Transcripts are auto-generated and may contain errors

Dr. Shamim Nemati: [00:00:00] We are on an exponential, you know, growth rate right now, and it’s just everything is moving so fast. So we are starting to see a sort of agentic AI framework where instead of having a simple like risk model, now you have agents for early prediction of sepsis. We have agents now to act as a ID doctor and agent to act as a hospitalist, and people are talking about getting these.

Agents to interact with each other, come up with a care plan for the patient. This is the 

Nicole Kupchik: sepsis spectrum, a podcast about antimicrobial resistance, sepsis, and how to expect the unexpected in your practice.

The beat kicks in that low synth hum. That pulsing rhythm feels like you’re walking into a club. But this isn’t disco or EDM, the flashing lights, they’re vital signs. The bass drop. That’s a crashing [00:01:00] blood pressure and that pattern looping underneath it all that’s data patterns. Only a machine might catch a subtle shift, a missed flag, a hidden clue, buried in labs notes, and case history.

It’s more important than ever for us to get out there and explore how technology like AI and machine learning are syncing with human experience. Not to replace it, but to sharpen it. From pathogen ID to therapy selection, these tools are already changing how we prevent infection. Stop sepsis early and personalize care.

Hi everyone, and welcome to the sepsis spectrum. I’m Nicole Kubic, critical care nurse, clinical nurse specialist, and your guide through the complicated and sometimes frustrating world of sepsis and antimicrobial resistance, or as we like to call this season Microbial Mysteries on today’s episode. We’re going to [00:02:00] break down the differences between artificial intelligence, machine learning, clinical decision support, and best practice alerts, and see how each one plays a role in tackling infections, sepsis, and antimicrobial resistance.

Dr. Shamim Namati is the direct. Of Namati Lab and associate professor of biomedical informatics at uc, San Diego, his research focuses on using machine learning to predict sepsis before symptoms appear, helping clinicians act earlier and personalized care. Alongside Dr. Nama is Dr. Gabriel Wardi, chief of the Division of Critical Care in the Department of Emergency Medicine at uc, San Diego, and medical director for hospital sepsis.

Dr. Wardi is focused on implementing these predictive models and practice supporting clinical decision-making, improving outcomes, and advancing antimicrobial stewardship. So let’s kick things off with Dr. Wardi and Dr. Namati to explore how together [00:03:00] they’re bridging prediction and implementation to outpace sepsis and push back against a MR.

Welcome to the show, Dr. Namati and Dr. Wardi, both from the University of California, San Diego. Thank you. 

Dr. Gabriel Wardi: Yes, thank you. We’re very pleased to be here. 

Nicole Kupchik: I am so excited to talk to you about what you’re doing in the world of AI and identifying patients who are septic. So currently in hospitals right now, a lot of us use CS as a criteria to identify patients who are septic.

So a heart rate greater than 90, respiratory rate, greater than 20. A white count that’s higher or low in a temperature that’s high or low. What’s the problem with the, with service criteria? Like why? Why do we need to think of different ways? 

Dr. Gabriel Wardi: Yeah. Well I think that’s for someone that you know, works in the ICUs and in the emergency departments, um, I think what everyone has been there has run into, is it almost.

All the patients [00:04:00] that you take care of will meet this criteria, and one of the things that drives providers nuts more than anything else is getting some kind of best practice advisory or some notification that someone might be septic. When there is a very clear pathology that’s going on, that means they are not.

They are not infected. So why is Sears frustrating? Well, if you think about it, if you were to walk up a flight of stairs, right, you’re gonna be huffing and puffing. Your heart rate will go up. You’ll meet Sears criteria. You do not need a notification that you’re septic. Likewise, for a hospital, a patient that is in the emergency department or the hospital wards, the vast majority of them.

Will meet those Sears criteria, even if you modify them. Um, and the yield of having some kind of notification for that is extremely small. This has previously been studied. It’s been studied. On top of the studies that have done the studies and what we’ve seen is this, maybe it might improve checking lactates, maybe checking blood cultures [00:05:00] for patients with sepsis, but it ultimately doesn’t improve how they do overall.

So it’s a lot of work for providers to click, click. Click. Does it improve outcomes? No. Uh, and so you can imagine, right? That is why, at least at UCSD, when we started years and years ago, using a Sears based criteria, the immediate feedback we got was that this is horrible and it needs to stop. 

