Loma Linda Anesthesia Critical Care Fellow Conference 11_14_2025

Uh, communicate.

Yeah, I'm here. Sorry, I just can't turn my camera on right now.

Um, oh, that's it, I'm offended.

Let me leave.

No, we're done. So let me get started by at least recording. Fantastic.

And just let me, um, I mean, I'm so glad that you are able to do this lecture first. I'm sorry about, um, we didn't have a couple of the fellows, but Dr. Bendjikan, I work with him at Desert Regional Medical Center, but he is, um, really big in AI. He actually, um, does international, like, meetings and, um, does a lot of, like, he has a whole podcast on AI. Obviously, he's critical care, but I think you also have a a business background as well, right? And has like 20 years of international experience as well. So I think you've trained at Harvard, and we're just really lucky to have you kind of here to lecture us on some of this stuff, because I know you do really cool things. And he's actually going to be at the conference, Tristan, which Tristan's actually there too. I mean, hopefully, I think you'll be there at the conference. So, No way.

Yes, I'll be one of the speakers. And actually, they asked me if I could do something cool. I am focusing next year of speaking on AI in nursing, and I've started already posting some items and things for nursing, but I'm getting more and more into AI in for clinicians, not just doctors. And that includes respiratory therapists, that includes pharmacists. When we had the Southern California SCCM conference, my session was about let's put on stage a nurse, a physician, a pharmacist, and let's just pin them against each other and let's see how AI could sort it out. And it was fun. So I like to make my lectures a little fun and quirky and unusual. Gamified education, so to speak. And so this is going to be hopefully interactive. There will be a didactic portion to it, and I hate the word didactic. I think it's more, let me equip you with what you're going to use in order for you to, for the second half of the lecture, sort of feel like you're not just being thrown in the water, right? So My— thank you, Dr. McIntosh. My background is I graduated from Morocco and I came in 2000. I did my residency up in New York and fellowship in Boston. I like to say who I did it with, not where I did it. It was with Bartolome Celli. And so COPD is, you know, I breathe COPD, pun intended, and exercise physiology, sepsis, ARDS. And I discovered a knack for conversations with families regarding end of life, regarding critical care. I was myself a critical care patient. When I was graduating from medical school, I had a terrible accident. I, I think I'm living on borrowed time. I should have been dead long ago. And so as a physician/patient, I saw my colleagues who I trained with and, you know, spend sleepless nights preparing for exams, and the trust was there, walk into the room, and I would study their tone and their face, and they tried to protect my feelings, and they ended up losing not trust as in lying or anything, but just as a traumatized person emotionally at that moment trying to survive, and they were not telling me the truth, and they had good intentions. And so I turned that into how would I want to be, and I treat my patients that way. And I speak about this and now I took it into AI, which is how AI can be there for patients as well. So some of my segments and lectures are for how do patients use AI and what do you as a physician could use as tips to answer your first patient that walks in the room with, here's what ChatGPT told me about what I have. Or with an AI scribe themselves. Why do we feel like the scribes are just for us? They are for patients as well. So there's a lot in AI that I like to speak about, including ethics and informed consent and so on. But today it's gonna be a lot of how can you use it on your day-to-day basis. For me, I also ask myself, well, This is, we've never seen something like this before. What I mean by that is we've seen where computers started, we've seen where the internet started, we've seen where electronic medical records started. This is all wrapped into one and we are going to blink and we're already there. So I think in 5 years we're gonna have robots around us. I think in 5 years the payment system as we know it is not gonna be there anymore. I think that we will think about how we as physicians work. And unfortunately, we think too much of ourselves. And recall-based knowledge is no longer what's gonna make a good physician because that knowledge is in the hands of third-year medical students and patients and physician assistants and nurse practitioners. So to think that our value as physicians is how much we have in our brain, I'm gonna challenge that. What is ours and cannot be separated is the context and the nuance that exists in the room and how we as physicians can use that knowledge that is at everybody's fingertips to interpret and execute at the patient's bedside. Because how we are right now evaluated for our work is about to change drastically. None of this is my own opinion. This is not philosophy. This is a, you're going to remember this 5 years from now and you're going to say a lot of what is said is actually already here. So let me get started by saying AI is our, AI is the new stethoscope. Every generation has sort of contended with something. Some generations have contended with the computers. Some generations have contended with the advent of Dr. Google. And now a lot of you expect that patients will come in with Google-requested information. We will be the generation that contends with AI entering healthcare in ways we've never expected before. And so before we start to dismiss it, we might want to learn a little bit about it. As a disclaimer, I am a Doximity Digital Health Fellow. In fact, my graduation is next week, and I'm a Perplexity Business Fellow. I graduated a couple of months ago. So those are two AI tools that can help us in ways that we've never seen before. So I went there because I was biased towards them. I'm not biased towards them because I'm with them. Now let's start by talking about what AI— most people think AI is like you put in data, some nebulous black box AI, and then we get value. In fact, AI has a lot in healthcare. There's the data, which is all of the sourcing from public health data, but also all the remote monitoring that we're seeing right now. We have the AI development. What is the engineering? What is the model? How is it trained? How is it validated? We have the regulatory. Is it software as a medical device or SAP and requires FDA approval? Or are they circumventing that by just creating as if it's a wellness software so it doesn't need FDA clearance? We also have the clinical and operationalization of it. And then after that, you see post-deployment. So when we're talking about artificial intelligence, we have expert systems, Planning, optimization, natural language processing. And when we're going to be talking about— no, this is not that kind of lecture. We are not going to be talking about engineering. We're not going to be talking about data, binary, non-binary. This is literally a literacy for clinicians because we are overwhelmed by data. We're not helped by it. And so I don't think that we trust AI, and I don't think we should. But we do rely on it, and we will rely on it. And that's not the same thing. We don't trust that a CAT scan gives us everything perfect, but we rely on it. So we learn the false positives, false negatives, the sensitivity, the specificity. We don't rely on troponins. I mean, we don't trust it completely. We rely on it. So we learn about all the parameters for it. Let me see if I can. I'm going to share again, but this time I want to make sure that the sound is included. So let me— tell me if you can hear the sound.

