Navigating AI in healthcare

0:04
Hello and welcome to today's program titled Empowering providers, harnessing generatives.

0:10
A generative A is potential in healthcare.

0:13
My name is Eric Whitman.

0:14
I'm the Senior editor for Innovation Technology here at Health Leaders and I'm going to be the moderator for today's webinar.

0:21
Today's program is sponsored by Optum.

0:24
Thank you to our sponsor and to you and our audience for giving us your time and attention.

0:29
This program is going to be about 60 minutes in length.

0:32
Note that an on demand version of the program will be available approximately 1 day after the completion of this event, and it can be accessed using the same login link that you used for this live program.

0:45
Before we get started, I've got a few housekeeping details here to go over.

0:50
First, to ensure that you see all of the content for this event, please maximize your event window and be sure to adjust your computer volume settings and or PC speakers for optimal sound quality.

1:02
Second, you'll find a resources list for today's webinar in the upper right side of your screen.

1:08
Here we've listed the webinar slide deck and additional materials for for you to interact with.

1:14
And 3rd, at the bottom of your console are multiple widgets that you can use to submit a question.

1:20
And we will have AQ and A part of this presentation towards the end towards the last 45 after the last 45 or 15 minutes to submit a question for that Q&A session, click on that Q&A widget.

1:34
It may be open already and appear on the left side of your screen.

1:38
You can submit questions at any time during this presentation.

1:40
However, be noted that we'll probably get to them at during that Q&A session.

1:47
Also.

1:48
Let me just shift up here.

1:50
There we go.

1:50
Should you experience any technical difficulties during today's program and need assistance, please click on the help widget, which is a question mark icon and it covers common technical issues.

2:02
Finally, it's my pleasure to introduce our three panelists for today's program.

2:07
We have Carol Schwinn, our VP Provider Technology Practice Lead for Optimum Advisory and our two provider panelist, Alvin Liu, MD, Director of the Artificial Intelligence Innovation Center at Johns Hopkins Medicine in Kristin Myers, Chief Digital Officer at Northwell Health.

2:24
Thank you all for joining us today.

2:27
And with that, I'm going to get right into the discussion with our first question.

2:31
It's an opportunity for you to tell us a little bit about yourself and your organization's.

2:37
Let's start with Alvin, your.

2:40
Hello everyone.

2:41
Thanks for having me.

2:42
My name is Alvin Liu.

2:47
I'm a practicing retinal surgeon at Johns Hopkins Medicine.

2:51
Outside of my clinical work, I'm entirely focused in artificial intelligence at Johns Hopkins in three different capacities.

2:59
First, as a clinician scientist, I lead teams of researchers in developing clinical AI tools also and I also lead the Gills AI Center at Johns Hopkins Medicine.

3:13
Second, I've been very involved in the implementation of AI tools across the health system, both in in a clinical domain such as autonomous AI for diabetic retinopathy screening and operational domain including Gen.

3:30
AI for revenue cycle management.

3:32
And 3rd, a few months ago, Johns Hopkins Medicine established an AI leadership team that has purview over all things AI related, Johns Hopkins Medicine, especially when it comes to AI governance.

3:47
And I serve on this committee as well.

3:50
OK, Kristin.

3:53
Hi, Kristin Myers, Chief Digital Officer and Head of Technology at Northwell Northwell's a Regional Health system around 28 hospitals in at New York and Connecticut.

4:06
And my role is to guide the organization through digital transformation and I'm doing that by integrating and harnessing the power of digital and data and advanced technologies to enable Northwell's objectives.

4:20
Thank you.

4:20
Great.

4:22
OK then Carol, good morning or good afternoon everyone.

4:26
Carol Chouinard I lead or Optum Advisory Technology practice, one of the largest advisory healthcare technology practice in in the US.

4:38
I've been doing this for doing.

4:42
I've been in helping healthcare organizations with technology related challenges for for quite some time here in the US and and abroad.

4:52
I've had the opportunity to to be part of some of the the largest and most complex programs again here in the US and abroad as well.

5:01
Looking forward to the discussion today.

5:03
Great.

5:04
Yes, this is going to be a good discussion.

5:06
It's certainly a very hot topic in healthcare right now.

5:10
AI in general is just taking healthcare by storm.

5:14
A lot of the early winds have been in administrative tasks, revenue cycle management, some of the background stuff.

5:21
AI in clinical care is is taking a lot slower to develop.

5:26
There've been some, some early winds, some questions about value about, you know, ROI in, in AI for in particular for clinical issues.

5:37
OK, I was going to get to the first question and here we have it.

5:41
And this is what I want to do is start to get the get the groundwork laid out to discuss how AI, generative AI in particular, is, is being explored in, in clinical care.

5:53
My first question, I'm going to start with Kristin on this one is, is to what extent do you believe generative, excuse me, do you believe generative AI has the potential to transform healthcare?

6:06
Yeah, I look, I think that it's definitely poised to significantly transform healthcare.

6:12
And I think it can do this by, you know, enhancing operational efficiency and, you know, personalizing patient experiences and accelerating research as examples.

6:24
And, you know, it's been driven by breakthroughs in large language models.

6:29
And, you know, there's such a growth of healthcare data and, you know, there's an urgent demand for scalable and cost effective solutions.

6:39
And, you know, we're always being told that we have to do more with less as providers.

6:44
And I think that Gen AI really can help enable that and can help us, you know, reshape how we work and engage and connect and make decisions and, you know, really assist with, you know, efficient care and enabling better access and experiences for patients.

7:04
And ultimately, I think it's going to empower healthcare systems and organizations to, you know, drive better improvements in care outcomes and value.

