Chapters Transcript Video AI Governance and the Transformation of Clinical Care and Healthcare Informatics An average mammogram reader, radiologist has been trained on, you know, maybe a few 1000 or even several 1000 images. The AI tool has been trained on several million images. So of course, it knows how to pick up those very, very Small lesions same thing for lung nodules uh we actually have uh uh uh internally developed uh research uh product that has picked up more lung nodules than the radiologist picked up and we are now looking to implement it internally here at UCI. Hi, I'm Lindsay Carrillo, director of business development at UCI Health. Hi, I'm Doctor Sunil Verma, Associate Chief Medical Officer for ambulatory care. I'm also a laryngologist here at UCI Health. Welcome to Physician Huddle by UCI Health. Today we're joined by Doctor Deepti Pandita, Chief Medical informatics Officer and vice president of Informatics here at UCI Health. Welcome, Doctor Pandita. Thank you. Tell us a little bit about your role here at UCI Health. Yeah, firstly, thank you for having me on this podcast. Uh, I'm an internal medicine trained physician by training. I'm also board certified in clinical informatics as well as internal medicine. I still practice here at UCI, but my bigger role is the VP of clinical informatics and the chief medical information officer, and what that means is I'm sort of the translator of tech speak to clinical speak and for um anything that goes on in the organization that has a technology component but a clinical implication around work flows or process or impact to people informatics steps in and sort of draws that. Um, line diagram between how the tech translates into clinical care. So tell us a little bit about the field. How old of a field is informatics, um, and, and how common is it? Is this an academic role or is this are we seeing this throughout health systems? So most health systems do have now an informatics role. It's, it's fairly new as a discipline. Um, for physicians, uh, informatics actually came on late to the landscape. Nursing informatics had been in existence since the 70s. However, for physicians, it came into play when EMRs were developed in, you know, the early 2000s. Uh, the discipline as an ABI ABMS specialty came. To play in 2014, so, um, as a curriculum as an actual subspecialty of medical medical um subspecialties consortium, uh, 2014 was the first administration of the board certification. I was fortunate enough to be in the 1st 140 in the country to take the board certification. And there was a grandfathering in pathway at that time if you had been in the field for some time. Um, now, uh, as of this year, you can only, uh, sit for the boards if you're done a clinical informatics fellowship. We are very fortunate here at UCI to have a clinical informatics fellowship. We take 1 to 2 fellows every year, and some of our fellows have, uh, graduated and gone on to do great work here at UCI. So, um, we are indeed lucky to have that, uh, support for having that academic clinical informatics structure here. So we've talked a lot about artificial intelligence in your field. Was this a recent, you know, huge explosion onto the scene? Is this something that you've actually been working on for quite some time and it's just new to the rest of us? So artificial intelligence as a. Field has been around since the 80s. I want to say it was mainly around machine learning and mainly around, you know, how can we make machines smarter and feed the data that we are generating from the machines to the machines so that the output is better. However, what happened around 2019, 2020 was this whole concept. Of what we call large language models um it actually was a very accidental sort of discovery of a concept called transformers where code could be generated using the spoken language, natural language, and that revolutionized the whole field and now we have large language models which you know make it so easy for even a lay person. To generate data and then that data trains itself into, you know, more machine learning and algorithms and all that and that is why artificial intelligence has become such a buzzword these days. Is, is how much of your role day to day is involving applications and use of AI? Help us understand where we are and where we're going. Yeah, so, um, we have had some kind of algorithmic AI within the tools that we use in our system for quite some time. However, now almost every vendor we work with every, you know, third party tool, and people don't understand. To deliver care you need a lot of different tools in the tool kit. People think of EMR as the major tool, but you'll be surprised we have about 300 or so different, um, third party systems within or outside the EMR that we need to deliver clinical care and in this day and age 50% of them have some AI component and if they don't, those vendors are actively working to create. Those AI components. So the simple example would be, say, uh, a hospital bed, you know, a hospital bed was smart enough to do weights and, you know, falls and all that. Now with an AI component into the artificial bed, it can predict falls. It can predict weight loss. It can show you trends. So now AI is becoming a component of something like a hospital bed. The same thing is coming into play into every system we use for clinical care. So I think to your point, because it's a little more consumer friendly these days and physicians are not immune to these trends and the things in the news, there's a little bit of panic, and you have even said if physicians are afraid, AI will replace them, then maybe it should. So talk to us about that. What do you mean and how could we debate this in a maybe more productive way? Yeah, absolutely. I mean, that's a little tongue in cheek statement, but it is a true statement. Um, physicians are very binary. They think of AI as, uh, human versus machine debate, and actually not that. I, I don't think AI is going to replace