MOLLY WOOD: Today I’m talking to Peter Lee, President of Microsoft Research, about what business leaders across industries can learn from the way that AI is transforming medicine and life sciences. He delivers a report from the front lines on the technological innovations that are transforming every aspect of medicine, from research to diagnosis to security and privacy, and even the fundamental way that doctors and patients communicate with each other. AI innovations are helping to evolve a healthcare system that is less siloed, less confusing, more thorough, more efficient, more secure, and even more empathetic. And if similar transformations aren’t happening in your industry yet, rest assured, they will be soon. Here’s my conversation with Peter.
[Music]
MOLLY WOOD: Let’s start with your shift three and a half years ago, when Microsoft CEO Satya Nadella asked you to rethink the company’s healthcare strategy. I want to ask you when AI sort of entered and became a major focus of what was already a pretty big strategy shift into healthcare, right?
PETER LEE: Right. Satya first asked me to take a new look at healthcare way back in 2016, and I was actually pretty confused by that. I was wondering, why is he punishing me? [Laughter]
MOLLY WOOD: It’s not considered like a fun field to try to transform.
PETER LEE: It isn’t, but I think Satya really saw the future and was understanding, you know, Microsoft is in literally every single healthcare organization on the planet. Everything from Kaiser Permanente and the UnitedHealth Group, all the way to a one-nurse clinic in Nairobi, Kenya, and everything in between. You know, his point was, the future is going to be a lot about AI and about the cloud and about health data, and are we doing enough there? And so that was the assignment. I joked that it was a little bit like him dropping me and some of my team into the middle of the Pacific Ocean and asking us to find land, because you just don’t know which way to swim. It took a little bit of time to kind of understand, what is it about Microsoft that gives us a right to participate here? What are the differentiated new things that we could offer? And the way that we ask that question is, If Microsoft were to disappear today, in what ways would the world of healthcare be harmed or held back? When ChatGPT was released in November of 2022, three days after the release I got emails from some clinician friends of mine around the world saying, wow, Peter, this is great stuff. And I’m using it in my clinic to do such and such a thing.
MOLLY WOOD: Immediately.
PETER LEE: Immediately. And so that really motivated us to try to study and also educate the world of medicine as quickly as possible, what this new technology is.
MOLLY WOOD: I mean, healthcare is universal. We’ve all interacted in one way or another, and it can be really personal and emotional, but it can also be super bureaucratic and complicated. What’s the potential you see for AI to improve the whole experience?
PETER LEE: Well, I think everyone who has contact with the healthcare system has moments of confusion and frustration. If you live and work in the United States, for example, and you have health insurance from your employer, let’s say, and you get some treatment of some kind, a few weeks later you’ll get something in the mail called an Explanation of Benefits form, an EOB, and that’s totally mysterious. At least for me, you know, I look at those things. I have no idea. Is this a bill? Um, you know, what’s being explained here? You have these weird codes, they’re called CPT codes. You shouldn’t feel bad about not being able to decode those things because I’ve actually interacted with quite a few C-suite executives in major American health insurance companies. And I’ve learned that they can’t even parse these things. And so a simple thing is when you get something like that, or maybe you get lab test results from a physical exam, you can show those things to GPT-4 or to Microsoft Copilot, and just say, take a look at this, explain this to me. So that’s really empowering. Last year, my father passed away after a long illness. And it was a struggle for me and my two sisters to look after his care because we all lived several hundred miles away from my father. And there were moments when the stresses of that would cause the relationships between me and my two sisters to fray. And what I’ve learned over the last few years is that so many people in our world go through this. And so the ability to give all the lab test results, all the notes, to GPT-4, explain the situation and explain that we’re going to have a 15-minute phone conversation with Dr. K, and then just ask the question, What would be the best two or three things to ask? What’s the best use of this time? The ability of that interaction to kind of bring the temperature down and really preserve family harmony and give us a way to feel empowered in interacting with a complex healthcare system is something that was very meaningful.
MOLLY WOOD: First, I’m so sorry to hear about your father.
PETER LEE: Oh, thank you. It was really his time and also, you know, he passed peacefully and with family around, so all of that was great.
MOLLY WOOD: I mean, those situations are so trying for families, and it’s really profound to think about technology helping to make experiences like that a little bit easier. It’s interesting how meaningful an increase in empathy can be in those situations, and you found that introducing AI into medicine actually can introduce more empathy. Was that surprising to you?
PETER LEE: You know, as a techie, I was guilty of thinking, when you think about medicine and healthcare, of immediately zooming in on AI. Diagnosis. So a technologist, traditionally, when they think about healthcare, will think, Oh, can we make an AI system look at radiological images? Can we get an AI system to pass the US medical licensing exam? All those things are good and important, but there’s so much more to healthcare. A big part of healthcare is the relationship between the doctor or nurse and the patient. Just a doctor being able to maintain eye contact and be present with the patient during an encounter instead of typing at a laptop, it matters a whole lot. A doctor being reminded by an AI system, oh, your patient is about to make her very first trip ever to France next month. Maybe it’s good to put an extra line in your email to her to wish her the best. Those extra little human touches. And so there are two things involved in making that possible. One is doing what I call reverse prompting. We always think about the human being prompting the AI system and then the AI system reacting, but the AI system can oftentimes prompt the human. But the other is just giving more time to doctors, to nurses, making them more productive. And so just the aid of an AI system that can say, listen to the doctor-patient conversation and unload most of the time and labor involved in, say, writing the clinical encounter note. These things, they add up and they really matter a lot for that human connection between doctor and patient.
