How AI Detects Sepsis Before It's Too Late

0:00 Sepsis is leading cause of death in hospitals. You can't just like throw in an algorithm and hope you're going to see outcomes, right? That's just like adding chaos to chaos. Clinicians are already keeping track of 40 plus patients in their head. No other industry we expect this kind of heroism. So, I don't feel like our current strategy is sustainable at all. We got clearance as the first AI sepsis monitor. This was a huge regulatory milestone. We've seen 3 to 5% absolute reduction in mortality. This means 20 to 30% relative reduction in mortality.

0:35 >> At some point it will be considered malpractice to not use clinical AI products like Bayesian. Is that a realistic vision? >> You know, there's three parts to this. The first is >> We are here with Dr. Suchi Saria, who is a globally renowned health AI researcher. she is an endowed professor at Johns Hopkins running the machine learning AI health care lab there, but more importantly is the co-founder and CEO of Bayesian Health, which is at the leading edge of clinical AI.

1:07 you were on our podcast 3 years ago and the world has changed in so many ways since then writ large, but also obviously specific to the health care AI space. So, Suchi, why don't we start broad and then kind of go deep over time? What is the state of the state of clinical AI from your vantage point? what is the kind of the full spectrum of clinical AI look like these days and where does Bayesian fit in that in that landscape?

1:34 >> Absolutely. First of all, thanks for having me, Julie. I love this podcast. let's start by just telling you a little bit about what Bayesian is, especially for the people on this podcast who don't know. So, what Bayesian does is it's a real-time clinical intelligence layer that is making care from what is today a very reactive system to making it more proactive. And the way it does that is it continuously reads the full clinical record, so not just text data, but text, labs, vitals, clinical history, medications, treatments. And it's really making sense of that massively multimodal data to do clinical reasoning really well, and we'll talk a little bit about what we do and how we make that better. And then really tying that with you know, AI-powered workflows to make all the follow-through steps really easy for the clinicians.

2:29 In terms of the big headline, so one big headline since we last talked was that we got clearance as the first AI sepsis monitor. This was a huge regulatory milestone. Today all tests that are cleared, so sepsis is leading cause of death in hospitals. A huge toll in terms of lives, but also in terms of hospital clinical variation, how it impacts day-to-day utilization within hospitals. So, it's a problem where I I'd say like over the last 15 years many have banged their head against the wall.

3:05 It's an example area where I've done over a decade of research, and over 3 and 1/2 years in partnership with the FDA we worked really hard to get through all of the pieces needed to really show with a great degree of rigor what was needed to get this, you know, in terms of the quality of the solution, how it works across diverse sites. and again, can go more into that. So, that's really was a very big milestone for us. Another really exciting milestone is that we've partnered with CMS because of our outcome data we got FDA breakthrough designation, and we were able to then partner with CMS in a parallel review process where they've made preliminary approval for reimbursement of this work. This is a huge milestone for the field, not just us.

3:51 and then just stepping back a little bit, you asked me, where do we fit? Like what's happening in the field of clinical AI? I'd say what's been so exciting for me to see is how much, you know, it felt from my early days inevitable. I felt it was imminent. It's taken a bit to get there, but in the last 3 years what we've seen is like really rapid integration of what I would call administrative AI. So a lot of work, you know, when you're looking at the patient journey in terms of, transcription, billing, coding, and really if you just go to a patient journey, they're sort of like preparing to see the patient, see the patient, and, you know, opportunities and use cases there are based on LLMs and text-based AI that people are showing a lot of, you know, summarization, transcription opportunities, as well as post-discharge coding.

4:44 And really thinking about billing and coding and cleaning that up. I'd say in the middle, which is when you're seeing caring for the patient, where you really have to be trusted enough to get clinicians to change their decision or use the AI's input to inform their decisions. To me, that's kind of where the holy grail is from a care standpoint, that gap, and that's what we That's where we've been pushing the, you you know, the world and where we are seeing a lot of progress, and I think, that's the the layer we own. so yeah, so the biggest story I'd say top line from a Vizgen perspective is really you know, really showing maturity in very hard use cases where we've gone from early results to showing it works across many sites to really showing beautiful outcome data to showing FDA approval to CMS partnership with CMS around reimbursement. And then second, taking the underlying platform concept here and really now starting to show these results across multiple sites in many use cases.

