From Code Blues to Digital Twins: How AI is Transforming ICU Medicine — The AI-Ready Doctor

Hello everybody and welcome back to the AI Ready Doctor podcast. Now, the ICU is a place where I work and life and death decisions are made there. Alarms, ventilators, pumps, each demanding attention, over 100 details per second that we integrate in those decisions. And people rely on us. We are their best and sometimes only chance. In the middle of that organized chaos, the real question is not what's happening right now, is what's gonna happen next and how can we stop it? So today we're gonna be talking about how AI in the ICU might help answer that question. As usual with this podcast, it's mostly about tools you can use immediately. So I'm joined today by someone who I'm privileged to call friend, lives and breathes critical care. That's Dr. Viren Kaul. Dr. Kaul is Chair of Medicine at Crouse Health in Syracuse, New York, and Division Chief of Pulmonary Medicine. He's an assistant professor at SUNY Upstate And I have absolutely seen and even voted on the award-winning in education and a leader in mechanical ventilation, airway management, ICU innovation. So he's one of the few intensivists that can pivot from talking about ECMO all the way to talking about foam med movement online. So sometimes all in the same breath. So Viren, welcome to the AI Ready Doctor podcast.

Azan, thank you. Thank you so much for having me. And I will echo the sentiment that innovation is something that brought both of us together. When I was a fellow, I loved talking to you about new things that are coming up and because I think that's our field. And I wanted to congratulate you on getting advanced certification in the field of AI in healthcare because Just like back in the day with social media, we need to start talking about this. We need to start understanding this field so we can use it for our benefit and our patients' benefit.

I appreciate you saying that. Yes. Hear, hear. Maybe louder for the people in the back. So I'm gonna start big pictures and you've got years of experience at the bedside. You led teams sometimes in most intense environments, code blues and admissions and intubations. So AI in the ICU, really, honestly, I mean, beyond the hype, what's real? Real talk.

So when I think about things I do daily, simple things, infections, right? Before people get sick enough to come into the ICU, there's a pattern to it. So there's studies showing that 25% of patients who get admitted for sepsis, for example, they've had an outpatient visit contact with healthcare within the previous week. And then 50% of patients who get admitted with sepsis, they've had an abnormal vital sign captured somewhere in the last 24 hours prior to showing up. So, I always keep thinking about how do we prevent things from getting this sick? I think for me, the best success of my profession and my job would be if I'm not needed. Which is a crazy way to think about it. But if you don't get to see me ever again, that would be great. That means humans are not getting that sick. So problem is these changes that occur before you get critically ill are tiny, they're subtle, and they're too small for the human eye. I sometimes joke that they're like the prolactinomas, micro prolactinomas of the endocrine world. You know they're there. You see that there's an impact on the endocrine profile. You just can't actually see them even on imaging. I would love for us to have some tool that can help us capture those changes early, help us predict them and their impacts early. And I do think that we are there and human brain just isn't there yet. So.

I love this. I think in the ICU when you first start, You're excited to do stuff, procedures, intubations, and the more experience you have, you start to do less stuff because the best ICU saves are the ones that happen even before anyone pushes the code button or an alarm starts. So let's talk about that. Early warnings. I spoke before about one of the strengths in AI is predictive analytics. I even joked that I want to do predictive analytics in fashion and see what's the next Devil Wears Prada going to show. But in the ICU where people's lives are at stake, what are the value of earlier alerts? Are they happening? Can they be done?

