This transcription is provided by artificial intelligence. We believe in technology but understand that even the smartest robots can sometimes get speech recognition wrong.

  Today on Town Hall

AI is not in the future. The AI is here. The AI is Here with us right now. a term at Amazon called undifferentiated heavy lifting work.

Repetitive. No value works if we could use a I to help freeing our resources, they could spend more time with our patients and improve patient outcome.

My name is Bill Russell. I'm a former CIO for a 16 hospital system and creator of This Week Health.

Where we are dedicated to transforming healthcare, one connection at a time. Our town hall show is designed to bring insights from practitioners and leaders. on the front lines of healthcare. .

Alright, let's jump right into today's episode.

 Hello everyone and welcome to Town Hall, a production of This Week Health and Al's AI Insights podcast. I'm your host, Dr. Al Valera. For this session, I'm honored to be joined by a collaborative clinical and vendor group using AI to solve a very important patient care scenarios. Let me introduce you to the team, Christina Silvers, who is a healthcare executive advisor with Amazon, Dr.

Christine Silvers Dr. Praveen Mecha, who's a hospitalist, medical oncologist from Dana Farber and instructor of medicine, Harvard Medical School, and King who is Senior Solutions Architect and Healthcare for Amazon Web Services. It's an honor to have you all on the show. Thank you very much for attending.

Please go ahead and introduce yourselves further. Al, thank you so much for inviting us to your show. My name is Chris Silvers, and as you said, I serve as Healthcare Executive Advisor with Amazon Web Services. And during my 25 years prior to joining AWS, I spent my time as a researcher in artificial intelligence, as an emergency physician caring for patients, And as chief medical officer at two healthcare technology startups.

My AI work originally focused on improving bedside alerts in the hospital intensive care units, which was a collaboration with MIT, Harvard Medical School, and Boston Children's Hospital. Currently, my AI interests are broader, really encompassing any areas that our healthcare customers are interested in pursuing, all to ultimately improve patient care.

That's great. Dr. Meka. Thank you, Al. Thank you everyone for having me here. Introduce myself. I'm Praveen Meka. I'm currently a physician at Dana Farber Cancer Institute. This is a hospital based in Boston. It's a cancer hospital. I've been a clinician in all for 15 years now in varying, positions and institutions I am currently a hospitalist.

So that means I take care of patients admitted to the hospital. And apart from my clinical work, I also love to work on digital health technologies. 1 of the things I'm working on is called console, but that helps interpret blood tests, especially the blood gas using a language model.

Excellent. King. Welcome. Thanks for inviting us. My name is Ching Liu. I am a solutions architect here at AWS. I have a passion in using technology to help conditions as well as improve patient outcomes. So I had the privilege working with Dr. Mecca and Dr. Silver's on creating this console bot. So I'm very excited to share what we have achieved.

We're honored to have you. I'm very excited to have you discuss it as well. So what is some of the most promising applications in AI and healthcare today? If you see it from your perspective and how do you see them transforming patient care in our near term and in our future? Al, there are just so many possibilities.

I would love to just share. A whole bunch of examples in the context of a fictitious patient scenario to make it easier to remember. If you don't mind, if we start, that'd be great. So let's say our patient has a newly diagnosed, likely lung cancer seen while getting a chest x ray. So our patient needs to be referred to oncology.

AI can help auto generate the referral letter. At the same time, the patient gets automatically invited to join an AI powered personalized cancer care app for navigating their care. And this app provides the patient with features like relevant educational snippets in the patient's native language and at their literacy level, streamlined self scheduling for doctor's appointments, and a support community with other patients and families.

Meanwhile, the oncology department starts receiving many fax PDFs of the patient records. AI can do intelligent document processing to group these documents and convert them using an electronic format. The oncologist then has AI review these records to provide a summary, so the patient can skip having to fill out those history forms in the waiting room on the clipboard.

And AI can be used for ambient listening of the patient provider conversation to then draft the clinical note for the electronic health record. This not only saves time for the provider, but also improves the experience for both the patient and the provider. Then let's say the oncologist orders a chest CT scan.