5:43 >> Right. No, it's been an incredible journey, to watch, sort of the leap that, you know, you and and the whole space has made in the last few years. Can you actually give us an example, Suchi, of when you we talk about you know, what was hard 3 to 5 years ago in specifically in clinical AI that you know, now is is tractable and addressable by by companies like Beijing? Was it the data side? Was it you know, the actual LLMs and how AI has evolved? Was it regulation? All of the above? Like what were the actual drivers of the progress, would you say in the last few years?

6:13 >> Absolutely. I'll start with the result, right? So, I'd say for 20 years people have believed that like we need some kind of a monitoring layer. We need some kind of a second pair of eyes or an oversight layer on patients because there's just too much going on. And if they had the ability to collect this data, to make sense of it in real time and to turn those into proactive clinical signals, people could do something about, right? Like companies like Philips have existed, but historically, I would say there were companies that focused on collecting this physical data.

6:47 And providing and measuring like raw data. And then the big gap was going from that raw data or signal like you know, electronic signal we're collecting and converting them into actionable clinical signal. And to close that gap, I you know, over the last decade, decade and a half, there's been a lot of CDS work, clinical decision support type research that has emerged, including from the electronic health record companies, right? Like Cerner implemented, for example, St.

7:19 John's as their approach to react, you know, improving sepsis, for example. And so, the issue with a lot of that work was that basically the quality in terms of the signal, in terms of signal to noise ratio, was just so low that effectively people learn to basically live in what called what's called a hard facts problem. It's like we know this is a problem, but we also don't feel like it's something we know how to solve because it is very hard. So, traditional CDS adoption has looked something like 3 to 5%. Pretty abysmal. Like if you go to leading health systems like with strong informatics groups and you interview them and you'd ask them, "What is your BP order set BP is best practice advisory alerts is what they call them?"

8:06 Like which is historically a very short form for clinical decision support type systems and you say, "What's your adoption?" They'll say, "The you know, they'll be like five, eight, 10. We are partners but they tried and tried and tried and they were like, 'Wow, we have 2% in conditions like sepsis.'" So, what was really I'd start from the result. The big solve we've made is really doing all the hard work that's necessary to drive adoption. And what we've been able to show is across many, many partners that we've implemented at, we are seeing 85, 89, 90, 95. In fact, the partner I told you where they had 1 or 2% adoption, they now have 86% adoption within 6 to 8 months of implementing our system. So, that but to do that, what did we need to do to get there? So, the hard solves we needed to make to get there was one, go from dramatically improve signal quality.

9:03 And when I say dramatically improve signal quality, like and I'll give go back to sepsis in terms of like you know, you're looking at today most cleared tests are in the 20 to 50% sepsis prevalence regime, right? All cleared FDA tests today prior to ours was in the 20 to 50% cleared regime. What that means is prevalence regime, which means is if you were to see a pool of population post clinician suspicion, one in two of those cases would have sepsis or one in three.

9:35 When we're trying to build something that is running in the background, continuous monitoring, you're really looking at a 2 to 3% clinical prevalence regime, right? 2 to 3% really means needle in a haystack. Mhm. So, the ability to get the same quality of signal in that very low prevalence regime is very hard. And so, that signal-to-noise problem it was very, very hard to solve and what we needed to solve for that we've been able to show a lot of progress in. The second then, of course, is the delivery of these, right? How do you deliver these workflows so that you drive trust? Cuz ultimately, even if you get really high-quality signal, it requires human-machine teaming or clinician-machine teaming.

10:16 And so, next was how do you deliver it in a way that builds trust and transparency? Third, of course, the pieces and necessary from a regulatory perspective, but in a lab, you can do it once and you're like, "Okay, this looks good." But as soon as you're going for regulatory clearance, they want to see, "Well, you made it work at a place like Johns Hopkins. How about LifeBridge? How about like a mid-size regional community center? How about a rural hospital?

10:43 How about different settings like the ED or ICU or floor? Like, you know, they really want to see that the results hold up in very different sites, very different settings, very different populations up front. So, that's a huge bar. And then the second, you also want a system by which this can be practically deployed and stay performing, which means over time as there are drifts and shifts or populations change, you have a full plan in place to re- in real time identify, track, monitor, close. And that level of rigor is very critical in driving physician trust and clinician trust for adoption.

