← Becker’s Healthcare Podcast
Becker’s Healthcare Podcast
Advancing Evidence-Based Care in the Age of AI with Dr. Peter Bonis
00:15:02
Loading summary
Transcript15 lines
- [00:00]
A
Hello everyone. This is Erica Spicer Mason with Becker's Healthcare. Thank you so much for tuning into the Becker's Healthcare podcast series. So today we're going to talk about advancing evidence based care in the age of AI and the opportunities, challenges and next steps for healthcare leaders that comes with that. And joining me for this conversation is Dr. Peter bonus, the Chief Medical Officer at Wolters Kluwer Health and Adjunct professor of Medicine in the Division of Gastroenterology at Tufts University school of medicine. Dr. Bonus, welcome to the podcast. Thank you so much for being here.
- [00:32]
B
I'm very happy to be here. Thanks, Erica.
- [00:34]
A
Well, we're thrilled to have you. And before we get into our conversation about AI, wanted to give you the opportunity to share a little bit more about yourself and your work in healthcare. Whatever you think our listeners should know.
- [00:44]
B
Well, sure. I I took the time to enter myself into Google Gemini, the latest version, the 2.5 Pro, and it was rather sobering because it reduced everything I did for health services in the last 30 years to just a few bullet points. So I will share what what what it shared, which is number one, as you said, I'm the Chief Medical Officer for Walters Klor Health, which is a global information services business. It also said that I had a role in the creation and the growth of Uptodate, which has become the most used knowledge resource by healthcare professionals now in the world and most trusted resource currently, as well, as you noted. It also said that I'm an adjunct professor in gastroenterology at Tufts. I actually still see patients, they go and supervise a fellows clinic every Friday for it. And finally, it said most recently that I've been helping to build and speaking about safe and effective AI, particularly in the high stakes domain in which I operate. And that was about it. So there's a, there's me in a nutshell, according to Google.
- [01:45]
A
So interesting and a humbling experience when we hear about ourselves from something like Google or AI. So thank you Dr. Bonus, for sharing and it's great to hear that you're able to keep a keep some of your work in the clinical realm as well as a practicing physician. So great to learn more about you. And I know technology and AI in particular is at the heart of what we'll be talking about today. And this shift toward using more technology in the healthcare environment is really changing how patients and providers find answers to health questions and the time it takes for them to find those answers as well. So from your perspective, what are you seeing as the biggest opportunities with this rapid Access to information, but also at the same time, what challenges should leaders be mindful of?
- [02:28]
B
Yeah, you're exactly right. So we are at an inflection point in how information is gathered. And so the challenge really is, can you do that and sort of benefit from the efficiency of new modalities of conveying information and stay accurate? And probably you can, but right now, at the moment, there's a trade off. And so sometimes that efficiency comes at the cost of accuracy. So the extent to which that matters varies. And certainly in a high stakes domain like healthcare, whether you're a patient or you're a healthcare professional, relying on some of these new modalities, like the large language models to gather information that you might apply at the point of care has its hazards. So I do want listeners of this to understand the stakes here. So if you just think for a moment at a time when you were sitting in your doctor's office, either for you or for a loved one, and that doctor gave you a piece of advice, how do you know that advice was correct? And so chances are you research it afterward. But there's a tremendous amount of trust now. There's a lot of data that supports how clinicians will get information. And as it turns out that about 30% of the time, if you give them information from an evidence based resource like Uptodate, but others as well, they'll actually change what they do. So now think back at that. You might have gotten a piece of advice and had there been some information available, it might have actually changed that piece of advice. So now map that to the availability of these efficient means to gather information for healthcare professionals. And what we found and others is that there are risks with them. And so we've heard of issues like hallucinations, they persist. But really the issue is how often will a healthcare professional recognize the hallucinations and then whether or not how often the healthcare professionals will recognize either false information or subtly false, grossly false information. Turns out that even specialists sometimes can get fooled by that. And then there's a matter of search