AI can do automated interpretation of the scan images and uses these results along with EHR data to help the oncologist decide on the treatment. And AI can automatically code the encounter to bill for reimbursement and fill out prior authorization paperwork required for the medication. Then let's say things go really smoothly for several weeks on this special medication, but one day the patient becomes really short of breath.

And they call 911. The emergency room doctor does a number of tests, including what's called an arterial blood gas that Praveen mentioned. And these results can be really confusing to understand. So AI helps with interpretation, which is critical to make the correct diagnosis and guide correct treatment.

The ER doctor then admits the patient to the hospital. AI can be used to check that the medications and procedures ordered make sense. given the patient's age, weight, kidney function, and so forth. And when the patient improves, let's say the patient gets discharged to home with a blood pressure cough and a pulse oximeter, AI can monitor these vital signs at home to look for any concerning trends.

Now, clearly I made this scenario up, but it's to illustrate that there are many ways in which AI can potentially help us in healthcare. Do we have currently an opportunity to facilitate any of what you said just now around patient care? Oh, definitely. There are already a lot of these that have been created.

Ambient listening, automated coding, multimodal analysis, and so forth. And for those that don't already exist, they can be implemented using AWS's existing technology building blocks. In fact, I think Praveen can describe one of these innovations that he's built using these building blocks. Thank you, Chris.

Yeah, I'd love to talk a little bit more about console bot. So, this is a tool that helps interpret blood tests a specific blood test called blood gas. So, blood gas a little bit background about blood gas for those of us in the audience who may not realize what that is. It's a little more complicated blood test than you'd get in the clinic.

It's usually done on the inpatient side. So, as a clinician at a hospital, I use it mostly to triage patients, especially patients who seem to be clinically decompensating and the problem with this blood test is, one, it's, of course, done in in a more acute setting, then, two it can take a complex blood test, like it has a var variation amount of complexity, a more complex blood gas with a triple acid based problem could take 5 to 10 minutes to interpret for a clinician.

And finally, when it gets up to a triple acid based problem, or even double acid based problem the risk of clinical misinterpretation by a physician goes up too. Like, for example, I was reading somewhere it's as high as 70 percent if they're the triple acid based problem for Misinterpretation so having a tool that does that for you would be very helpful. And that's why we start to work on it to begin with. There are 3 components to a blood gas usually, but they could, of course, be more. 1 is a pH. 1 is a CO2 level. 1 of the bicarbonate. A normal pH, if you go back to school, would be like 7, but for a human body, 7.

4. So anything lower than that, it's a static and higher than that alkalotic. And it's a complex interplay between the lungs and the kidney that pulls the acid base in one direction or the other. You can get more answers just by looking at a blood gas. For example, recently, I was taking care of a patient on the floor and he was trying to breathe fine, so I knew something was not right.

So the first question is, okay, what do we need to do about his breathing and does he need to be in the ICU? So I ordered blood tests that gives me an idea of how to triage a patient. In this case, there seemed to be respiratory failure. That's the term we use when the breathing is not very good and we've got to decide does the patient need additional respiratory support.

So we put him on a little bit like a BiPAP. It's a partial, like, support that we could do on the floor. And we checked the blood gas again in, in in 30 to 30 minutes to an hour. And we found it actually got worse. And on top of it, there was something known as lactic acidosis that points towards worsening sepsis.

So that's another life threatening condition. In the end, the patient ended up going to the ICU but I just wanted to give you an idea of the importance of a blood test, like a blood gas, that we use to triage and appropriately put the patient in the right right, bucket. By that I mean, like, right floor.

So yeah, that's, basically a little snapshot about a blood gas. And now what we built, a console bot, is tool that is a combination of language models, so we leverage language models, and we have little decision architecture that involves hard coded rules. Like, for example, I said the pH is less than this is a static pH, less than that, it's higher than that, alkalotic.

So we have a little more hard coded rules in addition to it. Thank you. Something we use known as RAG architecture. So that's also stands for Retrieval Augmented Generation. So what that means is when you ask a question of a language model to be able to get a more accurate answer, it goes to a knowledge base and gets a chunk of text that helps answer that question.

