ACM ByteCast - Suchi Saria - Episode 15

Episode Date: May 4, 2021

In this episode of ACM ByteCast, Rashmi Mohan hosts Suchi Saria, the John C. Malone Associate Professor of Machine Learning and Healthcare at Johns Hopkins University, where she uses big data to improve patient outcomes. She directs the Machine Learning and Healthcare Lab and is the founding research director of the Malone Center for Engineering in Healthcare. Saria has worked on projects with the NSF, NIH, DARPA, and the FDA and is the founder of Bayesian Health. Her many recognitions include Popular Science magazine’s “Brilliant 10”, the MIT Technology Review’s 35 Innovators Under 35, and World Economic Forum Young Global Leader. Suchi describes tinkering with LEGO Mindstorm and reading about AI and the future as a child in India and how, years later, she ended up at the forefront of applying machine learning techniques to computational biology. She explains how ML can help healthcare go from a reactive to a predictive and preventive model, and the challenge of making sure that the medical data collected is actionable, interpretable, safe, and free of bias. She also talks about the transition from research to practice and offers her best advice for students interested in pursuing computing.

This is ACM ByteCast, a podcast series from the Association for Computing Machinery, the world's largest educational and scientific computing society. We talk to researchers, practitioners, and innovators who are at the intersection of computing research and practice. They share their experiences, the lessons they've learned, and their own visions for the future of computing. I am your host, Rashmi Mohan. Each of us dreams of working on a meaningful project that makes a deep impact to the lives of people.

Our guest today lives that dream in the most tangible manner. Her pioneering work at the intersection of ML and healthcare improves outcomes for patients across multiple clinical areas. Suchisaria is the John C. Malone Associate Professor of Computer Science at Johns Hopkins University and has been working in the field of machine learning and healthcare for many years. She leads large groundbreaking projects with top government organizations like NSF, NIH, DARPA, and the FDA. She's been named in the World Economic Forum's Young Global Leader list and in MIT Tech Review's 35 Innovators Under 35, amongst many other accolades, and is also the founder of Bayesian Health. Suchi, welcome to ACM ByteCast. Thank you for having me, Rashmi.

You know, it's entirely our pleasure. And I'd love to understand, you know, a common question that I ask all my guests, because I get such fascinating answers is, if you could please, you know, introduce yourself beyond what I have already said, and talk about what you currently do, as well as give us some maybe insight into what drew you into the field of computing? Sure. Funny you asked that. Takes me back to my super early days. Actually, I grew up in a little town in India. And, you know, I mean, you're from India, you probably will find some of this familiar where it's not really traditional where I grew up as a, like, to, you know, like, little girls aren't, like, motivated or inspired to get into computing. It's something that you, you know, that's not sort of the typical path.

And in my scenario, what happened is I just had, you know, older cousins who were really into computing. And I was reading and I grew up, I was always a tinkerer, like always messing around with things, building things, unbuilding things, rebuilding things. And I happened to run into, I was reading a computer science book that talked about the next generation of AI and next generation of computing will be AI and these intelligent machines. And that was just fascinating. Like, and what was that going to be? And what would it take to build these smart machines? And that question fascinated me.

So I think that's really what motivated me into doing computer science. So most of the early part of my career was mostly being fascinated by AI as a field, the research itself, the technology itself. And my first experience was in robotics, trying to build, I don't know if you know, Lego Mindstorm kits, but basically using Lego Mindstorm kits, and then just my own programming to be able to build aspects of what would make a robot smart. And that, you know, one thing led to another, got into research very early,

happened to just be around really incredible, inspiring early superstars in the field who, you know, took me under their fold. I started doing research with them very early, got familiar with what state of the art was, started pushing it. And then machine learning as a field was just starting to gain traction because historical work in AI was very much around expert driven, more logic based systems. And, you know, machine learning as a way of building smart, you know, AI was relatively new at the time. And it was really fun to sort of understand the techniques that existed. I love stats. I love CS, loved engineering as a whole, physics. So it was just fun to put those ideas together to start, you know, just making progress through projects. So that was really sort of my early upbringing in

