that ten years ago, you know, 0:14:30.400 --> 0:14:36.240 twenty whatever, twenty teens, still hospitals were not sort of 0:14:36.360 --> 0:14:40.440 ubiquitous users of electronic medical records, right, like doctors were 0:14:40.440 --> 0:14:41.640 still writing on paper. 0:14:42.400 --> 0:14:45.320 Honestly, coming from computer science where I did you know, 0:14:45.360 --> 0:14:47.920 where I was involved in other areas of AI and 0:14:47.960 --> 0:14:51.440 computer science, like this was like the biggest like shift 0:14:51.720 --> 0:14:55.240 in mindset I felt every time I came back into 0:14:55.240 --> 0:14:57.040 the healthcare side of the equation, it felt like I 0:14:57.120 --> 0:15:00.040 was going at least twenty thirty years back, right. 0:15:00.120 --> 0:15:02.320 Like get a time machine going into the past when 0:15:02.360 --> 0:15:06.120 you walk into the hospital, which is particularly, I don't know, 0:15:06.200 --> 0:15:11.120 ironic surprising, given how in some ways healthcare feels very 0:15:11.160 --> 0:15:14.280 cutting edge, right, Like A central interesting thing to me 0:15:14.400 --> 0:15:18.040 about the work that you do is the way in 0:15:18.080 --> 0:15:20.840 which healthcare is. You know, you go get a whatever, 0:15:21.120 --> 0:15:24.280 a CT scan. It's this incredible machine and it uploads 0:15:24.280 --> 0:15:27.920 to a computer and a whatever AI radiologist can you 0:15:27.960 --> 0:15:31.520 know read the scan blah blah blah. And yet on 0:15:31.560 --> 0:15:34.480 the kind of data side, on the complicated patient at 0:15:34.480 --> 0:15:38.240 the bedside side, it's still very kind of old fashioned 0:15:38.240 --> 0:15:39.440 and almost artisanal. 0:15:40.520 --> 0:15:43.360 I mean, you raise like a fantastic point, which is 0:15:43.640 --> 0:15:48.120 I think when it comes to introducing and designing new medicines, Yeah, 0:15:48.200 --> 0:15:52.120 we've become really really good, but in terms of once 0:15:52.160 --> 0:15:56.560 the medicine is produced, in terms of actually accelerating the adoption, 0:15:56.800 --> 0:16:00.520 optimizing the update, yeah, designing who gets it and what 0:16:00.600 --> 0:16:04.680 does and when detecting early who would benefit from it. 0:16:05.120 --> 0:16:07.680 That's what I call the healthcare delivery side of the equation. 0:16:07.920 --> 0:16:11.240 I feel like there's a very very vast gap of 0:16:11.320 --> 0:16:12.800 what needs to happen to get better. 0:16:13.560 --> 0:16:17.480 So, okay, so you do this project. You see that 0:16:17.560 --> 0:16:21.960 it's too early to start a company because the world 0:16:22.040 --> 0:16:25.600 isn't ready yet, because hospitals aren't even widely using electronic 0:16:25.640 --> 0:16:28.240 medical records yet. Much less being ready to sort of 0:16:28.480 --> 0:16:31.840 expert the data and listen to the data, et cetera. 0:16:32.280 --> 0:16:36.560 And you take a job as a professor at Johns Hopkins, Right, 0:16:36.600 --> 0:16:37.440 is that the next step? 0:16:38.040 --> 0:16:41.040 That's right? And part of the move to Hopkins was 0:16:41.680 --> 0:16:45.840 realizing there's so much depth and breadth of medicine, not 0:16:45.880 --> 0:16:48.920 just around the on the actual devices or the engineering 0:16:48.960 --> 0:16:51.480 on the chemical or the drug development, but also on 0:16:51.520 --> 0:16:53.960 the delivery side, like how what does it take to 0:16:54.240 --> 0:16:57.640 scale ideas nationally? How do you design policy around it? 0:16:58.440 --> 0:17:01.800 There was sort of a whole institute dedicated to scaling 0:17:02.160 --> 0:17:06.280 ideas nationally, So to me that was extremely exciting to 0:17:06.400 --> 0:17:10.520 learn about what would it take to really build the 0:17:10.560 --> 0:17:13.680 foundations of a field like this. And moving to Baltimore 0:17:13.720 --> 0:17:16.919 was a big move, but I was just excited by 0:17:16.920 --> 0:17:19.240 the idea of learning it all and learning it especially 0:17:19.320 --> 0:17:22.040 as an engineer as ERNII research, as an outsider coming 0:17:22.080 --> 0:17:22.840 into healthcare. 0:17:25.160 --> 0:17:27.720 In a minute, Succi and her colleagues figure out how 0:17:27.760 --> 0:17:30.720 to use AI to detect when certain patients are at 0:17:30.800 --> 0:17:35.439 risk for complications and also how to get doctors to listen. 0:17:44.080 --> 0:17:46.800 So Succi is at Johns Hopkins in Baltimore and she 0:17:46.880 --> 0:17:50.960 has this big idea using AI to help doctors treat 0:17:51.000 --> 0:17:54.240 hospital patients, but she has to figure out exactly what 0:17:54.400 --> 0:17:55.640 to focus on. 0:17:55.640 --> 0:17:57.720 One of the big areas was this idea of like 0:17:57.840 --> 0:18:03.680 early detection of patients at risk for complications and diagnostic 0:18:03.840 --> 0:18:07.760 errors being the third leading cause of death. Like that's nuts. Like, 0:18:07.800 --> 0:18:11.600 so today, you know there are critical moments that are missed. 