11:24 Of course, if you have high-quality tools that show you know, proactive signal combined with adoption, you're going to see outcomes. And so, you know, we've seen some really beautiful outcome data that's come out of it that I'm happy to dive into as well. >> Yeah. Okay, amazing. So, you're basically you're saying a couple of things. One is that your system when you sort of move the paradigm from a downstream, you know, kind of post a physician having a suspicion that sepsis is occurring in a patient to one where you can be far upstream of that and actually detect across the entire population where signs might not actually be visible necessarily to the individual clinician. And basically, so that's kind of one paradigm shift. And then what you're also saying is that even if we might have had the best technology to address these challenges years ago, really the last mile adoption problem was one of the huge barriers to sort of dissemination of these kind of capabilities in the real life clinical setting and that's another thing that you've solved by doing the hard work of integration, workflow mapping, all that kind of stuff. So, you also talked about the just the the level of rigorous research required. It's a ton of work, a ton of time, a ton of investment on the part of a startup that, you know, otherwise could be doing lots of other things with your resources, but you chose to go down this highly rigorous, you know, super high bar path.

12:35 Why did did you decide to do that path and why do you think that it was a right choice for a company like Bayesian to make those extra investments up front to then commercialize as opposed to you know, taking any other approach where you could have, you know, come to market as just a CDS as sort of a standalone software product that didn't necessarily require FDA approval. >> Our conviction from day one was that we're really changing the standard of care and not just introducing like another analytic piece of data, right?

13:05 Ultimately, the fundamental thing that many people are missing when they're looking at these database products, especially as it integrates within care delivery, is you're creating a new care model. And in this care model, you're explicitly thinking about how the signal and the the machine is teaming with the existing clinical team and and a full-blown, you know, a RACI for who's doing what and how, and a very clear understanding of risk risk mitigation responsibilities, right?

13:38 So, when you're thinking of it that and that's so important because that's what allows us to reliably introduce these kinds of technologies. You can't just like throw in an algorithm and hope you're going to see outcomes, right? That's just like adding chaos to chaos. We need the ability to integrate this in a thoughtful way. So, when you're thinking about a care model, you really then have to think about evidence. You have to really then think about you have to be able to answer with a straight face, "Why is this better and why do you know it's better?" So, I think when you go there and you're and you're really ultimately looking to do this so that you can drive clinician adoption and clinician trust, that's when you go back to what other proof points you need to make that happen.

14:27 And so, you're going to need to run the studies and so, you're going to need to understand like first, does this actually work in the regular in the regime that you're expecting it to? Does it actually track and monitor and identify patients early and in a timely fashion? What is the rate at which it's identifying? How do you know who who are the kinds of patients it misses and when does it miss? How does it perform in the ED versus the inpatient setting versus in different critical care settings? So, you really need that kind of data. Now, historically, I'd say most people would just say, "We can do it." And they'd produce a marketing deck. I think the landscape as you're starting to think about care model and how fundamental it is, people demand, you know, the market is maturing where they're they they demand data. If they're going to get thousands of physicians within a system to change behavior, they need to know that there's a method to madness.

15:20 From having lived within three large health systems done it doing a lot of research to us it was really obvious that if we really wanted to scale that this is the way to do it. Now, it so turns out that FDA regulation changed along the way. There was guidance they issued and in these high critical areas which are time sensitive, they explicitly are now looking for regulatory approval. So, we did it because it was the absolutely the right thing to do in terms of this very rigorous approach, but by doing so it also allowed us to be very well prepared as the regulatory regime changed to go through approval. And then next in doing getting that approval getting FDA breakthrough designation, we also then became eligible for CMS and tap reimbursement which has been very exciting for us.

16:07 >> Tell us more about that Sujay because we all we often times say that, you know, payment is the tail that wags the dog in terms of adoption of, you know, any either new practice or a product in the health care market. you know, describe that process like the the parallel path process by which you not only get FDA approval but also get reimbursement sort of coverage and, you know, have you heard any feedback already from the market from your customers from prospect etc. that, you know, sort of represents how people might view Bayesian differently pre and post the FDA approval?

16:36 >> Let's talk about the process first. So, in the process what they're looking to understand is a very clear so, first of all, is this really solving an unmet need? That's they want very clear evidence for what is the problem, how big is the problem, is it unmet by existing technologies and what is the problem it's solving? the next is is there evidence of benefit? Like is it going to show benefit? And then along the way, you know, they also want a financial model to understand how would giving this kind of reimbursement incentive would, you know, obviously the goal here is as we see new the new technology add-on payment in particular focuses in technology that are breakthrough technologies that can solve an important critical unmet need where this kind of an incentive would accelerate adoption. So, that's sort of the program overall. In terms of health system perception and bio really there are two problems we're solving there. The first one I'd say the most important one is signaling.