7:16
OK, Alvin, same question.

7:17
You know where where will this transform healthcare?

7:23
So when it comes to Gen.

7:24
AI for healthcare, I like to think about it in three different buckets or three different major use case The first one first one I would say is using Gen.

7:35
AI specifically large language models for revenue cycle management.

7:39
It's already getting pretty popular and in fact a lot of the movement in this space is driven by industry, specifically startup of different sizes.

7:50
So my my guess is that out of all the application of Gen.

7:55
AI in healthcare, using Gen.

7:58
AI for revenue cycle management along the different stages of RCM will be the first application to really scale.

8:05
And honestly it is the easiest one to demonstrate ROI.

8:09
There's a hard dollar sign attached to each stage of the RCM.

8:14
The second application that I think it's also getting traction, it's in the ambient scribe space.

8:20
As you probably know, it's a very crowded space with a lot of companies trying to get into the space and really deployed in the real world, real world.

8:29
What I'm seeing is that there are several players in this space that are becoming more dominant.

8:36
So probably in about 5 years time, the ambient scribe space will look very much like the electronic health record space where you have a few dominant players and most other companies will likely not make it.

8:50
And the third application, which I think is the most interesting and meaningful one, especially coming from a clinician background, is using Gen.

8:58
AI for clinical purposes.

9:01
Eric, as you sort of alluded to earlier, the adoption of Gen

9:04
AI has been a lot slower in the clinical space.

9:07
And I think there are several reasons, one of which is the FDA made it very clear, they don't really know how exactly how exactly to regulate Gen

9:17
AI specifically LLM, because by definition you can get a different answer from LLM if you give it the same text prompts.

9:28
So this this is very different from what the FDA is used to regulating.

9:33
So I think the last application which I think is the most meaningful one, unfortunately will be the slowest one to be adopted.

9:42
However, the potential is tantalizing.

9:44
For example, it's already been shown that you can use large language model to analyze some of the emission notes, probably the different clinical notes in the emergency room and predict some very important metrics, for example, the risk of readmission within 30 days.

10:00
And these kind of LLM based clinical decision tools often times have been shown to be more powerful than traditional ones.

10:09
Now whether they actually get implemented or not, it's a big question mark because of the reasons that I mentioned.

10:16
Yeah.

10:17
And certainly we're going to get into both how, how AI is governed, governance of some of these models and some early wins where, where you're seeing successes right now.

10:28
We're going to get into that eventually.

10:30
Carol, I want to give you a chance, you know, what are you seeing from an industry wide point of view on this?

10:35
There are two two quick comments building on Kristen and Liu's answer one,

10:41
I well couldn't agree more that there's the the potential for really having a a game changing impact in our industry is is definitely there.

10:49
And a lot of times we, we forget the, the kind of we forget to acknowledge what has changed in the past two or three years and what the Gen AI capabilities where it has accelerated and has come to a point where it's, it's commercially usable at this point.

11:06
And I'm thinking of the ability of Gen AI to summarize to really taken to input a large quantity of, of unstructured information, which is really relevant in our industry where a lot of our healthcare data is definitely not at the level of standardization that we would like.

11:24
So the ability for for Gen AI to take into consideration a large amount of data and to summarize that amount of data into a specific context or to answer a specific question is really a tremendous capability.

11:37
Obviously voice recognition is, is an AI training AI capability that's being deployed and and recognized in I think as definitely you mentioned more very broadly and or very broadly and very rapidly in our system.

11:54
So the ability to to use a large amount of data to summarize it and then the ability to really sustain and manage a very interactive set of interactions with patients and with providers.

12:11
The fact that we are the Gen AI technologies at that level now is really game changing in our industry.

12:17
Maybe last comment about really Gen AI being applied to the clinical space as being maybe later down the road or being one of the the last domains to be exploited.

12:32
I would just, I was reflecting on the fact that a lot of our admin tasks also have an influence on care quality.

12:43
So although it's not directly linked to delivering care or ability ready to get patients into visits, into presenting to their visits when and where it, it makes sense, the ability ready to automate a lot of interactions between the providers and the payers.

13:01
There's a lot of administrative tasks that we can improve that we can automate in our industry that at the end of the day will have a positive impact on on our patients care certainly very good.

13:13
OK, OK, We'll we'll get to the some of the some of the win winning some of the positive cases.

13:20
But I want to talk a little bit first about some of the challenges and I'm going to start with Alvin on this one.

13:28
What are the biggest challenges in healthcare the generative AI can address and how can we prioritize them to make the biggest difference?

13:42
I think it depends on how you define a big challenge.

13:48
I would say going back to what I said earlier, I still think one of the big challenges we have in Healthcare is administrative work and depend on which study

14:03
you look at people quote numbers as high as 30% of the healthcare expenditure in US can be categorized as some sort of waste.

14:12
And I don't need to preach to the group that, you know, there's a lot of paperwork involved.

14:17
Well, when it comes to healthcare that can be simplified.

14:20
So I, I really think, you know, if we, there's always going to be clinical challenges, but in terms of challenges to the US healthcare, but one thing that's quite specific to the US is the fact that we do a lot of paperwork.

14:33
So as Carol mentioned, large language models really, really, really good at dealing with paperwork, especially when it comes to unstructured informations such as free text.

14:46
So I would say the number one challenge that Gen

14:49
AI can solve without having to deal with a lot of the medical legal risk is administrative work.

14:55
And specifically I would say revenue cycle management.

14:58
Because at the end of the day, you know, more than 90% of Americans are covered by some sort of medical insurance.

15:05
So the whole payment mechanism in the US really has to go through like payers and providers.

