Making Rounds: The hopes and promises of AI [Podcast]

With generative AIs like ChatGPT, Dall-E, and Bard crossing the line into practical utility in recent months, every industry is closely watching to see how what changes AI will bring. Healthcare is no exception. For some, the influx of news promises a revolutionary new way of obtaining and utilizing clinical and operational data. Others are more skeptical. Across the board, though, experts agree AI will change healthcare, and it’s imperative for leading health systems to begin thinking about how to use AI to transform care delivery.

On today’s episode of Making Rounds, Nordic Head of Thought Leadership Dr. Jerome Pagani sits down with Digital Health Practice Leader Kevin Erdal and Chief Medical Officer Dr. Craig Joseph. They discuss the recent strides in generative AI technology, the promises of it, and if the hype around it is warranted. They also discuss what it will mean for the future of healthcare, especially in light of the Big Squeeze, and next steps for health systems looking to leverage AI.

Listen here:



In Network's Making Rounds podcast feature is available on all major podcasting platforms, including Apple PodcastsAmazon MusicGoogleiHeartPandoraSpotify, Stitcher, and more. Search for 'In Network' and subscribe for updates on future episodes. Like what you hear? Make sure to leave a 5-star rating and write a review to help others find the podcast.

Show Notes:

[00:00] Intros

[01:27] Clinical thoughts on AI and current use cases

[07:44] Hopes and skepticisms around AI

[11:09] AI and the Big Squeeze

[17:15] Next steps for health systems leveraging AI

[22:25] The promise and potential of AI



Dr. Jerome Pagani: Craig, Kevin, thanks for being with us this morning.

Kevin Erdal: Happy to be here.

Dr. Craig Joseph: It's my pleasure, as always.

Dr. Jerome Pagani: We've been hearing a lot about AI and how it's being used now and what the potential uses for it are in the future with healthcare. From a clinical perspective, Craig. What do you hear people talking about and what are the use cases you see that folks are exploring right now?

Dr. Craig Joseph: Well, right now it's all generative AI and I think it all comes down to what we can see today with ChatGPT. Someone asked me yesterday, where is GPT on the hype curve? And I said, it's not because it's not hype, because it's real, because we can see it today. So from my perspective, AI has always been that thing in the future where it's going to help us identify trends and find patients who, you know, are about to get sick before they get sick. And that's still, I think is a seems to be a ways off. But today, what we can see with just generating text is are there are AIs that can summarize great amounts of information in a really smart way so that it's, it's pretty ready for prime time. So today we have a big EHR vendor that's already incorporated OpenAI's AI into their, into their EHR, into common workflows, they’re still testing this, so it's only a handful of clients, but what they're using it for now is for messages to patients. So hey, the in basket or inbox is a major cause of physician unhappiness because it's just overwhelming and most physicians don't have protected time to deal with it. And a lot of messages that used to be handled by others in the office are now going right to the physician. So how can we help with that? And, and so, they're using ChatGPT like functionality or actually ChatGPT to, to look through what the patient's asking to make a message in response and if the physician likes it, doesn't even require editing, off it goes. And those kinds of, and often the message that the patient gets back is better than what the physician would have would have written. More complete, more empathetic, and more up to date.

Dr. Jerome Pagani: And is the goal there really efficiency on the provider side, better engagement on the consumer health consumer side? What are the sort of the KPIs that folks are looking at at this point?

Dr. Craig Joseph: Well, I think they're still in the testing phase. I would say that the first goal is to reduce physician burnout. The second goal is to reduce physician burnout. And the third goal is also to reduce physician burnout. Now, will there be a ton of other benefits? I think you just mentioned them. Yes. You know, more, more patient engagement. I don't think most clinicians are excited about having more back and forth between patients and, and themselves outside of a formal office visit or telemedicine encounter. But I think that, hey, if I can answer questions that you haven't asked, because the AI suggested that this is a question that you might have in the future and it just kind of answered it ahead of time, that's a win for everyone.

Kevin Erdal: Yeah, that's a great point. And one of the things that I'm most excited about is, I get asked the question all the time, what are some of the quote-unquote clinical use cases? And we'll talk about some of that, I'm sure, here today. But when we talk about creating time for a clinician to enable them to work top of license, in a roundabout way, we're actually impacting clinical care. So that's very exciting. And hopefully, we'll be able to see some more of that.

Dr. Jerome Pagani: And what's the tech that sits behind this kind of generative AI that's represented by ChatGPT and other types of chatbot-y kind of things?

