Transcript
Speaker: Welcome to Critical Matters, a sound podcast covering a broad range of topics related to the practice of intensive care medicine.
Speaker: Sound provides comprehensive critical care programs to hospitals across the country.
Speaker: To learn more about our programs and career opportunities, visit www.soundphysicians.com.
Speaker: And now your host, Dr. Sergio Zanotti.
Speaker: In this episode of Critical Matters, we explore one of medicine's fastest evolving frontiers, artificial intelligence, or AI.
Speaker: From predictive analytics to decision support tools, AI is beginning to influence how we deliver critical care.
Speaker: But what does that actually mean for frontline clinicians?
Speaker: I'm a big believer that the best uses of technology are not defined by those who create the technology and have the most expertise in the technology, but those who use it.
Speaker: So today we're going to talk about how we can leverage AI at the bedside for our practice.
Speaker: Our guest is Dr. Sherrod Patel, a critical care physician with additional board certification in nephrology and echocardiography.
Speaker: He is a critical care intensivist at Cooper University Healthcare in Camden.
Speaker: He's also the assistant program director for the Internal Medicine Residency Program and an assistant professor of medicine at Cooper Medical School of Rowan University.
Speaker: Dr. Patel is deeply interested in applying artificial intelligence and other technologies at the bedside.
Speaker: Sherrod, welcome to Critical Matters.
Speaker: Thank you, Sergio.
Speaker: Thank you for having me.
Speaker: And I'm excited to talk about this topic.
Speaker: I could probably talk about this topic for hours, but I'm excited to spend this hour with you.
Speaker: Awesome.
Speaker: So why don't we start as a way of introduction with why do you think intensivists should care about AI in this topic in particular?
Speaker: Yeah, I think really to answer that question,
Speaker: um we have to kind of go back and look at what what are some of the the pain points in the icu and what makes the icu unique so the icu is unique in the sense that there's so much data there's arterial line there's the ventilator waveforms there's the labs the vitals there's there's so much data that it produces on a daily basis this is ripe
Speaker: to provide for an AI to learn from, which we'll talk about what this means about the learning.
Speaker: So there's so much data.
Speaker: So that's one thing.
Speaker: Two, I think intensivists should care because where in the hospital is there...
Speaker: more cognitive load.
Speaker: And I guess this could be argued, but as far as the cognitive load of the ICU, the noises, the sounds, the interruptions, the number of decisions you're making under duress, it's such a cognitively loaded place that I think AI is a great tool to not replace us, but to actually provide
Speaker: They'll fill in the gaps of where the human mind can't take over.
Speaker: So one thing about the human mind is that our ability to think about something, our working memory can handle about four to seven items at once.
Speaker: But in the ICU, we're trying to think about a lot more things at once.
Speaker: So our ability to do this is limited.
Speaker: There's a concept called ego depletion.
Speaker: which as you make decisions, your ability to make more complex decisions down the line gets diminished.
Speaker: And so AI could potentially help with this where it could take away some of the more simple decisions and automate some of the more simple things and then allow us that cognitive reserve of our pre-parental cortex to make the more complex decisions down the line.
Speaker: So I think that the applications of AI, the ICU is probably one of the top five places in medicine where it should be applied.
Speaker: Excellent.
Speaker: And when we've talked offline, I mean, you always refer to yourself as an AI practitioner.
Speaker: Could you share with us your path to making AI useful in your critical care practice?
Speaker: Yeah, yeah.
Speaker: And I was thinking about this, Sergio.
Speaker: I call myself a practitioner because this is the first time in my relatively young career that
Speaker: that I'm actually speaking about and actually applying in research things that aren't directly related to my credentials.
Speaker: So, for example, my previous research talks would be about things within nephrology or critical care and ventilators and echocardiography.
Speaker: But those stemmed from my training and my credentials.
Speaker: This I've kind of learned more as almost like a tradesman.
Speaker: I've picked up these tools.
Speaker: I've tinkered with them.
Speaker: I've practiced with them.
Speaker: And so over time, I got more, I became more proficient at using these tools.
Speaker: And I started looking under the hood.
Speaker: and looking to see how these tools worked.
Speaker: And so I call myself a practitioner because I don't have a degree that I've taken.
Speaker: I don't have a degree in machine learning.
Speaker: I don't have a degree in AI, but I'm a self-learner.
Speaker: I've taken many, many courses, these mini courses.
Speaker: I've read books.
Speaker: hours and hours about this topic and I've practiced and coded and made mistakes.
Speaker: And along that line, it's given me this degree of proficiency in it where I can actually think about this in a logical way and think about ways to solve problems.
Speaker: But before we dive into those applications and some of the things that are going on, maybe we should start with setting the floor with a little bit of AI 101 and really make sure that all our listeners are on the same page of what we're really talking about.
Speaker: So from a, I guess, high level perspective, what is AI?
Speaker: What is the difference between AI, machine learning, deep learning?
Speaker: Could you talk a little bit about that?
Speaker: Yeah, yeah.
Speaker: So AI meaning artificial intelligence.
Speaker: So you kind of think about it as creating a system that can perform tasks that usually that used to require human intelligence.
Speaker: For example, writing
Speaker: a poem or writing an epic like the Iliad, AI could potentially do things like this today and before it required human intelligence.
Speaker: So the artificial aspect of it is doing the things that human intelligence used to do that now AI systems could potentially do.
Speaker: When you think about machine learning,
Speaker: So machine learning is under the umbrella of artificial intelligence.
Speaker: So these terms often get thrown back and forth.
Speaker: And if you wanted to visualize a Venn diagram, that's probably kind of the best way to kind of talk about these things.
Speaker: But machine learning is a subset of artificial intelligence.
Speaker: And so let me kind of give you a simple example where everyone in their EMRs gets these sepsis alerts, right?
Speaker: These simplistic sepsis alerts based on their SERS criteria.
Speaker: These are explicit rules that are programmed to say when the heart rate is this, when the respiratory rate is this, and these two things exist, send an alert.
Speaker: This is not a learning system.
Speaker: It hasn't learned by patterns.
Speaker: But machine learning, it actually is a method that you provide a large amount of data, labeled data, to say, all right, algorithm, machine learning model,
Speaker: This is what sepsis is, and this is what sepsis is not.