Hey, Gemini, what do you see? Can you describe what image I'm scrolling through right now?

The image you're showing is a CT scan of the abdomen. It displays the liver, stomach, and surrounding structures.

What organ am I pointing to now with my cursor?

You're currently pointing to the spleen with your cursor.

Okay, Gemini, what am I pointing to now?

What organ?

You're now pointing to the liver.

Okay, so I'm scrolling down here now. Um, Gemini, what am I pointing at now? What organ is this?

You're now pointing at the pancreas.

Okay, and do you see anything abnormal on this scan?

Yes, there is a small amount of fluid around the spleen and the pancreas on the scan, which could be a sign of inflammation or injury.

Okay, um, anything else that you notice about the pancreas?

The pancreas appears somewhat swollen, which could suggest pancreatitis.

Okay, um, the patient has an elevated lipase level. Um, what do you think about that?

Given the elevated lipase levels and the appearance of the pancreas, it is highly suggestive of pancreatitis.

Okay, uh, what complications should I be looking for?

Based on the scan, you should be looking for complications like pancreatic necrosis, pseudocysts, and infection.

Any thoughts?

It seems pretty useful, that's for sure. You can just skim through.

So Gemini is to Google what ChatGPT is to Microsoft. It is powerful, multi-use AI that we are going to be using today. So now I would like to share with you this other. And these are not coming in the future. They're already here. This is Echo Prime, where instead of just doing an echo, it actually can read real time your echo with everything that you have been trying to learn With all of these conferences you went to and all of the certifications and all of it, it's not just an EF preliminary. This is literally the entire ECHO as it is being done. It is being read and somebody is going to sign off on it. Is it going to be trustworthy completely? Of course not. But do we trust completely everything that comes in a preliminary? We do know that there's going to be a reader of a cardiologist at some point, but at least we get what we need immediately. So Korea is trying to have 1,000 healthcare AI professionals by 2029. This AI has not only been the companion of a 13-year-old who killed himself but offered to draft his suicide note. There is the rise of patient-directed AI medical advice. Nurses are now prompt engineers, and they're learning about how to use AI as decision support. This is why we should care. We should care to be AI literate. There are 3 stages that I take my, I guess, those that want to consult me: AI awareness, AI fluency, and AI literacy. We should be AI literate by the end of this year, early next year, if we're not going to be behind, because AI is in the payment system. AI is in the notes. AI is in the ultrasound. I don't know if you know, but the butterfly actually, with a subscription, it can orient you to how to hold your probe at what angle, how to rotate it in order to obtain the best window. Everything that you are trying to do on your own, now you have a tool that can help you do it. The judgments remains critical. This is already here. This is not coming in the future. This is already deployed. Epic has already deployed Art, which is the physician copilot on Epic, where you can actually find notes that are from 5 years ago, see if the patient had a vaccination, Somewhere, what did Dr. So-and-so? It's an actual chatbot that you can, within Epic, ask about your patients. You don't need to click and scroll through every single note of consultants, the radiology that came. On the right lower corner, you would see Emmy. Emmy is the same, but for patients. Patients now can interact with the record. Can you tell me what is my TSH? Is it abnormal? What should I ask my doctor next about it? And in the right upper corner, this is now Comet, which is the ability to do predictive analytics. This worries me the most because it has predictive analytics, whether insurances are gonna be there or not to showcase if you have a 20% chance of lowering your A1C, 16% chance at 1-year risk of COPD exacerbation, 25% 3-year risk. This is already here. It's being deployed in 2026 globally. There are pilot hospitals that are already publishing this data. You'll never feel ready because ready is not a feeling, it's a decision. AI is cute, but is it for me? So you ask yourself, have you ever prepared a resume, a CV, learned a new language, performed a literature search, created a personal statement, researched or understood a new concept, had a presentation slide deck to make every day for something, a presentation? You're going on a conference. Have you had yourself looking for plane tickets and reservations and where to stay and how long, how far is it from the venue? Or drafted an email to an important person and you wanted somebody else to look at it to see if it is respectful enough? On the left-hand side, you will see How long it took the top apps to hit 100 million monthly users. Google Translate, 72 months. Uber, 70 months. TikTok, 9 months. ChatGPT, 2 months. This is where AI gets its information. On the left-hand side, what are the use cases in 2024 and now 2025? The top use, of AI in 2025? Companionship. You will find how people using it to organize their life, finding purpose, and healthcare. It is right here at number 10. On the right, where does AI, ChatGPT, Gemini, Claude, Mistral, where do they get their information? Reddit, Wikipedia, YouTube, because they have the highest of authentic usership. Not the accuracy of usership. So let's talk about what is this AI. We want to define what we're talking about. AI has multiple clouds under which each one, an umbrella. So we're talking about artificial intelligence, which is using or building machine that think like us. Under it, we've had AI for a long time, machine learning. That's what gives you those pop-up windows of DBT prophylaxis. That's what gives you the interactions between the medications you just put in. That's what gives you— and on and on. Machine learning is the ability to learn from experience of data without human programming. Deep learning are the transformers. That is what moved the AI curve. That is using, mimicking the human brain using artificial