the field. One thing that many people don't know about me, which I've actually only recently discovered people find strange and maybe young girls find inspiring was that when I, like growing up, I actually went to art school for 10 years on the side. Like, you know, I used to spend a good chunk of my life was, you know, painting and, you know, creating in charcoal and oil and, you know, spending a lot of energy perfecting art and my fascination with art and design. And in India, it was much more traditional and expected that because I was good at it, that's what I would do. And so when I was in, you know, 12th grade, you know how in India we have to take exams. So we took exams for like design and engineering and, you know, I happened to clear both. So then I ended up choosing to go into engineering, but

my other hobby and career path was to be a designer, which I, and I wanted to be a, you know, to do design across the board, like create new, uh, like fashions, create new beautiful places. So anyway, so that's, maybe I'll still get to that career at some point. But yeah, that's how I got into it. That's wonderful. I mean, it's two parts, right? One is, I think, finding that inspiration, like you were saying, you had cousins who had maybe exposed you to the field, the ability to actually tinker with a Lego Mindstorm early on. I'm sure that played a large part in just, you know, picking your interest in the field in itself. And yeah, I'm sure that the design bits also, you know, influence some of your decisions or the way you think about problems.

You know, I feel like every part of our education, you know, intentional or otherwise contributes towards where we are. But the other thing that you brought up was, you know, ML was sort of really, you know, on the upswing when you started out your career, but computational biology as well was really sort of gaining momentum in the 90s. You know, were you at that inception of that curve? And how did you get about applying ML into healthcare? I have to say, I've been incredibly lucky in my career with sort of, you know, being in places where things were on, like, you know, being early in a place. So it's possible to like really shape where things go. So, for instance, in the 90s, there was this, I'd say like in the 90s, there was this revolution of starting to measure more information in genomics, a lot of my field focuses more on the enormous amount of data we are collecting every day as a byproduct of daily

care, right? So like, when you go to a doctor's office, how you generate so much data, you generate data in the form of like, you know, they measure your vitals, they measure your labs, they measure like how you're feeling, what you're taking, how was your response? How are you, you know, doing on a daily basis? Now people wear watches. So there's all this information being collected there, how much you're sleeping, what's your daily activity. And then there's all this information that comes when, you know, for any person who has, you know, any health related challenges, they're under heightened observation where they're getting labs and vitals and other forms of data about them a lot more often. And to me, what was shocking was

the use of that data in driving care delivery decisions. So I'm not even talking about spending thousands of dollars collecting new data. I think that's going to just continually be the case in healthcare. We're just finding better and better ways to measure our health. But what we've lacked and still lack to date, which is what's shocking to me, is our ability to leverage all this data we do collect to drive decisions. We spend more money optimizing what color of shoe to show you on your Google ads page than like, you know, whether you will benefit from this particular medication or not, or should you choose a surgical option or should you choose to do nothing and go, you know, do physical activity, right? So the hundreds of decisions that you make daily and your doctor makes daily around your health and the use of

data today in making those decisions is, you know, very, very limited. It's super interesting to me, right? Because I feel like there's two parts to this. And, you know, I think the shoe reference was particularly sort of, you know, hit home because, you know, that's kind of what we do use on a daily basis. I mean, that recommendation that comes to me on a retail site is something that hits me immediately. But having said that, if I were to think about it from, you know, a patient point of view, how are we able to convey the importance of this to a patient? Like, I mean, I know that they record all of this information about me. How do I know that this is actually helping either improve my overall health, or in some ways, you know,

get me benefit in terms of maybe my spending, like how much I'm spending at a doctor's office, especially with a lot of the time since my insurance is covering it, that impact may not even be directly sort of hitting me, right? So how do you convey the importance of that to a patient or even to the, you know, to the caregiver. Absolutely. Let's think about something as simple as diagnosis. You know, diagnostic errors is the third leading cause of death. Delayed diagnosis leads to so much worse patient experiences all the way from anxiety to not getting the right treatment and your disease progressing where it's much, much harder to treat than if you got instead early and timely diagnosis and you could take the right actions in a timely way. Now, diagnostic errors leading the third, like it's embarrassing