0:18:11.760 --> 0:18:14.760 We get patients the wrong diagnosis or that they're developing 0:18:14.800 --> 0:18:17.560 something subtly and slowly. That's like a whole branch of 0:18:17.600 --> 0:18:21.960 diagnostic errors where you know, complication or a condition develops, 0:18:21.960 --> 0:18:24.840 but they don't get noticed in a timely fashion. And 0:18:24.920 --> 0:18:29.000 so these seemed perfect for AI to come in with 0:18:29.040 --> 0:18:31.439 the kind of data that exists to be able to 0:18:31.680 --> 0:18:34.440 flag patients that are high risk and make it easy 0:18:34.480 --> 0:18:35.600 to provide a second pair of eyes. 0:18:35.760 --> 0:18:39.760 Because it's basically pattern matching, right, I mean, differential diagnosis 0:18:39.840 --> 0:18:44.080 is taking lots of different variables from the patient and 0:18:45.200 --> 0:18:48.920 trying to put those variables together to match the patient 0:18:49.000 --> 0:18:52.520 to you know, thousands of other patients and say, oh, 0:18:52.680 --> 0:18:56.080 all of these, all of these variables, all of these 0:18:56.240 --> 0:18:59.199 health indicators suggest that the patient has disease X. Like 0:18:59.240 --> 0:19:02.679 that's fundamentally what a differential diagnosis is, and like machine 0:19:02.760 --> 0:19:04.439 learning should be very good. 0:19:04.320 --> 0:19:09.040 At that exactly. And previously people have attempted differential diagnosis 0:19:09.359 --> 0:19:13.159 with very coarse symptoms, like high level description of like 0:19:13.320 --> 0:19:16.120 you have cop your fever. What was different this time 0:19:16.160 --> 0:19:18.800 around is because of the HR, we had very detailed. 0:19:18.480 --> 0:19:22.639 Data the EHR, the electronic health record right exactly. 0:19:22.200 --> 0:19:26.399 And so it provided this brand new opportunity to do this. 0:19:26.560 --> 0:19:28.720 And then you know, naturally when you go down the 0:19:28.760 --> 0:19:32.239 list and start looking at problem areas, sepsis is a 0:19:32.320 --> 0:19:35.439 model disease. We chose to demonstrate the idea. 0:19:35.880 --> 0:19:38.720 So let's just talk about sepsis for a minute. What 0:19:38.880 --> 0:19:39.440 is sepsis? 0:19:39.720 --> 0:19:43.080 So let's say your patient gets infected. Your immune system 0:19:43.160 --> 0:19:45.840 is now going to do respond in order to protect 0:19:45.840 --> 0:19:50.560 your body, but in sepsis, it overreacts and starts attacking 0:19:50.600 --> 0:19:56.840 your organ systems, leading to organ failure and depth. And 0:19:56.920 --> 0:19:59.800 so the idea of its sepsis treatment is very much 0:19:59.840 --> 0:20:02.040 the earlier you can detect it, the better you are 0:20:02.080 --> 0:20:03.840 at like tackling it. 0:20:03.960 --> 0:20:09.720 Right, Okay, so I buy it. It seems seems like 0:20:09.760 --> 0:20:11.600 a big problem and it seems like one that might 0:20:11.640 --> 0:20:15.359 be solved or at least, you know, made less bad 0:20:15.600 --> 0:20:19.119 by with the application of machine learning. So how do 0:20:19.160 --> 0:20:22.000 you how do you actually do it? What do you 0:20:22.000 --> 0:20:24.600 have to do to build the model and see if 0:20:24.600 --> 0:20:26.159 it works and get people to use it. 0:20:26.280 --> 0:20:29.119 Yeah, so this is almost like what you're about to 0:20:29.160 --> 0:20:31.760 describe in two minutes what was almost a five year journey. 0:20:32.160 --> 0:20:34.680 So first, it's collecting a huge amount of data where 0:20:34.680 --> 0:20:38.320 you can identify both patients of suptic versus non septic 0:20:38.320 --> 0:20:40.239 and when they had it, and what other conditions did 0:20:40.240 --> 0:20:42.720 they have, and what else was happening in their life right, 0:20:42.920 --> 0:20:44.879 and you know, all the data leading up to that 0:20:44.960 --> 0:20:47.919 episode and what was done after the fact. So you 0:20:47.960 --> 0:20:50.000 get the data. Then the next part is, you know, 0:20:50.040 --> 0:20:52.639 you have to actually understand the biological process or the 0:20:52.640 --> 0:20:55.320 clinical process that's happening and layer that on top of 0:20:55.359 --> 0:20:57.240 the data to make sure you're going from like just 0:20:57.280 --> 0:21:00.240 bits and bytes to data that makes sense, okay, And 0:21:00.480 --> 0:21:04.280 then you implement lots of different learning algorithms to be 0:21:04.359 --> 0:21:06.879 able to experiment, you know, the thing that we first 0:21:06.920 --> 0:21:10.119 did versus the thing we do now. There's like lots 0:21:10.119 --> 0:21:12.399 of generations of improvements in order to get to a 0:21:12.400 --> 0:21:16.440 place where you're going from like, you know, not very 0:21:16.440 --> 0:21:18.320 good signal to very good signal. 0:21:18.840 --> 0:21:22.359 So you're building a model through trial and error, basically 0:21:22.400 --> 0:21:25.879 trying to get an AI model that has a high 0:21:26.440 --> 0:21:30.800 sensitivity and specificity that's good at issuing an alert when 0:21:30.800 --> 0:21:33.239 a patient has sepsis, and does an issue too many 0:21:33.280 --> 0:21:35.159 alerts when the patient doesn't have sepsis. 0:21:34.760 --> 0:21:37.320 Basically exactly, and also does it in a way that 0:21:37.680 --> 0:21:40.359