15:12
So even if we solve the inefficiency in the interaction between payers and providers, I think that could be a big win.

15:20
And the nice thing about this approach is, is relatively low risk, right?

15:24
Like we're not dealing with clinical decision.

15:27
And once you have, once you embed certain large language model capability into like a big organization like Johns Hopkins to initially deal with RCM, you can then relatively quickly you can then repurpose that set up to do more interesting clinical work.

15:47
Because ultimately if you are using large language models to do RCM, you need to be integrated with the electronic health records anyways.

15:56
So by definition, you can also get access to the clinical information and then repurpose the large language model tools to do something that that's more clinically oriented.

16:08
Yes, it's it's been said that AI can can take away the paperwork or take away the administrative tasks that that are befuddling doctors and nurses.

16:18
They spend too much time in front of the computer and not enough time in front of the patient.

16:21
The idea of the promise for AI is that it can take those tasks away and

16:24
It can do them better as well.

16:26
Christian, same question.

16:27
What are the biggest challenges in healthcare that generative AI can address?

16:32
100% agree on administrative burden.

16:36
I also think there's just such a rising operational costs in healthcare and they just continue to increase.

16:45
And, you know, all health systems are under significant pressure to manage expenses.

16:51
And, you know, I think that generative AI can really help optimize costs and, you know, in areas like revenue cycle, as Alvin was mentioning, and supply chain, as well as contact centers and even the technology organization.

17:07
I also think that it can help with gaps in patient engagement and access and, you know, really personalized communications and help improve self-service for patients.

17:19
But again, I, I think the focus has to be on high impact and low risk use cases.

17:26
And you know, it really needs to be focused on.

17:31
I think an organization needs to be intentional about what they're prioritizing and it should always tie back to the organization's strategic goals.

17:40
And what's important is to have that robust AI governance framework in place so that you can evaluate the risk and ensure the ethical use and guide the prioritization based on that strategic alignment and, you know, the input and the guidance with compliance and legal.

18:02
Yeah, I like what you said there, the idea that AI, there's so many positive use cases in administrative details.

18:10
Carol, do you feel that the industry is prioritizing the right things when it looks at Generative AI and what it can do to healthcare?

18:18
I, I mean, generally speaking, yes, with maybe a caveat and maybe acknowledging one of the areas where, where Optum has been placing a big bet and is in helping health systems accelerate their transition toward value based care.

18:34
If we're talking about really trying to solve the biggest challenges in our industry, access, cost, quality, etcetera.

18:41
I think there's a, a general recognition differently from from a federal policy perspective that value based care in different forms is, is definitely part of that solution.

18:51
And Gen AI I can help in multiple areas and multiple components of value based care from already understanding how some of the contracts and the incentive programs can be applied to or should be applied to providers delivering care and to patients engaging in their own care.

19:09
But also in improving the experience and the capabilities of the provider is really performing in that construct in the ability to further engage patients in their care.

19:21
And I know it's a, it's a big, it's a large topic per say, but in view, if you, if we view it three in the context, the broad context of value based care, Gen AI can do miracles in terms of engaging patients and guiding them to the right process in our system.

19:37
So using Gen AI to improve performance in the value of the base care in general to me is, is definitely one of the the the top opportunity that we have in improving our health system. So much potential.

19:52
OK, let's start digging down into the details now.

19:55
Kristin, how's Northwell developing generative AI programs?

20:00
What's your strategy for for testing these new ideas?

20:04
Look, I think that

20:06
You know, our strategy has to be aligned with our enterprise priorities and you know, we've identified high opportunity use cases and you know those have to be done with operational scale.

20:21
And you know, we've got to have strong data maturity and a clear business demand.

20:26
And you know, we look at where Gen AI can have impact and deliver value.

20:32
And so you know, we have a centralized AI intake process and, you know, governance framework that review every request that is being made by our stakeholders.

20:45
And you know, we're able to consistently evaluate risk and ethics and compliance and ROI and you know, the alignment with the organizational goals.

20:56
So, you know, I think in terms of testing, we make sure that we're conducting structured pilots and you know, we need to look at, you know, financial impact and operational feasibility and you know, making sure we've got success criteria and you know, we're looking at data usage agreements, total cost of ownership and readiness for broader implementation.

21:24
So that there's informed decision making before we're actively scaling.

21:30
And we're also looking at AI performance.

21:33
And you know, we have a strong emphasis on transparency and accountability and fairness.

21:38
And we want to reinforce trust in the environment and also continuous improvement.

21:45
So I think ongoing monitoring and governance are just essential to sustain responsible innovation.

21:54
Now, do you get a lot more ideas than you can test out?

21:58
And this is only the hype of AI right now.

22:01
Everybody seems to have great ideas about what AI can do.

22:04
So do you see a lot of that that you just say not now?

22:08
Yeah.

22:08
I mean, look, I think that you know where where we've got a lot of demand, but we're also looking at generative AI with the platforms that we're implementing as well.

22:23
And sometimes what we're finding is that, you know, there'll be a particular product that, you know, many, you know, some of our users would want to implement, but it's already part of a platform that we're implementing.

22:40
So you've got to really understand the road map of, you know, the technology platforms in your environment around Gen

22:47
AI as well to ensure that, you know, you're actively assessing where to place your investments.

22:56
And I think that's very important as well.

22:58
Yes, definitely very good.

23:00
Alvin, same question, you know, what, what generative AI programs are you developing and what's your strategy for new ideas?

23:08
So there are two live programs at Johns Hopkins Medicine that are related to Gen.

23:13
AI.

23:14
One is MB inscribing which I mentioned.