neural networks in order to perform complex tasks. And generative AI is under all of this, using all of this information to generate new text, and from there new audio. Then we did pixels, so now they are images, new videos, and even new code. So how does it work? Tristan, when you say you use open evidence, how does it work to give you exactly what you want is highly important. We took the tool and we trained it on data. Then we now create a probabilistic algorithm in order for us to know what is the next word. So I'm going to ask Dr. McIntosh, if I were to say peanut butter and jelly, that's how AI works. But I'm going to start using a bias teaching. If you were Dr. McIntosh in Malaysia, born and raised in Malaysia or in Japan, and I said peanut butter and? Okay, I don't know. Because that is American. That is not French. That is not Malaysian. That is not— so The data that you trained a model is important because it's going to tell you whether that next word, next video, next prediction of— it's not accurate, it's just more frequent. So there's an inference that is biased by the culture of that data that you trained your model on. So how is an AI tool created? Training data creates— we create an algorithm, and then we create the model based on that algorithm. Yes, but that is created in a lab. What about if you put it in the noise in the real world? That's the question. You will find me always having this human in the loop icon. In a lot of ways, because without us, deploying this is only going to be as good as the data it was created on and as good as the algorithm it was created and engineered for and the model that now is the result of it. One of my friends, I— she told me that she had problems with her spouse. And she went to an AI model and asked, can you help me resolve this conflict? And the first model said, leave him. He doesn't deserve you. You deserve better. And the second model said, the principles of conflict resolution are based on this step 1, step 2. Therefore, the action you could take, number 1, if it doesn't— And when she looked, the data training set of the first one was on love novels. The second one was on human resources booklets and material. So you can see that based on the training dataset, it's going to give you the output. And we know that as physicians, whenever we are given a study, a publication, first thing we say, what is table 1? What are the demographics? Are the patients studied here close to my patients to see if it's generalizable? Any questions so far? So I wanted to give you some linguistics and some jargon. Interpretability, accuracy, precision, sensitivity, and specificity. Explainability. I could have the best model, but if I cannot explain how it reached that conclusion, does it mean I can trust it? Would me as your attending say to you, I know because I'm your attending and I have experience, satisfy you even if I'm accurate in what I said? So the hallucination portion, just because you asked this a question, it's going to give you an answer. It does not know not to give you an answer. We spoke a little bit of a bias, and we spoke— we'll also talk about model drift. If I have that model of Gemini that understands the interpretation of CAT scans, will it be the same accuracy in a year? It's not. There is a degradation of the model performance. So I wanted you to at least have some jargon. So why so many AIs? We have ChatGPT, we have Deepseek, we have Gemini, we have Claude. Well, each one of them is competing with the other one and training in a different training dataset. We know ChatGPT trained on everything from the New York Times to the Bible to the neoclassic arts in the past. So it did have redefined the copyright infringement because a lot of people are suing it to say you used a lot of the stuff that was human-based, so you have a copyright infringement. And they redefined it by saying the output is not the same as the source, therefore it is not copyright infringement. And so it is redefining a lot of what we thought was known. One thing I want to give to you is A model could be a regular search model. That's the classic GPT. Or it could be a reasoning model which is using neural networks to actually reason through. It's going to take it a little bit more time, but it's going to give you more deep dive. And some of the models that you have, like ChatGPT, we have GPT-5, but you can see there a GPT-4o. So some of these were actually Don't ask what exactly is the number, just ask, is it a regular search model or a deep research or reasoning model? That will give you a difference in between both. Deepseek is one that is not based from the US, it's based from China. But why did it rise to prominence? Because it used reinforcement learning to build the model. With less expense, less time, and less resources used. While we like to jump to vilify everything that comes from China, in fact, this one is much more environment-friendly because the biggest cost, both financial but also in energy, in computing, in water to cool down the servers, All of these costs to the planet is much less than ChatGPT because the original training was done at a fraction of the cost in any way you see, and it is just as powerful as ChatGPT. So when you see these models, I wanted you to have an idea of what could be the difference between them. But why is it so complicated? Because we can't trust it. You can get a very confident answer, but you still need to verify it. Thus, the human in the loop. So how do we navigate AI? You need to be literate, and that's what I'm teaching you, some of the basics, its capabilities, but also its limitations. How do you interact with it? How do you talk to AI? How do you assess the output of something smarter than you? How do you supervise it? That's one of the problems we need to solve. So how do we talk to AI? If you learn nothing from what we talked about today, this is the core of what I'm going to be talking about today. We call it prompting. Prompting is how you ask AI. And Tristan, you shared with me, you asked it like you ask a colleague. But I'm going to ask you, Tristan, you have a medical student, you have a patient who is in septic shock. You say to the medical student, go start the patient on IV antibiotics. Is that enough to ask the medical student?