reliability. So if you search and then you enter the search again, you might get a completely different response. Then there's biases of all sorts. So racial biases have been well documented, but every bias, the cognitive biases that are present in the literature also appear in these distillations through large language models. So, for example, if you type in shortness of breath at the beginning of a prompt, you might get a different diagnostic suggestion than at the end of the prompt for it. Then there's the matter of automaticity, where you are so efficient that clinicians maybe just do what they say that the LLM says it's going to do, and that without really engaging their brain and thinking about and contextualizing it. And that can cause real hazards. So these things are not abstractions. They are really quite real. So again, think back at you being the patient for yourself or the loved one. And I'm going to give just a few examples of things that actually occurred in large language models that are currently used by doctors. So recommending surgery on the carotid artery on a patient who didn't require it, recommending that a specific antidepressant could be stopped cold turkey even though it was known to cause a withdrawal reaction. Recommended avoiding influenza vaccine because of an allergy, even though it was safe to give that patient the influenza vaccine. Recommended the wrong starting medication for rheumatoid arthritis. It recommended drugs known to cause fetal harm because it didn't inquire as to the context as to whether or not the patient could be pregnant as well. So in any case, the prospect that some of this information that a clinician might be looking at might sort of slip by and actually get implemented is kind of frightening. And so it's incumbent on those who are developing these applications to do everything possible to make sure that they are consistently safe and accurate. One other point, since you mentioned patients, patients are using this. We're recording this today on October 1st. So those of you who look at the Wall Street Journal, there was a very nice article written about this and saying that how patients are using this large language models to get health information, and there's the issue of it's practical because it's very hard to get employment or perhaps they don't have health insurance, or they have to delay. So isn't that better than nothing is sort of the argument for it. And the issue is can we have our cake and eat it too and, and actually be able to produce that information but make that high quality? It's very hard to do, but it can be done. Wow. Dr.
- [07:02]
A
Bonus. I really appreciate that overview and how clearly you outlined the risks of this immediate access to information. Those examples that you highlighted of what has gone wrong and what can go wrong is it's really compelling. So I appreciate you sharing that. And as you were speaking, it grows more and more clear to me that having information that a clinician is sharing with a patient or a patient is accessing, it must be trusted and of high quality. So from your perspective, what are the specific Strategies, tools, or even best practices that you've seen help organizations ensure that that information is trusted by both patients and clinicians.
- [07:43]
B
Well, there are a lot of considerations, but I wanted to spend a moment talking about governance because it's just so important. It's not an abstraction either. So having proper governance within a health system or any, any place where you're deploying technology is just critical and it's not easy to do. You need people who are well versed in technology and legal and regulatory matters, in workflow integration and understanding the financial implications and making choices and prioritizing it. And so not every system has that level of expertise, is still growing, and certainly AI has added a new dimension to it as well. But it's very important. And certainly health systems that don't have that will be relying on third parties, such as the vendors themselves or their electronic medical record system. But I want to make one point that may be overlooked even among those who do have these governance processes, which is to make sure that you involve the frontline. So what we have found in a survey that we did nationally is that only 18% of frontline providers were aware of any governance taking place around AI or other applications. And that's unfortunate because, number one, that if they don't know of the governance, they're going to be probably not adherent to the policies. So that's, of course, practical. But more importantly than that, their voice and their collective voices have to be understood to make sure that the workflow integration, the safety, the effectiveness of these applications are fit for their purpose. So that voice needs to be actively engaged, either by having representatives that fully represent that frontline workforce or finding other ways to communicate effectively and in an ongoing way.