Speaker: So you give them thousands, hundreds of thousands, and millions of examples of where sepsis criteria was met, where sepsis criteria wasn't met, or AKI criteria was met, and where AKI criteria wasn't met.
Speaker: And it's trained on this data.
Speaker: And based on that kind of pattern recognition and learning, it can identify sepsis.
Speaker: It can identify AKI.
Speaker: And it gets better as more data, clean data is provided.
Speaker: So there's a big difference between just a simple kind of algorithm that identifies SIRS
Speaker: versus machine learning, which actually learns from data and doesn't require explicitly programmed rules.
Speaker: For example, just identifying like SIRS in a patient.
Speaker: And then deep learning.
Speaker: So the deep learning builds upon machine learning.
Speaker: Deep learning is, and you'd have to talk about neural networks, and I don't want to get too much into the weeds of this, but to understand deep learning,
Speaker: It's essentially a type of neural network, and it's basically created based on the human mind.
Speaker: The human mind has layers of neurons.
Speaker: The same with a neural network.
Speaker: It has an input layer, it has hidden layers, and it has an output layer.
Speaker: data goes in to the input layer and there's lines.
Speaker: You can draw these arrows back and forth where what it's trying to do is learn these nonlinear patterns in the data.
Speaker: How do you identify sepsis?
Speaker: How do you identify AKI?
Speaker: And it has the label data of what sepsis is, what sepsis isn't.
Speaker: And it finds the nonlinear patterns, and it goes back and forth until that error rate is reduced as low as possible.
Speaker: And now it's trained to identify sepsis.
Speaker: So that's what deep learning in the neural network is.
Speaker: And these are powerful, powerful methods
Speaker: to identify these non-linear trends that the human mind actually will have a difficult time with.
Speaker: So that's kind of like basics of what AI, machine learning, deep learning, and what a neural network is.
Speaker: The other two terms that I think are worth maybe throwing in there and you can maybe give us a simple explanation is algorithms, which you already mentioned, right?
Speaker: Which I still think is important part of AI and large language models or LLMs.
Speaker: Yeah.
Speaker: Yeah.
Speaker: So an algorithm could be basically a recipe is an algorithm, right?
Speaker: So you look at a cookbook,
Speaker: It tells you how to make something that you're basically you're following a predefined method to solve a specific problem.
Speaker: So that could be an artificial AI doing that or a human doing that.
Speaker: So that's the basics of an algorithm.
Speaker: Now, when we talk about kind of the evolution of AI, and I mentioned kind of like the neural network already, the big thing that when people talk about AI today, they're probably talking about large language models.
Speaker: They're talking about Chasie PT, they're talking about CLAW, they're talking about Gemini.
Speaker: So back in 2017, so what led to this point where there's this big, huge AI explosion?
Speaker: In 2017, a paper came out, and it was by Google, and it was titled, Attention is All You Need.
Speaker: And it introduced this new type of neural network, and it was called a transformer model.
Speaker: And again, we could easily get into the weeds on this, but to kind of just keep it high level, what this did that improved upon the previous neural networks, like recurrent neural networks,
Speaker: LSTMs, these are just examples of neural networks that existed before, is that the previous neural networks did well with sequential data.
Speaker: So when you trained it, you would have to train it sequentially.
Speaker: And it had a difficult time remembering what came before it.
Speaker: So for example, if you're trying to predict text, it had a difficult time predicting the next word because it didn't have the context of all the previous words before it.
Speaker: But with the transformers, now it has contextual understanding.
Speaker: It's something called attention.
Speaker: And this attention made it better at identifying that, let's say, a patient who is a 74-year-old patient with COPD had pneumonia six weeks before.
Speaker: It's coming back again with dyspnea, completed antibiotic course.
Speaker: And so this kind of context...
Speaker: a transformer model would be able to keep in context the temporality.
Speaker: When did the patient come in before?
Speaker: What diagnosis does it have?
Speaker: So all these things are kind of the attention is given to this, and it's better able to predict the next word than the previous neural networks were.
Speaker: And that's what kind of really made these chat TPTs and things like that more powerful.
Speaker: The second thing that really changed beyond the attention and the contextual understanding
Speaker: was the ability to train it.
Speaker: And so you could train these models using something called parallelization, which now it can use
Speaker: a bunch of GPUs to train the data and break the knowledge or the information into small chunks, and they could all run at one time.
Speaker: So you were able to train a lot, the models on a lot more data than you were before.
Speaker: So that's what kind of the two big things that led to this explosion of what we think of AI today.
Speaker: So most people, when they speak about AI today, they're probably talking about large language models.
Speaker: They're probably talking about ChatGPT, Gemini, and things like that.
Speaker: And the other thing I wanted to ask you, Sherrod, is if you could share with us some things that AI is not and maybe share some common myths or misconceptions about AI commonly held by clinicians.
Speaker: Yeah, I think it's easy to anthropomorphize ChatGPT or Gemini, and it almost feels like you're speaking to a human.
Speaker: Really, what ChatGPT and these LLMs are, they're essentially really complicated calculators.
Speaker: They're prediction machines, and they're thinking in numbers.
Speaker: They're not actually thinking in words.
Speaker: They're thinking in numbers.
Speaker: And so...
Speaker: The fact that they don't understand, like when we think about morality, emotions, and words like pneumonia and MI, all these words, this means nothing to the large language model.
Speaker: These are all numbers and it's trying to predict basically the next word.
Speaker: So it can do things like create something out of thin air.
Speaker: And this is called, one concept is called hallucinations.
Speaker: And so a concept could be is that, wow, these outputs are so beautiful.
Speaker: I can trust these, this reference.
Speaker: It looks so real.
Speaker: I can trust this.
Speaker: And so trust is something that is going to need to be earned in this.
Speaker: And it can hallucinate and create references out of thin air.
Speaker: It's getting better.
Speaker: But this still exists.
Speaker: So really, for one of the misconceptions for AI is that you could trust all the outputs.
Speaker: We are not there yet.