So think back to the time when we didn't have qSOFA scores or SIRS scores and we were out there looking for how do we find patients before they're getting sick? We know they're getting sick, even when they're in the hospital. Now just take it a little bit further. Some of these scores are based on 2 or 3 or 4 variables because that's what our current computational and brain power could handle. Now think about this. If you can expand this to other metrics that change in a patient, not just the basic stuff, a little bit further, right? Skin color or variations in blood pressure or variations in temperature, not the absolute temperature, the variations, things of those nature. If we pick them up even earlier, can we prevent that deterioration that you talked about? The chall— and then there's a number of scores that are already out there. So people who use Epic for their EHR know that Epic has a deterioration score, which for those of you who are not on Epic quickly, it, basically looks at trends every 15 minutes in a patient's chart and gives a score from 0 to 100. And of course, higher the score, more chances you have of experiencing an RRT. The patient potentially can end up in the ICU or experience even mortality. So the question becomes, yes, you can pick up early, maybe a little too early, but is that helpful? Will that change how we intervene And will that be something that we can manage? Will it, or will it add cognitive load? Because cognitive load is a very, very real thing.

Real.

And I don't know the answer to that. And I worry, and you're right, I worry. Because when EMRs became a thing, clearly there were advantages, right? We could read each other's handwritings. We could track, we could make graphs, we could, and we do all of that. And it does help, it's more portable, but we also know what happens when you get 50 alerts. So I don't know the answer to that, but also why we need to be involved early. Can we train this tool to tell us exactly what we need when we need a little earlier than we do right now and doesn't bog us down?

So it's like that friend that tells you, or that colleague that tells you, hey, watch that patient. It is helpful, unless it is said 50 times a day. At that point, it's like crying wolf too many times. So there are validation studies for sepsis scores to actually get those patients attention early. There is even a study about cardiac arrest and predicting it a few hours early, I would say if I can tell you that the earthquake is in a few hours, it's better than not, but it's even better if it is 3 days before. So the prediction score, I would agree that it's not ready for prime time. It needs to be validated so it's not just a solution looking for a problem, the next shiny object. We need meaningful clinical outcomes. So I also want to throw in a futuristic example. In my lectures about AI in medicine, there are companies that already are leveraging machine learning, which in verbiage for AI, it's one of the types of AI, and that is the machine learning which gives you predictions that you can be actioned upon later. They pick up on what the naked eye cannot see. And in this, there is one company specifically that I think about, and I've given talks about this before, that is, they are in the ARDS space. And in the ARDS space, on the ventilator-human interface, where based on the CAT scan that is done for the patient, based on the vital signs for the patient, they actually create precision therapy ventilator settings through high-fidelity digital twins, and it integrates into the interface of the ventilator. And so from the Building digital twin, they can actually look at regional stress in various areas of the lung, and they run models on a digital copy of the patient. This is precision medicine at its best. It's been validated now, but where do you see this in your skeptic but not cynic view of AI in the ventilation of ARDS?

So I just wanted to really quickly go back to the concept of a digital twin, if that's okay. So for those people who may not, this is not actually, again, I feel like a lot of words are fancy currently, but really what it is is a simulated high-intensity but high-fidelity environment. So think about autonomous cars. So before you put them out on the road, if you want to see how would they perform in a real-life setting. So, you know, you create a track with dummies, with obstacles, with slippery surfaces, right? And you put that automated car in that setting and see how it performs. So it's a little bit similar in that you're recreating in real time, not learning later, but in real time, your patient's physiology and pathology. So, what happens is we know and we have unfortunately because the recent pandemic understood and we know that different causes of ARDS behave a little bit differently. ARDS at baseline is a very heterogeneous disease. It's heterogeneous in how it involves the different parts of the lung. Then extrapulmonary factors come in. Your habitus, your positioning, your hemodynamics that change how you respond to this ARDS pathology. And on top of that, now you have ventilators. You have variety of different modes. You have a variety of effects of those modes. It's a lot just for one patient. It's a lot of data to think about today morning, and then it's going to change in the next hour. And it's gonna change in the next hour. Now, we don't know what's gonna happen in the next hour, and there's no takesies-backsies with real people. That's the thing, right? You always think about what if I had done this differently? You can now. What I think the promise is, you can do that now and say, this is what I have, this is all the data, this is this patient, this is how we are ventilating you. Put all that information together for me based on literature that we have over all this time. Critical care is not new. It's been there since the polio pandemic, right? The iron lung. What would potentially happen if I was to set a PEEP of 20 versus if I was to put you on APRV? Am I really gonna compromise your preload or am I gonna get the benefit of recruitment? Today, you make a choice and you get a result and there is no going back. And to me, that is absolutely the gem of the future. Give me a chance to do better for my patient. So I think it is the humble counteracting factor. I think it's funny. People worry AI is going to compromises, I think it may humble us in our hubris in critical care.