Um, no, I feel like, because you would need additional collateral information to specify, well, what is tinnitus of infection, right?

Or yes, completely. And if you came back and found that your medical students started the patient on IV penicillin, why would you be mad at them? You did not give them neither specifics nor context. That is the core of how you talk to AI. And if you talk to your medical students and said that, and they said, would you just say, I don't know anything, I'll just go do it myself. First time you'll do it. Second time, at some point you're going to say, I need to invest in the medical students so they understand what I mean. And that's prompting. You need to actually say to your medical student and now to AI, I need you to act like a cardiologist. I'm going to ask you about Swan-Ganz catheters. So you give it a role. You also give it context. I am a fellow in Loma Linda University. Why? Because it knows the internet, so it'll know what Loma Linda University and what a fellow is. I am going to be presenting a Swan-Ganz catheter presentation to Engineers, or to an audience of nurses, or to an audience of patients, or to an audience of medical students. So you give it context. And based on that context, it narrows the probabilistic answers it's going to give you from wide range to a narrow range, and that's how it's going to give you exactly what you want. If I was to just go to Google and say, When do we COVID vaccine? Versus if I was to say, when do we COVID vaccine? And then go down the possibilities and all of the links that we have and choose the ones that are scientific, skip the ones that are patient-oriented, or maybe the ones that are from some conspiracy theory, or the ones that are for the advertisement of something. I think in 10 years, people are going to laugh. You mean to say that in your generation, you used to go to a box and put something nebulous, and it's gonna give you hundreds of thousands of results, which you have to skim to. Nobody goes to the second page. The first page is curated by the financial gains of a search engine. That's how you gathered your information. I mean, it's almost— now that we know AI, what it's capable of doing, it's laughable. So when you are asking Open Evidence, or you ask any other tools including ChatGPT or Gemini, you want to give it context, you want to state the purpose, and you want to verify the output. And the second thing that is important is you need to iterate. If the medical student didn't understand you, you will send back that question in a little bit more refinement. So the anatomy of a GPT-5 prompt is you give the role, you give the task, you give the context, you give the reasoning behind it. You want to tell it, do you want it in a table? Do you want it in a narrative? Do you want it as an email? Do you want it as bullet points? Do you want it in a 5th grade language? Do you want it in expansive? Do you want it in a funny tone? How do you want it? And then you put the stop conditions. For example, I want you to act like a travel agent. I want you to find the top cheapest plane tickets to go to Orlando, Florida for May 2026, where ATS 2026 is going to be held. I want you to find— to give them to me in a table comparing the number of stops, the where I'm going to buy quiet, and I want you to only look at reputable travel sources. Do not look on Google. Do not look on Yelp. Do not look on— Am I making sense so far? Allison? Tristan?