23:16
And 2nd, we deployed Gen

23:21
AI tool for small part of RCM specifically in obtaining prior authorization. In terms of a broader AI strategy that's not limited to just a Gen

23:34
AI, but just AI overall similar to what similar to what Kristin mentioned at Hopkins, we also have a formalized AI governance structure and I serve on that.

23:48
Essentially, what it is is that every AI vendor who wishes to interact with Johns Hopkins have to go through the same intake process with specific questions related to such as IT infrastructure and cybersecurity.

24:02
After each vendor goes to the initial stages, it will also need to find an internal sponsor from Hopkins, a business owner to that advance that application along different routes.

24:14
And there are three subcommittees, 1 is imaging, 1 is clinical, and 1 is operational.

24:19
So depending on the nature of the application, it will go into one of these committees.

24:24
And a big emphasis we have at Johns Hopkins is that these committees have to be led by clinicians such as myself, of course, with other stakeholders on those committees from various parts of the organization.

24:38
But there's a very strong emphasis that this whole AI governance structure should be led by clinician because ultimately we're interested in the responsible and safe deployment of AI for patients.

24:51
And then once the application goes through the committee, it will be reviewed in a very specific manner.

24:58
Every application we have to go through the same evaluation rubric and as a committee then we vote yeah or nay or

25:05
But oftentimes what ends up happening is we'll have some back and forth between the vendor itself and also the internal business

25:13
Hopkins business owner.

25:16
And I like to also, you know, echo what Kristen talked about, which is the importance of the AI platform, which is a catch 22 because from a health system perspective, like naturally, we're more inclined to work with vendors that provide AI platform that can do multiple things.

25:37
In general, we don't really want a very specific point solution.

25:41
However, the the real life challenge is such that each of these Gen

25:46
AI application pretty difficult to deploy.

25:50
It requires you know like modification of workflow, change management and also different kind of fine tuning.

25:58
So what we notice is that on one hand, we like to work with AI platforms.

26:03
On the other hand, it's impossible for an AI platform to do everything well.

26:07
So there are certainly some AI platforms out there that can claim that can do many, many things.

26:14
And what we notice is if you claim you can do many things well, you just end up doing none of them well.

26:19
So the trick for health systems is to find a platform that's wide enough to be useful, but also why that it just becomes too general or another way to say is find a point solution that's important enough for us to invest in despite being a point solution.

26:39
I like what you say there, especially the idea of a catch 22, because in technology and innovation now it's people are kind of moving away from niche solutions because they only solve one small problem.

26:53
But that's where this industry makes its advances and in finding these small solutions and then expanding upon them.

27:02
Are you seeing the same things, Carol?

27:04
Is this is this a challenge in in how AI is is governed and and developed?

27:10
Absolutely.

27:12
And it's definitely is it's a different beast, if I, if I can call it that way, from a lot of the, the technology programs that we we've been working on for some time.

27:24
Really in the sense that first and foremost in, like most people, most health systems are building a list of their top 10 to top 100.

27:33
But on that list, you have some small things, you have some some bigger things, you have some of the initiatives that are almost personal.

27:40
So really the adoption of AI app and at the enterprise level, but but all the way down, it happens also all the way down to the personal level.

27:47
So much more challenging from a governance, from a testing, from a quality control perspective, really to govern and to monitor.

27:56
So activities such as really offering education and building awareness for all of the users potentially would be using Gen

28:06
AI as part of their work, whether it's again, individual personal work or as part of process or workflow in the organization.

28:14
It's very important to add that to the programs or the the list of activities that should be performed as part of rolling out Gen AI more broadly. It's definitely a, a different perspective, different shape from compared to a lot of the, the enterprise programs that most of us have have had to manage in the past.

28:38
Yes, OK, here we go.

28:41
We've we've talked about this a little bit now.

28:43
Now let's let's talk a little bit more specifically about it.

28:46
I'm going to start with you, Alvin on this one.

28:48
You know, how do you address Gen

28:50
AI governance and do you manage AI differently than you do other innovative technologies?

28:57
So I already briefly talked about the AI governance process, Johns Hopkins that is for both Gen

29:04
AI and non Gen

29:05
AI.

29:06
I would say, you know, one additional consideration for Gen AI is that oftentimes the dominant players in the industry are very large language models hosted on a cloud outside of the enterprise IT environment.

29:24
So if you are a hospital or health system that cannot really host your LLM, then you really need to think clearly about how to send data out of your own like firewall and interact with these like large language models that are hosted on the cloud by some bigger tech company.

29:44
So I would say, you know, in general, the consideration for AI technology should be pretty uniform and that's why we have a very uniform governance process that governs both Gen

29:56
AI and non Gen AI.

29:59
However, there's an additional level complexity when it when it comes to Gen

30:03
AI for the reason that I mentioned it.

30:06
You spoke a lot about vendors,

30:08
Does John Hopkins deploy a lot of AI vendors and do you are you able to develop anything in house?

30:16
So I think very similar to Northwell as a health system, we got a lot of requests, inbound requests both solicited and unsolicited.

30:25
When it comes to different kinds of AI solution. I do not have a good sense of what's the median number of solution deployed across the country,

30:36
So I can't speak to the fact that whether you know, compared to the other systems, whether Johns Hopkins it's experiencing more inbound requests or not, I would say you know, but Eric, going back to your question,  we've had in terms of deployment AI, there's one area that we've we have a lot of experience in that which is in autonomous AI for diabetic retinopathy screening.

31:03
So this technology was first approved by the FDA in 2018.

31:07
In fact, when it was first approved, it was the first ever fully autonomous AI device in any medical field to be approved by the FDA.