Absolutely. When you think about how far we've come, but so far it has mainly tested theories and clinical trials over a long period of time to give us answers that you can just extrapolate and judge on. But if you had the ability to do testing on models that are digital, without affecting the patient to see how, at least theoretically, in a simulated environment, which is not the real environment, it gives you an extra layer, which is why AI will never replace physicians. It would just be our copilot when we do this. In the ICU, we have decision-making, we have rounds, we have integration of a lot of details. How do you think some of the tools that we have— in my previous episode, we talked about some of the tools like Open Evidence, maybe ambient scribes, whether it's Abridge or Freed AI, or even the new Doximity Scribe. How do you see this helping us, or is it just a new shiny object in our rounds in the ICU?

So, I think there's two elements for me on this. One is the clinical application integration, but the other side of what I love doing is teaching. And I often talk about the perfect round that would start at 8:00 AM and at 10:00 AM, we would see all the 30 patients. I would get complete histories and physicals from everybody because I want to know what you know. I would teach you on every patient, but I would also see all the patients myself and we would yet be done at 10:00 AM. And we all know that it's just not possible. So I think about why it's not possible. So for me, number one, it's interruptions, right? They're just the nature of the beast. And the reason interruptions occur is because we have to get— take that phone call. We have to answer somebody's page. I am hoping that a lot of that can potentially be automated. So that is one part of EMRs that I think is untapped. So if you put in an order and there's an error or something needs to be changed, can we automate some of that and automate it in a validated manner, decrease interruptions? That would be my absolute vision and dream. But what's already happening? Let's talk about that. Bedside things, bedside tools, ultrasound, we all have it. I'm hoping we are all using it. And even when we are not using it in, say, austere settings, how do you use it and how do you teach people to use it? So there's already tools out there. So for example, Butterfly has an AI-associated product, right? And it can help you to learn the two hardest things with ultrasound, acquisition and interpretation. So, that AI helps you say, change your probe position, tilt 30 degrees a little bit, tilt 20 degrees a little bit, so you get the perfect parasternal long axis view. Oh, you got the view. Now, let's interpret. It's another friend that in a high fidelity manner teaches you application of ultrasound, right? So that's a great example of doing both. So it improves your clinical application of a very impactful tool. It also helps you teach. And I think that would be great. And then as far as rounding tools for learning, you've talked about OpenAI. I jokingly tell people this. I was like, back in the day, we used to have to use paper, then we use calculators. Now you have MD Calc. Correct. What if, just saying, you know, all those multiple blood gases that get sent out, and this is a pet peeve of mine, we'll fight the battle even in 2050. I know this. They're still going to be getting sent. But at least maybe now this tool will put together and tell you, hey, the AA gradient's changing. It's changing right in front of you. You gotta act on it.

And then it'd be nice to have it real time.

Uh, real time. And now you teach about AA gradient. So that's what I see it. So tools you already have, uh, taking in all the information that's out in all these books, putting it at your fingertips and putting it all together and, uh, using that to make rounds effective. I think that is where I see this going.

This is one study where they actually show that there is a language model, but for echocardiography. So as the echocardiography images are being acquired, your preliminary is no longer just an ejection fraction. It is an AI interpretation subject to the human model of overlay and human in the loop where they proofread it in the end, much like we do now, but it is much, much more real time. And so for me, like you, I would like to see this enter our ICUs to help us with clinical decision support.

Yep.

Do you use currently ambient scribes in your ICU?