Yeah, you're just basically specifying the parameters, right? The more information you give, the more you can tailor the response. Input equals output.

You are correct. I want to push back a little bit though. There is a method to the madness, and I'm introducing you to that method. There's an anatomy to the prompt because that anatomy changed when we started with ChatGPT-3, moved to GPT-4, moved to GPT-5, or that anatomy changed when we were with Gemini Flash to Gemini Pro. So as these models are improving, you improve Also, the prompt. What can I tell you that we already do? The way we used to think about CAT scans 5 years ago is not the same we think about CAT scans now. There's a high resolution, there's the cut, there is the vascular components, there is whether or not we can swap the density, whether we can go. So there are some characteristics as the technology has evolved that we have continued to learn to refine the way we think about that CAT scan interpretability for our patients. Yes? Still with me?

Yes.

Yeah. Beautiful. And then context engineering, memory, and tools that you use. I gave here some prompt examples. I'm spending a lot of time on the prompt because it's important. Provide me with an explanation of Swan-Ganz catheters. I want you to act as a professor in cardiology. Provide me with an explanation on Swan-Ganz catheters. I want you to ask that the professor in cardiology provide me with an explanation of Swan-Ganz catheters. Use 5th grade language to explain it to me. Even better, I want you to act as a professor in cardiology. Provide me with an explanation of Swan-Ganz catheters. Use 5th grade level language to explain it to me. Organize it in 10 bullet points and in 250 characters or less. Use as many examples and analogies as you need. Here you can actually start to use some clever ways. For example, you can say, use a C score of 1 to 10, 1 being low creativity, 10 being the most creative. Provide the above explanation as a C score of 7, and so on and so forth. You can just adjust your prompt in order for you to get more specific. You can also Go to ChatGPT, go to Gemini, even to Open Evidence, and you can say, I'm going to upload 3 examples of my emails that I wrote. I want you to use the same tone of my emails to give me the output. Or you can use this in emails that you are writing, or in personal statements, or if you are providing putting together a slide deck. This is an example of a poor prompt: "Tell me about sepsis." Good prompt: "As an ICU physician, I need a differential diagnosis for a 67-year-old diabetic patient with fever, hypotension, elevated lactate. Include sepsis considerations and immediate next step." You give it a clear role, specific scenario, or defined purpose. What is one of the prompts that I found useful? I am an ICU attending in Desert Regional. I'm about to walk into a room where there's a 62-year-old female who's dying of multi-organ failure from septic shock due to urosepsis. Unfortunately, she has 5 or 6 system failures. She has a delightful family, and I have been very busy today. I would like for you to give me some statements that convey compassion yet honesty in order to have that family meeting with them so that I can appear compassionate and empathetic, but at the same time be honest and not let my tiredness color the way that I speak to them. And I was pretty impressed. I'm going to see my loved one. I had a family member who got into an accident. They are a teenager, and I would like for you to help me find ways where I could be a good family member, supportive, but at the same time not fake, and give them some words of wisdom to say what they could have done to prevent that accident on my way to pick up my loved one. So it helps you and augments what you do. You're still the one who speaks to the family or to your loved one, but at least it gives you some supportive statements. Now, this is an important lecture I gave for CHEST where what are the top tips to spot a bad research article? You'll recognize some of the things that we teach. What is the question of the publication? Are the subjects like my patients? What is the sample size? What is the effect size? What happened to the dropouts? Is the study design biased? How good are the results? How are missing data addressed? Does the measurement matter? Am I measuring O2 sat or am I measuring PaO2? Check the figures. Are they truncated just to show an amplified effect? Did— how did the authors manage confounding biases? And I wanted to create something similar for how do you evaluate outputs of AI. On the left, you have the same list that I went through. On the right, I'm looking at AI. What problem is that AI tool solving? Is it clinically relevant? Was it trained and tested on patients like mine? What is the size and the diversity of the training data? How does it handle missing or incomplete data? Is it just hallucinating? Is the model biased? Is it fair across gender, race, language? If someone tells me, here's an AI that can read all of your ultrasound or CAT scans, or he's an AI scribe that can do your medical notes, I'm gonna ask all of these questions. Otherwise it's nebulous and I cannot trust it for my patients. How reliable are the results? How accurate, how precise, how calibrated are they? How transparent is the method? Explainability, reproducibility. Are the outputs clinically meaningful? How is the performance reported? Are there metrics? Is there, is my institution looking at those metrics? Are they validating them in real-world data? And is there human oversight, the clear role for the physician judgment? And so this is where some of the ICU benefits of AI, predictive analytics, early detection of sepsis, clinical decision support, open evidence, resource management, staffing for length of stay, et cetera, real-time monitoring, and in documentation. Any questions so far?