to me that, you know, today diagnostic errors is still the third leading cause of death. And from a patient point of view, it's so real. Like when you're struggling, you're trying to figure out like, you know, what, like if you could use your data to start understanding more clearly, what is the problem you're having? Why is it you're having? Today, you go to a single doctor and they say something. In only a small number of occasions, people have the luxury of even getting a second opinion to then figure out if that's correct. And you're really like in that scenario, sometimes it's like mad obvious. Other times it's not mad obvious. Like an example of this is my nephew. He was in the

hospital. He was hospitalized for something totally different. In the process of his hospitalization, he ended up getting sepsis. It didn't get detected timely in, in an early, timely way. And it progressed to septic shock. And I, we unfortunately lost him. And like my, of course, this is, you know, this is my nephew. He was, so it means very personal things to me, but the reality is this is not unique. Like, you know, there are the number of preventable deaths today that are actually avoid, you know, that we could have avoided if we only had the ability to detect patients at risk and treat them in a timely way is just mind blowing to me that we, you know, AI could like the use of data, the use of smart technology to be able to, you know, identify patients in need for

like tiny patients in need and at risk in order to give, you know, give more proactive care is absolutely what we should be. And we're not because, and again, wouldn't it be great if we weren't using these cycles for like, we've become so good at using these cycles for every other thing. Wouldn't it be great if we could use these cycles to actually improve patient outcomes? Absolutely. You know, I think what you're saying makes immense sense.

I mean, looking at your journey as well, right? You started out with predicting outcomes for premature babies. Now you're looking at chronic illnesses and how to manage them. A lot of your work is around improving these health outcomes. What would you say are like the key problems that you see today that you really are like burning to address? Yeah, so I'll answer that. I'll answer

that in two ways. So the first is from the healthcare standpoint, and then from the technology standpoint. So let's start from the healthcare standpoint, because, you know, that's sort very reactive, right? When you have the ability, you have foresight, when you have the ability to use your data to predict what's going to happen, where things are trended, you can be, you know, you can then put your resources together to give what the patient needs in a timely way and get them to feel better quicker. So this notion of like reactive, and why is it reactive today? The reason it's reactive today is because the whole healthcare system is built around, you know, one, you don't, like patient comes to you when they have a problem.

You look at the patient when they have a problem, you have literally 20 minutes to make a call in the best case, right? If you're in a hospital, you're even busier than that because you're trying to do 20 things at the same time. You have so many patients to look after. There's constantly urgent things that are happening that are drawing on your attention that are causing you to context switch. There's interrupts. There's constant escalations that need your attention.

So the point is you have very limited time. It's reactive because you have limited time, patients coming to you when they have a problem. And then your job is in that very, very limited time to quickly make a decision about what to do next. And then the patient goes away and then that's that. And you're busy and you've moved on to the next thing. And so that paradigm is just not suited to being able to catch hard problems early enough. Now, today, patients end up, but if we could flip it with data, with the right use of machine learning, you could flip it, right? You could have the ability to continuously record data, which is actually happening today. So, you know, one thing that may be helpful for the listeners to hear is, you know, in 2009, there was the High Tech Act in

2000 telling the Affordable Care Act and, and like that, like those legislations were incredibly powerful in accelerating the adoption of electronic medical records in the US, for example, which means we suddenly meant from having homegrown electronic systems where we are partially recording data and disparate systems across each health system uses its own disparate system. And many, many providers just use paper. You come in and they're taking notes in their notepad in doctor's handwriting. And so the challenge was that infrastructure didn't allow for us to have the ability to really take advantage of that data because it was not in a place where we could liquidate

it. But today, because of the legislation 2009-2010, we went through a five to six year journey where across the board, health systems went for implementing electronic infrastructure that allows them to record all these interactions with patients, record data, like all of the information that's collected around clinical information, social information, historical information can now all be in the digital electronic form that can be leveraged. So now imagine putting all that data together with a service, right? That's just basically continuously watching, pulling the data, integrating it, you know,

stitching it together across your past, and then putting it all together to identify what are the things you're at risk for and then surfacing it to the right people, which are your providers or you as a patient or the family and making it very easy to act. I mean, doesn't that make, I mean, why aren't we doing that? Because if we can do that, we could identify so many diseases, the breadcrumbs for those diseases are in the data way, way earlier than our current system. You know, in our current system, you know, we recognize or we get to it. And part of it is because patients come to doctors when they're already aching and hurting and things are already way down the, you know, pike. And wouldn't it be great if we could

have detected some of these much earlier? And if so, we could, you know, prevent it or treat it. For sure. And I think the point that you bring up about, you know, predicting, and this seems to be such a classic marriage between, you know, the use of machine learning in this specific domain, because of the value that you could get from predicting some of these issues ahead of time can literally change a person's life. So I completely resonate with what you're saying. You also mentioned technology challenges. What kind of technology challenges are you running into? Absolutely. So I think this makes sense now, right? So the infrastructure for collecting data now exists. The data now exists.