So it's answers more detailed and more specific. So, creating that, we created console bot, and in the most recent iteration, we had accuracy of 98 percent on a sample of 100 blood tests. So, we gave the language model, like the bot 100 tests to interpret, and 98 percent were accurate. We, of course, need to validate a little bit more on real clinical data, but this is the preliminary data that is very encouraging.

It's fascinating. Make Praveen, is your is the access to the technology within your EMR is a work within that drop down? Or do you have it portable? Or I guess both within the EMR and on in your devices as well. So, as of now it needs more clinical validation, so we have it like an app, like a web app.

You go to that. You can start entering the blood test, and you get an answer. In the future, once it's clinically validated, we will work on integrating it in the EMR, so the clinician does not have to enter each and every value. As you mentioned, there are like several values, like sometimes it can be as high as five to seven values you have to enter in.

So, if it could do it automatically, that would be great. So, once we know it works well. On clinical real patient data, then we'll look at the next step at integrating it as fast today because I remember as a fellow emergency physician, always had that one little nugget back in my head. I gotta figure this out.

And how do we do this? And you sit by yourself in the corner and try to do the math on, figuring out metabolic, acidotic, etc. So, so wonderful work. Thank you so much for really solving a problem that one is very important to get it right every single time for the patient, but also needs to be done in a timely manner.

It was incredibly helpful for diagnosing patients with COVID, and they processed over 400 x rays on the first day and 65, 000 x rays in the first six months. And moreover, their AI solution impacted the clinical decision making about 20 percent of the time. And another example in terms of improving diagnostic accuracy, GE Healthcare wanted to embrace machine learning to drive improved patient outcomes.

So they collaborated with clinicians at UC San Francisco to create a library of deep learning algorithms for improving imaging technologies like ultrasounds and CT scans. And they incorporate data like patient reported data. sensor data and other sources into their scan process to help in distinguishing between normal and abnormal results.

And thus far, 82 percent of their healthcare decision makers surveyed say that the use of data is already resulting in improved patient care, and 63 percent report lower readmission rates. This is the dawn of penicillin being introduced to health care years and years ago. We're actually helping patients now, not just watching them suffer from sepsis.

So moving along, AI has really come a long way to helping us drive better care for accuracy. But understandingly, it isn't just clinical. It isn't just informatics. I know there's a whole technical evaluation that's done. And Ching, you can talk about that. I mean, what kind of considerations do we bring in when we're implementing A.

I. And the health care, especially around security, cloud based on data priority. Yeah, of course. Because we're in health care security is always on top of our mind. We need to safeguard the patient data. And in the United States, this means that we need to protect the patient data according to HIPAA requirements.

At A. W. S. We have a list of HIPAA eligible services. Which means that as a customer, you can use any of these services to process and store P. H. I. Data in a secure way. for example, one of the services that Dr Meta has used in his solution is called Amazon Bedrock and Amazon Bedrock is one of the HIPAA eligible services that we can leverage for bringing generative AI into your application.

So, and there are a number of things you need to know in terms of the security aspects of Amazon Bedrock. First is none of the customer data is used to train the underlying model or shared with the model providers. And secondly, All the data is encrypted in both at rest as well as in transit. And the third thing is from the data sovereignty perspective, the data stays in the region that you specify, so it's not copied to other regions or locations without your awareness.

And not only Bedrock is HIPAA eligible, it also meets other compliance program requirements such as GDPR. Sock ISO and so on. And the last thing I would say is that implementing AI in health care is no different than implementing other health care applications. You still need to worry about data access control.

So the data is only accessed by authorized users and with patient consent. And you should also think about how do you enable patients to maintain control over their data. So they should be able to opt out if they would like to figure it. You're on a roll. I'm curious, just like implementing anything new in life.

People have to gain trust in that. Where do you see the limitations and let's just say the challenges that you've supported from a clinical perspective, the clinicians using your design, your technology in creating that trust with them, as well as we all clearing to trust with the patient. How do you bring all that together?

Yeah, so we do face a lot of challenges and I would also like to share how we are addressing those challenges as we are working together with our customers. So I think we, we already touched about security, right? So it's always everybody's concern, how do we.