31:17
And a recent study that looked at the real world deployments of medical AI technology across the board, this particular use case is the second most applied use case in the country right now.

31:29
So it's starting to be becoming at scale.

31:32
So at Johns Hopkins we started deploying this technology back in 2020.

31:39
So we did not develop this, but we as an organization probably has one of the most extensive longitudinal data set when it comes to real world deploymentistic technology and we have been able to look back and see what happened.

31:54
So very briefly, diabetes is a very common condition and diabetes involvement of the retina is a leading cause of blindness in the working age population around the world.

32:06
It is recommended that every adult with diabetes get an eye exam once a year.

32:11
And historically you go to your eye doctor and that's your annual check out.

32:14
But now with autonomous AI, you can actually get the screening done within minutes when you go to your primary care doctor.

32:21
So it really saves one visit and it's a lot more convenient.

32:24
What we show at Johns Hopkins is that by deploying this technology, the deployment was associated with general increase in adherence for this actually HEDIS metrics again going back to value based care and also when we looked at different patient subgroups, the improvement in adherence was particularly pronounced in historically disadvantaged patient groups such as patients covered on Medicaid and African Americans.

32:53
So I would say this is a good example of what we can learn from real world deployments, AI tools, but of course there are other ones out there that we have not gotten a chance to deploy yet.

33:05
Yeah, certainly there are a lot of options.

33:07
Kristen, same question, you know, how do you address Generative AI governance and do you manage your AI projects differently than you do other innovative technologies?

33:18
Yeah, so all AI related initiatives go through a formal AI intake and governance process that we set up last year.

33:28
It's part of a broader technology intake framework, but we do have heightened scrutiny as it relates to AI.

33:35
So we look at the ethical implications, data usage, regulatory compliance, and, you know, just the technical complexity and, you know, the process really begins with centralized intake.

33:50
There's a triage meeting held two to three times weekly.

33:54
They evaluate, the team evaluates business needs and technology alignment and fit, looking at enterprise platforms as well.

34:04
And then the initial reviews go through risk and security and an ethics assessment and you know, partnering with, you know, our cyber team and then it's reviewed by multidisciplinary governance committee.

34:20
And you know, we look at opportunities with the standardized framework.

34:25
So we have a rubric that goes through risk and business value and the maturity, technical feasibility and alignment with organizational priorities.

34:37
And then you know, we also have finance reviews biweekly looking at ROI, total cost of ownership and also some of the success metrics of these programs.

34:50
And then final decisions are really made through cross functional collaboration at the executive governance committee.

34:59
So I really believe strongly that having, you know, governance framework really can address financial and also the clinical needs.

35:10
So you have to be able to drive the AI adoption and you have to be able to ensure that there's transparency into what is in the environment, you know, from a technology standpoint, what's being utilized.

35:27
And then you have to be able to monitor the AI models and also optimize the models where appropriate.

35:35
So, you know, I, I think that it warrants a separate process, a separate and distinct process than, you know, other technologies.

35:48
Yeah, I like how you mentioned to the the idea of financial and clinical value.

35:53
And that's that's one of the bugaboos right now of AI, of ROI is right now in this economy with health systems struggling really to maintain the bottom line, financial ROI is is crucial for success, for scalability, for sustainability.

36:11
And yet with AI, there's so much potential to move the needle on clinical outcomes that doesn't relate to the financial part.

36:21
Is it a challenge to to look beyond financial ROI and think about clinical ROI?

36:28
Look, I think that, you know, the organization tries to take a very balanced approach when looking at ROI think that, you know, there's the tangible and then the intangible, right?

36:41
And sometimes, you know, we make investments in programs because, you know, we think it's going to, you know, improve clinical care or also be an experience play.

36:56
So again, I go back to it has to be intentional, it has to go back to overall mission and strategy of the organization.

37:06
But you know, clearly, you know, given some of the constraints that all health systems are operating under, you know, the financial ROI is going to be a huge driver I think in the future as well.

37:21
Certainly.

37:23
Yeah.

37:23
Carol, your thoughts on this is it, is it still all about financial ROI and our health systems doing the right things with governance?

37:32
Well, I think I mean the short answer is yes, just by starting by leveraging or expanding the existing governance structure,

37:40
That's the place to start.

37:42
And that's a place where policies are are being established on what the organization wants to do and how it wants to do it.

37:49
But building on what Kristen was saying, being able to deploy some of the tools that actually allows for the application of these policies, being able to monitor really the behavior of elements in the number of these Gen AI tools is critical both in terms of what we get out of such an AI tool and, and the existing LLMs and in terms of how and being able to monitor how they will evolve over time.

38:19
Even in in some of the early successes in ambient listening, for example, we have seen already a lot of progression, a lot of enhancements in most of the models out there just in, in a matter of a few months.

38:33
So the, the ability to monitor how they these models will, will evolve is going to be key.

38:40
The other aspects that we've been hearing a lot this, that kind of sensitivity to really to sensitive information.

38:48
So as we question, as we use the these models, a lot of times we provide, we share information and that has the potential of, of being public or being used as feeding these these large language models.

39:02
So having the ability not just to have the policy, but having the tools to control and limit really how, what information is being shared with, with which providers or which models is going to is going to be key in the future.

39:14
And that's going to require the whole new set of infrastructure and set of tools that will may, may not have been present in, in, in the past environment.

39:24
Yes, certainly good answer.

39:27
We have zoomed right past the half hour mark of this as I knew we would.

39:30
Eric, can I can I just address what Carol and Kristin said really quickly?