No, that's actually a very interesting model.

I'm glad you found it useful. What I'd like to do for the rest of this is to introduce you to some of the tools. And this is recorded, so you can just follow along. And please, this is not a lecture. I want you to stop me. I want you to challenge me. I want you to tell me stuff that you find maybe biased, or you don't like the sound of it. So I'm going to start here by using something that you know. You, you know ChatGPT, but I'm going to use something else. I'm going to use Gemini. Gemini.google.com. And in Gemini, what you see here is this. So now you understand that you can ask the questions that are in a prompt. So Alison, if you're there, do you have anything that you're working on right now? Maybe a slide deck, maybe you are needing to look up something you saw on a patient today.

I guess one recent thing was a patient, a TBI patient with subdurals and also venous sinus thrombosis. Um, whether or not we should start a heparin drip. And the context of this is that we have various services recommending for and against the heparin drip in this patient, and the risk of bleed versus risk of stroke. And last night, the patient decompensated and actually did have a worsening head bleed requiring decompressive craniotomy. Um, so the question is the risk and benefit of starting anticoagulation in a patient like this.

So let's ask the question: are you looking for the evidence? Are you looking for advice? Are you looking to put a visual comparing risks and benefits? What are you looking for?

I think the evidence.

Okay, provide you. So we're going to give it a role. You are a, um, a world-renowned expert in anticoagulation in the neurocritical care field. That is the role. Then we're going to say, My ask, provide me with recommendations on anticoagulation in this particular patient and with the evidence. Don't make up any new evidence that led you to those recommendations. And I'm going to say output provided to me in a table with comparative risks and benefits of heparin infusion versus no heparin infusion. And just for the kick of it, we're going to say provide me with a scoring system that objectively compares the risk-benefit ratio so I can make my decision. Do you— is there anything you want to limit it? Do not look at something or look at these guidelines. Is there anything specific you want to restrict it with?

I don't think so.

Do not use— I'm going to just do, do not use any evidence that is not peer-reviewed. How's that? Now, this is a general retrieval. If you click on tools here, you will find that there is deep research. That is a reasoning one. It'll take longer, but it'll give you a much, much, much more in-depth article in the end, or whatever output you asked from it. Another thing that while it's doing it, you can see here, you can create a video. You can create images. That's what we call Nano Banana. That's how people create images of them with Drake or them with Michael Jackson. And this is guided learning. If you wanted to learn something new or you wanted a learning guide or learning partner for your boards or whatever, you can click on this and it opens the context of learning. So this is one of the most challenging scenarios in neurocritical care. Here's the evidence. Evidence of anticoagulation, TBI post-surgical. Here's a risk-benefit. Initiate versus withhold. You can export it to Sheets. This is the decision-making framework in lieu of a score. Instead, as an expert, I use a framework on sequential priorities. First, 0 to 72 hours, confirm hemorrhagic stability, then reevaluate after 72 hours of stability. And this is the summary of the recommendations. Now, do not anticoagulate, focus 100% on managing ICP, CPP, and Achieving hemorrhagic stability. Next, post-op 3 to 7 days reassess. If and only after 72+ hours of CCT stability, this aggressive bleeding TBI has proven they are on the extreme end of the hemorrhagic, etc. Thoughts? Feedback?