Leverage generative AI or other AI technologies in a secure way without exposing patient data or any organization's sensitive data. So we have AWS services and open solutions to help our customers set up that secure infrastructure. So I think another, other. Challenge or another common question people ask is how do we get started?

Do I need AI skills? And that's where amazon bedrock service comes in, right? The service is designed to democratizing access to cutting edge foundation models as well as large language models to provide easy access to a wide range of foundation models in a secure way. One of the probably easiest way to get quickly started using generative AI is by logging into AWS console and start using the playground feature and you can start, experimenting with different large version models and see how it performs with your particular use case.

And I think, one of the common. Phrase we hear in the field of generative AI is hallucination, right? So it means the generative AI is making incorrect and false statements. But these statements are presented in a way that sounds very factual. And because of the nature of generative AI, this is not something we can get rid of it completely.

But there have been Techniques that we are leveraging to, reduce the hallucination. For example the retrieval augmented generation that Dr. Mecca mentioned is a common approach to bring additional context to the large selection model. So it can generate a more accurate response. There are some other techniques such as chain of thoughts.

You can adopt in your prompt to, to make, to kind of guide the large action model to come to the right conclusion. And it is a rapidly evolving field. So I think we're going to see more and more interesting approaches in the future to find different ways to improve the performance of large ion models And I think the last thing I want to mention is that when people are thinking about building generative AI applications or bringing generative AI capabilities into your existing workflow, one thing to, to think about is how do you protect that generative AI application? Especially from undesirable content.

What I mean by that is there could be topics that you would like to avoid. For example, if you're building a generative AI powered patient appointment booking chatbot, you probably don't want this chatbot start giving out medical advices, right? So that's a topic that you would like. the genetic AI to avoid discussing with the patient.

There could be harmful content such as violence hate speech, and so on, and you would like to filter those out. And there are some interesting ways to Break generative AI called prompt attack or prompt injection, which is a way to kind of trick the generative AI to do something that's not supposed to do.

So, how do you kind of safeguard your application against all these different undesirable content? We have tools such as Amazon Bedrock guardrails that you can leverage that you can kind of creates this layer of protection around your generative AI applications so that you can safeguard your application against the those contents that you want to avoid to you.

You create an incredible environment that we can practice medicine in safely and ultimately take care of the patient. So looking at the future or near future, as we see it today, personalizing that medicine, how do we personalize it for the patient so no longer the patient have to talk to one of us to get information and go to the next step, whether it be creating a call for an appointment or getting advice on medications or giving just some challenge question to say, Hey, how do I do something with my, Take care.

You know, A new albuterol inhaler I received those kind of questions don't need us anymore. They can use AI. That's personalizing the medicine just like we're personalizing shopping on Amazon itself. We're personalizing the same experience, I believe, which is wonderful in the same understanding that we're here for the patient and patients will benefit.

When you do so, how does that enhance through AI, Christine your treatment plans, your patient outcomes, is there better satisfaction, and what have you seen from that aspect? Yeah, well, there's a lot of different ways in which we can personalize medicine, both in coming up with diagnoses. Is Figuring out optimal treatments or in educating the patients.

For example, I can be used to analyze the patient's individualized results, such as from the H R and their genomic data, and then in conjunction with the published research that helps the providers figure out what their diagnosis is, like the Undiagnosed Diseases Network can help patients with.

symptoms of unknown origin, or which treatments might be more effective for a particular patient might depend on their particular biomarkers, like in the case of breast cancer or lung cancer. A different example is in personalizing medicine is with the clinical practice guidelines.

Which I'm sure you, remember that, or are familiar that these are often hundreds of pages long and updated every year or every few years, and there are guidelines for heart failure, for diabetes, for chronic kidney disease, and so forth, and it makes it really difficult for providers to keep up. This AI can help providers identify which specific individuals are aligned with which particular practice guideline recommendations.

And one of the startups that I previously worked at, in fact, focused on creating these algorithms. I'll just give an example on personalizing education. The T1D Learning Camp game is an application created in collaboration with Harvard Medical School which uses generative AI to personalize the conversations that teach children with type 1 diabetes how the specific foods that they say they're eating will then affect their blood sugar.