39:34
I think the, the question of clinical versus financial ROI is it's a very important one and what essentially I think the industry's been, you know, healthcare industry is doing is trying to balance both in a very delicate way.

39:48
And the truth is right now in 2025, there's quite a bit of misalignment between clinical ROI and financial ROI.

39:56
But I'm cautiously optimistic that as the US healthcare system moves like towards value based care specifically, you know, the CMS made

40:07
a decision that by 2030, the goal is CMS is to have all traditional Medicare beneficiaries and the majority of Medicaid beneficiaries to be under some sort of value based care arrangement.

40:20
When that happens, I think the clinical and financial ROI for deploying Gen

40:26
AI will be a lot more aligned.

40:29
And going back to an example that I gave earlier today, which is it's been demonstrated that we could use large language model to predict which patients are at a high risk of being readmitted in 30 days in a purely fee for service set up.

40:46
There's not too much financial incentive to do that to improve the clinical outcome, right.

40:52
But once you pair with some kind of value based arrangement and the CMS does have some new program in this area and specifically you know 30 day remission rate is one of those metrics that are involved in payment calculation, then you can see all of sudden the two really start aligned.

41:10
And I think until we see the more widespread adoption of the value based medicine, there is always going to be some tension between clinical ROI and financial ROI.

41:21
But hopefully the two are going to be aligned as we move towards Value Based Care.

41:28
Very well said.

41:29
Yes, definitely.

41:32
Yeah.

41:32
We zoomed right past the half hour mark and I wanted to mention that we will have that Q&A section coming up.

41:38
I only have a few questions left on my side, but we've got some really good questions coming in from our audience.

41:44
So I, just as a reminder, you've got that Q&A button.

41:48
Use it and we will, I'll, I'll start asking these questions as we, as we move along.

41:52
I, from what I'm gathering, we only got about 20-19 minutes left anyways.

41:55

41:58
In fact, I'm going to jump into a question now I've seen a couple of times and it, it kind of addresses some of the concerns that that doctors and nurses have about AI and that this kind of goes into the hype surrounding AI.

42:13
Oh, there's a lot of concern that AI is going to replace doctors and nurses because of, of, of what it can do.

42:21
How do you address that concern?

42:24
I'll start with Kristin on this one.

42:27
Yeah, look, I think it comes back to, you know, AI literacy and also education.

42:34
And, you know, there's, there's the promise of AI in terms of really being, you know, like a, a human assistant, right, in terms of being able to assist, you know, with a lot of the administrative tasks.

42:51
And I think that, you know, being able to demonstrate that, that, you know, you always need from

43:00
And, and, you know, I'm interested to hear what Alvin says about this, but you know, from my perspective, having a human in the loop as it relates to, you know, AI in clinical care or clinical use cases is extremely important.

43:15
So I, I don't see the, that physicians and nurses should be concerned about being replaced.

43:23
I but I do think that AI is a tool that they can incorporate into their tool set, you know, long term, especially around, you know, getting time savings and, you know, ensuring that they can spend more time at the bedside, quite frankly, with their patients rather than some of this administrative burden, which we know is is considerable, quite frankly.

43:49
Nice.

43:51
OK, Alvin, same question.

43:54
Yeah, that, that's a great question.

43:56
And that's something that I.

43:58
I get asked all the time.

43:59
I think my short answer is it's very unlikely that AI is going to replace doctors and nurses and, you know, healthcare providers.

44:10
However, what's for sure it's going to happen is the nature of these jobs going to change significantly.

44:16
And at a high level, I would say, you know, AI, I always say AI would not replace doctors, but doctors who use AI will replace the doctors who don't.

44:26
And to a certain extent, I think that's going to be true.

44:29
I mean, AI is not magic in the sense that it's just like any kind of technology, right?

44:33
It's simply, you know, makes us a lot better at what we do.

44:36
And I'll give you an example.

44:38
So I'm a retinal surgeon.

44:39
I look at like I examine patients all the time. For the same kind of patients

44:44
20 years ago, a retina specialist may spend 20-30 minutes just looking at the patient trying to figure out what's going on.

44:51
Now with much better technology, specifically imaging, it takes me a lot less time to figure out what's going on.

44:59
So the nature of my job is different now, even compared to 10 years ago.

45:03
Now I can spend more time, you know, providing an intervention or spending more time talking with patients, explaining the condition.

45:12
And I think AI is going to play a similar role.

45:15
I, I do agree with Kristin that because of a variety of reason, one of which is medical legal reason, more likely than not in the next 10 years, AI will be deployed in an assistive manner, meaning there's always going to be a human in the loop.

45:31
We got to have an AI assistant help us to do a variety of things.

45:35
There will be a small amount of applications that that that are going to be fully autonomous, meaning with minimal to no human intervention.

45:43
But at least for the next 10 years, I think that will be the exception.

45:47
What's really interesting is that this is also very context specific.

45:53
There are now many studies trying to compare, you know, for certain workflow or clinical diagnosis or just a clinical scenario, which one is better human versus AI only versus human plus AI.

46:06
And it's actually surprising that depending on the clinical context, the answer is a little bit different each time.

46:13
More likely than not, the purely human team end up losing.

46:18
But it's still like a debate whether only AI is better versus AI plus human because that's very context specific.

46:25
So at least in the academic sense, more people are doing their vigorous study to investigate different kinds of clinical scenario workflow to see which one's better, like AI only versus human

46:38
plus AI. I think the future lies in how we define AI as well.

46:44
I mean, AI used to be artificial intelligence.

46:47
Everyone called it artificial intelligence.

46:49
More and more I'm hearing augmented intelligence, yeah, as as the true definition.