17:36 They've just heard so many people come and tell them, "Oh, I got this. I have results." Almost all companies that talk about results stop with the marketing deck or stop with a white paper that is a two-pager that they've generated. But, when you go underneath it, you start to see how did they do clinical ground truthing? How did they measure the metrics? What metrics did they measure? 99.999 false flat. They don't have any results on diverse populations. They did clinical ground truthing all wrong. They didn't measure the right metrics. So, really for us, you know, this kind of heavy lifting was really useful for us to be able to demonstrate we've done all the work to get you know, to show the proof. Of course, we treat we treat FDA not as the ceiling but as the floor.

18:27 When we What we mean by this is it's not like the be-all and all. For us it's the floor. Like it just demonstrates that you know, the level of end-to-end rigor. And I'd say devices that got cleared by the FDA 5 years ago, 10 years ago or devices that got cleared on a predicate that were started 10 years ago is very different from what the kind of clearance we're doing today because 10 years ago there wasn't an understanding of post-market surveillance the way it is today. 10 years ago there wasn't as much understanding of drifts and shifts and the kind of framework you need to really get it to work in the real world.

19:03 So, that's the kind of work far more rigorous and high bar we had to go through to establish that we have an end-to-end plan. And and that kind of is very relieving for our health system partners because they're seeing that we're not just you know, we're we're really leaders here. We're we're framing we've we've we've partnered with you know, agencies to really help define some of how what great looks like Mhm. And that we've cleared that bar, and it's signaling both from FDA, but also other agencies like CMS as we get you know, as we make more progress, and and then of course there's the financial piece of it. Turns out in many of the problem areas we're working, there's already a very high financial ROI component in addition to the clinical component. So, in some sense there's already incentive to some extent to do this work. I'd say the N tap just de-risks it one more layer, right? It provides like for every patient that's monitored, they get per patient reimbursement. The recommended payment by CMS is about $62 per patient that's monitored. So, that's pretty meaningful reimbursement for a system to and this is you know, to incentivize adoption of this technology. Yeah. And of course the goal is as we adopt, we will continue to create data that then shows you know, downstream efficacy then continues to substantiate the case for more payment.

20:30 >> Yeah. I want to unpack what you talked about on the on the drift and shift side, which you know, I think both applies to the underlying population. Obviously the the human population definitionally is is dynamic, and when you move from either health system to health system or even between departments, you might have different underlying characteristics of the of the humans that are being evaluated. Plus obviously the you know, definitionally AI models themselves and software, part of the benefit is that you can update them as we learn more you know, to kind of tune the model you know, and/or the new new new the foundation models come out that you can take advantage of. what is Can you actually just explain to to folks who might not be aware like what is the framework that the FDA is using today to be able to even contemplate products that have those dynamic characteristics in them as opposed to the historical regimes of more static products, like you you were talking about devices and more traditional labs and things like that. Those are obviously much more static in nature than the products that we're talking about today. And where do you think there's still work to be done to evolve the FDA frameworks to a point where they can most effectively contemplate products like what Bayesian has?

21:34 >> In terms of thinking about sort of what the process looks like today, there's several different questions they're trying to answer. The first question is is it safe? Next question is is it effective? And really in answering both safety and efficacy, they've dramatically expanded the scenarios in which they're testing and evaluating your product, especially if you're doing something new. So, not something where you know, single image, they've been doing single image for like 30 years, 20 years.

22:04 And you're using an old predicate. So, more straightforward. But when you're trying to do this multimodal, lots of inputs, that kind of data, there are lots of new considerations because there aren't as many cleared predicates there. So, when you're doing this and really everything I'm describing should apply more broadly to a lot of AI devices, but the point is the way FDA works, there's always a notion of like, you know, there are different vehicles for clearances and every time there's something new, there that's when all the scrutiny, right? Like you get to push.

22:36 So, the the thing they want to know is first of all, you know, show that what is the indication for use? Who are you? What population are you serving? What is the claim you're making? And then show me evidence that it's doing what you said it was doing, right? So, in this particular case, if you wanted to show early detection prediction, we had to show in high large-scale clinically grounded cases, and clinically grounded means panel of expert physicians have gone and evaluated these cases. You're showing me what your performance would look like in the real world as you would implement a system like this.