Before we go on that topic, actually, as I was thinking, about education and AI. Just wanted to go back really quick. You know where I really, really think AI is going to be super helpful is in high intensity but low sort of occurrence situations. So, as I said, I do OB critical care. It's an absolute lifelong passion of mine. So glad I get to do it. We're talking about really young moms, Usually it's no-fault injuries, right? And that's why it bothers me, right? And I want all of them to do well.

So this is an obstetrics ICU, correct?

Yes, right. Obstetrics in the ICU. And as you know, we— it's such a, it's such a tough area that a lot of studies don't even include pregnant moms.

You are correct.

Okay. So, we are now in this tough area evidence-wise and clinically. And now I want to teach the next generation of medicine, critical care, obstetrics, and MFM trainees, right? About amniotic fluid embolism.

Yeah.

How do you do that? So, you can simulate it. You can simulate it. The tough part there is that it's, when it happens, it's so drastic and it's so infrequent. You're either ready or you're not ready. And the problem is that if you end up being there for that patient at that time, right? You're gonna have such a different approach to that situation depending on who you are. So if you're an anesthesiologist, different approach. If you're the obstetrician who's already treating the mom, doing the C-section, for example, you're in a different cognitive skill space. If you're the intensivist who's been called, if you're the resident who had to respond to that code in the OR, everybody's coming from such a different background to address this one problem that is so severe, so no fault, and so rare, but needs everybody to work together. How do you teach that in a in a simple single simulation. You can't. For me, that's where I do think AI will shine. So you do run the same sim. Everybody is part of the same simulation, but the AI can give personalized feedback. Excellent. So as a medicine, let's say as an intensivist, my fellow critical care fellow, when the mom's blood pressure dropped, your thought was XYZ. You were thinking hemorrhage. You were thinking this because that's your training. But you know, that's a natural gap. You as an intensivist need to keep this in your mind. Whereas as an OB, you were more worried about making sure the mom doesn't have surgical complications. So you maybe need to focus on that. And as an educator who was monitoring that simulation, how do I focus on this wide variety of trainees all in one go? I can't.

And it does have the empathy in medical education as well. It's literally like you are at the gym and you have your spotter. They cheer you on, but at the same time they do give you tough love and feedback in a way that is personalized to you. Yeah, phenomenal, phenomenal. I do see where this could be saving lives basically, because it prepares you for something that you'd never be otherwise prepared. In the ICU, and whether it's education or decision-making, we also have a tool that you've worked before. You and I worked in the neurocritical care unit, and when somebody comes in— we talked about that in the recent podcast— with a stroke, that CAT scan read is absolutely crucial to be done in a significant increasingly small amount of time for those that make decisions and take action to do it within a shorter and shorter frame of time because time is brain. Tell us about these tools that you used. One of them we presented before, Viz.ai, but maybe others that you have now.

Yeah, so we at our institution have had experience working with Viz.ai in the past. We have recently moved moved on to another tool called AI Doc. And some of them were just organizational decisions. It was not one over the other. So what I will say is that number one, this technology picks up something you already have, a CAT scan. It doesn't take that long to scroll through it. So you can't argue that it's so much looking at the scan. It's picking up the abnormality.

Yeah.

Right? Is that, no, this is bad. We have to act on this. But you know what's the hard part is we talk so much about multidisciplinary care, MDDs.

Yes.

Now MDDs are great, more brainpower, more expertise, more interests, but you know, they also suffer from sometimes what I call the challenges of too much democracy. Because someone does have to ultimately call it.

Yes.

Someone has to say, we are doing mechanical thrombectomy.

I think in residency we call them the captain of the ship. Who's the captain?

Who's on first? Right. So who's on first? Now with these programs, the AI is not on first, but once the AI calls it and says there's a large vessel occlusion and having worked with your team and previous interactions with the team members, it says, You know you are going to activate your lab. Okay. Now there's not a lot of confusion. You know you're gonna almost certainly end up activating the lab.