Speaker: We might get there one day through careful implementation, but you'd have to be very, very careful and really be hypervigilant of things such as hallucinations when you're doing research and even eventually getting to the point where you're using these tools for clinical decision making.
Speaker: The second point I would make about this is that AI is not going to replace physicians.
Speaker: I think it's kind of humans go through this kind of black or white thinking, right?
Speaker: So it's that that's one of the cognitive biases that exists with humans.
Speaker: is that we think in black and white terms.
Speaker: And it's not going to be AI versus doctors.
Speaker: The best implementation is going to be an AI that augments what humans do in clinical practice.
Speaker: And through like looking at some of the cognitive limitations
Speaker: load of the clinician so we can make our complex decisions and be better with the other aspects of being a doctor, like the humanism aspects and things like that.
Speaker: And so that's a second thing that I think is a misconception that I think we really need to rethink.
Speaker: And then another misconception might be is that the AI needs to be
Speaker: perfect or the data that we're feeding it needs to be perfect.
Speaker: I think this is where kind of we're going to be running into some problems.
Speaker: Like how good does the AI need to be for us to actually use?
Speaker: So I'll use the example of an autonomous car, right?
Speaker: So an autonomous car, if we all drove autonomous cars, even today, the accident rate would probably be much, much lower.
Speaker: It'd be one in a million.
Speaker: but that one in a million would probably be scrutinized very, very heavily.
Speaker: Whereas humans, the fallible that we are, we may have an accident rate of one in a thousand, right?
Speaker: The AIs could be much, much better than us and there'd be less total harm to humans in general, but our ability to let go and just accept that one in a million is more difficult than the one in a thousand for the human.
Speaker: So in the medical aspect of it,
Speaker: So one kind of, I think, like, I guess a bias that we may have against AI is that it can make mistakes.
Speaker: Well, I agree AI can make mistakes, but humans make lots of mistakes, right?
Speaker: So you're like the medical errors is one of the biggest, biggest reasons for morbidity and mortality in medicine.
Speaker: And so if the AI is properly applied, there could be a small error rate.
Speaker: But I think overall, it would improve the error rate of the overall medical system.
Speaker: So I think those are some of the things that come in kind of myths and misconceptions about AI that I think we should really rethink.
Speaker: Perfect.
Speaker: Let's move on to applications of AI and critical care today.
Speaker: And I would like to do this in kind of like two steps.
Speaker: First, I would like to just get your overall assessment of some applications where we are in general.
Speaker: And then I would like to go more specifically in how do you use AI in your day-to-day, whether it be clinical or non-clinical, and maybe dive a bit deeper into that.
Speaker: So in terms of the applications that,
Speaker: I kind of hear all the time.
Speaker: I hear about predictive analytics.
Speaker: I hear about natural language processing for clinical documentation.
Speaker: I hear about decision support tools.
Speaker: I hear about AI-powered imaging analysis, right?
Speaker: I mean, at the end of the day, x-rays and MRIs are just pixels with basically, like you said, number information, right?
Speaker: I hear about workflow optimization.
Speaker: And I also very interested in how do you apply it to teaching?
Speaker: So maybe we can go just like I said, Sherrod, at a high level.
Speaker: What do you know what's going on in this world?
Speaker: And we can start with predictive analytics.
Speaker: Yeah, I think the predictive analytics, and this was before the transformers.
Speaker: This was before.
Speaker: Sergio, I've been in this field for like six years, but six years in this world is, it just moves so fast.
Speaker: that when I think about what existed four or five years ago, it seems so rudimentary and simplistic compared to what is coming out today.
Speaker: So the things that exist and have existed for a while but are getting better, one is sepsis detection, right?
Speaker: I think the big problems and pain points in critical care especially is that we have these conditions that are time sensitive.
Speaker: Strokes, sepsis, AKI,
Speaker: of which if we get to it early, we can change the trajectory of that patient.
Speaker: And so what AI can potentially do very well is to find and identify these things earlier than the human mind might be able to.
Speaker: And let me give you an example of how it might do this.
Speaker: It's even a simple machine learning model where it's trained to identify sepsis.
Speaker: Now, this model could have something called feature engineering in it.
Speaker: So again, not to get into the weeds of where we might say, all right, the lactate's elevated.
Speaker: I'm worried this patient has sepsis and the patient may not do well.
Speaker: Well, the machine learning model could have something called feature engineering input programmed into it, where it not only just looks at the lactate, it looks at the fact that the patient's albumin on admission was 1.7.
Speaker: And a little known fact is that if you look at most sepsis predictive, like mortality prediction tools, albumin is one of the biggest predictors of death.
Speaker: And I think of it as like a biomarker of your fragility.
Speaker: And it might say, all right, well, the combination of the albumin and the lactate being elevated and the creatinine went up by 0.1, which we might not even blink at.
Speaker: The probability of sepsis has gone up from 35% to 75%.
Speaker: It notes the low-grade fever, 100.6, which we might not even bat an eyelash at.
Speaker: And then it changes the probability to 78%.
Speaker: And so these sepsis detection tools find these nonlinear patterns, and it can potentially look at all these features, the lactates, the vitals, and everything, and look at them through different mathematical representations.
Speaker: that in a way that it allows it to catch these things at a much, much earlier level so we can actually modify the trajectory.
Speaker: And I think sepsis and kidney injury are low-hanging fruit for this where it can potentially, there are models that exist for AKI that can identify AKI 24 to 48 hours before the creatinine rises.
Speaker: Now, these aren't being applied directly
Speaker: at full scale yet, but if you look in the literature and you look at papers, and I've been a part of one paper where we used the neural network and we were able to identify AKI before the creatinine rose.
Speaker: And this is something that needs to be solved because creatinine, again, it doesn't rise until 24 or 36 hours after the insult.
Speaker: And at that point, you may be too late to actually correct the process that led to the AKI.
Speaker: And we both know when a patient has AKI with any other condition,
Speaker: your trajectory in that ICU has changed.
Speaker: Your probability of death has gone up significantly.
Speaker: So sepsis, AKI detection tools are, I think, big, big.
Speaker: There's a lot of investment going into that because you can potentially change the trajectory of these.
Speaker: And I think that my understanding is that we might slowly not even be aware that a lot of our EMRs are incorporating better models for prediction, right?