We're getting ever-increasing amounts of data. So the question is, what are the bottlenecks? Why aren't we seeing this today? And we obviously know there are companies that have, or there are efforts that have wanted to do things like this, but have underestimated what are all the bottlenecks we need to solve for. So in my view, I mean, at the highest level, if I had to say, there are like a couple of big challenges that are really, really important.

So first is in healthcare, unlike some of the other fields like advertising, you're collecting not or, you know, say genomics or say image recognition, you're not just collecting one type of data from one sensor, you're really collecting data from hundreds of sensors, right? Like think of heart rate as a sensor, blood pressure as a sensor, like social needs as a sensor. Some of this is human recorded data, right? Like you're, right? Like when you're interacting with the nurse and they're doing a full head-to-toe evaluation, they generate data around what they saw and how did you feel in your pain levels. And some of this data is very qualitative in nature. And so a big part of this is the ability to integrate very diverse type of data that come in.

And integrating that well really requires deep understanding of the data itself, the way the data measured, in technical speak, what the measurement error models are underlying this data, whether it's missing at random, missing not at random, missing completely at random, missing not at random, missing completely at random, for example, you know, and building technology that embraces the fact that you're doing heterogeneous, multimodal data integration in order to actually draw high quality outputs. So that's one core part of the challenge, right? You have to figure out a way to integrate this data well. And that partly where you have data with very different amounts of noise and messiness. And so you've got to embrace that challenge.

The second is safety, right? In many fields of ML now, like, you know, when you're using it in problems like education, healthcare, it's extremely important to understand that somebody is going to use the output to make very important decisions. Therefore, you have to understand that somebody is going to use the output to make very important decisions. Therefore, you have to understand safety, risk, bias. And there are various ways in which you can do this not very well and generate outputs, which maybe look like from an accuracy perspective, the way we are used to naively measuring accuracy in other fields, that the accuracy is very, very high. But in reality, those metrics give you a very limited part of the picture. In order to make these outputs actionable, they need to be

actionable, interpretable, safe, bias-free, or bias-mitigated to the degree possible. And that's crucial. So how do you do that? And what techniques promote that? So that's, I think, a big... And we've spent a lot of time doing research in these first and second challenge I spoke about. And then third, over the last couple of years that I've sort of becoming intimately familiar with through my experiences at Bayesian is the notion of what I call human machine teaming, which is how do you go from. And I'm sort of leading a big grant in this space with the National Science Foundation as part of the frontier of work.

They have a big frontier of work program. What is the frontier of like human expertise looks like as new technologies come to bear to augment humans? And so here, what we're, you know, what we're trying to understand is basically when in a field like medicine and healthcare, you have experts at every level, right? So you have care coordinators, you have case managers, you have nurses, you have physicians, you have specialists, and even physicians, you have specialists at every level. And how do you build systems where it is possible or AI, where it's possible to partner, to team, to collaborate? And the reason those ideas are fundamental is because in one sense, if all you're really doing is automating a rule, the provider already knows, like, hey, when the temperature rises above this, I want an alert. You could do that. That's what historically

people have done in healthcare. You get extremely limited value out of it. It just gets to a lot of false alerting. People don't really trust it. They don't get a huge amount of value out of it. It just gets to a lot of false alerting. People don't really trust it. They don't get a huge amount of value out of it. They can see a lot of value in what I described earlier as true AI, where basically you're able to integrate data, continuously watching the background, identifying these early signs and symptoms reliably, and then surfacing it because now you're making the life better, right?