Kevin Erdal: Yeah, a tremendous amount of data that's trusted, with a lot of velocity, in which we can start to return the answers in a, not just a scripted fashion, but also a way in which somebody can consume, which is something that we haven't seen a lot of. So let's let the Microsofts and some of the big EHR players be great at what they're doing, bringing the technology to the table, but then enabling that workflow within the EHR or the ancillary system to be able to interact with the patient and the clinician. So there's a lot of different components, but I think one of the piece of the puzzle that is most appropriate to talk about in today's world is the trust and the accuracy in which we've been able to start to see in the early stages. Like Craig said, there's a lot of testing going on. There's a lot of things that we're going to learn in the next six-plus months. But the integration is one of the core components that we're starting to see some core success around.

Dr. Jerome Pagani: So Deep Blue is an example of just a huge look ahead predictive model that was really successful at, at, finally, at chess and was able to beat even grandmasters. Is something like ChatGPT just that on steroids?

Kevin Erdal: To a degree, but also with additional context and around, again, I'm going back to this trusted data concept, a means in which we can analyze, right? So if you start to look at all the different patterns for chess, it's going to be a little bit different than all the different patterns in which a patient might start to ask questions of our health systems, not just the physician. So “kind of” is my response there, but to be able to have a robust knowledge base to a degree that somebody can ask a question of via ChatGPT or something similar, and then return a response that is, in fact, accurate and in line with what an expert such as a clinician or even an operational expert is going to not only agree with but understand, that is one of the core components that is, is really making some advancement right now.

Dr. Jerome Pagani: And adding algorithms to do NLP and, is, I'm guessing, sort of the next step where this is going to make it so that interacting with these types of AI are going to be, feel seamless and natural.

Kevin Erdal: Yeah, so, and we've been doing NLP for a long time, so you kind of hit it on the integration of AI, NLP, analytics, data, EHR, the whole gamut, right? So as we start to bring these things closer together, that's fantastic. But we started talking about NLP about a decade ago already to try to figure out how can we get more information from a clinical note, for example, and how can we do that in a speed in which that's going to support the particular use case? In this case, kind of what we're talking about is the interaction between the patient, and I’m going to say the health system, maybe specifically the physician or specifically, you know, somebody at the front desk, but just the speed in which we can start to accommodate that is really quite exciting.

Dr. Jerome Pagani: Craig, how are physicians feeling about the promise of this kind of technology, is there a resistance?

Dr. Craig Joseph: No, this is a hallelujah moment, I think, you know, we've been talking about, one of the one of the major things that have been contributing to burnout is, is documentation, right? In the olden days notes, and by olden days, I mean 20 years ago, notes were small and written on paper and didn't include a lot of things which were taken for granted in and through a number of, of regulations and interpretations here in the US, our notes have become mammoth. So we have note bloat and people expect a lot more in their note. And so it's taken physicians a lot longer to create those, even if they dictated them. It still takes a long time. And I think that we see now that, again, this doesn't seem to be hype. This is what seems to be capable today. If we can simply record the conversation that's happening in the exam room and, and accurately get the words down and know who, who spoke them, that the AI can summarize in such a great way, that the idea that five years from now that a physician will be creating documentation seems foreign to me. I just can't imagine it even happening. I think maybe again in the United States for, for legal reasons, one could imagine a full transcript of the conversation that's easy to store would go somewhere and be difficult to find, but would be there if it were ever needed. But that summary is going to have everything that, that is required for care of the patient, for billing, for other reasons that, you know, quality checking, those kinds of things, that it just, it just exists. And, and again, there have been third party, there are vendors today that have been working for a long time on, one imagines the physician saying, okay, I think we're going to order some augmentin for your child for that recurrent otitis media on the right side and then just saying aloud the dose and then looking at the computer, seeing that it's correct and pressing a button to say send or signed. Right? And we've been talking about that as the future and it's kind of been coming in fits and starts. But it seems like the functionality is really there and I can use it today to do crazy things that don't involve personal health information. And so if I can do that kind of stuff today for free, it doesn't seem like it's going to take very long for me to, to get to a point where, you know, this is really going to free us. That's always been the goal. That's why scribes have been popular. I can just have a conversation with the patient, I can say things, and they just magically get written down.

Kevin Erdal: Well on that concept around transcripts, right, that happens elsewhere. And this is where we like to talk a lot about how are we bringing experience from non-healthcare industries into healthcare? People have been automatically storing, and it is quite easy to store that transcript and then to use that data to inform further research within their business might be research centered around buying patterns and habits and things of that nature. So we're doing some of that, and then people are using that to train, again, outside of healthcare, some of the frontline staff members. So it's really cool to be able to think of how we can take those same exact concepts to your point, Craig, and then start to infuse them into healthcare. And it's something that's already been done before from a technology perspective, not from an operational component necessarily.