23:10 The next thing they want to see is that show me how this would vary across different populations, right? People with different comorbidities, people with different age groups, people with different kinds of you know, racial races. So, that's The next thing they want to see is how does it vary across different settings? So, like ED, critical care like all of the settings in which it may be applied. The next thing they want to see is how does it vary across type of site? Like rural, community. So, they're really asking for a fair number of deep dives to understand performance. The next thing they're asking is how does this degrade? So, like how much data has to be missing? Can you show me things about different settings in which things could change?

23:54 And some of some of the times they even are pushing it to create change that feels pretty artificial. It's unlikely in the real world. But they just want to push it to see how will performance change and vary so that they understand the limitations of the device. So, those are all like so they're pushing in a variety of ways. They're also talking about what if this type of thing was missing, that type of thing was missing and so on and so forth.

24:17 The next thing they do is then say, "Okay, now that we understand how it would perform in very, very settings, I want a post-market plan for surveillance." So, they want to know "How are you going to measure at every site what your performance looks like? How are you going to know that it meets some test criteria? What is your criteria? How are you going to measure that? How much data are you going to measure? Then, depending on sort of what the operating points are, what are you going to do if it doesn't meet?

24:43 What are you going to do to risk mitigate? They need a very clear plan for all of these scenarios. Then that So, that's very real, right? So, it's like pre-launch ongoing now monitoring. Tell me all the different ways in which this can degrade and what is your plan for catching each one of these proactively, and what is your plan for acting on these? All of these have to have a very prescriptive plan. So, that's a fair amount of work, but it is actually the work you really truly need to do in order to deploy these systems reliably. It's not a let's see what happens kind of strategy. It is a we have a strategy.

25:14 >> Mhm. >> And then finally, they also introduce this tool called the PCCP, which is predetermined change control plan. I'd say that's like the newest tool in the quiver from an FDA perspective where they can approve a PCCP. What the PCCP does is allows the company to partner with the FDA on what are the kinds of things where you could take your tool apart from what has already been approved. There are new things you could do, changes you could make that as long as you follow a particular protocol or process, you could do without a full submission.

25:48 And so, that's the next thing. And so, you know, and I'd say their comfort with what is allowed under the PCCP is where I think there's opportunity for more openness. Ideally, there's lots of things we can do in a very rigorous fashion using this kind of a protocol for improving and learning over time. But today, to get a PCCP cleared, you have to get your review team to get comfortable with pre-approving all those kinds of changes. And that's where I think they're a little bit you know, it's a little bit baby steps.

26:16 Like you get them to approve comfortable with some baby version of a PCCP, and then you're getting more and more and more rather than free rein. Like we trust you. >> Yeah. No, it's they're effectively trying to strike the balance between the the like sort of the point of using a software AI-based approach is that you can be dynamic while, you know, trying to balance the the safety side of it, which makes complete sense. I mean, you and you're talking a lot about what effectively rolls up into kind of governance of AI, which is obviously a super hot topic these days, especially with everything happening with Anthropic on the on the kind of the broader generalist model side. but, you know, obviously much more consequential in the healthcare care W- and you were one of the founders of Chai, there's there's a lot of sort of organization activity around, you know, how do you create these governance models that, you know, implement oversight when clinical AI is being used in the real life setting to ensure safety and and and the appropriate level of reporting etc. What do you think you know, first of all, I guess are there examples in the real world that you have seen where governance has been implemented in what you would call, you know, a pragmatic way? And, you know, what's your point of view, Suchi, on where does the responsibility lie? Like you could say individual clinicians have a responsibility to govern themselves and their use of AI.

27:28 you could say the institution, like the hospital systems that they work for, are really the ones who hold the burden. You could say vendors and you know, companies like Bayesian. You could say the government needs to play a role. Like where where do you think the the center of gravity of responsibility lies? And again, have you seen any examples of implementation in in the wild that you would point to as a great example of how to do it right?