And it's all in a game format. So all of these ways are Personalizing treatments and education that can hopefully then improve the patient outcomes access. Terrific. I think we've AI has come a long way in terms of really communicating and joining the patient outcomes with our technology. So allow a collaboration to care rather than a one way informative process where the patient now has more information.

They have more information. ever could imagine, but also has now ability to ask questions and get the answers immediately through the A. I. Component. That's one patient. Let's talk about hundreds of patients. Talk about medical research. Where's medical research, drug discovery and new therapies coming through the advancement of utilizing a I within the research room.

So recently I was reading like the Nobel Prize in chemistry was for AI modeling of protein folding and apparently the program these two researchers created it was called, I think, AlphaFold was able to predict protein folding in 200 million instances. So that's a huge mind boggling number and such protein folding can be used for identifying protein too.

Figure out which drug to make, what the target should be, or to predict antibiotic resistance. So in a lot of healthcare applications. So, so the fact that they were recognized at that scale tells me is already widely being used in general. Like, we focus on language models, but they've also the traditional AI previous to language models came up.

So it's been ongoing for a while. I'll let Chris Wayne to. Chris, go ahead. Yeah, and besides the drug discovery area, AI can analyze data from the patient's records and match them to clinical trials to increase enrollment and potentially the diversity in the studies. So one example is this company called Thread.

They've developed a cloud based platform for running decentralized clinical trials. And they estimate that they can support up to five times more inclusive enrollment compared to the industry benchmarks. And then another example is sort of gets back to the clinical practice guidelines I was just mentioning is the people heart study out of Harvard Medical School.

This is a research study interested in understanding how data can empower individuals to evaluate their own heart health. And it's trying to look at whether individuals. who gain digital access to their lab tests and their medical history from their care providers can then produce personalized heart health recommendations based on the current guidelines from the American College of Cardiology and the American Heart Association for the, prevention of heart disease.

And then, Praveen mentioned, the work in with the Nobel Prize Pfizer is another example working with AWS. They're using generative AI to help identify new oncology targets, previously a manual process to aggregate the information, but Now, AI can help to identify and correlate the relevant data and speed the whole process and improve the probability of their success.

That's terrific. I think, we have a new tool that we all have to learn how to use appropriately and to its extent. There's little spears going in and out. Youth group has created a very strong spirit to clinical outcomes and understanding how to utilize it for patient individual care, but also for research.

But we know technology is only as good as access to that technology. So where do you see AI working around the underserved and remote areas in healthcare? Do we use. Telehealth around that. How does this work for those patients who cannot get to you, but you need to get to them through a I there's a there's one example.

I like to share. So, this is actually a project. We collaborated with New South Wales in Australia. It's a government entity called telestroke service. So this was probably a couple of years ago Deep learning A. I. Two. bring to medical imaging. So, what we have done here is we created a virtual platform that uses to identify types of stroke.

So, whether it's is chemical or Hemorrhagic based on brain CT scans so the AI can also annotate the images further by segmentation. And there are two major benefits that this AI assisted telestroke service brings. 1st is this AI. Assisted labeling of medical images can greatly reduce the time required for manual labeling, right?

It provides a condition with useful data at the time of patient management. The second benefit is that because it's a virtual platform, it enables staff in remote areas, access to a network of specialists. So helping to save lives and improve patient outcomes. Terrific. Any other examples of utilizing this for remote access to care?

Yeah, I have an example. As part of AWS is 60 million commitment to improving health equity. We provided resources to a startup called your own AI to help build and scale AI powered applications that enable oncologist to provide remote patient monitoring and tele oncology care. So they offer their solutions to patients in Nigeria, Kenya and Rwanda.

If you take Rwanda as an example, there are fewer than 15 oncologists for a population of 13 and a half million people. And Huron AI created a digital application that allows doctors and patients to communicate frequently and easily through auto generated text message prompts about symptoms and side effects.

And so their work is increasing the ability to provide care to more patients in more places.