Dr. Jerome Pagani: I think those are some good examples of how these kinds of technologies are going to affect clinical operations and workflows. We've talked a bit about the Big Squeeze and the factors that are contributing to them, particularly sort of the financial freefall, the labor pains, and that risk of disintermediation. How can technologies like AI come in and begin to help with alleviate some of the stresses that those factors are creating sort of outside of the clinical workflow space?

Dr. Craig Joseph: Well, it's a good question. You know, specifically looking at disintermediation, having smaller groups who are maybe more technically advanced than some of our incumbent health- large and in health systems, you know, they've been at a disadvantage. The, the, the big systems. And I think that AI is really going to be able to kind of help bring everyone up to potentially very similar levels. You know, the idea of omnichannel communication, that I can have a relationship with the patient the way that the patient wants to have that relationship, and maybe that someone who's got a phone, maybe that someone who's got a laptop, maybe that someone who actually wants to talk versus just communicate via chat and to be able to have the technology that's now at the level where it's just a utility, it's just there, it's not cutting edge anymore, really does kind of add a level of equilibrium there that I think that we're getting to very quickly.

Dr. Jerome Pagani: So one day, will hospitals have algorithms or recommend surgeons for me based on what music they play on the OR and what music I prefer?

Dr. Craig Joseph: That would be awesome, generally you’re sleeping. But ...

Dr. Jerome Pagani: Yeah, but you could still be. I mean, it's, this is important. I don't want to be listening to the Bee Gees while somebody’s cutting it, that makes me doubly unhappy.

Dr. Craig Joseph: Wow. And see, that's funny because I would want to be listening to the Bee Gees. Your point, though, is well taken, and there are times where there are data points that are very important. For instance, what language does my physician speak or, you know, are they very comfortable in taking care of the elderly or not? And those kinds of things are kind of difficult to put your fingers on. Certainly some of those you'll find in the physician's biography or something that you can search on when you're looking through a list of physicians, but often not the case. And so if it can be suggested because the AI can look over your social media that's public or other things that you've said that are important to you and, and connect you with the right folks or even in the right, in the right location, some people are going to be very comfortable having an outpatient surgery at an ambulatory surgery center out in the community. And there are going to be others who are like, boy, you know, I know someone who had a bad experience like that and I really want to be at the mothership with thousands of physicians around me and difficulty in parking. And that's the experience I want because I feel safer. Wow. Well, that's, that's helpful to know.

Kevin Erdal: And it's a really interesting perspective. I haven't thought about a whole lot, but we by capturing more of this information, we can not only learn a ton about the patient and what they're going to need to receive the best care possible. But we can also start to understand where physicians are most successful and start to align that with the patients coming in. That's a phenomenal use case that hopefully we can all work towards advancing.

Dr. Jerome Pagani: Kevin, what about on the sort of operational end of things?

Kevin Erdal: Yeah, so this is, the operational side of the house right now is where we actually see a lot of people starting with their AI journey, which is really exciting. So we've seen a lot of nice use cases around things like claims denial. We're starting to see some very successful use cases specifically within supply chain. And really, you know, for a supply chain example, just to give a quick kind of use case, we're starting to see how people can manage the POs coming into a health system and use less resources, right, less humans because right now it's very hard to find a number of people to support any amount of activities going on within the health system. That goes back to the big squeeze and all the things that we've talked about related to that in the past. So if we can automate and we can, with confidence, start to understand how we're purchasing supplies, when we need to purchase some of those supplies, create a frictionless environment in which we can all operate, that becomes a very, very compelling area to start in your AI journey, but also have tremendous impact to the organization. And we can also start to look at some of the buying power activity that's coming along with that within supply chain. And we've started to see people save up to ten plus million of dollars just by being more efficient and more effective with the existing processes that they have, but automating some of that activity.

Dr. Jerome Pagani: So I think those are some great operational examples. What about on the people front, the sort of talent perspective?