Like they don't have the time to be watching all their patients, but they have, but if you can surface early, you know, you can identify the ones where things are going downhill and reliably bring it up. Now you're partnering with them and partnering with them in a way where you're making the job easier and you're making it very easy for them to take the right action, make the right decisions, where otherwise they might have missed it. But to do that partnering and teaming well, there's a role for how the technology has to be built to enable teaming. And I think that's also sort of an important core challenge that we need to continually solve for. We can't just assume that the way we traditionally built AI to drive advertising is the same AI that's going to drive, you know, is going to be the right partner for human experts like physicians and providers. Yeah, that's an excellent point. I mean, there's multiple points that you brought up in that last section, which I want to dig deeper into, right? One part

is just the huge responsibility associated with this kind of an application in comparison to many others, right? Just the impact of this. When you talk about data cleanliness and biases, et cetera, one of the things that, you know, in many other applications, you have a continual stream of data for a certain user. For example, if I'm browsing a retail site, you know on a daily basis what kind of products I'm clicking on, how long I'm spending reading a piece of content. In your case, when a patient comes in to a health clinic, it's not continuous, right? I mean, you can't really follow that one patient and their particular pattern. How do you sort of,

you know, is that a challenge at all where you have to sort of generalize over a larger sort of population and still get accurate results? Oh, yeah, absolutely. So I think there are a couple of things you noted there. So the first thing is the data is more complicated. The reason it's more complicated is because it's not your classic every second something is measured continuously. It's that, you know, things are measured very frequently. Let's say you have a device on, you're going to measure it frequently. Now you're done with the device, it's off. So like, let's say you have a heart rate monitor, you know, you just had a big surgery, you're being monitored, people are worried about your recovery. So you might imagine putting a heart rate monitor for like a 20 day period after the surgery itself and you're going home. Over that period of 20

days, we're getting very good high quality measurements. And now, let's say you take it off, you're not using it anymore. You know, a month later, you're going in for a checkup, you go in, you do a full set of labs, you do a full set of vitals. So you're getting a collection of deep data, but not, you know, like instead of getting a lot of heart rate data, you're getting, you know, 20 different kinds of data, but a snapshot in time for these 20 pieces. And you go on, go forth, you might have another checkup. In some cases, maybe it turns out the patient actually recovery didn't go so smoothly. They needed to be admitted back for a follow-up procedure.

So you'll get a huge amount of information back in again when they get admitted. So what this tells you is basically instead of like the predictable cadence that you might see on a website when a user enters a website or is browsing on the internet, what you're going to see is more like irregular and, you know, diverse, like deep and wide data. But that's okay because, you know, at the end of the day, even if you measured my, you know, like my creatinine on a day, like, you know, two hours, every two hours, it's not going to matter because the reality is my kidneys don't evolve that fast. My creatinine once a day is a huge amount of information, maybe twice a day at best. On the flip side, something like blood

pressure, if I am experiencing shock, it's going to change over the course of a couple of hours. So what is important here is the, like, you know, we're getting these different forms of data at different cadences, but, you know, we know how to think about taking these, you know, we just have to be mindful of embracing the fact that the data that are being measured is a bit more intentional here. It's not just, you know, happen to be, it's intentional in the sense that physicians are recording something when they are worried about you about in that manner. So when, and so that intentionality itself is information, right?

And how do you take intentionality and the information together to start making inferences that, you know, that say something? Yeah. And when we talk all this talk of data, Suchi, you know, of course, we're going to talk about privacy, right? I mean, data privacy is critical in any field when you're using consumer data, user data. What is there anything special in the healthcare field? I'm even more sensitive. Is there anything special that you have to do while you're, you know, using this data? Yeah, I think there's like technology solutions

and then there is philosophy, right? Like, and you need both. You need to understand sort of the culture in which we want to exist. And then there's the getting the technology right to make it happen. So the technology piece is much easier to think about. It's, you know, today we have access to,

you know, very secure cloud-based infrastructure where you can do monitoring in the most granular way to understand and encrypt data and to understand, you know, and secure your data. Let's just say at a high level to figure out both the way to secure the data, but also make it so that the access of the data itself is controlled in a very granular way, right? So you're not sort of just giving carte blanche access of that data to anything. It's only being used in a way that in our example, for instance, I'll give you something concrete at Bayesian, beyond sort of the very rigorous security measures that people think about, also thinking about like, what is the data being used for? And in our scenario, for example, we use all the data to drive and power all of the, you know, machine learning applications that are being used by providers in order to