physicians anytime soon. However, if physicians are not savvy enough to embrace AI, they will lose out on all the cool stuff that AI will bring, uh, into their armor, which is basically using AI as a collaborator. As a helper, as a sidekick to sort of drive more efficiency. So I think clinicians need to get over their sort of initial reaction, which is, oh, I don't want AI because it's going to replace me and actually look at it as where it can augment their intelligence instead of just think of it as artificial intelligence. So. To that end, how can we, you know, what are ways currently that it helps with decision making, you know, the clinical decision making that we do every day when we see a patient. Are there, do we have applications of that right now? So this is the strange part because we have a lot of decision support tools that in the background are using AI you know, we have had cognitive computing tools for like sepsis decision support or um you know false risk predictions and all of that in our system for some time what AI will do the current state of AI what it. We will do is put it more into a predictive manner. So we have had algorithmic AI telling you if this then that sort of decision making, but now it'll also say if you don't do this today, day after tomorrow, the patient can have this, and that is where that predictive modeling is going to play a bigger and bigger role. So you mentioned sepsis. Can you give an example of that? Like, what, how, how is it? Where do you see this going? So a patient's in the hospital, you're managing them, like, take me, take me from there, how AI is gonna help us with that patient. So again, if you take an out of the box sepsis predictive algorithm, so say it was developed at Epic Systems in Verona or something. That itself we have used and we found that it was not very useful. So it's supposed to tell you from the data coming in from the vitals, the labs, the notes of the provider, the, the prior history of the patient, everything is in the EMR, so it's gathering all that and synthesizing it and giving you a synopsis. And based on those variable data points, it's going to tell you that this patient does have sepsis, you know, so it's taking your sort of thought process out of it when when you're in a busy situation and telling you the likelihood of sepsis. Now once it tells you likelihood of sepsis, that's not enough, you know, what are you supposed to do then? So now it'll tell you. There's not only a likelihood of sepsis, you have not prescribed this antibiotic which should have been given in the first, you know, you know, 12 hours or 8 hours. And by the way, now if you have ambient. It'll also tell you, do you want me to order that antibiotic and uh you know, you can the Ambien can place the order. Once it places the order, pharmacy can see the order and then they can have AI systems on their end to say, is this patient's insurance covering this order or is there something that we need to do differently in the very. Your future we'll have genomics embedded into the EMR and then the pharmacist can see that oh this genome sequence for this patient does not merit this antibiotic there's a better one or there could be an adverse reaction with this antibiotic and the AI is telling them all this by synthesizing the chart because remember we are in an information overload state. We have so much information that a human mind cannot synthesize or get a synopsis of that information. This is where AI is so smart that it can synthesize all that information, do a sort of chart biopsy, and present it to you in a manner that you just get those bullet points that are relevant to that patient's care at that point of time. But how do we, how do you deliver that care and, how do you teach care to the doctors, right, that are gonna be taking care of us one day. The how do we teach that use of that, the responsible use of that, and not to just trust the answer that the computer spits out of you, and this is where education is key and I think it's the responsibility of academic institutions like ours to start them young. So whether they are medical students, residents, fellows, first being aware, you know, I, I have asked people how much AI are you using in your system and they tell me, oh, we don't use AI. And there are numerous tools today that we use that are AI and people are not aware of them, so awareness through education is #1. 2 contributing to that and picking up on things that may not be sounding right. So I was giving you the example of the sepsis algorithm if that algorithm is not trained on our patient population. Then the algorithm can go very wrong and being aware of those kind of things that oh this was an out of the box some company told us this is the best sepsis algorithm we bought it we implemented it and it's not quite showing us the results being aware of those kind of things that this could be. A possibility that we didn't feed the right data, or maybe there's a bias in the algorithm. So having that mindset of coming from a place of knowledge, you can actually train the AI instead of the AI telling you what to do. So how do we account for that? You talked about having a very diverse population. Orange County is, you know, filled with diverse populations, but we don't have everybody and not everything has ever been seen here. So how do you make up for that and make sure that the systems you're using isn't biased in some way or, you know, missing some majors where AI governance is key and we bring various states. holder so any AI tool that's coming into our system, we have a checklist and we ask the vendor certain set of questions. Is it FDA approved? Is it um what kind of algorithms were used? Are the algorithms tested in populations like ours have other academic institutions published on it or, you know, do we need to do this on our own and then the governance body has. Ethicists, you know, who weigh in into the value of the tool and then we have a life cycle