Unlocking the Potential of Generative AI in Healthcare with Kaveh Safavi, the Senior Managing Director of Global Health at Accenture

[00:00:00] Hey everybody, Saul Marquez with the Health Matters Podcast. I want to welcome you back

[00:00:13] to today's episode where recording a lot of great interviews from Health 2023 in Las Vegas

[00:00:19] and today I have the privilege of sitting here with Dr. Cabe Safavi. He is a senior managing

[00:00:25] director at Global Health Accenture. He helps organizations harness the promise of technology

[00:00:31] and human ingenuity to humanized healthcare. I'm really excited to have you on the podcast,

[00:00:36] Cabe thanks for joining me. Good to be yourself. And so to kick things off,

[00:00:40] you know, I always like to better understand what inspires our health leaders. So what is it that

[00:00:45] inspires you in your work in healthcare? Well, I tell you, I started out becoming a doctor because

[00:00:50] I saw primary care doctors, pediatricians and I grew up around doctors and I just thought that

[00:00:55] was a fascinating way to make a living. I got into healthcare and I was really struck by

[00:01:01] the power asymmetry between doctors and patients and I felt like healthcare would be better if there

[00:01:06] was a better balance and trying to figure out how to empower patients as part of the care system

[00:01:12] was important to me. I'm now in a different phase of my life. I am part of the baby boom population

[00:01:17] and we are facing the next decade where there will be far more of us than there are caregivers

[00:01:23] and we are all going to have a really big problem getting access and waiting times are going to be an

[00:01:28] issue. And it's actually literally a personal issue and I'm very much interested right now in how

[00:01:33] we're going to solve that access problem. And I have done enough research and looked at the numbers

[00:01:39] and I know that it's not physically possible to solve the problem by finding more people to do

[00:01:44] the work because there aren't any or not enough and therefore we're going to have to think about it

[00:01:48] differently and having spent the last two decades thinking about how technology and people work

[00:01:53] together. I can't see any other path forward other than taking the work that's being done by our

[00:01:59] caregivers and figuring out how to scale their capacity by using technology and people together

[00:02:04] and the alternative is a serious crisis in waiting times. Yeah, it's certainly a big issue and one

[00:02:10] that needs solving and it's a personal thing for you and you're on a mission to make it happen.

[00:02:15] I love that. I'm an impatient man. That needs results as all of them. And doesn't like to wait.

[00:02:21] And we need it now in healthcare and you're not alone. We are in that time where we like our Starbucks,

[00:02:26] we like our Amazon Prime and things need to be there today and our healthcare is no exception.

[00:02:31] I see from Accenture's research that nearly 100% of healthcare leaders agree that Genai is ushering

[00:02:38] a new era and enterprise intelligence. How will this technology transform the future of healthcare?

[00:02:43] Well, Genai is a relatively new technology. We've known about artificial intelligence in general

[00:02:48] and largely it's been through the lens of knowledge and prediction but Genai is a very different

[00:02:55] kind of technology. There are many use cases but the use case that has caught our attention

[00:02:59] particularly for healthcare is language and the reality is that a lot of what takes up human time

[00:03:07] is the acts of reading and writing. I'm not talking about reasoning or judgment. I'm just talking

[00:03:12] about the literal acts of reading and writing. And if we could find a way to take some of that off

[00:03:18] of the human and give that to tech then people have more time to do the things that humans do which

[00:03:24] is really much more about the judgment, the conversation, the things that only the humans can do.

[00:03:31] And particularly when we're facing a shortage of doctors and nurses, we're going to have to figure

[00:03:37] out a way to do this differently. Yeah, I couldn't agree with you more. And so just pointing to

[00:03:42] the research again that you guys have done it also shows 40% of working hours in healthcare could be

[00:03:48] supported or augmented by language-based AI. How will this help the industry's current workforce

[00:03:54] channels? Well, so the 40% is actually of all jobs. So if you look at all jobs across all industries,

[00:04:01] 40% of what is considered work today are language tasks that are amenable to large language models

[00:04:09] that have been enabled by generative AI. When you go into healthcare specifically and I'm going to

[00:04:14] ask focus on let's just talk about healthcare delivery for a second. Around 30% of the tasks can be

[00:04:21] impacted by language models. 20% of the tasks the technology can actually substitute for what the

[00:04:28] human does. 10% of it can help the human do a better job, but the human still has to do the work.