Yep.

And all those backs and forths, should we come in? Should we not come in? Should we, you know, who is coming in? Who needs to know? Does the ICU need to be called? It, it just gone.

Yeah. Streamlines it.

Just streamlines it, right? Yes. And it takes the guesswork a little bit out there. And it also, I think, interestingly creates very committed partners to care because you're on another whole nother platform besides your EMR, paying attention to just that one patient with your colleagues in real time at that moment.

Absolutely.

It's almost like it trains you.

This AI is moving into the COPD space. What's your take on that?

The way I think about it is it's somewhat similar to Epic, right? So when Epic first came out as an EMR and just for, we should probably put this out there. I personally have no conflicts with any of these companies. I'm just curious as a person. So I remember when Epic was first launched a long, long time back when I was a resident, it was one program. And over time, there are modules. You can add Beaker, you can add Slicer Dicer, right? You can— so as companies grow, as the technology grows and it evolves, it gets more capabilities. So that's how I see it, right? So Viz.ai has come out into the COPD space. A lot of other such companies are going into the PE space, for example. Port activation. And I think that's just natural growth. What I skeptically worry about is I hope they don't expand so much that everything is sort of, yeah, we can help you pick up the PE, but then like you mentioned, there's no application of it. And then a lot of clinicians will just lose interest is what I worry about. So I would love to see these companies, including this, Pick that product, fine-tune it to the point of deliberate application validation, then move to the next tool, then move to the next tool, and then move to the next tool. Ultimately, it'll come together. But I think it's exciting. The COPD module you specifically talked about, the goal there is to identify essentially disease burden, right? You have You know, when we talk about how many exacerbations have you had, how many inhalers are you on, ability to use and adhere to inhalers, again, it's a very complex decision-making process. And with how burdensome the disease is and how prevalent it is, it's very easy for things to slip through the crack or for you to fall into a pattern. And we know that sometimes can defeat the process of individualization of medicine. So, I think it'll actually tell you, hey, no, no, this is a high-risk patient. High risk because not just that they had a lot of exacerbations, but once they had an exacerbation, those 5 days of prednisone, you know, go-to did not work. So, that's where I'm excited.

That's where I see them getting into the process. But Speaking of the process, this came out recently, a couple of days ago, speaking about a scan which was actually published by Google Health model that the report said something about basilar ganglia, which there is no such thing. It's the basal ganglia. And so there was a few months before anybody picked up on it And they did correct the report, but they didn't correct the study. And so there is a lot of chatter about, did you just make up a body part? Or did you just make up a name? I want to call out these hype articles and these hype news. It's likely to be a typo, but this is AI. Why would they make a typo? So trust in the ICU, or in healthcare in general, you can't deploy without trust. It's like the oxygen of our healthcare industry. So thoughts about these hype kind of news?

Yeah, I mean, ultimately it is new. I think we have to accept that it is in relative nascency for what the potential Could be it's in its relative nascency. So their mistakes will occur. We also have to remember it is artificial intelligence. The artificial part is it's not independently intelligent. The source of that intelligence is us. Is it's feeding off of data it's learned from us. And we are flawed creatures. Right? So when you train an AI or you create a language model, your biases will creep into it. Your errors will creep into it. Guess what? Your spelling mistakes may creep into it.

Very much so.

Right? So the way I see it is that it's it's like using any tool is you have to watch for errors. The same fear we have that AI is going to take over, it to me is the comfort. It's only going to take over as much as you let it.

Yes. It's working in partnership, but still the ultimate decision is yours. Right. So the ultimate verification is yours.

The way I see it, yes, errors can occur. As I'm sure you've heard about the now famous report with non-existing references. So, you know, errors occur, you just got to check them and redirect it. If you don't redirect it or fix them, they will perpetuate, which is essentially technically called hallucinations, right?

Yes.