Nicole Kupchik: So they’re too sensitive but not specific to sepsis.

Dr. Gabriel Wardi: Exactly. 

Nicole Kupchik: Yeah. And I think the patient population we classically saw that would, that the serves would alarm is like pancreatitis is a classic example, right? Or a patient who’s having pain. 

Dr. Gabriel Wardi: Oh, anything, right? Yeah. I was, you know, I just worked in the emergency department last night. Um, and again, you hit the nail on the head.

Pancreatitis, uh, let’s say there’s some kind of traumatic injury, heart failure, some kind of substance intoxication. The list goes on and on and on about conditions that will cause some degree of abnormalities and vital signs [00:06:00] that will have people screen positive for serious criteria. And that’s why, again, it’s, you know, now that we have better tools, uh, that moving away from that.

I think is far and away the most appropriate approach to take care of patients that have sepsis. 

Nicole Kupchik: Yeah. You know, and a lot of work that’s been done out, the University of Chicago has shown that really the respiratory rate is probably one of the most. Predictive criteria of deterioration. But yet I’ll be honest, I’ve worked bedside for 30 years and we don’t measure the respiratory rate.

We make it up a lot of times, you know, and so, and I mean it’s kind of, we’ve got memes made up about us, so we own it. Right. You know, but um, but truly I always try to instill in people, respiratory rate is one of those things you gotta pay attention to. All right. So let’s switch gears and let’s talk about what you are both working on.

So you’re using artificial intelligence. And machine learning to identify patients who are septic. So first of all, can you talk about the difference between AI and machine learning, or is there a difference? 

Dr. Shamim Nemati: Yeah, there, there are some differences. So artificial [00:07:00] intelligence broadly refers to ability of machines to mimic human cognition, and there are different levels to that.

Anything from pattern recognition to reasoning to uh, what we call sequential decision making. If you, you know, wake up this morning and you, you feel like having certain food, you have to make a number of options, including, you know, which restaurant I need to go. Once you get the menu, you have to select, you know, which menu items.

So all of these decisions, uh, allows you to get to your final, uh, destination, what you know, what food you want to choose. So sequential decision making is a big part of artificial intelligence. And then, um, within artificial intelligence, we have different ways of achieving that level of sort of intelligence.

One of them was the old system expert based systems, and then they were symbolic logic. And then more recently we have, uh, machine learning, which is about learning directly from data. If you have examples of, um, data that is labeled, let’s say, for [00:08:00] sepsis or other conditions, then you can learn from that data to recognize patterns.

And then, uh, once we have, um, this type of, um, uh, cognitive systems, then we can incorporate into, uh, clinical practice. That’s where we get into clinical decision support systems. Uh, these are the type of either rule-based or ai, uh, based systems that, uh, assist care emissions with early recognition of sepsis or with, uh, picking up the right antibiotics.

And then from there we go into, you know, implementation. That’s where we have best practice advisories. Uh, these are systems that generate, uh, popups based on, again, rules or outputs from machine learning systems. 

Nicole Kupchik: Now, can you talk about like, what have you learned? You, you’ve studied, you’ve done a lot of studies at your facility.

What have you learned? Like, so what’s different about AI versus what we’ve all clinically been using, which would be like CS or news or maybe like a, a. [00:09:00] EHR alert, what’s different? Is it like a vital sign or is it a trend in vital signs? Like what have you learned? 

Dr. Shamim Nemati: Yeah, I think that a, a AI systems, uh, we talked about highly multidimensional data.

So you are now going beyond just your vital signs to all the laboratory measurements, but also all the history of medications the patient has received. If you have diagnostic imaging, so we talked about multidimensional data, but also multimodal data. AI has the ability to look at interaction among risk factors, um, as well as temporal, uh, trends in, in this data.

Like the prime example that we use. You know, if a patient arrives in the emergency department and the temperature actually is low. In the setting of immune system compromised patient or elderly patient, it has a different interpretation than, you know, a young, uh, uh, young patient. Similarly, if you have, let’s say, you mentioned the serious criteria, [00:10:00] if the serious criteria is elevated in a patient post-surgery, so that’s a temporal pattern, it has a different interpretation.

So AI systems are really good at looking at infraction among risk factors as well as temporal patterns in the data. 