No, that's so cool how you can be so specific with like your inputs to kind of direct the answer with it. I never thought about using it in that capacity.

I love it. Now let me ask you, Dr. McIntosh, I was going to ask you—

I'm sorry, go ahead. You put in like, don't make up any new evidence. Would they make up new evidence?

Like, yes ma'am, yes ma'am, they will, because they do not have— it doesn't have something that is called, I don't know. And so this leads me, thank you so much for segueing me to the topic of hallucinations. And I'm going to try to share with you hallucinations are not really hallucinations. That's what they called it, the engineers. We as physicians, we call it confabulation. And so it confabulates, and I call it lying with confidence. So this person, this is the prompt that it gave ChatGPT, and it gave it an answer about this richness and operational taxonomic units. And these are the references. None of these references exist, yet all of these names are in PubMed, and all of these names and all of, all of them All. In fact, the word of the year of 2023 was hallucinate because of that. I don't hallucinate. I just believe what I'm told. So what do we know about hallucinations? Well, guess what? There is so much that happened around hallucinations. There are lawyers that actually sent in front of a judge something, and the judge says, I can't find these cases. And it turns out that they had put their— all of their case on ChatGPT, and they just copied and pasted it. It actually happened to me. It happened to my colleagues. There are now— all the journal editors are required. They're looking at all the references that you send with every publication to make sure that they exist. And another limitation I wanted to share with you today before I sort of I thought that with visuals, it's actually much more than just telling you slides. So this is a video that was made by various people in France about an AI that creates images. It's one of the most powerful and popular and the best ones, and it's called Midjourney. It gives you almost realistic images and pictures with just a text prompt. And I'm gonna take you to something similar. If you were in English, you would say— if I would say to you the word the projects, if you just learned English yesterday, the project is an English word. That's it. But if you lived here in the US, the project has a social Connotation. Well, in France, the, the outskirts of town are called banlieue. And in banlieue, it's a French word. If you just learn French, it doesn't mean anything to you. But in France, it means something with a social connotation, much like the projects are to English. So this Midjourney, these people, here's what they do, and it is in mostly, um, subtitled. C'est abusé. So what meant was one word that has a bias has infiltrated AI. And what that meant is, unfortunately, we're going to have to make sure that every tool that we use, including open evidence, that for the sake of our patients, we ask, where was it trained? What biases inherently does it have? Because AI will magnify our biases into a decision tool if we come to over-rely on it. This is a study that was published in Nature Medicine earlier this year. And what you can see is when large language models recommended what to do for a demographic group, it was very different. If you asked, give me a decision about this particular patient, and all you did was change 35-year-old with dyspnea, woman versus man, or you did a Middle Eastern versus a Black patient, you started to see differences. Black patients, more frequent recommendations for invasive procedures. And mental health evaluation exceeding clinical necessity. LGBTQIA+ patients, mental health evaluation 6 to 7 times more than control with identical symptoms. Unhoused patients, psychiatric interventions. Low-income patients, advanced diagnostics or specialist referrals less likely. High-income patients, more likely advanced diagnostic tests. So we need to be aware of those hallucinations and biases in healthcare in every tool that we use. The last thing I want to share with you today that you will hopefully find useful is a tool called Gamma AI. And Gamma AI is a tool to create a slide deck. Dr. McIntosh, do you have a presentation you're going to make sometime in the near future? Yeah. What's the topic?

Um, mechanical ventilation.

So we're going to click on this and say presentation. Paste here. I want a presentation on mechanical ventilation. I am a critical care attending. ICU attending at Loma Linda University. The audience is whom? Fellows?

Yeah, sure.