Speaker: I mean, you talk about sepsis, the famous sepsis alerts, right?
Speaker: But I think that even that is getting better and better without us even being aware of it.
Speaker: So I think it's a great example of how AI is permeating into our practice, whether we are aware and like it or not.
Speaker: 100%.
Speaker: I think one of the biggest problems with the sepsis alerts was the false positives, where you lose trust in these alerts.
Speaker: The trust is lost.
Speaker: So as these models get more advanced, the false positives will most likely be less.
Speaker: And so when you actually see the alerts and your mind puts together, it's like, wow, I got this alert.
Speaker: and actually did predict the patient's deterioration, the trust will increase and we're more likely to be reacting to these alerts in the future.
Speaker: So I'm excited about the AI applications for these early detection tools.
Speaker: Then the other kind of big thing is that you mentioned is the natural language processing for clinical documentation.
Speaker: I think this might be the lowest hanging fruit.
Speaker: summarization.
Speaker: So summarization might be the lowest hanging fruit.
Speaker: I think one of the biggest pain points for me, so let's say I have 16, 17 patients in the ICU and I'm rounding.
Speaker: When I have a new patient and they're complex and they've had multiple admissions and they have resistance to multiple antibiotics and they have been on steroids in the past, but they're no longer.
Speaker: For really getting a good
Speaker: medically tuned AI system or large language model to really kind of give you a nice summary of the relevant things that you're looking for would really reduce my cognitive load.
Speaker: And then I can actually think critically about the information that I have in front of me.
Speaker: And I think the best examples would be is the patient comes in with septic shock.
Speaker: The urine is dirty.
Speaker: There's pyuria.
Speaker: There's bacteria.
Speaker: They get put on the usual antibiotics.
Speaker: But the AI, the summarization, might note to you that they've had resistance pseudomonas in the past.
Speaker: So the sethpeme that you started them on
Speaker: it may not be adequate.
Speaker: So I think the natural language processing, clinical summarization is low hanging fruit.
Speaker: Decision support tools.
Speaker: This is kind of where I'm actually working on one decision support tool.
Speaker: And I think this is probably one of the more difficult parts and the one that's going to be scrutinized a little bit more by the FDA and things like that.
Speaker: So when you're actually using AI to make actual clinical decisions,
Speaker: Like, for example, differential generators, next best test.
Speaker: Like if the AI can help you decide what the next best test is.
Speaker: De-escalation of antibiotics.
Speaker: These are high-risk decisions that you'll be making.
Speaker: And there aren't many AI tools at this point that are actually applied.
Speaker: And I think this is more...
Speaker: because of the worry about the mistakes and the errors that might come with this that they haven't.
Speaker: Because I make tools like this.
Speaker: And I'll tell you, Sergio, is that just kind of like when I create these synthetic cases to put in there, it's better than us as today.
Speaker: It's better than us today for differential generation to think in a statistical probabilistic way.
Speaker: rather than just more of, I've seen a few cases like this, so this is what I'm going to do.
Speaker: And I'll give you an example, is that, and this might go into kind of the next question you're asking, is that how do I use some of this?
Speaker: I have a tool that I made, it's a multi-agent tool that I kind of, I don't use it to
Speaker: for clinical decision-making, I go back and I retrospectively my difficult cases, I'll go back and I'll put the cases in then.
Speaker: And so I had a case of a young lady, she was in her 40s, came in with altered mental status, had a seizure, and the MRI showed temporal lobe enhancements.
Speaker: I'm like, okay, this must be HSV encephalitis.
Speaker: The patient's on acyclovir.
Speaker: The patient had renal dysfunction.
Speaker: So the acyclovir is not benign in that situation.
Speaker: And so, but the LP fluid, there was nothing.
Speaker: It was completely clean, no cells, no protein.
Speaker: The HSV PCR was negative.
Speaker: And then we got into like a multidisciplinary discussion and the consensus was to keep the acyclovir going.
Speaker: And so...
Speaker: I did.
Speaker: I kept the cyclivir going because at that point, that was a consensus.
Speaker: But I went back and this tool that I created, I call it BayesBuddy.
Speaker: It uses Bayesian reasoning and using Bayesian probability.
Speaker: It finds likelihood ratios of tests and it gives you probabilistic differentials of what's the likelihood.
Speaker: And I gave it the data, obviously, without any patient info.
Speaker: I gave it kind of a similar case and it gave a probability of this being HSV encephalitis to be like 0.003%.
Speaker: And when I think about the actual risk and harm of the acyclovir in the setting of AKI,
Speaker: I could argue that the risk of the acyclovir and the AKI was higher.
Speaker: So in my own personal use, just for my own kind of like creating my own probabilities in my head, I often take my difficult cases, things that I'm not sure about, and I'll put it into my own multi-agent tool that I created.
Speaker: And I'll look to see how is the performance and how does it look objectively compared to what my decision-making was.
Speaker: So that's an example of a clinical decision support tool that
Speaker: I can't put into real practice as of yet because there's just a lot more scrutiny for it.
Speaker: And so that's an example of something I use and then potentially an example of where you could use as a clinical decision support tool.
Speaker: AI powered and imaging analysis.
Speaker: So this existed even before where we are.
Speaker: So there's a lot of vision models that have come out where specifically the pre-training is for vision purposes for x-rays, MRIs, CTs, and things like that.
Speaker: There were neural networks that existed before this, one of which was called convoluted neural networks, which even at that time,
Speaker: It was as good as junior radiology attendings.
Speaker: And the only, the most experienced radiologists caught things that the AI didn't, even at that point.
Speaker: But this is getting better and better.
Speaker: So this is also one of the lower hanging fruit
Speaker: is that if you don't have an on-call radiologist at night or you're in a rural setting, potentially the AI has gotten to the point where an application would be is just x-ray interpretations, which could be overread by the attending the next day.
Speaker: And this is not in prime time yet, but this is a potential application.
Speaker: Another one is for point-of-care ultrasound.
Speaker: where in the ICU, I use a lot of point-of-care ultrasound.
Speaker: But how can you make this more accessible to someone who hasn't had as much training?