Nicole Kupchik: So it basically isn’t saying all patients are created equal. Let’s use this, this, this, and this. It’s saying, okay, you’ve got this, you’re this age, you’ve got this history, and now let’s.

Figure out if you’re septic. 

Dr. Gabriel Wardi: Yeah, I think so. You know, a lot of this depends on, you know, kind of how the, and some of the terms we’ll use here, we’ll kind of define as we go throughout them. Uh, and I think, uh, you know, Dr. Nemati gave some excellent, you know, definitions of artificial intelligence. But what we’re able to take advantage of is, you know, with the advent of big data.

Very rapid processing times right now is that the amount of data that we can use to train these machine learning models to make predictions is vastly superior to what we had in the past. And by taking advantage [00:11:00] of novel computational techniques, many of which Dr. Namati has pioneered, you know what we’re able to do is ingest that many more input features.

Predictors that we then put in the model that then generate the output. And we’re also, like you mentioned, able to take advantage of multiple layers of these predictions that then mimicking how the human brain works, spits out a prediction about what’s going on with these patients. And so how have we done this at UCSD?

Well, we had a wonderful. A PhD student who just finished, who is actually able to allow us to set up a system where once a patient checks into the emergency department at UCSD, we start abstracting data. And as Dr. Dati mentioned, this means demographics, vital signs. Lab results, past medical history, imaging results, and we’ve even moved to the point right now of abstracting clinical notes and it all gets analyzed looking for these patterns that suggest someone could be septic.

And if it crosses the [00:12:00] threshold, then we basically send the notification to our providers that, Hey, consider sepsis in this situation. So it’s a very elegant way of doing it that really minimizes the number of false positives. Uh. And false negatives that we see compared to something like a Sears based system or an older rules-based system that, uh, that you know, that many institutions use.

Nicole Kupchik: So have you been able to calculate the sensitivity and specificity of what you’re using now? 

Dr. Shamim Nemati: Yeah, so, so we have a number of publication on this domain. Our 2021, um, new Eng England, uh, um, nature Digital Medicine paper reports, uh, a UC of around, uh, 94%. Uh, that means, um, if you, you are, that’s amazing. 

Nicole Kupchik: Wait, we need to stop there.

Hold on. A point 94 area under the curve, correct? Yes, sir. That’s very predictive. 

Dr. Shamim Nemati: Yes, although, uh, since the prediction of sepsis is a rare event prediction task, you can have lar high AUCs. But in terms of positive [00:13:00] predictive value, which essentially says. You know, among the number of alerts that I’m sending to the physician, how many of them are through alerts?

You know, you could have high a UC, but your positive predictive value may not be very high because the pretest probability of the event is very, very low. Uh, so I think we reported about 20% positive predictive value in that study at ity of 70 or 80%. Okay. And since, since then, um, we performed, um. Over a year of real time chart review with Dr.

Abia and the rest of our team. And that has been really, really fun, uh, to sit down on a weekly basis, go through cases, and our system is saying that this patient is septic. Dr. Wardi looks at the chart. You know, he would say, actually this patient tends to have like, uh, cirrhosis. This is not sepsis. And then we would look into the case, try to understand why that is.

And so through, uh, capturing a lot of sort of underlying, um, um, you know, reasoning that [00:14:00] doctors are doing clinical context by looking at clinical notes, uh, various type of data elements, we were able now to get the, um, positive predictive value up to 60%. So we have a new paper coming out in, um. Nature Digital medicine, demonstrating that by use of large language models.

And better capturing of clinical notes and clinical context of patients. We can get much, much better PPV positive predictive value now. 

Dr. Gabriel Wardi: And I think just one thing to emphasize right, is I do think that with the advent of large language models and generative ai, where um, that we will see a tremendous increase for institutions that have the capability of kind of linking the sand to any kind of predictive rule to see.

A very, very nice bump, uh, in positive predictive value, a UC. And just to give you some other numbers that have reported about some other sepsis models, I think probably the most famous one out there is the epic sepsis score, and there’s a very famous, um. You know, a description from the University of Michigan where [00:15:00] they took a look at the positive predictive value there, and it was below 10% for that model, and there was a lot of false alarms and there was a lot of missed patients there.

So that was, you know, a. You know, obviously the Epic sepsis score has a second iteration that’s coming out there, but certainly, you know, some of the, the research showed that it, you know, might not be performing as well as we thought. And then using some of these rule-based systems based on Sears or modified Sears typically have, you know, a positive predictive value.