27:48 >> I'd say today feels a little bit like the wild wild west, which is most health systems are being left to themselves to kind of own the risk, manage risk, and figure out what needs to happen. So, I'd say today's model puts pretty much all the burden on them to figure out what needs to happen. The absence of which is either they're overwhelmed and scared, or they're relying on some third-party external vendor. But, I think there is a little bit of a knowledge gap between what done great looks like versus what most governance vendors are able to provide because none of them themselves have gone through any kind of FDA approval to really be able to bank on they understand what governance is. I'd say today what's been really interesting for me to watch, just from having done a ton of research in this field, from having written a you know, partnered with various federal agencies, one thing I find really funny is I feel like there's a lot of governance players today that are like offering a layer of protection and governance, but I I'd say they're pretty like thin in terms of their own understanding of what excellent looks like. So, a little bit they're like flying and building the plane, flying the plane as they're going.

29:03 I'll step back and say what I believe needs to happen. I do think the burden needs to shift way more on the vendor. Mhm. So, when you're going, I'm not saying this is a practical strategy that scales to every solution, but let's start with the easy things, things we know. In many First First thing you want to do is you want to match your governance to the type of product and the type of risk. So, as soon as you're doing anything patient-related, care model-related, you want to make sure there's an element of good understanding of risk, good understanding of governance, and good understanding of a risk mitigation layer. Now, many people confuse administrative applications that do ultimately have impact on patients like denials as administrative, but I would say those very much have impact on the patient. So, I'd put it traditionally less as administrative clinical and more thinking about is it impacting patients and patient care. And if it is, then risk needs to be thought about a little bit more carefully. Now, in that area, not every application is today regulated by the FDA.

30:07 they are they I don't believe they're in a place to be able to do that. On the flip side, there's a very good opportunity to learn from what they're doing. So, I think the ability this kind of rigorous framework and chassis you need to be able to do this diverse set of evaluations upfront prior to go live, and then post go live having an end-to-end chassis for really deeply understanding, you know, how do you monitor and then tune and retune and risk mitigate. To me, all of that needs to be a requirement for every vendor that is implementing. It It so happens that in our case, well, we we built the plan. We have a plan.

30:48 But we had to an honestly get an agency like the FDA that is very very rigorous and detail-oriented to then verify and validate a plan, right? So it was exciting for us that we could do that. And that helps us in building trust. But not every area is FDA regulated. So I do think there is an element of every vendor needs to have that level of accountability. And in these scenarios where they don't, basically the health system is assuming the risk. Because ultimately when there's going to be a malpractice risk case, somebody has got to cover for it. Now, I do think a lot of health systems today are very much a little bit overwhelmed.

31:23 There's too much going on. So there's an element of let's try it. And I think this is where there's an opportunity I mean, it'll be interesting to see how things evolve. I do think for Bayesian, what's been very exciting is by having this end-to-end chassis, you know, expanding to a whole host of really high-value use cases, we're we're really becoming sort of this you know, continuous real-time clinical intelligence platform that isn't just about algorithms. It's really about an end-to-end solutioning for driving adoption and outcomes in key core areas that they can like, you know, deploy with like sleeping at night, not needing to worry that crap, what am I doing?

32:04 >> You're also getting at an interesting point and there's a sort of line that we talk a lot about in health care in particular where at some point it will be considered malpractice to not use clinical AI products like Bayesian. What's the is that a realistic vision of, you know, what could what could ultimately happen with these products? Like what do you think is the what needs to be true for us to get from where we are today to a place where literally it would be, you know, considered not standard of care to to not to be not using these products?

32:33 >> You know, there's three parts to this. The first is like in many conditions like sepsis, like a number of other, you know, these emerging critical care like clinical decline areas where there's opportunity for proactive rescue. What what's very interesting is when you go back and look at the clinical data, the signs were already there. And then the way malpractice claims lawsuits happen is basically someone is arguing that there was you know, a scenario where the signs were already there.

33:05 >> And so it was missed. >> Or the you know, what would be considered standard of care was not applied. So now, if you have a system that doesn't really work, that like effectively is primarily ignored because maybe the false alerting rate is so high or it's mostly noise and physicians don't trust it. And in fact, there was a really interesting case at Mount Sinai that was like that. Where the system, it was written up in the New York Times. I happened to make meet the parent of this child who this happened to. Where you know, the the EMR alert went off for sepsis.