And now once it makes an error, It's like a drill. Imagine a drill. You're drilling at home and imagine if the, the drill, you know, trigger got stuck. It'll keep going.

That's a good metaphor. That is a good metaphor.

So you don't stop using drills, right? You stop and you figure out or you take the battery out.

All right. We're going to take that as a metaphor in So when we talk about process, when we talk about the power of predictive analytics education, the power of what it can help you prepare for the rare but high stakes, I guess the last question becomes about what will it take for AI to earn the trust of the critical care team? Transparency, explainability, integration. Where do you see the missing part? If I was to be an AI entrepreneur, where, how can I earn the trust of the ICU team?

I think ultimately for anybody to accept something new, they have to believe in it. And for those who do start believing in something new, they have to know that their input counts. So what I mean by that is I don't need to be an AI guy. I don't need to know how to make AI happen. I don't need to be the one writing the models. I don't need to be the one testing the models. I just need to know How it is done. Help me understand. Talk to me like I'm 5, like I sometimes say, okay? Because I can't be good at everything in life, but I'll learn to pick up new things. Just talk to me like I'm 5. Make me understand what it is. And then if I have questions, if I have problems, if I want to change something, bring me in. Let me sit at the table. And say, hey, it's not your area of expertise, that's cool, but you take care of humans. Just listen in, tell me, does this work or not? And that's the buy-in, right? But that's nothing new. That's, I think, Leadership 101 is you only, you go so far together, right? You want to go quick, go alone, right? And AI can go pretty quick, but then it'll be alone. If you want to go far, you just got to take everybody together. And I think that's really all about it's going to come down to. Like, don't, you know, you can't isolate people and say, yeah, you don't, you're not going to get it. You know, you're, what is that? Okay, boomer kind of, you know, thing. We can't, we can't do that.

It's just never leave anybody behind. No. I think that, um, overall the the ICU, while we think that there are high stakes and that AI may not belong, we've shared how it belongs in rounds, it belongs in decision-making, it belongs in predictability and early warning signs, it belongs in education, ultrasound, and perhaps even in empathy and conversations with the families. I teach my residents how to use it in coming up with sentences to perhaps drive a tough message. And the fact that it needs to be transparent in what model and what dataset did you train it on so we know what to expect in the output.

Right.

Viren, thank you so much. Any last thoughts before we conclude today's episode that was absolutely inspiring, and I felt heard, and I felt I'm with my tribe.

I'll give you a story if you want to include it.

Please.

So the way I think about all things with AI is a conversation I have with my wife, who's a horse rider. Okay, she likes riding horses. I am very unlikely to get on a horse, right? And the way I tell her is, I said, if you knew that your 1,500-pound truck could just start doing things on its own, would you get into it? And that's my excuse for not doing it. Okay. And her answer, in summary, always is, if I could talk to the truck, I would.

I love it.

And her horse and her have a relationship. She feels confident that the horse will do things for her. It may not do them for me. It'll do them for her. So AI needs to be that horse. It's got to have a relationship with its rider.

Personalized.

If it's going to be the shiny new truck, a runaway truck, it's going to get a Verane. If it's going to talk to its rider, it's going to get Laurence. And that's how I feel about it.

I might just perhaps insert, if you're not gonna get on a horse, maybe once on a trip in Morocco, you could get on a camel.

Oh, I got on a horse. I got over my own fears. I was bold with it, just like I'm with AI. So I did. I did. I love it. And I would love to come to Morocco.

Fyran, this has been fantastic. Thank you. I think we've gone from the bedside reality to digital twins and back again. And we landed on just an important truth, and that is AI in the ICU is not about replacing the human factor, it's just about protecting it. So we can think clearly, we can act decisively, we can be prepared for the rare but fatal, and maybe just catch trouble before it catches our patients.

I love that.

Thank you so much. Curiosity is infectious, so catch it here at the AI Doctor, Ready Doctor podcast, and we'll see you next time.

Thank you, Azad.