Speaker: Well, there's AI tools now being integrated into a lot of the newer ultrasound machines, which will calculate IVC collapsibility.
Speaker: It'll identify pulmonary edema on a lung ultrasound.
Speaker: It'll calculate the ejection fraction.
Speaker: It can automatically calculate cardiac output based on some Doppler analysis.
Speaker: And so that's a potential application as well.
Speaker: And then workflow optimization.
Speaker: This is probably another area that AI could potentially do very well in.
Speaker: And for example, how can you triage
Speaker: patients quickly in the ER where they automatically get their assignment for step down versus ICU rather than the ER doctor calling and pleading to the intensivist to accept the patient.
Speaker: If the calculated severity and acuity scores are high enough,
Speaker: then the patient will automatically get put towards an ICU bed as opposed to a step-down bed or a floor bed.
Speaker: So this is potentially possible as well in the situation of like kind of workflow type of things.
Speaker: And so these are kind of the examples, a few examples of how AI can be applied in the intensive care unit.
Speaker: Sure, Rob, before we go into more specifics of your use of AI for solving problems, one of the things that a lot of clinicians come to me and ask about is related to natural language processing for clinical documentation.
Speaker: But
Speaker: thinking more of what they have to document on a day-to-day with their notes and patient visits.
Speaker: Now, I think for the ICU, I haven't really found the great tools.
Speaker: There's obviously a lot of kind of AI supported scribes that I think are great if you're having a conversation in an office, right?
Speaker: But I think that when you walk into a patient's room that's intubated in a mechanical ventilation and sedated, right, there's not a lot of conversation going on.
Speaker: Any thoughts in terms of AI helping us take care of that menial work that, or what I call like low impact, right, work that we all kind of don't enjoy that much?
Speaker: Shallow work?
Speaker: Yeah, yeah, yeah, 100%.
Speaker: I think like one potential, like,
Speaker: Sarah, because how do we generate the HNP in our head for an ICU patient?
Speaker: It's often done through our clinical routes.
Speaker: So we're talking about, we're hearing about the patient.
Speaker: So one potential option in that situation is to give an AI context.
Speaker: The audio ability of a lot of these large language models is getting better and better.
Speaker: You could potentially have an AI where you click the button on, or it might be in the room, or it could be wherever you're around it.
Speaker: and it's listening to rounds in context.
Speaker: And these, this AI could potentially be fine tuned.
Speaker: So the fine tuned meaning, so when the AI was trained, it's trained kind of in a general way, but you can do something called fine tuning, which you could like, all right, AI, I'm going to change your neural architecture and weights to be specifically good at writing HMPs, notes and developing differentials and writing assessments and plans.
Speaker: And so that could be a fine-tuned AI where that's what it does.
Speaker: It may not be as well, do as well with math calculations or writing poetry, but it has specifically become better at that specific task.
Speaker: And so there are probably ways to write good notes and HNPs and progress notes.
Speaker: And it could probably take context if this was a patient, it's a reoccurring patient,
Speaker: You could look at previous notes.
Speaker: It could have the context of your progress note that you wrote the day before.
Speaker: And so it could probably give you good skeleton outputs of H&Ps, progress notes, assessments, plans.
Speaker: And the rest you can kind of just fill in and add in.
Speaker: Because we have to remember is that whatever comes out isn't the final product.
Speaker: You're still the human in the loop, and you're still going in.
Speaker: and adding and changing things.
Speaker: So, and then the AI could potentially just learn your practice pattern even better, how you write notes as well.
Speaker: And so that could be potentially added as well.
Speaker: So I think it's very much possible for you.
Speaker: So, but it would be a little bit of a different
Speaker: kind of workflow than in the ER or in the outpatient.
Speaker: I think the outpatient would be really kind of, it's much simpler to do because there's less ambient noise.
Speaker: If there's an audio system in the room, it can just hear the conversation between the clinician and the patient.
Speaker: So I think that's a little bit easier to apply.
Speaker: The ICU might be a little bit more difficult, but the audio is getting, the features are getting so good.
Speaker: I don't think we're that far away from that.
Speaker: Yeah, and I think that there's already applications that kind of take this principle outside of maybe a clinical conversation, but like there's all these applications and even devices that are AI powered that you can use for a meeting, right?
Speaker: And then it will create a summary of the meeting and to do and to follow ups and stuff like that.
Speaker: So I think eventually that is something along those lines, but much more clinically oriented.
Speaker: Cool.
Speaker: Yeah.
Speaker: So let's talk about how you use AI in your day to day.
Speaker: And I would like to maybe have one or two examples in the medical arena with some detail.
Speaker: But I also want you to share with us an example that's non-medical.
Speaker: I'm definitely interested in that as well.
Speaker: So go ahead, Sherrod.
Speaker: Perfect.
Speaker: And so I program my own agents.
Speaker: So let me kind of first start off by just identifying and just really defining what an agent, AI agent is.
Speaker: What is agentic AI?
Speaker: Because you're going to hear this more and more going forward.
Speaker: So if you were to look at 10 different sources, you'll probably hear a slightly different definition.
Speaker: So I'll give you my definition.
Speaker: An agentic AI or an AI agent is essentially the large language model.
Speaker: Let's say a chat GPT model and the GPT 4L model.
Speaker: And now it has the ability to almost autonomously, semi-autonomously,
Speaker: make its own decisions.
Speaker: So for example, if you asked it a question about the latest guidelines for the initial, the latest guidelines for tidal volume titration and ARDS, it could rely on its own kind of training database.
Speaker: And that is such kind of
Speaker: information that's in the literature and it's been cemented it's unlikely to probably hallucinate about that but it could say all right just to be sure i'm going to search the internet i'm going to search cochran i'm going to search up to date i'm going to search pubmed to get a better research report on this and so it could potentially just connect to the internet so you have these api calls to the internet connect to pubmed connect to cochran and make these decisions and pull that information
Speaker: and actually be able to have a separate agent that communicates with this first agent and it'll organize that data.
Speaker: And then you could have a third agent that talks to these other agents as the fact checker and the hallucination checker.
Speaker: And so the output you get is kind of a multi-agent type of thing.
Speaker: So this is what I primarily use.