improve patient outcomes. And, but, you know, like you might see on any other, you know, like the, the Facebook scenario where then people think about secondary uses of the data and what are the secondary uses of the data and what is the policy around secondary use of the data. And I think that's where there's philosophy matters and like the culture matters. And at Bayesian, we take sort of a very particular point of view, which is we want to use the data to improve patient outcomes. And that sort of, that drives our ethos of access. But more broadly in healthcare, I think a lot of concerns arise from not having very clear philosophy and understanding of use. Like when the data is getting collected, who's using it? What are they using it for?

Who else has access to it? And I think people are becoming more aware. People are cleaning this up. And I think there's a lot of fear that sometimes the data is just being used in order to treat it like, you know, an asset for generating extra cash or revenue, you know, without very clear dotted line to how is it impacting outcomes? Yeah. I mean, I think the point you bring up, you know, the ethical considerations of how you use this, if that responsibility lies with, you know, the creators of this technology,

how do you, you know, is there a need for regulation there? Yeah, I think this is such a hard and interesting question. I don't actually, I don't have, I mean, how are we doing it today? I think really today it's a bit chaotic, I have to say. I think like it's so new and it's all happened so fast. People are trying to sort out what's the, you know, what's the right approach. Is it through regulation? Is it through, is it through responsible partnerships? Right. Like, so you've seen the AI partnership group where they've brought in industry leaders from across the board to come together, to think together about what are responsible AI practices. We did the same thing in healthcare,

where we got a collection of like the top leaders, researchers in the field to come together and think about what are the responsible machine learning and health or AI and health practices. And, you know, certainly, I mean, regulation if done right can be very effective. Regulation if done wrong can be very, very wasteful. So the question is, you know, I think the jury's out. I think we're figuring it out, figuring out what is going to work at scale. But along the way, one thing that's very important is we need more leaders who think hard about it, who take a responsible stance, use that to lead by example, right? We need that because that's the basis for forming anything that scales.

And we need more research leaders. We need more practitioners who are leaders, you know, clearly stating policy, clearly, you know, transparently discussing policy. I think transparency is also crucial. I think today there isn't a whole lot of transparency. And so just sort of bringing transparency itself so that we can actually have a discussion out in the open, like, is it okay for large health systems

to be making money from the data they're collecting, where the only purpose of the company you know, company is to be able to monetize data to get another stream of revenue. Maybe it's okay because, you know, the revenue they get is actually then put back into improving patient care. So maybe it's okay. Maybe it's not okay because they're using it to monetize and they're seeing this as a revenue stream.

And it's not really like, you know, and that's preventing them from, you know, liquidating the data for other use cases to other companies because they're worried by sharing that's reducing the value of the data they're able to monetize. I don't know. But I think these are examples of types of discussions we're having. We ought to be having more of and we need more responsible leaders taking, you know, voicing the opinion transparently. No, absolutely. I think that, you know, having that dialogue and transparency, as you call out, is super critical. And I think also, like you said, you know, having those examples of people

who are taking a stand and understanding what their journey is like, just having that access to that information will help drive a lot of these choices as well. One of the things I want to go back to in the few problems that you had called out earlier was really about partnership. So, what was really interesting to me about your journey is how you've made this transition from research to practice so seamlessly. I mean, you seem to sort of, you know, that's the theme of this podcast as well, but you seem to make that transition, you know, back and forth easily. You make it look easy, but I can't imagine that it is. I'd love to hear your thoughts around that.