[00:04:34] It's that 20% that gives us the headroom to start solving the supply and demand problem. When you go

[00:04:39] to things like healthcare insurance, that looks a lot more like the overall job market where you're

[00:04:45] in the more like 30-40% of tasks are language tasks that can be substituted. The key here is

[00:04:51] that technology has the ability in some cases to automate, that's replace and in some cases augment,

[00:04:57] which is to do better. Most of our artificial intelligence experiences pre-generative AI have

[00:05:04] largely been about providing information to humans to make better decisions, so augment. We've

[00:05:09] thought about it really as a technology to make care better by augmenting human capacity,

[00:05:15] not automating. But in the environment we're in right now, when access is a fundamental problem

[00:05:21] in the UK, in the NHS right now, you're going to wait at least 18 months for an elective procedure.

[00:05:28] In Ireland, you'll wait two years for an elective procedure in the public healthcare system,

[00:05:32] and in some urban areas right now, elective procedures depending on what you're looking for could be

[00:05:36] three, four, five, six months. That is untenable long-term. It's going to go from a nuisance to a

[00:05:42] political problem. We have to automate. Augmenting isn't going to solve the problem, and it's the

[00:05:47] automation component that I'm most interested in. That's the more novel part right now,

[00:05:51] but there's a real opportunity there. Yeah, I would say so, and some of these wait times are just

[00:05:56] unsustainable as we think about how public health unfolds and just the epidemic that could happen

[00:06:02] as a result of it. Healthcare organizations are looking at using Gen AI for learning purposes

[00:06:09] and planning for pilot cases this year. What would you caution these organizations about when

[00:06:14] really experimenting with Gen AI? I'm going to answer that question first by describing what are

[00:06:20] the uses that are likely to be the best uses in the near term. And generative AI is in its early

[00:06:27] stages, so what it will be able to do in three or five years, the models are going to get more powerful,

[00:06:32] so we can't fully conceive of it. But like what's in front of us right now, we can actually look at

[00:06:37] non-healthcare industries to see patterns. So common patterns in healthcare, common patterns are

[00:06:43] using the technology to help people find information inside of their own organization.

[00:06:49] The lingo is enterprise surge, but when you are trying to find something in your hospital or

[00:06:54] insurance company, it could be anywhere and it's hard. And if I can find it, that's all that time

[00:06:59] saved looking for information. The second thing is what we'll often describe as creating first draft,

[00:07:05] often describe this content creation for content creators. So it could be, for example, in the

[00:07:10] clinical setting now, what's happening, which is preparing first draft email as an example,

[00:07:16] or document summarization so that it reads, understands and synthesizes in a reliable way. So

[00:07:22] instead of me having to try to look at a 200 page 10 year record, one page pulls everything

[00:07:28] salient. And that I'll never get through 200 pages as a doctor anyway. And I won't miss anything.

[00:07:33] So that kind of stuff really helps us do our jobs more quickly. Recognize as you can tell, none of

[00:07:39] these are making judgments or substituting for doctors, judgments. All of these are tools to

[00:07:44] help clinicians do their job better than the big one that's really going to come up is the ability

[00:07:48] to create a written document like a note, whether it's a nurse's note or a doctor's note and convert

[00:07:54] a narrative into structured content so that you don't have to type. If the typing problem, which

[00:08:00] is a 13% drag on physician productivity, is a huge problem. And generative AI will eliminate the

[00:08:06] need to touch the keyboard. It will eliminate the need. Most observers think that within three years,

[00:08:11] we will have technology that widely available that will not require us to use a keyboard to interact

[00:08:16] with an electronic health record. The first versions are already being shown to the market. So

[00:08:20] that's a lot of time back. None of that was possible without generative AI. What I didn't describe

[00:08:26] is advice machines to patients. So the leading organizations, the first thing they did before

[00:08:32] they had any generative AI was say, generative AI is not an answering machine. It's a language

[00:08:37] machine. So you see organizations putting prohibitions about using the technology to face a patient and

[00:08:44] give a patient advice even though the common lay press often talks about that as the thing.