management process so just implementing the tool is not enough we have to sort of validate we have data scientists who are looking into the tool to validate it and then we are also looking at the algorithms that you know were brought in to develop the tool are they representative of population. In our area and not just orange or I mean now we have the community network hospitals in our midst and we need to include that population so we need to be very cognizant that an out of the box off the shelf AI tool will not work in every situation mainly because it has to be fed the right data to give the right output. So two things that I found, one thing in the past that was difficult was just reading an EKG. And the computer would spit out the thing, you know, at the top and every cardiologist said that the computer was wrong. Has AI been able to figure out how to read an EKG yet? Actually, AI is surprisingly smart when it comes to those kind of um machine output um sort of tools uh one is radiology uh images and the other one is EKG type images so. So tell us what that means like. What what is the radiologist gonna do or the cardiologist gonna do with AI because that's really, you know, a minute to minute thing that's in front of you and that's something that we're doing every day. So yeah, tell us more. So AI um has been shown to be of tremendous value in very, very small lesions. So think of mammograms. So 95% of lesions. In mammograms can be picked up by artificial intelligence tools whereas a lot of those very small lesions were missed by the human eye, you know, because there's a finite limit to the human eye and the reason the AI can do that is say an average mammogram reader radiologist has been trained on, you know, maybe a few 1000 or even several 1000 images. The AI tool has been trained on several million images, so of course. It knows how to pick up those very, very small lesions. Same thing for lung nodules, uh, we actually have a uh uh internally developed uh research uh product that has picked up more lung nodules than the radiologist picked up, and we are now looking to implement it internally here at UCI. So there are these, um, very subtle things where AI excels the humans, um, retinopathy is another one, you know, um. This is where you know large scale retinal scans cannot be done because you know only so many patients can be seen by an ophthalmologist. There are machines that have been developed using AI where you can walk up to a kiosk, you know, like look into that kiosk and it'll tell you if you have retinopathy with diabetes or not and that's. A win win. I mean, can you imagine in a third world country, you have, you know, thousands of patients who are being left out. You just have a retinal, you know, scan kiosk in some public area and people can find out their results. Going back to the lung nodule then, when is this going to go from developmental to implementation. Both here at the academic center and then more importantly throughout our county with our physicians in our county, what sort of time frame is being predicted for that? So typically any cognitive computing model that you develop, you have to study it for at least 1 year to 18 months just to make sure that. It is functioning the way you designed it. It's giving you the data you wanted it to give you that the inputs are actually what you want it to do. So for example, if it was studied on a very small sample at Orange, now you wanna input, you know, the Irvine data and the community network data to make sure that there it's free of bias. And then, um, algorithms have this tendency to what we call deprecation, you know, they lose their sort of smartness over time so. Models need to be managed for deprecation and that is a very, very sort of uh intense process where you need to look at it every few months to make sure there is no risk of deprecation and once it's stable, only then do you implement it and even after it's implemented you have to constantly manage these models. Can you give us an example of deprecation? Yeah so deprecation means that remember just feeding the initial data is not enough. There is machine learning that needs to happen. On the fly, so for example, a patient may have had a stroke and the MRI is using AI to detect that, but in the meantime, the patient had another stroke and now, you know, the, the machine gets a little confused and uh that's where you need to sort of um offset the, the algorithm to say not just one lesion, two lesions, you know, so those are the kind of things that if you're not aware of. If it's a single lesion algorithm, it will deprecate, you know, because it's not um designed to have a multi lesion sort of approach. How is it gonna make um the life of our physicians that are listening easier? I mean, two of the probably more annoying things we have to do is chart, and then also just look through the chart to find the relevant information. Now that everything's linked, it's almost too much for me. To identify what's important and what's not and uh you know, it's hard to search. So how does what where's technology gonna go to make these things easier for us? So that's actually the more exciting part of AI. So you know, leave aside the cognitive computing decision support which is a more labor intensive. The administrative um branch of AI is actually very robust and it's very ready for prime time so as you know we are. Doing pilots with ambient scribing and uh and so for the listener what does ambient scribing means so ambience scribe means that basically there's a large language model behind the scenes. You use a vendor we have head to head pilots going between DAX, which is a Microsoft Nuance product and a bridge, which is, uh, its own company, and both use large language models. What you do is you're recording. Conversation of the provider and the patient and it transcribes that conversation into a note in the very near future, it will also transcribe anything you say in the conversation. Hey, Mr. Jones, I'm