Was it hard? Is it continuously hard? Yeah, it's actually both hard and easy at the same time. So I can tell you what about it is hard and what about it is not hard for me. So the part that is what I'd say, like the part that was actually not as hard for me was making time. Like it's been absolutely phenomenal to be able to bring, you know, we work so hard in research. Like I have not known, you know, my PhD students, my postdocs, my scientists, like they stay up till 3am sometimes working on these papers because they truly deeply care. We

care about impact and impact at scale. So why are we working so hard? Because we truly deeply care about impact. And the reality is that's how some of, that's what like the life of most researchers are. They want to make an impact, but to make an impact turns out sometimes, you know, we can write all the papers we want. The papers alone don't get you to impact. The papers are like an amazing starting point for what doing it well and doing it thoughtfully would look like. And to me in healthcare, one of the big gaps I saw was, you know, on the one hand, there was enormous need, but on the flip side, the challenges, like how we were going about an industry, addressing those challenges, coming at it from a research expertise point of view,

didn't make sense to me. And on the flip side, I saw there was state-of-the-art research that could really, truly tackle some of these challenges, but they weren't really making it out because the current mechanism of publications alone doesn't get you there. remodeling the algorithms, that's crucial. But there's also product, delivery, integration, you know, experience, delight. All of that matters to get adoption. All of that matters to get to outcomes, right? Because obviously if you're, you know, if people don't use it, you're not going to get there. And in order to use it, it's the end-to-end solution. And so to me, the part that was easy is

we're all motivated by impact. And when you're motivated by impact, working backwards, it seemed easy to think about how the two worlds fit together from bringing state-of-the-art research with state. And this is my first company experience was actually another company that was spun out of Stanford that was also in the data space called Aster. And it was just like, it was just really fun to see when you take state-of-the-art research, you marry it with state-of-the-art, like, you know, you go use that to then build solutions

that are tackling hard problems in industry, like what's possible if done right. And so here it was the same experience, bringing state-of-the-art research. And then in healthcare, there are so many problems. We're not short of problems. You know, marrying the two and doing it well has just been really, really fun. And really, like that part has not felt what's hard.

What has felt hard is culture. What has felt hard is going against the grain because there's an expectation in academia that when you're an academic, what the only part you should care, like, you know, there's traditional academia and there's like, and traditionally in academia, you know, it's very easy to get suckered up in situations where people think the only thing that matters is the papers you publish. And it's the bean counting of papers you publish. Fortunately, I'm in an environment like Hopkins where, you know, my dean is just super, like, he's just sort of like a visionary leader, like really care to do impact at scale, why sort of the typical currency of academia, typical academic research currency of just, you know, writing papers alone isn't adequate. I think that that's where it's hard. So I think it's the culture bit. That's what's what I find the hardest, and kind of helping

people understand, like, remember why we were here to begin with. It was impact at scale. Yeah. I mean, I think that you hit the nail on the head. I think we all look to do that, whether we're in industry or academia, but what would be your advice to maybe, you know, researchers early in career, or even students who are thinking about a career in, you know, in computing research, what should be some of the considerations that they should, you know, think about as they're entering this phase of their life? Yeah, I think, so I guess that three, like, you know, the things that I really benefited from is find mentors who energize you, like find people who inspire you. It was really helpful for me very early on

to find people like, or, or, you know, or readings like reading that inspired me. And because then you don't have to fit the mold. You don't have to be like everybody else around you. You can, you know, you can find the people who inspire you and sort of, and, you know, maybe if they're taking non-traditional paths, you know, you can chalk your path. If somebody else can do it, you can do it. You don't have to be around. So, so to me, a big part of it, and maybe this is partly coming from India, you know, where there's so like, there's a norm and there's an expectation of a norm and fitting a norm. And to me, that was very burdensome. I loved the, like, I was so excited that I could, I had the freedom, was given the freedom. And I don't know exactly how, but, you know, was stubborn enough to like break out of the mold and be able to be inspired by people who, you know, I used to read a lot of biographies and, you know, found people that inspired inspired me felt like if they could do it, I could do it. And it was just really fun to then chart my own figure out chart my own course what needed

to be done, and then find people who could help me make it happen. And then along the way, the most important thing, you need to have a lot of fun, because what's the point of doing it all if you're not having fun. And, you know, when done right work, I mean, work is fun for me like this is like when I have my day off I do exactly what I do on a work day I was just going to ask you what do you do outside of work yeah my life is all very neatly stitched together into things where I love creating things I like making the world about like it's's going to sound so cliche. It's like Silicon Valley, maybe what people might call the S, but like, I just love the idea of using engineering superpowers. Like, you know, being engineers are creators, engineers are value creators. Like at the end of the day, what the field gives you is a lot of like, like ability