[00:08:48] That's the one thing we don't want people to do. Now that doesn't mean it can't do it eventually.

[00:08:53] But the challenge there is that the generative AI is really good at the language fluency. That's not

[00:08:58] the problem. The problem is the actual answer. Generative AI doesn't create knowledge. It doesn't

[00:09:03] create an answer that never existed. It will generate an answer right or wrong. So if you're going to

[00:09:08] use generative AI to give a patient, you better have a bullet proof body of answers that it's going

[00:09:15] to be looking for the information and the models that we're seeing today that are built on the open

[00:09:19] internet are no different than if you would have told someone two years ago, go look it up on the

[00:09:24] internet and take care of yourself. If that wasn't good enough two years ago, it's not good enough

[00:09:27] today. Generative AI didn't solve that problem. It just makes that information more compelling in

[00:09:33] its presentation, not more accurate, more compelling. And so that's one of the big areas is to keep

[00:09:38] that clear. There are a few other areas that are really important. One of them is to understand how

[00:09:42] models are built. Many of these models take information that you put in and use it forever to learn. That's

[00:09:49] how generative AI is supposed to work. Great, but that means you better put information that you want

[00:09:54] essentially make public and lose control of. So you cut and paste patient information into chat GP

[00:09:59] 3.5, you just released HIPAA information without consent. You put proprietary information in there,

[00:10:06] you released it forever, you've lost copyright and protection, all of those things. So getting

[00:10:10] people to understand that that's how it works isn't an important issue because we know for example

[00:10:16] let's say a large hospital system. The IT organization will put policies in place that say you have

[00:10:22] to go use our generative AI tools, but the independent medical staff are well meaning and they might

[00:10:27] start hacking. Now I can use generative AI to build something to do X, Y and Z. They have no idea

[00:10:32] what they're doing. And they're exposing themselves and others to risk. So we actually think that it's

[00:10:37] really important to get principles out early that people understand what generative AI is, understand

[00:10:42] issues of privacy and protection of confidential information, understand the concept of advice versus

[00:10:47] fluency in use cases. Those are all big issues. They are. And it sounds like you've given this a lot

[00:10:53] of thought. Have you guys documented this? Is a guide like this available to the listeners? So

[00:10:59] for example on the provider side there's actually a coalition of health systems. Mayo was one of the

[00:11:03] leaders to promulgate rules for delivery systems to think about this. There are a number of

[00:11:09] organizations that provide guidance around responsible artificial intelligence and guidance.

[00:11:15] And there's also a whole body of policy concepts that are being put forward. In the EU there's

[00:11:21] actually full regulatory standards, and it's based on the premise. And we'll see some of this

[00:11:26] bleed its way into the US at some point. The thing about generative AI that makes it wonderful

[00:11:31] and scary is that it is constantly doing things that the people who programmed it didn't think

[00:11:37] were possible. That means we can't fully control it. And I'm not talking about necessarily

[00:11:43] creating a sentient intelligent being with agency, but think of it like a nuclear reactor that

[00:11:48] goes out of control. So because it has the potential to do more than what we think one could argue

[00:11:53] that there are uses of the technology that probably should be regulated or classified or limited

[00:12:00] because they pose a societal risk. And you're starting to see it regulatory frameworks being built

[00:12:06] to allow us to make decisions about the wrong uses of generative AI because of the fact that

[00:12:12] it can get out of control pretty quickly. Thank you for that. Yeah, it sounds like right now

[00:12:17] there's work being done to navigate to provide guardrails and best cases around the use of the

[00:12:24] technology and more to come. So I really appreciate this discussion a lot. And I know it's top of

[00:12:29] mind for a lot of leaders, Kavya. So I appreciate that how do you take a responsible approach to

[00:12:35] Gen AI that reduces the dangers to judgment and diagnosis? Yeah, first thing is what we describe

[00:12:40] before which is understanding that it's not an answering machine. It's a language machine.