Speaker: And so up until very recently,
Speaker: When you use ChatGPT, you use Claude or Gemini, it wasn't an agentic process.
Speaker: But now in ChatGPT, there's this deep research tool, which works very, very well.
Speaker: And so what I used to do until very recently is I would have a research tool.
Speaker: So if I wanted to write a new paper on fluid management in ARDS, I would have my agents search the web, compile all the information, fact check it,
Speaker: And so there would be four or five agents each with its own role, and I would get all the information back.
Speaker: And then I would look at that and then I could potentially start writing at that point once I had the information.
Speaker: The second way I use this is for hypothesis generation.
Speaker: So I actually use a really cool audio feature on the Gemini app on the phone where if I'm thinking about non-invasive ways to identify intra-abdominal pressure elevation using ultrasound,
Speaker: I would use this Gemini tool and we would just hypothesis generate.
Speaker: And I would say, all right, so signs of intra-down web elevation, what would they be on ultrasound if I was thinking about it?
Speaker: And it might say, okay, it might be a smaller than usual IVC or collapsing IVC despite a high CVP.
Speaker: And so we'd go back and forth and we kind of really refine this thought process and hypothesis.
Speaker: And then once I've gotten to the point where it's clean, it sounds good, then I might go and apply this in a research fashion.
Speaker: And so hypothesis generation is probably one of the biggest use cases I use this for.
Speaker: And again, most of the time I'm programming my own agents to do this.
Speaker: But now just the foundation models are so good, you could just use the foundation models to do things like that.
Speaker: So these are some of the medical use cases that I do for like, but non-medical.
Speaker: Our private conversation before, so you know, I was telling you that my love, my first love is philosophy.
Speaker: And I love to just take these books that I loved and I thought I knew, and I'll put it into something called Notebook LM.
Speaker: It's a Google product.
Speaker: And you put it in there, and this tool can create a podcast.
Speaker: um and i could listen to it there's actually it's a really cool feature where it's it it has two people talking and it sounds like a real podcast and so the it this gives me kind of like a high level understanding of that book again and then it has auto prompts on there to ask the next next best questions on this and say all right how can i understand this book more deeply um and so my own learning
Speaker: as like having a personal tutor.
Speaker: That's my kind of non-medical kind of use case for AI.
Speaker: And that actually has probably been my favorite use case is being able to learn difficult material in a more efficient way than I was before.
Speaker: And I think that one of the uses that is very, very important
Speaker: simple and like you said i think because you know what you because you're feeding it kind of like the the information it's less like to hallucinate is just to identify like giving them like a pdf of a paper of a article and just can you summarize the five more important points here right or can you tell me what they did in methodology and it does i mean takes like two seconds right
Speaker: Now, I think that works very well when you've already read the paper.
Speaker: It gives you other insights as opposed to doing that instead of taking the time to read the paper.
Speaker: Because we were talking about the effort, a lot of the effort that we do for things that maybe AI can do quickly are part of our building blocks for expertise.
Speaker: So in certain areas, it's probably worth using it as a co-pilot as opposed to just the only source.
Speaker: But those are all great, great uses.
Speaker: I was going to ask you also in terms of when you say you program the agent, can you give us a little bit more specifics about that?
Speaker: Like for me, what does that really mean?
Speaker: Is that you program with something like Python or you program with prompts?
Speaker: Yeah.
Speaker: So I and to that, I think it goes back to how I kind of learned all of this is where like a few years, like about six years ago, I became frustrated with kind of the
Speaker: end-of-life care with patients and I felt like every intensivist might have a different impression of what their kind of trajectory would be and different specialties would have a different trajectory and and I didn't think we were doing right by the patient by providing all of this kind of variable information so I six years ago I taught myself Python because I wanted to create this mortality prediction tool and then we did and I actually
Speaker: worked with a few engineers over at Rowan, which is associated with our medical school, and we created this tool.
Speaker: And I learned a lot.
Speaker: I made a lot of mistakes.
Speaker: I learned so much from it.
Speaker: And then I also learned SQL as well to kind of
Speaker: pull data from these huge data repositories.
Speaker: So I learned this building that tool before.
Speaker: And so I primarily program in Python specifically.
Speaker: And so when I'm programming these things, I'm using using Python and I use these kind of agent framework.
Speaker: So there's a lot of these out there now.
Speaker: Microsoft has one which is called Autogen.
Speaker: There's another one called CrewAI, and there's another one I use which is called Agno.
Speaker: And these are all agentic frameworks that use, and one of the languages you can use is Python.
Speaker: But what they do is where you may have needed 300 lines of code before, they've created these kind of agentic, like almost-
Speaker: simplified way to create these agents with Python, where it would be 100 lines of code as opposed to 300, 400 lines of code.
Speaker: And so for myself, as a programmer, I'm 201 level, but I know enough code and then I know enough how to prompt an AI that I can kind of create more advanced code and be able to troubleshoot things.
Speaker: And so this is how I create more advanced code
Speaker: apps and tools that are beyond my coding abilities.
Speaker: But I primarily code in Python and that's how I kind of create these kind of agent tools.
Speaker: Thanks for sharing that.
Speaker: And just a quick question.
Speaker: If I would talk with family, young kids going into college, I would always encourage them, like, you know, learn to code.
Speaker: Now, if you are young and starting or if you're old like me and want to move forward, do you think that learning to code is the key or just learning more about how to prompt AI?
Speaker: Since I keep reading that AI probably in the very short future will be as good coder as any human.
Speaker: Yeah, no, that's a great question.
Speaker: And I think I'd have to answer that in kind of a twofold way.
Speaker: One, should you learn to code for the practical applications?
Speaker: Maybe not.
Speaker: Should you learn to code because it can teach you how to think cleanly?
Speaker: Yes, I think.
Speaker: For a young person, learning code really teaches you how to think logically in a way that may not have been there in your initial primary college education.
Speaker: It's really cleaned up my thinking.
Speaker: And now I think in terms of almost like Legos, where these code blocks in my head, when I see a problem, I'm thinking of code blocks in my head that I would use to solve that problem.
Speaker: So I think of a young person coming up, 100%, I think, learn new code.