Two, 3%, right? With, as we talked about earlier, a significant number of false positives, uh, that really limit the utility. So it’s something that we’ve been very proud of here, uh, as kind of this continual, these continual improvements to the model that we have live, uh, in our emergency departments. 

Nicole Kupchik: So I, well, I love that you’re doing reviews of act real patients and humanizing all of this and, and asking the question like, what?

How applicable is this? And you know, what patient populations are maybe is it not getting it right? And then adjusting. [00:16:00] So what, what are you doing on the back end then? Are you just, are you tweaking the system, um, frequently, or what does that look like? 

Dr. Shamim Nemati: Yeah, I think, uh, post implementation monitoring of this system and continuous improvement is just so important.

There is actually entire debate, uh, within the field of artificial intelligence about, uh, it’s called the embodiment problem. The idea that. Human beings, they, you know, achieve their intelligence by interacting with environments. But a lot of these AI systems that are isolated in a computer with a, you know, limited data, they don’t have that ability.

So here, uh, we essentially try to do something of that nature, you know, by sitting down with Dr. Wardi and a number of other, uh, clinicians on a, on a weekly basis, we try to learn from their feedback. And incorporate that back into the algorithm. Sometimes that’s in terms of adding additional, um, logics for suppressing the alert.

You know, the obvious one is if a patient is on [00:17:00] antibiotics, you don’t want to send an alert. Right. But also, Dr. Wardi, uh, he would tell, tell you that, you know, if this patient, um, is, um, this belongs to certain units, we don’t want to send an alert for various reasons, you know, in Yeah. If the algorithm hasn’t seen enough examples of patients for surgery, then um, maybe we want to, uh, be told alerts in that patient population.

So the current system is really focused and fine tuned on emergency department, uh, patients, and it really has gone through many, many iterations of, of fine tuning. 

Nicole Kupchik: Now how? I would just think in the ED it might be kind of tough ’cause you don’t have as much data, right? Because the patient’s just presenting.

So have, so Dr. Ward, you’ve worked in the emergency department for a long time. Are you finding it to be helpful at the bedside? Or how do you feel? So 

Dr. Gabriel Wardi: I. I obviously am very biased about this. Uh, what I can tell you though is we did study how our model did when we went live, [00:18:00] um, at UCSD, and what we were able to show is that with our best practice advisory, using the deep learning model described by Dr.

Namati, what we were able to show was a 17%. Relative decrease in sepsis mortality from patients when we went live. And what we’re also we found is providing a nice biologic mechanism for that. Is that. After we went live, we found that patients, uh, were getting more prompt antibiotics and more likely to have their sepsis bundle completed.

So it provided an appealing, right, uh, an appealing mechanism on why we were seeing these improvements. Now, does that mean that every patient, the alert fired on, needed to have, you know, blood cultures, antibiotics, and lactates? Absolutely not. But the way that we framed this, um, was that we worked closely with our bedside nurses.

They would get the alert and if they thought sepsis was possible, they would then discuss with the physician. And I think that was a very, you know, it was, we specifically chose that based [00:19:00] on the way that our electronic health record worked because we, the physicians, if the chart was not open at the time, would not get the notification.

And so you can see how that is inherently problematic because oftentimes in a busy emergency department, an attending physician might have 15 or 20 patients. They’re not going to have that open, whereas the bedside nurses with them a lot more of the time, and it worked beautifully. We’re one of the few institutions to actually show that some of these newer technologies, using these deep learning models can actually have an impact on how patients do, which is very, very exciting.

Nicole Kupchik: Well, and how do you feel, how have the nurses embraced it at the bedside? 

Dr. Gabriel Wardi: So we’ve, we’ve taken a look at this too. And so we did a number of surveys, uh, to see how people felt about this. Uh, on average what we saw is that nurses tend to. Like this, about 75% of people had either a positive or a neutral, um, you know, kind of perception of this.

Something that we also noted, uh, is that there was a little bit of a generational divide, [00:20:00] meaning that people that were closer to finishing their terminal degree, either their. MD or their nursing degree found it much more useful than those have been in practice for quite some time, and I think that makes sense, right?

Usually, if you think about it, right, the younger generation is on their phones all the time. They are much more ingrained into what AI can do. These are the ones that love the Gen AI stuff playing with chat GPT. Whereas a lot of our senior physicians in particular would tell me upfront, they said, Gabe, this is cool stuff, but you know what?

