[MUSIC PLAYING]
Welcome, I am Herve Balzano, President of Health and Benefits at Mercer and Global Leader at Mercer Marsh Benefits. We as Marsh McLennan are delighted to be partnering with the World Economic Forum in establishing the Digital Health Action Alliance.
The alliance aims to bring together a diverse group of stakeholders across the value chain from tech and finance to healthcare, pharma, employers, and insurers as well as important actors in the public and social sectors. Together, we are exploring how our digital capabilities can improving health outcomes. And in support of these efforts, we are very excited to be launching this video series on advancing health access and affordability through technology.
So for today, I am thrilled and honored to be joined by Dr. Bill Weeks, Director of Microsoft's philanthropic AI for health program to share his view on the impact of AI on health and care. Bill is a well known physician and economist. And alongside leading Microsoft's AI for Health team, he also conducts research with non-profit organizations and academic centers on how to improve health and health outcomes. He also serves as the medical director for Bing, Microsoft's search engine. Bill, thank you very much for joining me today.
It's great to be here. Thanks for having me.
[MUSIC PLAYING]
So to get us started, AI and generative AI is all the talk across the media today. But for you, I'm sure this is something you've lived in breast for a long time. Algorithms and electronic medical records have existed in healthcare for a while. But can you explain briefly why people like Microsoft's founder and health advocate, Bill Gates, are calling this new technology revolutionary?
Sure. I think particularly the large language models, the generative AI is revolutionary because it kind of transforms how we interact with health data. Historically, as you point out, we've had electronic medical records and providers, like me and other physicians and nurse practitioners, would review records for a patient and understand-- kind of glean from that review a summary of the patient.
And sometimes, when you got a new patient, they maybe had a chart this big or an electronic medical record that was hundreds and hundreds of pages long. And it would be very hard to go through it in your 15-minute allotment to see a patient. So what generative AI can do is very good at summarizing information and kind of going through those hundreds, if not thousands, of pages to provide a coherent, brief summary that's accurate that captured what the patient has experienced.
And, importantly, I think can look at some differences in what has been reported over time. A lot of times, what's happened in electronic medical records is a copy-paste function has been used quite a bit. And the challenge there is if an error is put in, if someone says that I'm allergic to penicillin and I'm really not, it can perpetuate indefinitely.
So generative AI summary can say, in this one note, it said that Bill is allergic to penicillin. But in the others, it doesn't. So reconcile that. Get that right so it doesn't kind of perpetuate and we can kind of focus on the problem, focus on accuracy, and make sure that we understand the patient pretty clearly and pretty quickly so that we can kind of spend more time treating them and interacting with the patient.
[MUSIC PLAYING]
Well, thank you, Bill. This is very insightful. And by the way, it sounds like you and the other Bill have very similar views on the potential for AI in health. It's quite revolutionary. But given your experience and role, what do you think is the biggest opportunity or impact that you are expecting to see from generative AI, especially in health?
I think the biggest impact that I can anticipate is-- there are kind of three. One is improved care quality. One is improved care access. And one is reduced care costs.
And I talk here not just for kind of the developed world but also for the developing world. So from a care quality perspective, it's a challenge. I went to med school 40 years ago now.
And so it's tough to keep up. There's a lot of data that comes out every single week in journal articles and new knowledge that's generated that might be able to impact the patient that's sitting right in front of me. So it's very, very challenging to keep up.
And the reality is that there's quite a difference in the quality of care that one provider might give compared to even the one in the next room over. So what I see generative AI being able to do and AI models in general is to really close that quality gap, improve care quality for kind of everyone who comes in by providing easy access to the best information that's possible at the moment and not kind of relying on historical training programs that have led, according to the IOM, to 17-year gaps between when knowledge is kind of known and when it's actually implemented at the bedside.
So that gap in knowledge transfer will close. And the gap in quality across providers will close. And I think that will improve patient safety, improve patient outcomes and things like that.
[MUSIC PLAYING]
We are also excited to see our generative AI might facilitate the shift to safe care as well. In our recent health sentiment survey of 17,000 employees across 16 markets, we've seen a big shift in employees wanting to be more involved in self-care. And they are very receptive to digital solutions, which increasingly use AI. I want to get back to chronic diseases or NCDs, such as heart failure, cancer, or diabetes. Are you already seeing AI being used to impact and advance patient care and particularly in chronic disease?
Chronic care is more around understanding the relationship between social determinants of health and these chronic health care conditions. As you may, know there's a long literature that indicates that actually the health care system impacts 20 or so percent of health, whereas the other 80% is impacted by socioeconomic status, social determinants of health, and genetics essentially.
And so the health care system-- historically, we've looked to the health care system to address health. But it really can only address a fraction of the health. So by better understanding the social determinants of health using AI models and predictive models, we might be able to impact population health better, particularly for these chronic conditions, and reduce the frequency and the prevalence of those chronic conditions, which would improve health care costs and reduce pressures on the health care system and things like that. So that's one area kind of in the public health arena.