Speaker: you may not be able to use those skills directly, but just being able to think in a code type fashion will give you a leg up in a competitive, whatever competitive market you're going to be in.
Speaker: And I think it was just as far as the pure benefit of your brain architecture changing and being able to think more clearly, I think that benefit is just, it's beneficial no matter what the practical application of it.
Speaker: And then if you're older,
Speaker: and you're thinking about learning the code yeah i think it's great but let's say you don't want to learn the code that the no code tools are getting better and better so the inertia the the the blockades that existed to get into being an ai practitioner are getting smaller even in the past two years
Speaker: where the things I needed to code before, there are so many no-code tools now that you can get something in an AI app up and running in a few hours that's quite complicated.
Speaker: And it's really just kind of a plug and play thing.
Speaker: So as long as you actually know what things you want to put together, the more important thing is you understanding the actual details of what connects together.
Speaker: And so if you're learning to code and you want to get into AI, I would be reading a little bit more about what is a transformer model, what is embedding, what is attention, what is a vector database, all these things.
Speaker: You understand it more, you'll learn how you can connect these things and almost in like a Lego type of fashion and create the things that you'd like to create.
Speaker: Excellent.
Speaker: So as we move forward to closing, obviously, you did mention some of the challenges and cautions, right?
Speaker: You talked about hallucinations and trust.
Speaker: Could you just tell us, I mean, from your perspective right now, what are some of the biggest risks of using AI in a high-stakes environment like the ICU?
Speaker: Yeah, I don't think we're ready as of yet.
Speaker: It's just using the foundation model.
Speaker: But there are different approaches to doing this.
Speaker: So one, you can use an architecture called RAG, which is called Retrieval Augmented Generation, where the large language model is connected to the ground truth, the guidelines, the antibiogram of your hospital, the vitals, and the best methods to wean a ventilator.
Speaker: So it's connected to that.
Speaker: and the system prompt of the large language model is instructed to only rely on this external information, but it can do have its usual kind of decision-making, calculating, prediction ability, but rely on this outside source of information to make decisions.
Speaker: That's been shown to reduce hallucination.
Speaker: It doesn't get rid of it, but it reduces hallucinations.
Speaker: And so I think one of the biggest risks is if a clinician were to use ChatGPT today for clinical decision making without these additional kind of fail-safes and guardrails, there could be issues and you could potentially apply things to your patient that may not even exist.
Speaker: It could be hallucinations.
Speaker: The other things is just ICU research, right?
Speaker: So you creating...
Speaker: writing papers for ICU research and not fact-checking things and not fact-checking references.
Speaker: That can be a risk as well.
Speaker: So hallucinations still remain a big problem for large language models, but there's ways to mitigate and reduce this risk as well.
Speaker: And I think that that's where really I feel that a lot of it is right now that you still need, like you said, to have the human in the loop.
Speaker: And for high stakes environments, you need experts in that loop, right?
Speaker: I mean, and your expertise is important there.
Speaker: I think there's a great case of a lawyer who generated some information
Speaker: or he had basically a couple of cases generated by AI as to support his argument.
Speaker: And then they were all hallucinated.
Speaker: None of them were real.
Speaker: Eventually they figure it out and he lost his license, right?
Speaker: He was disbarred.
Speaker: And even though he didn't know they were unreal, I think the judges told him, look, I mean, it's your responsibility where you present.
Speaker: So he was disbarred.
Speaker: So obviously that you have to be careful, right?
Speaker: There has to be a human in the loop.
Speaker: What are you most excited about in the short term regarding this whole area and critical care?
Speaker: Yeah.
Speaker: So can I add a detail, ask a detail about that?
Speaker: Do you mean the AI in general?
Speaker: Because in general, I would say the new reasoning models and things like that are coming out.
Speaker: Or most excited about applications in the ICU?
Speaker: I would say both.
Speaker: Both.
Speaker: Okay.
Speaker: So...
Speaker: The things that I'm most excited about in just AI in general, which will have applications in medicine, is the AI agents is number one.
Speaker: Number two is the new reasoning model.
Speaker: So these reasoning models are incredible.
Speaker: And one of my best use cases for this, I think, going forward is that these reasoning models are able to
Speaker: almost they used to something called chain of thought reason and almost mimics human type thinking.
Speaker: Um, and it's able to really kind of put together a large amount of information,
Speaker: and think about it non-linearly and solve really complex problems.
Speaker: It has its issues still, but the outputs are incredible.
Speaker: The ultrasound example I gave you earlier of coming up a non-invasive way, I used a reasoning model for that.
Speaker: And I told it to use all your knowledge about the basic science
Speaker: of how ultrasound transmission works, how intra-abdominal pressure works, and how the best put together, how do you think elevated intra-abdominal pressure would affect the size of the IBC through the ultrasound lens.
Speaker: And so this type of thinking, these reasoning models are incredible at.
Speaker: So I'm very, very excited about that.
Speaker: And I think taking these reasoning models and AI agents, I think their application in the ICU are going to be incredible.
Speaker: So as the reasoning gets better and better, the outputs will get better and better and more accurate.
Speaker: And so I think the application of both of these in the ICU will be really, really interesting.
Speaker: What would you recommend or how would you encourage other clinicians to start their journey or become better AI practitioners?
Speaker: to practice, to just get in there, get your hands dirty.
Speaker: There's great courses on course.
Speaker: Sergio, I wake up every morning and I'm very excited to wake up because we live in an era where you can learn anything without actually going to a college campus.
Speaker: And so there's websites called Coursera.com or Udemy.
Speaker: There's open source MIT courses.
Speaker: There's open source Harvard courses.
Speaker: There's open source Stanford courses on all of these topics.
Speaker: I suggest just getting in there, taking a course on Coursera on there's an AI and healthcare course in Coursera, which would be a really by Stanford, which is a really nice kind of primer course.
Speaker: on things you need to know if you want to get into this field.
Speaker: Once you have that kind of primer, then you can actually just start practicing and trying to find these no-code tools.
Speaker: If you want to program, great, take a programming class or Python course.
Speaker: But if you don't want to learn how to program, there's so many no-code tools for which there's courses as well on Udemy.
Speaker: There's a...