I’m not using it. 

Nicole Kupchik: I, well, yeah, I, I get it though, right? I, I get it. Yeah, that makes sense. I mean, I’m using chat GBT to help plan my vacation to Italy. It’s pretty amazing. Well, I’m, I’m really excited about that. So I think what we’ll do is let’s take a quick break and when we come back, we’re gonna dive in even more and talk about ai and then we’re also gonna talk about antimicrobial resistance and how perhaps this may be helpful.

Are you a nurse infection, [00:21:00] preventionist, or healthcare professional who wants to stay ahead of the curve? Visit sepsis podcast.org to learn how you can receive free nursing CE credits by listening to or watching the sepsis spectrum. It’s our way of supporting you and together better understanding the ever evolving world of sepsis care and a MR.

And now back to the show.

Welcome back. All right, so we’re gonna dive in a little more on ai. Can either of you think of a clinical situation where your program you like, gimme an example where you feel like your program made a difference in a patient. Can you think of a specific. Patient scenario where you feel like your program really helped?

Dr. Gabriel Wardi: Oh, there’s, yeah. I think there’s a ton of examples Okay. Of this, and I think, you know, you know, is. Anyone that’s worked in an emergency department knows that there’s usually what [00:22:00] I call two flavors of sepsis that come through, right? There’s what I call monkey sepsis, right? Where someone is febrile, tachycardic.

Hypotensive, right? That the first year medical student or nursing student says, Hey, this person is septic, right? They are coughing a lot. Their vitals are horrible, right? They need immediate intervention. Versus the patient that has much more of a nuanced presentation where it’s not immediately cleared at the providers what’s going on.

Their symptoms are much more vague, and we know time and time again that patients that have vague presenting symptoms with sepsis are usually going to be the ones that have delays in the initiation of antibiotics, the delays in initiation of recognition, and those that get in trouble. And that is very specifically what we focused our model to and to take a look for.

So as simple as in the past week or so, we, you know, going through with our model, uh, we’ve been able to identify probably about five or six patients that look like they’re getting missed by the emergency department physicians. Just, again, their initial thing, we’re not quite sure what’s going on. Some [00:23:00] diagnostic uncertainty.

The antibiotics get started. Of what the model has been able to do. And just to put that in context, you know, we talked about the outcomes that we have is since we’ve gone live, it’s on average, there’s probably been about five to six lives saved every month at UCSD Health since this model has gone through.

So typically right, that’s not gonna be, again, what. I used the term monkey sepsis. I’ve used that before at, you know, presentations in frontal lots and lots of people. But I think that’s not what we’re looking to get. And those are the ones that have these vague symptoms of who we’re trying to catch. And that’s where I think we see the value.

And this ranges from either, you know, very elderly patients that come in with maybe a little bit of abdominal pain that later end up having, let’s say, pyelonephritis, uh, to younger patients that have lots of co comorbidities. Bad cancer, right? That are hiding some kind of pneumonia that takes a while for the treating physician to kind of start.

So we’ve seen that entire spectrum of patient populations, uh, that we’ve had a very nice impact on how they’ve done, uh, by [00:24:00] using our, um, you know, our deep learning model, UCSD, 

Dr. Shamim Nemati: one of the revelations for me as, uh, we went through this exercise of a year of. Real time chart reviews with Dr. Wardi is just realizing how much diagnostic uncertainty doctors they have to deal with.

I mean, it, it’s just amazing that under limited, you know, availability of data in the earlier phases of sepsis, and then the tremendous amount of diagnostic uncertainty they have to make this. Decisions. And so I think from my point of view, what AI is doing is really aggregating multimodal data in real time and providing them with one additional pieces of information to incorporate into their decision making.

Nicole Kupchik: Yeah. Well, and I think, you know, I, so many of us who’ve have gray hair and have been around for a long time feel like, you know, the EHRs or data go to die, whereas now you can actually take all of that work we’re doing and bring it together and figure out which direction we need to go with patient patients.

I think this is exciting and really is [00:25:00] gonna change the landscape of healthcare. 

Dr. Gabriel Wardi: I do too. And I think the one thing just to emphasize, uh, that you heard from Dr. Dma right, is we’ve thrown around the term multimodal data, right? And it’s really taking, you know, what’s available in the EHR, but adding additional things on top of that, right?