gonna order a CBC and uh uh comprehensive blood panel today and maybe we should also do a chest X-ray for your calf. It can transcribe that into an order as well. And in the very near future it'll also transcribe so you have community acquired pneumonia and I think you also have uh vocal nodule which I'm going to refer you to ENT and it's gonna put that code in as well um so those are all future developments right now we are just. Using it for the documentation alone, but the coding component, the order component, all that is coming very soon. So this is the exciting part where AI will take off a lot of burden from the clinicians and our own ambience scribe studies internally have shown us that across the board physicians are being able to save at least 20 minutes a day. Now it may seem very trivial 20 minutes a day, but you add it up over weeks, months, years, it's a lot of time saved, um, and that is time you're giving them back for their wellness, uh, you know, or whatever else they want to do. And by the way, that's 20 minutes in the current form, right? You more time will be saved when it has additional capabilities in the future. So, so what is the appetite of physicians for all of this? I mean, I mean, this is crazy exciting stuff. Uh, I think Lindsay, you sort of started at the beginning saying there's a lot of doubt about it. So where do you see it? Is this the newer docs are more excited, the more senior docs like us, like what what what's the. You know it's, it's very interesting to me that it's all across the board and embracing technology is not an age defined uh sort of uh demographic it is more around curiosity and excitement um so people who are naturally curious will embrace it, um. We have also seen a lot of peer pressure tactics working so they see a colleague using it and they're like what is that? And uh the colleague is saying oh this is the coolest thing and and then they ask us like can I use it so there's a lot of that um. You also have to understand that the tools don't work across specialties very similarly. I mean, think of uh ICU doc the patient is not participating in any conversation there of course this tool is not going to be useful for them. So you know there are some nuances, but for majority of the patient provider interaction, if it's conversational, the tool works really well. What are some of the myths that you've heard or questions you've, you know, been asked from physicians that you're just constantly having to dispel? Yes, um, a lot of concern about security like, you know, my voice is being recorded, where is it going, you know, from patients too. I mean, you know, we are in a. Charge state, you know, sort of, um, federally and, you know, state wise where patients are concerned like, oh you're recording my voice? are you recording me? uh, what is, what are you gonna do with this information? So there is a little bit of apprehension there and then we have to reassure them that we are obtaining. Consent. This is a HIPAA compliant tool. It, it's not going to leave our environment, you know, once you reassure people, um, people are more likely to adopt. And then, you know, you have to explain them, explain it to them in a manner that you're using these tools anyway. You know, if you use an Alexa or a Siri or anything, it's the same sort of technology. Your voice is out there on the internet anyway, you know, uh, like my podcast I'm doing here, my voice is now going to be on the internet. Yeah, I can pick it up and you know, GPT tomorrow can tell me you did this podcast in this manner and my voice files are there. What's funny when you go to the airport now they take a photo of you and I say decline the photo as if I'm really preventing my information from being out there my pictures out there anyways. And by the way, I use the same technology when I come in. Internationally and Global entry recognizes me. So it's funny how we make decisions and we rationalize it in different situations. 100% the DMV already had your photos out yeah, of course the DMV technology systems are very archaic. They're probably not able to upload that to the internet. Well, you have been so generous with your time here. Any parting thoughts? What are you excited about? What keeps you motivated and curious? Well, I, I think the whole spectrum of AI is very exciting. I am also very excited about um the future of health care in terms of genomics and. And uh you know how precision health will evolve with us having genomic data in the EMR and that with machine learning and AI algorithms actually driving patient care to a more sort of molecular level instead of the, you know, population level we are doing today. Yeah, it's just wild to think how we learn medicine, how it's being practiced, and how it will be practiced. Uh, we're a really crazy time where all three of those are gonna are very, very different. Exactly. And then my, my second optimism and excitement is reducing the administrative burden, not just of physicians but also of our nursing colleagues. I mean, if we think of physicians having a lot of cognitive burden of documentation, nurses have it even more and the reason the AI tools have not. Been very good yet is because they don't write in prose they write in flow sheets and you know sort of line items and things and AI has not been smart enough yet to sort of populate it in that manner, but there are now pilots going on across the country where they have, you know, succeeded in doing that and I want to bring that to our nursing colleagues here at UCI very soon. Amazing. Thank you so much for being with us today. This has been Physician Huddle by UCI Health. Thank you for joining us. Thank you. This was an episode of the Physician Huddle podcast by UCI Health, produced by Brett Shaheen, Angelica Yugubi, and Victor Ting. For more episodes, information on clinical trials at UCI Health, or to refer a patient, review the show notes or visit clinicalconnection. UIhealth.org. Created by