Students? Fellows of critical care. And we're going to go— some of these are, you can choose default or traditional or tall. You can choose generate from lines or summarize the text or preserve this exact text. So We're just gonna go into a presentation. You can make it into a web page document. And we're going to try 10 slides, 15 slides. How many do you want? 15 slides. And then we're going to put in it. Do you have a specific thing you do not wanna miss? Make sure to include pathophysiology and all the different vent modes. Vent modes, asynchrony, maybe. Okay. Generate outline. It's gonna give you your outline. You can change whatever you want right here. Or if you like this outline, we're going to say that it's— we like this outline. It's already asking you, do you want it minimal, concise, detail, extensive? We're going to do concise. What color theme you want? We're going to choose this color theme because why not? Do you want AI images? You want stock photos, web images? We're going to leave AI images. Do you want which model? We're just gonna— I know the models, but for right now, what kind of art style do you want it? Abstract? Do you want line art? Do you want illustration or photorealistic? We're gonna leave it photorealistic. And any other thing we wanted? Vibrant, gradient, peachy. I want it professional, sophisticated, not basic. And we're going to generate. And we're going to watch the magic. Airway protection, respiratory failure, clinical triggers, primary goal. We have normal versus mechanical breath, positive pressures, diaphragmatic contractions, understanding peak pressures, plateau, auto-PEEP. Here are the modes, key characteristics, advantages, disadvantages. Pressure-controlled ventilation key characteristics, pressure support, spontaneous mode, cycling criteria, requirements, APRV and PRVC. Then you have the ventilator waveforms, visual interpretation, asynchrony recognition and management, common types of asynchrony, flow asynchrony, cycle trigger, double triggering. Sedation requirements, inspiratory time, cycling criteria. Thoughts?

That's really good. I mean, I've used ChatGPT a couple times to like help with presentations, but this is visually very different than what they, when the, what they usually produce.

I don't know.

So maybe I have to use that other area or like presentation specific.

Yep. And you could download it into PowerPoint.

Wow. You call this, this one was called, what was it? Gamma?

Gamma.app or just put Gamma AI. There are a few. There's Beautiful AI, Decktopus AI. Those are the ones I use the most. Sometimes what I do is I go to Gemini, which people are sleeping on Gemini. They go to ChatGPT, but Gemini is powerful. And I take some of that and I take the output and I put it in text. So no longer a prompt, I actually put the text and say, put slides for this text and it will turn it into visuals based on the text you have just generated from a generative language, large language model.

Are these free ones, or is this like a better version?

Because you can have some of the, like, you know, much like ChatGPT or other, you have the free version gives you like one image a day or one thing day, or they limit the number of slides. Same thing with Open Evidence is staying free. Doximity GPT is staying free. We didn't talk about scribes today, but that is a full conversation on its own. Feel free to go to my podcast. I talk about these and I bring in guests and we speak about publishing. There is— there are AIs that can actually modulate your manuscript towards a journal so you don't have to do it. There are AIs that can help you with your scribes that I introduced in my rounds with my residents, which gave me back a lot of time. But at the same time, the output— I'll give you one thing about Scribes. The hallucinations in Scribe could happen. You would say the heart rate is normal and it gives you EKG is normal, or you would say breast exam is normal and it'll tell you mammography is normal. So the human in the loop is essential, but it does give you a lot of power and it gives you back time. And that is what I'm going to end today. I hope that I gave you The Gift of Time, where most of our burnout is coming from. And I hope that you found something useful here. Please feel free to reach out to me. I'm mindful of the time. We are just a couple of minutes over, so do not hesitate to reach out to me. And I look forward to a time where I get to meet you in person. Thank you so much.

That was really good.

That was— there's a lot of good insight to that for sure.

It is my pleasure.

Really helps to learn how to actually use this, and I really appreciate all the tips on all the pitfalls on where, you know, those who are inexperienced would blindly follow some recommendations and know how to at least recognize it and start practicing how to avoid them. I really, really appreciate it. Wonderful.

It is my pleasure. Do use it in your own lives, please, because that is inconsequential, so that when you're about to use it for a patient, you're already trained and know how it works. I'll give you an anecdote. I received a text from my loved one saying, we have 2 tomatoes and a potato and a cilantro and this. Tell me what you want for dinner. And I was rounding, and I just grabbed all that and put it on one of the language models, and I said, give me— you are a renowned chef. I want you to give me a Mediterranean recipe with these ingredients to make within 30 minutes for a soup tonight. And then I sent it back and now it's called the ChatGPT soup in our household.

That's pretty funny.

Please take care and I wish you a fantastic weekend and don't hesitate to reach. My podcast is the AI Ready Doctor. You don't have to pay anything. You can just listen to it. And I have the Medical AI Academy. You can reach out to me. I'm happy to walk you through one-on-one. None of that is for money. It's just to bring the clinicians to protect our patients.

Thank you.

Awesome. Thank you so much.

Bye-bye.

Bye.

Thank you.