Speaker: There's a tool called NAN, which is basically it's kind of almost like you're kind of just connecting things together in a visual way to create AI applications.
Speaker: So there's courses on NAN.
Speaker: And so really, I would say just take a beginner's course on any of these websites, find these low-code, no-code tools, learn how to use one or two well, and just start practicing and tinkering with it.
Speaker: And I think the same applies, right, to your daily use outside of medicine.
Speaker: I keep playing and you start learning, right?
Speaker: So, for example, what used to be like a prompt on Google, best restaurants in whatever city you're visiting, right, now it can become a much more detailed prompt for ChatGPT and it gives you actual some recommendations that are usually pretty good.
Speaker: Definitely, 100%.
Speaker: Excellent.
Speaker: Well, I would like to close the podcast with a couple of questions that are unrelated to AI.
Speaker: And I hope that you do not use AI to answer these.
Speaker: Would that be okay?
Speaker: That sounds good.
Speaker: Number one is, and I know you talked about philosophy, but you talked about books.
Speaker: So I like to ask our guest, is there a book or books that have influenced them significantly or a book that they have gifted very often to other people?
Speaker: Yeah, I think so.
Speaker: There's a few.
Speaker: One is Meditations by Marcus Aurelius.
Speaker: I love this book because, and so Marcus Aurelius was a Roman emperor and he had his own diary and he wrote, he practiced something called Stoic philosophy and he basically just wrote in it.
Speaker: And this was never meant to be read by anyone.
Speaker: But eventually someone got their hands on it and they published it.
Speaker: And it's a really nice way to look at this guy who was the most powerful man in the world, how he approached just waking up and how he's got his mindset in a correct way to be able to function in the most virtuous, ethical way possible and to maintain his ability.
Speaker: wellness and his mental health, as well as functioning well and creating the best Roman empire he could create.
Speaker: So I think a lot of the lessons in this book hold true today.
Speaker: Another one is thinking fast and slow.
Speaker: So this book might be one of the
Speaker: two or three books that changed how I view the world.
Speaker: So Thinking Fast and Slow was written by these two Nobel Prize, now Nobel Prize winning behavioral scientists and psychologists from Israel, Amos Tversky and Daniel Kahneman.
Speaker: And it basically,
Speaker: looked, it showed scientifically and experimentally showed the fallibility of the human mind, right?
Speaker: So a lot of economics research assumed a rational character, a rational human being.
Speaker: This kind of flipped it on its head and saying, well, you know what?
Speaker: Our decision-making actually isn't rational.
Speaker: We're not very good at statistical thinking.
Speaker: We're not very good at making money decisions.
Speaker: And so when I read this, it really, really changed the way
Speaker: I viewed decision making, I had more compassion for myself and I gave myself more slack.
Speaker: And it really directed
Speaker: my interests going forward.
Speaker: And I honestly, without this book, I probably wouldn't have picked up coding and AI and AI applications in medicine.
Speaker: So that book probably is one of the two or three books that changed the way I view the world.
Speaker: Excellent.
Speaker: I think they're both phenomenal books and I would definitely put them in the show notes.
Speaker: So thanks for sharing that with us.
Speaker: The second question, Sherrod, is could you share something with us that you changed your mind about recently?
Speaker: Oh man, I, I think one big thing for me is that I am always trying to get better at something.
Speaker: Um, so for example, I I'm 41 years old, but my goal this year is to dunk a basketball.
Speaker: So I'm doing jump training and I'm like, my wife's just watching me.
Speaker: It's like, I'm on the couch doing, uh, split squats and doing hops and kettlebells and all that stuff.
Speaker: And, uh, but
Speaker: In the same way academically, I have intellectual ADD and I'm always trying to learn something new.
Speaker: But what I've found is that this really fed my anxiety.
Speaker: It wasn't great for my mental health because I always felt like I needed to be on the go.
Speaker: So the one thing I've changed my mind on in the past few years is that there's nothing wrong with stillness.
Speaker: And in fact, actual stillness and just sitting there and doing nothing and just being present is
Speaker: has made me more creative and actually more, more my output has improved as well.
Speaker: Not to mention improving the relationships between my wife, my family, my friends and things like that.
Speaker: So I have a long way to go with this concept, but the concept of stillness is something that I've changed my mind about in the last few years.
Speaker: Excellent.
Speaker: And the other thing you could do is just lower the rim.
Speaker: Always work.
Speaker: That's correct.
Speaker: That's what I did.
Speaker: That's what I did to solve that itch anyways.
Speaker: So that's, I think that's what I'm going to do.
Speaker: Yeah.
Speaker: Well, let me know how it goes.
Speaker: So the last question is what would you want every intensivist to know?
Speaker: It could be a thought or quote or a fact as we close.
Speaker: I think the thing I wouldn't want the intensivist to know, especially in regards to this, um, the, this oncoming, um,
Speaker: AI wave, it's a tsunami, it's coming.
Speaker: And what I would recommend is just surf it, ride the wave.
Speaker: And for that, we're just going to have to be accepting of these changing workflows and these tools coming along.
Speaker: What the understanding is that
Speaker: If you work with it, your patient care, your accuracy, your cognitive reserve, all these things could potentially get better.
Speaker: So I think that would be my parting message for intensivists on this topic.
Speaker: And I think this is a perfect place to stop.
Speaker: And, Rod, thank you so much for sharing your expertise, your time, and your enthusiasm.
Speaker: I think that you're right.
Speaker: I like what you said.
Speaker: So I guess our closing comment is jump in the water, start paddling, and ride that wave.
Speaker: 100%.
Speaker: Hope to have you back soon.
Speaker: I'm sure that in a couple of months, everything that we talked about will be changed.
Speaker: But we'll definitely have you back and maybe to talk about other clinical topics as well.
Speaker: Sherrod, thank you very much.
Speaker: Thank you, Suryodh.
Speaker: Thank you for having me.
Speaker: Thank you for listening to Critical Matters, a sound podcast.
Speaker: Make sure to subscribe to Critical Matters on Apple or Google Podcasts and share with your network.
Speaker: Sound's transforming the way critical care is provided in hospitals across the country.
Speaker: To learn more, visit www.soundphysicians.com.