[00:12:45] And so whatever is the underlying scientific knowledge that you need, that's the thing that

[00:12:51] you ultimately have to be relying on. It's not a bad way of communicating. That information can

[00:12:56] be very compelling. There's data that says it can be even more empathetic in its communication than

[00:13:01] human can be. Awesome. That makes people maybe more responsive to it, but recognizing the limitations

[00:13:07] of it are it doesn't create knowledge that never existed. It doesn't create the right answer if

[00:13:11] it didn't exist. It's not an analytic tool. It's a language tool. That's by far the most important.

[00:13:17] And at this point, the challenge we have is that most people who haven't studied generative AI's

[00:13:22] only experience as chat GPT and the problem with that is simply that the way we experience generative

[00:13:28] AI is through a question answer experience as a consumer. But the purpose of it is not about

[00:13:34] the answer or the accuracy of the answer. It's about the ability to understand and comprehend

[00:13:38] the question with its subtlety and deliver an answer in a rhetorically compelling way. Not about

[00:13:44] the answer being right, but we don't have any way of experiencing the language fluency without

[00:13:50] a question and answer experience. And we get confused and think that what we're really doing is

[00:13:54] judging the answer. So in many of us did this when GPT was first in the market and we were

[00:13:59] admonishing people to be careful, and I would fall into that trap. We'd say here's how to prove you

[00:14:04] shouldn't rely on it. Have it write your own biography. Okay. And that's a, obviously set up because

[00:14:09] it wouldn't know your own biography. That doesn't prove anything. It shouldn't be able to write the

[00:14:13] biography if it doesn't exist because it doesn't have answers. So it doesn't prove anything. What

[00:14:18] you should ask for is assuming the facts and the biography are true. Is it a well-written biography?

[00:14:23] That's really the only thing that GPT is doing. I love that. Now great context and great points here.

[00:14:30] Kavit helping us understand Genai, its promise, its limitations, the framework of thinking

[00:14:36] that we should be using around it. So can't thank you enough for a phenomenal discussion on this topic.

[00:14:41] What call the action would you leave the listeners with?

[00:14:44] I think we cannot underestimate how important this is going to be to our work. And it's going

[00:14:49] to be driven in the short run by the necessity for us to get time back. And we're going to have

[00:14:55] to use this technology to get time back so we can give it to ourselves and to other people. So

[00:14:59] there's a lot of discussion about this scene through the lens of hype. I think that if you really

[00:15:04] step back and think about what it does and think about the long art, this is an inevitable technology

[00:15:08] that's going to have a pretty profound effect, you just have to place it in the right context to

[00:15:12] understand its impacts. Totally. Now I love that great closing thought. Folks, just a phenomenal

[00:15:18] conversation with Dr. Kavit Safavi from Accenture. Check out the show notes for just the details and

[00:15:25] hit rewind because this is one that I think you want to hit rewind on a few times and share with

[00:15:29] a friend. I really enjoyed our time together. Kavit, thank you so much. And oh, the last question,

[00:15:34] if somebody wants to reach out to you and learn more about the work you and the team at Accenture

[00:15:38] are doing, how can they do that? You can either find me on LinkedIn, Kavis Safavi or Accenture.com

[00:15:44] and just look me up there. Outstanding. Folks, make sure you do that. It's in the show notes,

[00:15:49] take advantage of it, and really appreciate your time. Thank you. Nice to see you. Bye, guys.