So the utilization of this technology is fantastic. It is available to everyone, but not everyone is mastered it. So, how do we close that gap in terms of what opinions are. You can do. And what I can understand is the goal here, but a majority of people are asking for it, but don't know how to get started.

How do you close that gap to really bring them along in terms of technology at the bedside and in their clinical practice? So on the medical school that associated with the Harvard Medical School, the students already increased to use language models to get a feel of it.

Because, we know the patient's already starting to use it, so, you might as well learn how to use it, too. And simultaneously, I think clinicians are still very skeptical about using language models. I think creating more opportunity for clinician industry collaboration or clinicians having a more say in how language models are developed as they're used within healthcare.

Or even like there has, there have also been some concerns flagged around ethical AI, so if the creating data was created in such a way that was free of bias and made it more equitable, I think that would help people get on, get more on board then more testing before we, like, even my project before we roll out to to clinicians to use on the ground.

We have to, of course, test it and vet it. in that process, we need more clarity by the regulators to like when you have a, I being part of the workflow for clinician. Yeah, of course, question about HIPAA and privacy and let's say. Just a question, like if the language model gets updated, which they are every few months, how does that affect the FDA approval?

So I think there are several steps we need to take as all the stakeholders have to take before we get to the state where everybody's using AI. Ching you're the expert on technology here. How do we learn what's inside there and utilize it for our patients? Yeah. So, I like to mention about an Amazon initiative called AI ready.

And above the resources I like to share today is called AWS skill builder. And the website is skillbuilder. aws. So this is an online learning center where you can learn a lot of not only AWS, but AI technologies for free or at a very low cost and it's geared towards both non technical and technical audiences.

So, I would like you to try it out. It's free to register. And again, the website is skillbuilder. aws. That's fantastic. Well, first of all, thank you for having Amazon and be able to help the clinical world understand. Your expertise and all of you has been there, a tremendously valuable platform. to get initiated, but also to grow with in terms of as your advancements grow, as your expertise grows, your experiences grow, so does the advancements of patient care.

Moving ahead, where do we go here? Two years now, five years from now, where will AI be in terms of our clinical practice? And where do you think Are the best areas in healthcare that will be advancements be made in the future of healthcare itself? Well, I could see AI permeate every aspect of healthcare.

It does not mean if the ability for it to permeate would lead to it being like from if I were to see it from a patient's view, like I use devices all the time to manage my health. So I could see a I being a big part of device monitoring even behavior modification like people are trying to lose weight or quit smoking.

I think I could be a big part of that too. Then alerting before it should become critical using on device monitoring. Let's say if you have an a fib, you didn't realize you had you're doing it. What you could tell you, I think it's already in play. I think in the future, more conditions could be monitored.

And also, as she had mentioned about scheduling appointments once you have Agents like a agent that could do it make it very easy to look through and find an appointment that works for you And even preparing for appointments, you know when you're going to see your doctor what questions to ask and follow ups.

Of course keep your schedule follow ups with the doctors. That's the way I see it from a patient's point of view. Like, it could affect every aspect of healthcare from a clinician's perspective. Like, I'm a clinician. I try to think about. And in my daily day, I could see a help in every aspect.

Like, for example, burnout is a big concern by clinicians, anything that can help with burnout documentation or or coding because another thing to do, like, I can help do that for you. Then research and education, it's already playing a big role over there. I think Bush press and Shane talked about how is contributing big time to research and faster drug development.

And also, for clinicians, as part of burnout the tool I'm working on it addresses a cognitive workload. Like, when you're busy at work, you already have the pagers going off, documentation, and then the labs can sometimes have, like, hidden things that, you can overlook. So having a tool, a clinical junk that helps you manage your cognitive workload would actually be pretty helpful too.

So that's the reason I came up with console box. So it's like a decision support like like your Sidekick, who's helping you manage your your daily tasks. And also errors. Errors are a big part in healthcare that adds to a lot of cost. So having tools that can help reduce errors, it could be pretty helpful over there, too.