So one thing that we’re very interested, right, is how does, let’s say the sepsis predictive model says this person could be septic. Well, what’s the value then of adding on. Some of these novel diagnostic tests, uh, that are out there, right? To see how that improves your sepsis prediction. And what we’re seeing right now, uh, having started a study on this is it really does a pretty amazing job of kind of giving you that diagnostic certain, uh, that you’re looking for.

So as we get more and more available data, reaching out, be be, um, outside of what’s the EHR again wearing additional biologic data. Wearable devices are something that there’s a lot of excitement right now. Right. How does that impact the prediction? Uh, again, you know, clinical notes ingesting those through large language models to kind of help refine these predictions, [00:26:00] I think is where we are going.

And some institutions are taking advantage of this right now. 

Nicole Kupchik: Yeah. Very exciting. All right, I wanna ta get a little more complicated. Now. I wanna talk about patients who have antimicrobial resistance. So what role, ’cause we know those patients are extremely complicated. What role do you see AI playing in a patient who has a MR?

Dr. Gabriel Wardi: So, I can hop on this first. I know Dr. Nemati has, you know, who’s been doing some work on this with some of our collaborators here. But I think, you know, if you. Take a look. There’s, I think there’s two ways to think about it, right? One is the patient that has a known, documented resistance to specific antibiotics, right?

To make sure that they do not get those. You don’t need a fancy AI system to do that, right? A rule-based system, or at least a clinician that is aware of kind of how to use the EHR. Uh, where I think this gets really interesting though, um, is to make that prediction about what the. Optimal antibiotic is, uh, that will minimize any kind [00:27:00] of harm from that antibiotic.

Cure the bact, the presumed bacteria, and also lead to the least amount of resistance that could be possible. And there is emerging data, uh, that machine learning applications can do very good at actually choosing the right antibiotic for patients. Now, how that translates to a physician taking the AI.

Recommendation is a much different story, and that’s something that we work with quite a bit on the implementation side is those human machine interactions about what, you know, how that impacts decision making. Because as anyone that’s worked in the hospital know there’s often big opinions. Uh, from physicians about they think they are right and it doesn’t matter what anything else is saying.

But Dr. Nani, I’ll let you take it away a little bit on some of the work you’ve been doing on in this domain. 

Dr. Shamim Nemati: But disclaimers, I’m not the expert in this domain, but I do work with some experts at UCSD. In particular, I work closely with Dr. Eli Arnov sponsor, who is an infectious disease doctor, and he coined the term I biogram.

So if you are, you know, [00:28:00] familiar with hospital antibiograms, you know, these are. Essentially a statistics on, uh, organisms that has been observed and then you know, what type of antibiotics they res respond to. And so what Dr. Arnov sponsor has done said, let’s actually go deeper and let’s personalize that for a given patient.

So patient arrives in their emergency department, we are able to pull their historical data. Based on all avail available data on this patient to say what is the likelihood of any given, um, you know, bug or organism in this patient? And also what is the likelihood of resistance, you know, for to certain, uh, anto antimicrobial therapies, uh, you know, for this patient.

So that level of personalization now allows us to go beyond existing. Guidelines. Uh, just to give you a sense, uh, we looked up our institution’s guideline for, um, you know, selection of antibiotics, and it’s pretty simple. It’s looking at allergies, but not necessarily even [00:29:00] at, you know, history of what, what type of, uh, organism this, this patient had in the past.

The cultures, you know, antibiotics. Uh, I know that physicians. At the bedside. They do consider those type of factors. But again, just imagine in a very busy emergency department how often they can go deep enough and look at all of those, those factors. So that’s where I think AI can comes into play. 

Nicole Kupchik: Yeah.

Well, I think if, especially if you have a provider at the bedside can say, see that A, this is getting you to the right therapy, and B, it’s saving you time. Maybe at that point they can be convinced, right, that this isn’t a bad thing. 

Dr. Gabriel Wardi: I agree. You know, I think that’s the tricky thing, and a lot of this, right?

It depends on the, in my experience at least, and maybe others have different experiences, but a lot of it depends on the type of physician that you’re working with. You know, and that’s something we’ve seen throughout, you know, with using, you know, some of our predictive models, right? Certain physician groups are a little bit more pliable when it comes to taking recommendations.