[MUSIC PLAYING]
I think the whole-- when we look at health care and we think about health care and what the goals of health care are is really to achieve the so-called quadruple aim, which is improvement of patient satisfaction, provider satisfaction, and then care quality and outcomes all over health care costs while reducing those costs. So we want to do those things, improve those numerators, those three numerators, and reduce health care costs all while also kind of addressing equity at the same time.
And I think AI in particular can focus on those particular areas, from the obvious of kind of increased care management for people with congestive heart failure or coronary artery disease or something like that, diabetes, to actually the more mundane things, like just helping providers write notes better and be more responsive to email requests and be more responsive to consultations by just allowing them to dictate something very quickly that can be immediately put into the system and immediately addressed without taking a whole bunch of their time.
So there's been a number of studies that have actually indicated that by using generative AI, you can reduce nurses' workloads. You can reduce physician workloads. And so that kind of decreases the per capita cost because they can spend that time that they were spending typing and fooling with scheduling and things like that. They can see patients, which is then more efficient and more effective, I think.
[MUSIC PLAYING]
Totally agree. And this is a remarkable example of the power of AI. So thank you for sharing it. So, Bill, you just shared one of the great examples of advances. But what do you see as the key barriers to adoption? What is likely to stop or delay success? First, perhaps, of course, you could focus on your top three barriers.
Your acceptance by providers of the technology. To be honest, providers feel, I think, a little burned. And they are burned out from electronic medical records, which have consumed a lot of time and haven't necessarily delivered on the promise of kind of increasing their effectiveness and their efficiency in providing care.
So I think they're a little reticent to immediately embrace AI and GPT. And so that's obviously a barrier. I think ways to kind of address that particular barrier are to make it very clear that GPT and these AI models are not there to replace providers. But they're there to support them so that they can improve care quality and improve care outcomes.
And I think when many physicians understand and use these models, they see that it's actually quite helpful to them. It helps them use their time more efficiently. It helps them focus on the area that they need to focus on.
[MUSIC PLAYING]
There have been reports of fabrications or hallucinations that have come from some of these models that people will very rapidly write about. And I think the challenge there for me is to think about what's a fair comparison because I think sometimes we expect perfection from an AI model.
And the models kind of are-- it's kind of garbage in, garbage out. So whatever you put in is what you're going to get out of the model. For a variety of reasons, medical errors have been accounting for about 100,000 deaths in the United States a year. There was just a recent report that found that misdiagnosis was responsible for 750,000 either deaths or disabilities or problems of care per year.
And so I think that we need to be fair in our assessments. I just read a report that said that radiologists would expect a 95% accuracy rate before they accepted an AI model. And I'm not sure that that's quite a fair rate.
I think the fair rate to be compared to would be, what would your average radiologist-- how would your average radiologist perform? And if your average radiologist performs at 95%, then great. I seriously-- having done a lot of work in quality and safety over the years, I doubt that's the case.
I think that as long as we're improving on kind of average provider performance, then that should be kind of an acceptable thing. We can't expect perfection from a model. It's unrealistic. And I think it's important to recognize that providers are not perfect either.
And then finally, there's the issue of kind of responsible AI and bias. And Microsoft has a whole series of responsible AI reviews that they do with any of these models, as do other companies. So that's really critical to understand that these models might be biased, that they're based, again, on what you put in them. And so they may not be overly-- they might not be generalizable to an entire population.
The harsh reality, though, there too is that neither is medicine as it's currently practiced. Medicine as currently practice is fairly biased. And there are many, many studies that show inequities and disparities across racial groups and across gender groups and across orientation groups.
And so I think we can use AI to actually reduce those biases. And, actually, you can run multiple AI models, one on top of another, that can look for biases in the process of care that's being delivered.
Well, thank you again. If we have time maybe one last question, which I like to ask all of our guests, are you optimistic about the role technology can play in the future of health? And why or why not?
Entirely optimistic about that. I was on a panel not too long ago. And people said they were cautiously optimistic. I said I'm totally optimistic about this.
I've been in health care and studying it for 40 years or so. And what we've not seen in health care that we have seen in other industries is the role of technology in kind of improving care efficiency and care delivery efficiency. We just haven't seen the efficiency outputs.
And, usually, in other industries, when you have technology inputs, you tend to have a reduction in labor or improvement in the efficiency of labor. And like I said, we just haven't had that at all in health care-- not very much in health care. So I see this as really being transformative in improving that efficiency, which it goes back to the first point I was making around that will improve access.
That will improve the patient experience. That will improve the provider experience. That will improve care outcomes and care quality because it'll be a very high, consistent quality of care that people are getting and not variable. And all of that together will help reduce health care costs.
So I'm entirely optimistic about this technology like I just haven't been in 40 years. We've been dreaming about these kind of technologies being available and the ability to integrate all the types of different data that come in, from social to medical to individual kind of heart rate monitors and things like that. And I think now it's accessible. It's doable. We have the compute power to do it. We have the models to help provide individualized, personalized care for every single person that can be, again, facilitated by health care systems and providers to co-create health for the general population.