33:40 Cuz you know, clinicians by and large were ignoring these alerts. So they just dismissed it. But really the question from a parent's point of view was this patient, my son, had signs. >> Mhm. >> My son even triggered these systems to to say, "Hey, you got to do something." But it was ignored. And so how does a system like scenario like that get resolved? And I think from a system perspective, from the clinician's perspective, they say, "Well, it is a standard of care to ignore these because we think they don't really they don't really work." But that's not a very good answer, right? Like you the fact that even a simple system saw that something was wrong, prompted it, and you ignored it, that's problematic. So to step to step back a little bit, really malpractice is all about what is standard of care. What is expected to be reasonable practice? I'm not a lawyer by training, so I may not be using all my words correctly, but roughly like what would be considered a good doctor or a good care team? What would they have done and did you do that? And in this example, if the signs were already there and maybe even a system managed to flag it and you ignored it.

34:51 That's bad. So, the way to think about this practically from a health system's perspective is if they have a system that is really just like going left, right, and center flagging every case, that's really bad because now you're mounting your case for malpractice risk because now if your clinicians are ignoring, that actually makes you look worse. Right? So, that's interesting to me. The second is that alternatively, if you have a system that actually has very high signal to noise, in other words, it's very good at catching these cases, catching them precisely and not crying wolf all the time, now cases that you would have previously ignored or signs you would have previously ignored now are getting caught.

35:29 Because your clinic it's manageable, clinicians are coming in and doing the right thing. The great news is if they're going as you get high adoption, they're doing the right thing. Great. You're improving patient outcomes, you're missing these cases more less often, but then even better, in the scenarios they disagree because the machine is not always right and and that's just not true for you know, in most cases there's still opportunity. This teaming where they can document why they disagreed and you know, and make that make a system that's very easy to do. So, really it's a much more robust way of practicing where you have a second pair of eyes, it allows your clinicians to have peace that they did everything they could.

36:06 They you know, critical moments weren't missed. They were able to real-time rationalize what was going on. And that now allows to both improve patient outcomes, but from a malpractice risk standpoint dramatically improve the outlook. And in many of these cases, these are massively expensive lawsuits. Mhm. >> Yeah, and that brings us to maybe our last question is just on like what is the impact of clinical AI will be on the clinical workforce. So, you just described a number of scenarios in which the way that a clinician might operate in a in a sort of point of care setting might need to evolve to ensure that, you know, these tools are being incorporated appropriately and that they're being responded to accordingly. And in in general, you know, we're seeing across a number of aspects of the economy this notion of, you know, jobs being unbundled and rebundled in really interesting ways. Like even in the in the tech world, right? We see the jobs of the product manager and the software engineer and the designer, you know, really evolving in in pretty pretty interesting ways because of the introduction of AI into these internal company operating environments. What would you say, Suchi? What what do you predict, I guess, in terms of how clinicians' roles will evolve and are maybe are evolving in in real life if you're seeing anything on the ground related to this through your work at Bayesian that, you know, will that will need to happen to sort of appropriately set up our health care system to be able to absorb these tools in the most productive way possible. And you know, what is the role that companies like Bayesian are playing in that? You know, a lot of this is stuff beyond just the technology that we all need to get right in order for our health care system to be able to fully take advantage of said technology. And you're clearly at the tip of the spear of that. So, what are your thoughts on on sort of the workforce implications?

37:42 >> I think let me start with today clinicians are already sort of monitoring or keeping track of 40 plus patients in their head. No other industry we expect this kind of heroism. So, I don't feel like our current strategy is sustainable at all. Like we need a new way. And now the question is why why AI and can we get it to be at a place where it's doing this well? So, I think the to me a future is one where the hospital becomes proactive instead of reactive, right? There's a continuous intelligence layer that's really looking at every patient longitudinally stitching together data, understanding what is typical or normal or baseline for them, understanding what looks anomalous relative to their own baseline and is able to then in time with high-quality flag that and make it possible for clinicians to act.

38:30 The next thing is that by doing this well, quality now goes from becoming just another campaign to a durable strategy. What I mean by this is today a lot of like, you know, patients are becoming only more complex harder to take care of. I'm going to go into hospital finances because, you know, no margin, no mission. So, ultimately, like hospitals aren't getting paid enough to take care of the patients they need to take care of. And so, you do really need the ability and like these cases that are missed really turn into a lot of excess days in the hospital. So, not only are they bad for patients, they're also bad for the hospital bottom line in terms of excess ICU utilization, lots of readmissions. Basically, a lot of excess cost, which is not only burning out physicians, burning out the care team members, it's just ultimately not a good durable way to provide care.