Let’s say the choice of antibiotic. Other [00:30:00] physicians though, are gonna fight you tooth and nail, and I’m not gonna list out any specialties or make any generalizations. I think most people can probably think off the top of their head where they see, you know, kind of the, the different opinions coming from.

But, uh, I think that’s, again, that’s a fascinating way, is again, how does the, you know, how does the physician, how does a nurse interact with the model and use that to inform patient care. Uh, and I think Anna. Anti antimicrobial stewardship is a huge one, uh, because I think we all recognize the, you know, the potential harm that could happen through inappropriate use of, uh, of antibiotics.

And I think machine learning and AI applications give us a great way of potentially mitigating some of that possible harm. 

Dr. Shamim Nemati: And we talked about, you know, early prediction of, of sepsis, but they are also able to track patients over time and see if in response to. Antimicrobial treatment, you know, is the risk score actually improving or not?

And that might be a signal to use then to say, okay, maybe we should reevaluate this patient for appropriate [00:31:00] antibiotics, 

Nicole Kupchik: which is super helpful at the bedside, you know? Uh, so, all right, so let’s say the year is 2040. Oh, wow. Okay. Not that far away. It’s not, it’s 15 years. Right. Okay. What does AI look like in the world of sepsis?

So we’ll start with Dr. Nama. Like what? Where would you like to see this go? And this is of course not retirement. 

Dr. Shamim Nemati: You know, it’s, it’s interesting because back in 2020, I was actually teaching a course here at uc, San Diego on clinical natural language processing. And we were teaching, uh, our students about AI and, uh, some of these modern, uh, AI architectures like transformers, which is the underlying architecture of large language models.

And. Even with some of us, uh, being so intimately involved in the field, we could not predict the sort of all the advances with large language models. [00:32:00] So, you know, even if you tell me 2030. I would hesitate about prediction because we are on an exponential, you know, growth rate right now, and it’s just everything is moving so fast.

So we are starting to see a sort of agentic AI framework where instead of having a simple like risk model, now you have agents for early prediction of, uh, you know, sepsis. They have, they, we have agents now for, to act as a. ID doctor an agent to act as a hospitalist, and people are talking about getting these agents to interact with each other, come up with a care plan for the patient.

So this is like really, really rapidly evolving area. On the top of that, as Dr. Wardi mentioned, all of these novel biomarkers are coming in, you know, the ability to distinguish between bacterial versus viral infection and then continuous monitoring sensor coming, um, you know. You mentioned the problem of missing data.

Just imagine if on every patient on a minute by minute, we could have, [00:33:00] you know, lactate and other types of analytes, you know, uh, recorded and agents, uh, AI agents are ingesting all of that data. It’s just really hard to predict the future right now. Gabe, what do you think? 

Dr. Gabriel Wardi: Yikes. Again. Yeah, I, you know, one thing that, um, that, uh, you know, one of my mentors has continuously said is that, you know, we’ve probably under promised what AI can do in the next few years, uh, but it’s probably going to be over prompt.

We haven’t really considered what it looks like 10 to 15 years down the road. And I think, you know, probably in an ideal state, uh, is what we’d see is symbiosis between the artificial intelligence application and the providers at the bedside, uh, that are. Using whatever is out there, because again, to Dr.

NA’s point, I don’t think we have any clue, uh, at the rate things are changing on this exponential curve right now, what that’s gonna look like in 15 years. But I think what we will see, uh, some, again, this symbiotic relationship that is allowing us to provide much more precise care, uh, for these [00:34:00] patients that starts to expand outside of the hospital to maybe predicting who will develop sepsis in the community to help.

Guide a lot of these decisions, uh, that hopefully might be able to stop sepsis from happening in the first point. Uh, you know, as we get better predictive capabilities, as we’re able to ingest more data, uh, as we’re able to use, again, some of these wearable devices and things like that, uh, to really kind of make a huge dent by stopping sepsis from happening, not just.

Providing early treatment through identification once people hit the hospital doors. But again, you know, 15 years from now, yikes, it’s gonna be, you know, who knows what that’s gonna look like? 

Nicole Kupchik: Yeah, well, I, I’m just excited for what could happen on the inpatient side because we know patients get septic under our nose and it’s very challenging to pick that up.

I mean, we know sepsis is the leading cause of death in US hospitals and like, I’m, it’s excited to see where that goes. All right. Well, I just wanna thank both of you. Any parting thoughts before [00:35:00] we, we part ways today? 