to make things that know we didn't exist before. And so to me, I love the idea of using our superpowers to figure out what can we do to make the world a better place. And that's, that's my, when I have free time, that's what, like I sit and think, and I love reading. I love making, you know, I used to do a lot of art. I don't get time to do it anymore. But I love reading. I love, like, being able to think about where there are gaps. And how can we then, you know, use our superpowers to close those gaps? I love how you say, sorry, I didn't mean to. I love how you call it a superpower. I feel so much more energized today, because you're

absolutely right. The ability to build and create is something unique that we have as engineers. And to use that in a meaningful way is truly, you know, is very joyful. Absolutely. You know, we all need our engineering capes, you know, and I wake up in the morning, I'm so excited. Like I wear my little cape, and I'm like, today is going to be the best day. Fantastic. Sushi, I would love to understand, you know, when you have so many, you know, ideas, where do you find your inspiration? And like, how do you determine what you're going to focus on next? I think for me, I feel like I've been on like, what feels now like a 10 year like path to trying to think about like, how do we bring, it is inevitable that, you know, I think five years from now, people will say it's crazy. I, maybe even three years from now, hopefully two years from now, but five years from now,

people will say, it's crazy that we weren't using data to drive decisions every day in healthcare in a more proactive way. And so to me, how do we make that happen? And how do we make that happen faster has been my focus. That is what I think about. That is what captures my imagination all day long. Like when I read, I think about how are we, you know, what are the barriers? Why isn't it happening? Is it financial? Is it administrative? Is it technology? Is it people?

Is it the way we're talking about it? Is it not enough use cases? Is it like not high enough important use cases? And then in terms of like, it's impacted skills. So you have to really understand impact. Don't, I mean, one thing I did a lot as a kid was I used to love doing just hard things because I wanted to prove to everyone I was badass. And I apologize for using that word on your podcast, but, but that's true. Cause then, you know, when you're in a kid, like in India,

like, you know, the it's survival of the fittest and, you know, the survival of the fittest. And, you know, it was all about being solving the biggest, hardest thing around me. And at some point in my life, I realized like, okay, that was really cool, but I need to move from solving biggest, hardest thing to, to really, to make an impact. It needs to be, you know, it needs to be not just hard, but important and important means understanding the relationship of your creation with the world. Like if you create it, what will it do? What will it impact? How will it impact? How do you measure impact? How do you know it'll succeed? So I think I really use impact to drive my focus, you know, not just sort of like, it'll be fun to build, it's hard to build,

but also is it actually important? And then if it's important and hard and fun, fun is just easy, then that's sort of really what's driven my focus. And for me, really last 10 years has just been all about the central question I raise. The data is there, the infrastructure is there, the technology is there now, or is very close to getting there. In many, many use cases, it's already there. So how could we mobilize this technology to move to a world where we're not operating the way we did 200 years ago? Our healthcare system today is just the way it was 200 years ago. You go in, you go in the doc, they listen to you, they record what you say, and then they give you what you need to do, all within that moment. And that's what we did 200 years ago. That's what we do today.

And that's kind of a little bit like earth shattering, embarrassing to me. Like, and I think we could be doing things differently. And we ought to be doing this faster. And now, that is shocking to know that and you're absolutely right to not leverage all of this knowledge, technology and resources that we have would be a shame if we weren't able to make that impact. So what would you say for our final bite, Sujit, what would you say is most something that you're most looking forward to in this field over the next five years? I would love in the next five years to be able to show that maybe we cut diagnostic errors by 50%. Wouldn't that be remarkable? I think if we move fast enough, I think that's possible.

That would be incredible. Yeah, absolutely. And I think the point that you bring up about the fact that we haven't made a whole lot of progress in the last couple of centuries. And the time is now. And really, I think it would be a marvelous outcome for the world if what you said actually became true. Thanks so much, Rashmi, for your very thoughtful and insightful questions. I really enjoyed them. No, thank you, Suchi. I think it was an absolute pleasure to have you on ACM ByteCast. And good luck for the future. We're super excited to see all the magical things that you're going to do

in the years to come. ACM ByteCast is a production of the Association for Computing Machinery's Practitioners Board. To learn more about ACM and its activities, visit acm.org.

For more information about this and other episodes, please visit our website at learning.acm.org. That's learning.acm.org.