39:24 So, the ability to go from like yet another campaign or another checklist or an another education, which very much fades away after a few months, really hardwiring this kind of infrastructure to make the right thing to easy to do proactively, I think is where, you know, we need to go. And I think ultimately we do have to think of it as not just software technology, we have to think of it as care model redesign with AI as a part of it.

39:52 >> So, Suchi, we've been talking a lot about sepsis, which is obviously a, you know, significant use case and one that we're already making a tremendous impact. but your the vision behind Bayesian is obviously much bigger than just sepsis. you've now moved into a couple of new areas. You just had a big announcement about some new use cases that you've deployed with with Mayo around deterioration, palliative care, etc. Can you just talk to us about what is your rubric for evaluating new use cases to incorporate into the Bayesian platform.

40:20 What makes for a good use case to apply, you know, clinical AI to versus versus not? >> So, let me tell you say a little bit about a couple of examples just so it helps understand like what's what's a good use case, right? So, first is we really think about it as largest drivers of mortality and utilization for health systems, right? So, mortality and utilization. So, what are some areas and it turns out there are 150 plus areas like this where if you could have identified this case more proactively and put individualized targeted smart workflows based that was targeted to patient context, you could have dramatically changed the patient trajectory.

41:01 And in doing so, not only are you improving patient outcomes, you're changing the utilization trajectory for that case, too, in terms of ICU use, readmissions, length of stay, capacity. So, biggest drivers of mortality and utilization and that's sort of really where we work backwards to figure out. Almost every use case we have has three key primers or levers of benefit. First one is quality clinical outcomes. The second is financial understanding the financial benefit. The third and financial benefit to the system. And then the third is frontline clinician can we streamline workflow?

41:38 Can we build something in a way that the right thing becomes the easy thing to do and we are unburdening them of what are very complicated workflows today. So, we're saving them time along the way. So, those are the three levers for every use case we go after that we're looking for. The next thing is we're often looking for conditions where these are easy to miss or were missed. And so, some examples and I'll give you some very different examples, right? So, we spoke a lot about sepsis, super hard problem, everyone's banged again their head against the wall. Super exciting to see results where we're now consistently seeing it big, small, large, diverse hospital settings, we've seen 3 to 5% absolute reduction in mortality.

42:16 This means 20 to 30% relative reduction in mortality. We're seeing you know, 1 to 2-day reduction length of stay for this population. Really, really exciting results on average. And then, you know, of course, some more, some less. We're also seeing totally different use case, unmet palliative care need. So, patients with advanced illness very easy for them to not get the proactive care they need from you know, palliative care clinicians. The hard part is they're everywhere. They could be anywhere in the system. It's really easy to miss them. It's very hard for the frontline clinicians who are really, really busy to, you know, and in the moment kind of treating what's happening right in front of them instead of thinking of the longitudinal trajectory of this patient.

42:58 The fact that this patient's been bouncing back over and over again, maybe the thing we're doing isn't working and we need a different strategy. So, what we've seen this is a multi-year collaboration with Yale where we recently announced some really exciting results. We were able to see within 60 days 25% relative 25% reduction in readmissions. Huge. They had almost 200 bed days you know, capacity that was soaked up from these patients who kept coming back, coming back and wasn't even the right side of care for them. It makes the patients unhappy, it makes their families unhappy. It's It's just bad care. And so, this was a great scenario with this kind of proactive clinical monitoring and clinical intelligence that led to the workflows dramatically change patient outcomes and system outcomes.

43:43 yet another very, very exciting area where we have super cool results at what I would what we call proactive rescue. There are 98 different odd complications that lead to a patient you know, emerging sudden decline that can lead to an ICU escalation basically, you know, or some kind of organ damage, organ failure leading to often hospital you know, death. Terrible. Turns out our data shows that in many of these cases you can identify these patients almost 12, 14, 20 hours earlier. And if you can identify them, and based on patient care context give the right proactive supportive therapies, you can basically change the patient's course.

44:24 You know, not only are they living, but they're living happier, healthier, and it's very impactful for health systems that are implementing strategies like this. So, really cool to see how diverse the opportunities are. >> Well, Suchi, it it is incredible the work that you're doing. thank you for everything that you're doing to you know, both enable physicians and clinicians to practice more proactive medicine, to you know, enable health systems to deliver more high-quality care to patients, and obviously for the patients in terms of of the of the impact that you're making on their on their health. great to have you here as always.

44:57 >> Thank you, Julie. This was super fun for me.