And finally, there was this thing I was reading about in basket management, because a lot of primary care doctors are getting burned out from in basket management. So having AI help you with that would be very helpful, too. If it's okay, I would like to demo my tool and at this time and show how it would help a clinician interpret a blood test, but that'd be great.

Thanks for me. So I'll put in a couple of blood gases which I, Have a layer of complexity to them, and then we'll see how console bot interprets them. So let's say the pH is 7. 36. PCO2 is 44, bicarb is 24, so they're all within range. The only wrinkle is an ion gap is elevated.

So sometimes this can slip through that when you look at the initial 3, oh, blood gas is normal, but an ion gap isn't. So this is actually a complicated blood gas with a layer of complexity to it. So, what the tool is doing at the back, it goes through a hardcore bunch of rules, takes a chunk of text from a knowledge base, and sends it to the language model.

But this is also this high anion gap metabolic acidosis with mixed metabolic alkalosis. It's actually pretty common if someone has like, in this case, a diabetic ketoacidosis, so they're static, you do anion gap acidosis, but they're also throwing up, vomiting a lot. So, that actually leads to metabolic alkalosis.

It's actually more complicated than what it looks, when you see the blood tests initially. Let me give you another example. Let's say this patient is really sick. PH is 7. 05. So that means patient is pretty static. PCO2 is 50. Bicarb is 15. And, in this case, we have an oxygen PO2. That's 180. So we know the patient oxygen.

And to.

It's 0. 4. So this is a more complicated blood test that we would see during a code like code blue. That's a pH of 7. 05 is 40. we have by current 15. I just want to make sure I got all of them. And I'm gap. Let me take that out. That wasn't the last one. Pio 2 is elevated to that's higher than usual.

Let's go through with this.

Yeah, so, as of now, at the back, in the back, it's going through an algorithm with hard coded rules and sending it to the language model. And this is what we get, uncompensated mixed doses. It tells you like, there's probably a multi organ involvement, metabolic doses. Possibly, we lost the respiratory esters, the hypercapillary respiratory failure.

And on top of it, if you see the PF ratio here, which we have to calculate, usually clinicians can sometimes forget to see if the patient's ARDS is calculated for you. So, it tells you at FIO2 of 0. 4, like 40%, oxygen level of 180 is actually normal. yeah, so that's all I have to share for this. I'll refer to Xing.

From my point of view, I would like to see more and more of such sidekick applications in the near future. AI is not in the future. The AI is here. The AI is Here with us right now. And there are a lot of things we can do with the help of AI, especially there's a term at Amazon called undifferentiated heavy lifting work.

So that's what we mean by repetitive. No value works that our we see our conditions and providers spend a lot of time doing. So if we could use a I to help freeing our resources, spending time on those types of work they could spend more time with our patients and improve patient outcome.

That's what I would like to see more. in the near future, in the very near future. I'm sure you've got it on the road map for this week. Christine, anything last words from yourself? Yeah, I agree with both what Praveen and Qing shared. I think AI and other innovative technologies will further empower both patients and clinicians with the information that we need to make better and safer decisions.

And with these continued advancements, I think we'll make much more progress towards that quadruple aim. really to improve patient outcomes and improve the experience for both patients and providers, while also decreasing healthcare costs and doing all of that in an equitable way. Excellent points.

Thank you very much. Christine, Praveen, Ching, it's an honor to have you participate in today's AI and healthcare discussion and how you're using clinical evidence and clinical experience and technical design around artificial intelligence to really enhance the current state. It's no longer future.

We're here. We're living the innovative current world that healthcare is evolving from. So I want to appreciate all your time here today. And we'll have this video available to everyone shortly on the web and through one, we'll share it with you all as well. Thank you all for joining us today.

We look forward to seeing you. Next month Town Hall by This Week Health and Al's AI Insights.

 Thanks for listening to this week's Town Hall. A big thanks to our hosts and content creators. We really couldn't do it without them. We hope that you're going to share this podcast with a peer or a friend. It's a great chance to discuss and even establish a mentoring relationship along the way.

One way you can support the show is to subscribe and leave us a rating. That would be really appreciated.

Check them out at thisweekhealth. com slash partners. Thanks for listening. That's all for now..