46:55
And there's also like legal like question in that because more like now there are more and more use cases where we know if you augment human with AI, the outcome is better than just human.

47:08
So let's say 20 years from now, if you're a human doctor and you don't use AI, you actually provide substandard care.

47:16
And that's I think that's a question that will need need to be addressed in a relatively soon future.

47:22
Yeah.

47:22
And it has been said that a patient could eventually sue a doctor for not having AI, which is an interesting thought.

47:31
Carol, your thoughts?

47:33
I couldn't agree more with everything that was just said.

47:35
Maybe a follow up question for Doctor Liu

47:38
actually, I, I, my, my bet is that exactly as you said that the, the, the work of our clinicians is going to change drastically.

47:47
And I'm intended to compare to going from paper to DHR and how much of a change it has been for clinicians.

47:55
I, I don't know if you would compare Gen AI as having a similar level of impacts in terms of how much it's going to change the way medicine is being practiced.

48:11
You may have gone to mute there.

48:27
No, we're not hearing you.

48:28
Oh, there we go.

48:30
OK, there we go.

48:32
Thanks.

48:33
So Carol, I think the impact is going to be even bigger because if we think about it going from paper to electronic, we got a little bit more efficient, but that's pretty much it.

48:46
And before the rise of AI, we had been able to leverage electronic health record in terms of big data and gain some additional insight.

48:57
But the insight really has been very limited for a variety of reasons like each out the data is pretty dirty.

49:03
And before the rise of LLM, you can only analyze the structured data anyway.

49:07
So I would say when you look at the actual kind of things, so insight we can derive from electronic health record versus just paper, we did not, we have not done that much better.

49:19
But I think Gen

49:19
AI is going to categorically change the game, not just in text.

49:24
Because for the newer like newest form of your Gen

49:28
AI models, it's really, we're really going multimodal now, meaning you're combining text with image plus voice and everything.

49:35
So it's a completely different universe that we're living in.

49:38
So I think the impact is going to be much bigger.

49:41
And going back to the question of whether, you know, AI is going to replace doctors, I think another thing we need to remember is there's a chronic shortage of healthcare providers.

49:50
And that shortage is going to get worse and worse.

49:54
So I think there will be plenty of problems for us to solve and plenty of patients for us to take care of.

50:00
Like I, I'm not worried that doctors will be out of a job anytime soon.

50:07
Perfect.

50:08
We are 10 minutes left in the hour.

50:10
And I'm definitely going to go into the questions I'm getting from the audience.

50:12
Now, one person here had mentioned and and we touched on this here and there, the idea of data, data, so much of AI and pretty much all of AI is data, data and data analysis.

50:27
How do you ensure that this data is protected and that your patients data is being secured?

50:36
Kristin, we'll focus, we'll start with you on this.

50:40
Sure.

50:40
I look, I think firstly, from a saga perspective, making sure that, you know, you're, you know, looking at the maturity and the investments being made in the organization to protect data is paramount, quite frankly.

50:57
So, you know, I would say, firstly, addressing from a cybersecurity standpoint.

51:02
Secondly, data governance is especially important and making sure that separate from, you know, your AI governance, even though it's, it's somewhat interrelated and can move from committee to committee, but important that, you know, there are appropriate standards and controls in place to manage data.

51:26
And thirdly, I would say data use.

51:29
We also, you know have a data use team looking at all of the provisions that are in some of these contracts.

51:40
And what we've noticed over the last few years is just the fact that many vendors are expecting healthcare organizations to provide them all of the clinical data and you know, for them to be able to train their models on.

51:59
And I think that making sure that you're putting provisions as an organization to protect the patient data is tremendously important.

52:10
So that's how I look at predicting patient data with those three areas.

52:18
Nice.

52:18
OK, Alvin, same question.

52:20
Privacy and security of of data.

52:23
I think that's obviously an important issue.

52:25
And I agree with everything that Kristin talked about, which is, you know, having a robust cyber security infrastructure and also AI governance.

52:33
I think that's in 2025 and going forward those are going to be table stakes and really what's expected of us, especially when it comes to integrated health systems.

52:43
I would say on the technical front, there are two interesting developments that that that can also address some of this issue.

52:50
One is, you know, everyone knows that AI is driven by data.

52:55
So in general, more data the better.

52:58
And historically what people have done is, you know, send vast amount of data back and forth between entities to combine the data to train the AI models.

53:08
There's now there's been like this is a this is a development that started a few years ago, essentially a framework called Federated learning or swarm learning, in which we don't exchange data.

53:22
Each entity and organization would train the model independently and we exchange the model weights and details instead of the data itself.

53:31
So in that way, we can actually pull data indirectly without actually exchanging data while taking and at the same time being able to train model in a better way.

53:43
The second development something called differential privacy.

53:48
At a high level, it's a way to generate synthetic data from real data.

53:55
But this is a very particular approach where it's been proven mathematically that one cannot reverse engineer the actual identity of a particular patient from the synthetic data set.

54:09
So that's a mathematical, this is a mathematical construct where it's literally impossible to reverse engineer like what's involved in original data set.

54:18
So there's some kind of something called differential privacy.

54:21
It's still relatively early, but it's been proven that this is possible in the tabular data or structured data format, where you can create a synthetic data set that cannot be reverse engineered but still preserve the characteristics of the original data set.

54:37
So these are two interesting technical developments that could make a difference in the next few weeks.

54:44
Yeah, this is interesting.

54:47
Excuse me, Carol.

54:49
Are our health systems and hospitals paying enough attention to privacy and security, I'd say yes,

54:58
although I think the, the part of the answer is to take responsibilities to, to some extent away from, from the hospital and the health systems and put it in the hands of, of the patients themselves.