- [09:29]
A
Dr. Bonus, I appreciate those points so much. And governance, from my understanding with conversations among other healthcare executives across roles, across health systems, those governance programs are actively being built. And to your point earlier, that can mean everyone's at a different stage of that. And at the same time, organizations are rapidly adopting these technologies. So it sounds like finding that sweet spot of really being diligent about governance, but also innovating at the same time. It's a critical combination. And so I think as we close our discussion today, I wanted to look ahead more to the immediate near term, the next six to 12 months. I think that's a timeframe that perhaps would seem way too short to make any kind of anticipated remarks about from a healthcare lens. But with AI accelerating so quickly and advancing so rapidly, I want to get your perspective on the next six to 12 months. What would you encourage healthcare leaders to really prioritize when it comes to advancing evidence based care? And you know, I know we've already talked about some practical steps, but any others that come to mind to help them prepare that you would advise.
- [10:38]
B
So let me describe this from the lens of some of the challenges in health services which existed way before AI was as common as it is now, which is that care varies tremendously across zip codes, across regions, across a number of factors that no one would really wish has a bearing on the actual care that you receive. And if you're a patient, you are entering a healthcare system if you have access to, and you ideally want to make sure that the system will exit youth from that system with the best possible decisions and the best possible care. So that's the goal, but we can see across systems is there's kind of a bell curve of what people do and in some cases it's average. You want to drive everyone to the right side where the care is really optimal and certainly not in the left side where there may be things which are less than optimal to do that. So if you are a administrator of a health system, what you ideally want to do is make it really easy to do the right thing. And technology and workflow tools are enabling that in ways that were never before possible. So the real challenge is can you find that happy nexus between workflow tools which are safe and effective and make it easy for patients to get the right care and providers to do their job well and to take some of the friction out of the system. And of course, if you are paying attention, you have to realize that no margin, no mission, you have to be able to be solvent. So part of this is also to invest into technologies and workflow processes that keep the lights on. So there is that, that happy intersection between the two. And the challenge is finding that finding the best way to deploy and use the technology like EMR systems that you've already invest to and third party applications and other approaches that again insert into workflow in thoughtful ways that are safe, effective and help every patient get the.
- [12:33]
A
Right care, such important aims. And Dr. Bonas, this has been a great conversation. I want to make sure, before we close, is there anything maybe we glossed over too quickly or any final takeaways or thoughts you wanted to share with listeners?
- [12:48]
B
I think it's interesting to look at kind of where we are not just in healthcare, but as an industry in adopting AI. And I think there's a few more shoes to drop here. It's just not going to be a linear progression of getting better and better and better yet. Hopefully it will. But if we look at it purely for how much money has been spent. So I was interested in, there's been a lot of reports on this. I was interested in one that came out From Bain & Company, the consulting group, and they pointed out that the spend on computer now for the frontier models is around $370 billion per year. And to sort of monetize all of that in a rational way, you'd have to essentially be creating $2 trillion of new AI revenue by 2030. And just to put that into perspective, that would be more than the 2024 combined revenues of Apple and Amazon and Microsoft and Meta and Nvidia combined. So there is a lot being spent on compute, and the applications that are built on those backbones have to be fit for purpose and drive both value and the domain that I operate in, clinical value. But they also have to have meaningful economic value as well. So I think there's going to be a period of time in the coming months and years where we start to understand how that's going to all shake out. And, and I think that is an interesting perspective. But as we conclude here, the point that I want to emphasize again is that if you're doing things within a healthcare domain, stakes are very high. Being fit for purpose and making sure that you're doing everything possible to deliver something that is consistently safe and effective and achieves its purpose is critically important. That's where a real value is created and real human value was created.
- [14:38]
A
Such an important message to end on. Dr. Bonus. I want to thank you for spending time with Beckers today and for sharing all of your insights with us. It's been a pleasure having you.
- [14:46]
B
It's been my pleasure.
- [14:48]
A
And we'd also like to thank our sponsor for today's episode, Wolters Kluwer. Listeners, please be sure to tune into more podcasts from Becker's Healthcare by visiting our podcast page@beckershospitalreview.com.