Kevin Erdal: Yeah, we're starting to see some of our HRIS colleagues start to leverage AI in a much more innovative fashion as well. How can we start to align expectations in terms not just a job description, but what kind of skills do we need, what kind of people are going to be successful, and then starting to apply that to the pool of candidates that some of the organizations are already talking with in some cases, and/or pools of candidates they are not talking to? If you start to look at where technology is going, we already started talking about other industries and how we're starting to leverage technical skill sets in healthcare that are coming from outside of healthcare. We are now starting to see some very robust algorithms inform the HRIS type department. And so some of those recruiters and some of the talent acquisition individuals supporting not only healthcare systems but people around the healthcare community is surface opportunities for folks right to say, hey, this job description might align really well with what Craig wants to do based off of the conversations we've had going all the way back to transcripts, we're documenting a little bit differently in terms of when we talked to Craig from an initial conversation, what aligns with his interests, what aligns with his skill sets, and then maybe three or four months later, we actually have a net new job description that would then match those two very well. So we're starting to see some creativity around that. Again, something that's been done outside of healthcare in terms of the recruiting space, when a recruiter is listening in, they're going to say, Hey, we've been doing this for a long time in Sector A, B or C, or we're starting to see that applied to the healthcare ranks, not just for the clinical roles or things of that nature, but really broadly across all the different kinds of skill sets that are needed within a health system.

Dr. Jerome Pagani: Kevin, we've talked to a number of clients who may be just starting or a little ways into their journey with AI, and most of them, as we discussed, have started with those sort of back-office use cases because they're low risk and it's easier to measure the ROI. And so now we're having conversations about how to transition some of those efforts to the clinical front. And so, what should folks be thinking about as they're sort of either just starting out, making that transition to the clinic side, and then do you treat AI the same way you would treat other innovations that you're beginning to roll out within the enterprise or are there a unique set of characteristics that you need to be thinking about to really support the rollout and use of, of those AI products?

Kevin Erdal: Yeah, there is some uniqueness in the place to start. I always go back to talk to the owner of the outcome you're aiming to support. So in this particular, you know, just in a specific use case, I would go to Dr. Joseph and say, "Hey, you're trying to improve patient outcomes for specific populations of patients. What does that exactly look like? What kind of KPI are we starting to impact?" And then work with the people that are intimately involved in the specific workflow. So as we start talking about a specific KPI, start working backwards, don't try to go right at the data or right at some very exciting Python code to try to solve world hunger, for example. Let's truly understand in detail what the outcome we're trying to support, then talk to the experts and then we start looking at, do we have the data to support that particular use case? If not, we got to go back and do the heavy lifting. We have to do some of that lovely data wrangling and we have to make sure that we have a scalable environment in which that data is residing so that it's going to be supported long-term because anything we build here and now say within 2023 we expect or we at least hope in today's world that it's going to be applicable in 24 or 25, 26 and beyond. And it has to be able to learn, the models have to learn over time and adapt over time, but it still has to be in line with that particular workflow that's going to support the outcome that we're after. So it's not just about tech. I say that all the time. I've probably said it on other podcasts as well. Truly understand the outcome you're shooting for. Start talking with the experts early and what impact it's going to have on their day-to-day operations, or the way they care for patients, and then get into technology and start developing.

Dr. Craig Joseph: Let me emphasize what Kevin just said. I think we all agree I'm the technical expert on this episode today.

Kevin Erdal: Correct.

Dr. Craig Joseph: Thank you, Kevin. And the fact that the data might exist doesn't mean that they're correct or accurate. Right? And so really emphasizing talking to the end users, the folks that are going to use it, what I have seen before is people who don't do those steps say Ah, well, we have the fields right here. The information is right there. Let's, let's do a ton of work, and then you find could take the outcome of that work, go to the end users and they say, no, this is the wrong list. These are not the patients that we really want to prioritize. Well, of course, they are. We? Oh, no, you know what? Because of the two big insurance companies, we don't have the information for them. So those fields are blank. So I'm not sure what you just did, but you've only found the bottom 30%. And then we still have the top 70% because the fields are empty, because we never got the data, because for whatever reason. And so knowing that ahead of time will probably be helpful.

Kevin Erdal: Yeah. And when you can articulate some of those gaps and then we turn around and you know, as a clinician in this use case, Craig, you might say, well, I don't document that information in the EHR. I use this other ancillary system. And then we realize, oh, we don't have interoperability between that other system, therefore we don't have access to the data yet. We just have to integrate that on the back end a little bit differently to support the use case. I couldn't agree more, but you don't know that until you talk to the expert. So that's why I always go back to starting there first.

Dr. Jerome Pagani: And on the governance side, I mean, there are some unique aspects to AI, things like drift and, and sort of off targeting and other things. How does the governance need to change for products like AI?