Dr. Shamim Nemati: These are really, you know, exci exciting times. And, um, I think as a, as a data scientist, you know, working closely with clinicians at at UCSD, it has been just amazing experience.

And I, I encourage, I think all of our divisions, departments to sort of break down through some of the existing silos and I think work more closely with domain experts. Um, because, um, again, as I mentioned earlier. Building AI just instead of databases can go. So, so far you have to implement these systems.

You have to, uh, study them in an actual clinical, um, environment and gather feedback from clinicians, from, you know, patients. And incorporate those, I think is only through this iterative process or refinement that we can get into systems that are reliable, uh, and useful. 

Nicole Kupchik: Well, and I love that you both have partnered and worked so closely together.

’cause I think you just hit the nail on the head is like you’ve gotta bring in that [00:36:00] human clinical component to really. Make a difference in with these programs. Yeah. Right. Dr. Wardi, any parting thoughts? 

Dr. Gabriel Wardi: I think, you know, one thing is that oftentimes ai, machine learning, deep learning, generative ai, it’s a little bit scary to a lot of people just because they don’t have that background, uh, in it.

And I think the advice I would, if that is you, uh, if you are that person that has, you know, is. These uncertain about this dive into it because you’ll see there’s a number of medical journals right now that are publishing all kinds of amazing things. Uh, that’ll bring you up to speed pretty quickly.

That’ll demystify a lot of these things that sound quite frightening. You might not understand the true nuts and bolts about what makes, you know, the neural network work. Uh, but at least you’ll have a very good understanding of kind of the broad terms, the potential applications, and ultimately how you can use these to improve patient care.

So, uh, you know, just dive in. Like I, you know, some of the journals I’d recommend if you’re interested in stuff. Uh, there’s [00:37:00] NM AI that I think does a great job with it. Uh, there’s also, uh. NPJ Digital Medicine, fantastic journals that really kind of bring clinicians, uh, into this, that are very digestible.

After you get through a few little, a few of the pieces there. 

[music]: Yeah. 

Nicole Kupchik: Well, I just wanna thank both of you for joining us today. I am so excited about the work you’re doing. I really think I can’t wait to see where both of your careers go as well. ’cause you’re both just, you’re making a huge difference in medicine, so please keep doing what you’re doing.

So. All right. Well, thanks for joining us today. Thank you, Nicole. This was 

Dr. Gabriel Wardi: fun. Of course. Thank you for the invite.

Nicole Kupchik: Well, that was a super fun conversation. I am so excited to see where the future of AI goes. You know, if it can make our work at the bedside easier, improve patient outcomes, and just really if it can be accurate and decrease the workload at the bedside. I, I’m all game. I’m 100% all in on [00:38:00] AI and machine learning.

Well, I just wanna say thanks for joining me on today’s episode of the Sepsis Spectrum. If you like the show, we wanna hear about it. Please leave a review wherever you’re enjoying this podcast. It helps a ton. You can also reach me and our awesome team@infoatsepsis.org or visit sepsis podcast.org to share any stories of your own questions.

Concerns or episode ideas. To learn more about Sepsis Alliance, visit sepsis.org. The sepsis spectrum is brought to you by Sepsis Alliance. I’m your host, Nicole Kubic. Our executive producers are Allison Strickland, Hannah sas, Claudia Orth, and Alex Coleman. Our producers are Erin, corny, Rob Goldman s Shahnti, Brooke, and me Nicole cch.

Our post-production producer is Tim Scott. Our editor and engineer is Jason Portizo, and our music is by Omer Ben-Zvi. To learn about Sepsis alliance’s podcast, legal disclaimer and compliance policies, you can visit sepsis [00:39:00] podcast.org/disclaimers. The sepsis spectrum is a human content and sepsis alliance production.

Thanks for watching. I hope you’re enjoying the sepsis spectrum. Leave a comment below and let me know if you want to binge some more episodes. Just click that playlist right over there and if you’re feeling super generous today, give this video a like, subscribe if you haven’t hit the bell. All the things.

And of course you can also listen on the go wherever you get your podcast. Bye.

Episode 6 of The Sepsis Spectrum: Microbial Mysteries features Dr. Gabriel Wardi and Dr. Shamim Nemati. Click on their names to learn more about our guests.