55:10
And we're seeing that in, in many countries where we have the federal policies, government policies really points to doing exactly that.

55:19
Because a lot of people don't realize to what extent there is fluidity in, in how data is being shared in our system currently between providers, between payers and providers between payers, There's a lot more fluidity than than people realize.

55:35
And as long as we really limit the ability for patients to control themselves, put some some controls in what they share and who they share it with, I think we're going to we're going to, we're going to keep seeing these challenges in more security and privacy protections for each of the participants in our health systems and payers and providers.

56:01
Definitely is going to be better is going to improve the situation, but the amount of data being shared really exposes really personal data at the level that we we've not seen before and that will only accelerate with Gen AI.

56:14
I nice OK, we've only we've got less than 4 minutes left here.

56:18
Certainly something conversation we could have all day or fodder for future webinars.

56:25
But I I did want to ask one last question and this one's totally on me.

56:30
The idea of the evolution of AI, I mean we're still very early in, in how AI is being developed and used, especially in clinical care.

56:38
So I wanted to ask each of you where you see this going where, where will be the biggest advances in say the next couple of years?

56:48
You know, what do we get to look forward to?

56:50
Alvin, let's start with you.

56:51
I would, I would say the two things that you know, I'm quite excited about one is, you know, called agentic workflow and this is more towards the administrative side of things.

57:09
So we already talked about and we all agree that some of the most, you know, scalable use cases initially for Gen

57:16
AI is for administrative purposes, whether it's RCM or supply change management and whatnot.

57:23
And right now you look at most of these solutions right now they are presented as a copilot is is essentially in a system to humans.

57:33
I think in the next two years we are going to see a lot more so-called agentic workflow being deployed.

57:41
And fundamentally the technology is pretty similar.

57:44
It's still based on large language models.

57:46
The main difference is that these agents really are autonomous and can do things for humans involved planning.

57:54
So an example I would give is right now you can use Gen

57:57
AI to help you with obtaining the necessary documents to approve medical necessity when it comes to submitting a prior authorization.

58:07
The next evolution of it is an AI agent that can physically download the information from the electronic health record, log onto the payer portal, upload the documents and physically submit it for you.

58:21
So I think that's going to happen very soon because these are pretty low risk tasks.

58:25
We're not asking these agents to take care of patients.

58:27
It's going to happen.

58:28
So that's more like agentic workflow.

58:30
And the second part is on the data part, which is, you know, ultimately I think a lot of people's fantasy is to have an AI doctor of sort, right?

58:40
And then you probably see a lot of studies trying to ask, you know, that looked at whether these large language models can answer board style questions.

58:48
And typically they do pretty well.

58:51
But I think there's a limit to this approach because ultimately, when you go see a doctor, you're not trying to look for someone who can answer a bunch of board style questions or multiple choice questions, right?

59:00
Like you want someone you can talk to.

59:02
You can talk to someone who would say, oh, like based on your, your condition, I've seen some of the patients before five years ago.

59:09
So this is, this is why I'm, you know, offering this advice.

59:14
So what I'm saying is that a lot of the really crucial clinical knowledge is not captured right now in the way that these models and products being, you know, trained.

59:24
And a lot of the really good clinical clinical information comes from informal conversations between doctors in medical conferences where very difficult cases got presented and world's expert debate about how to manage certain cases.

59:39
And this is the kind of really good medical knowledge and data that's not captured at scale.

59:44
So my contention is that in order to really go to the next level, we need to start generating good data that's not currently captured and that that would be a good development in my in my opinion. Very good.

59:59
Nice, Kristin, same question.

1:00:02
What excites you the most about a the future of AI?

1:00:05
Yeah, I mean very similar.

1:00:06
I think agentic AI definitely for, you know, utilizing it for automation of, you know, a lot of the administrative burden, clinical efficiency.

1:00:18
But you know, I also think we can be,

1:00:21
You know, potentially amazing in the customer experience space and you know, getting these virtual assistants to handle, you know, more complex tasks and automate triage escalation, proactively resolving issues and, you know, ultimately creating an integrated and seamless and efficient experience for our patients.

1:00:41
I think, you know, we're going to see that evolve over time.

1:00:46
And then Alvin also mentioned multimodal AI for diagnostics, right, bringing together text and imaging and genomics.

1:00:54
And you know, all of this clinical information for, you know, comprehensive unified clinical pictures and ability to really personalize the treatment plan with greater precision and efficiency.

1:01:07
So I think there's there's lots to look forward to in this space.

1:01:13
OK, Carol, you get the last word on this.

1:01:15
I think we're, we're almost out of time.

1:01:18
But I, I see Jenny and I enabling value based care and pointing, starting by pointing to things we should be doing better things that we're, we're not doing that we should be doing in, in along the spectrum.

1:01:33
And then I, I see in the context of agenda AII picture each of us having a digital twin against which agents are really monitoring all the time and taking action when it when it makes sense.

1:01:49
So we're, we're, we're a lot of things point to us moving in that direction and I'm super excited about it.

1:01:57
Yes, yes, definitely.

1:01:58
OK, thank you everyone.

1:02:00
Fantastic discussion.

1:02:02
It's all the time we have today.

1:02:04
I want to thank our our panelists for once again for great discussion about generative AI and I want to thank our sponsor Optum for making this program possible.

1:02:15
Finally, thank you to to our audience for participating today and offer some really good questions.

1:02:21
We hope you'll join us in the future for other health leaders events.

1:02:26
And this concludes today's program.

1:02:31
Thank you.