Kevin Erdal: Yeah, the most successful organizations we've talked to date, Jerome, have governance first and foremost specific to AI. This is a different kind of governance to exactly what you hit on. And what we need to look at is how are we not only managing things like drift, but how are we also ensuring that we're connected with the clinicians that we've worked with in stage one? And it might be two years later, we have to stay engaged. You can't just say, hey, this is a one-and-done kind of activity. Like Craig mentioned, things change, all of a sudden, documentation pathways might change and therefore we have gaps in data two years from now that we wouldn't have had when we started the project. So step one, start governance specific to AI, you got to bring in the clinical practice. You have to bring in some of your operational leaders, you have to have some of the technical leadership all in one and be open about it, talk about use cases, and please be willing to say no. That's what I don't hear a lot of.

One person might get very excited about a particular model and what it could do in the ambulatory setting or in a specific specialty. But if the data is not there to support it or the technology isn't implemented in today's world to, to support the use case, then maybe that's not where you need to start and that has to go at that governance level.

Dr. Jerome Pagani: And as folks are getting started with their AI use cases, what does that look like sort of end to end, and timeline wise?

Kevin Erdal: Yeah, it varies drastically. And I hate having to say that, but when we get into some of the clinical use cases, there's going to be a lot more analysis that has to go into where is that data that Craig and I were just talking about? Do we have the technology to support it? Do we have the skill sets internally to support it? That's also why we start to see some folks starting in the operational areas, because a lot of that data is a little bit more transactional and tends to be a little bit more available. So as we start there, as we have success, we can not only build up our governance structures, but we can also represent to the organization where AI has been successful, creating a comfortable environment for us to then go into the clinical settings, or even starting with some of the front desk activities, starting to help with patient engagement, things of that nature.

Dr. Craig Joseph: So let me summarize what I'm hearing from you, Kevin. A couple of days.

Kevin Erdal: Sorry, I intentionally did not put a day or date to that.

Dr. Craig Joseph: Three days.

Kevin Erdal: Four days.

Dr. Craig Joseph: Four days, and everything will be good. All right, great.

Kevin Erdal: Now, some of the variation, though, in all seriousness, is, it goes back to is the data position to support the use case. If that answer is no, we're going to spend several weeks positioning the data, and then we're going to start building the AI to support that use case. So then we can start talking three plus four, five, six months, just depending on how much heavy lifting has to be done. For the organization that already has a clinical data warehouse, for example, that they know and they trust and the data is well positioned, now we can start talking about a three-month timeline that is realistic, not necessarily at scale, but at least we can start to bring things forward much, much quicker.

Dr. Jerome Pagani: Well, I for one, welcome our robot overlords, but I'm wondering if you guys have any last thoughts about this topic, which I know will probably not the last time we'll be discussing it, but anything you want to leave us with?

Kevin Erdal: Well, I think from my perspective, AI is here. One of the things Craig teed off with right? We don't need to wonder if we're capable within healthcare anymore. We're already seeing positive use cases. We're seeing positive impact across the spectrum. Right? We've talked about the operational side. We've talked about the clinical side. We've talked about how we're reducing some friction in the day-to-day operations throughout a health system. The secondary component I would say is go start talking about those use cases proactively, whether you have a robust team or not. What is your organization talking about? What are your clinicians excited about inside the institution that you might want to start to double down on and just start having that dialog and conversation around? What would it take for us to do this right now?

Dr. Craig Joseph: Yeah, I think the excitement is that there are things that don't, that most physicians don't think actually contribute to their success or to, you know, patient health. And if, if we can get something like an AI to take on the role of some of these important yet not really requiring a human with a lot of experience to do, then that's a win for everyone. And we're there today. The future is now. And so I think the excitement is warranted and real and I can't wait to see what happens.

Dr. Jerome Pagani: Agree, and I'm really excited that JeromeAI is taking off and beating out CraigGPT’s market share already. So this is really exciting.

Kevin Erdal: You both are much more sophisticated than I, I have to spin something up here.

Dr. Craig Joseph: I'm actually sleeping at home right now. And this is a generated voice that you're hearing.

Kevin Erdal: Sounds great.

Dr. Jerome Pagani: Kevin, Craig, thanks so much for joining us and we'll look forward to seeing you again soon.

Kevin Erdal: Appreciate it.

Dr. Craig Joseph: Thanks.

Topics: featured, digital health, podcast

Module heading text

Get the highest quality chemistry and microbiology testing services aligned closely with current good manufacturing practices (CGMP) for all types of products across all phases of development.

Subscribe to receive blog updates