MOLLY WOOD: You said something a bit counterintuitive, in a way, at a conference recently about how that time that’s freed up that should allow doctors and nurses to do the work, you know, not offload the technical work to AI, and that AI can, as you just pointed out, actually be the more empathetic communicator.
PETER LEE: Yeah, I have a colleague, he’s a neuroradiologist, Greg Moore, and he had a friend, a vibrant, very successful friend, and she unfortunately got diagnosed with pancreatic cancer. And using Greg’s connections, he got her into the specialist clinic, at Mayo Clinic, really one of the top places for that particular kind of cancer. And being the go-getter that she was, she was insisting on a cutting-edge immunotherapy. But these specialists, these are the very best people on the planet in treating this type of cancer, were dead certain that that was the wrong approach, that they needed to start with a particular chemotherapy. The patient was insistent in disagreeing, and so there was a conflict that eventually led the specialist to come back to Greg and say, We’re having a problem interacting with this patient, can you talk to her? Greg, not knowing what to say to this absolutely desperate patient, consulted with GPT-4. GPT-4, interestingly, came to the same conclusion as the specialist. And they had this conversation, GPT-4 and Greg, on how to talk to the patient. At the end of that interaction, Greg, in a weirdness about AI today, thanked GPT-4. And GPT-4 said, you’re welcome, Greg, but let me ask, how are you doing? Are you holding up okay? And are you getting all the support that you need?
MOLLY WOOD: Whoa.
PETER LEE: Again, it’s in this idea of reverse prompting that just got Greg to just take a step back and reflect on his own mental state and on his own psyche and ability to cope with the situation of such a close friend in such a desperate situation. That’s very extreme, but there are lots of smaller things as well. The largest manufacturer of electronic health record systems is Epic, and Epic has been rapidly integrating GPT-4 and GPT-3.5 into various applications in their EHR system. And they’ve been then working with academic medical centers to do controlled studies to see if it works well, if it’s not making lots of mistakes, patient satisfaction, doctor satisfaction, and so on.
One of the things that they’re finding is that when GPT-4 writes the after-visit summary email to a patient, the patients are consistently rating those notes as more human than the notes written by the doctors themselves.
MOLLY WOOD: Wow.
PETER LEE: And of course, it’s not the case that they’re more human. They’re written by a machine. But when you’re a busy doctor, you might not just take the time to, say, congratulate your patient on becoming a grandparent. Those extra little touches, it just shows that somebody remembers and cares. It can just make so much of a difference in the connection between doctor and patient.
MOLLY WOOD: I mean, that’s fascinating and kind of heartbreaking that AI clearly learned from the data it was trained on that empathy is a key part of medicine, but our medical professionals are so overtaxed that they can’t take the time to do it. I also love this kind of reverse prompt idea, like AI as an assistant taking some of the load off so medical professionals can get back to fundamentals, which are about care.
PETER LEE: Well, it’s such an important point because right now there is this crisis in the US, but there have been numerous studies that show over 40 percent of a clinician’s day, on average, is spent on clerical work, documentation, and note-taking. I really love my primary care physician, but every time I see her, her back is turned to me. She’s sitting there at a computer, typing while she’s talking to me. And the reason she’s doing that is she has a life. What I mean by that is if she didn’t take the time to write those notes during the encounter with me, she’d have to take that work home with her. That’s called, in the profession, pajama time. Some doctors don’t want to do that while they’re with their patients and they take that work home and jump in bed with a laptop and spend two hours doing that documentation and clerical work. And so what if AI could reduce that by half or by 80 percent? So much more would be possible.
MOLLY WOOD: You speak about this topic globally, and I’m curious about how your findings apply to doctors and nurses across the world. Is it just in the US that we have, you know, burnout and clerical loads that are untenable? How do you find that this technology is translating to doctors in other parts of the world?
PETER LEE: It is a global issue. However, it is worth emphasizing just how extreme the problem is in the United States. Over the next five years, there is projected to be several-hundred-thousand-nurse shortage in the US healthcare system. And then if you go to the UK, the National Health Service, it is not unusual outside of London to have a multi-month wait if you need to see someone for primary care. There are huge parts of Africa where people still might live an entire lifetime never seeing a doctor. And then in China, the caseloads on primary care physicians in China is now approaching 80 patients per day.
MOLLY WOOD: Whoa.
PETER LEE: For a single primary care physician. And the kind of burnout and, in some cases, violence fueled by just frustration that people have. It really makes headline news in that country. We also have something called the “silver tsunami” that is coming. There are demographic changes where the aging population is reaching a point where there will not be enough young healthcare workers to care for an aging population. And so all of these things are about to really become extreme issues. And all of that leads to fewer and fewer bright young people wanting to enter into the profession. Now, the US healthcare system is reacting—for example, there’s a whole slew of new medical schools that have sprung up. In fact, I’m on the board of directors of a new medical school, Kaiser Permanente School of Medicine. But that’s just one of a dozen new medical schools that have sprung up in the US just in the past three years, in an attempt to produce more doctors and nurses. The fundamental root cause is, can we make being a doctor, being a nurse, the kind of satisfying profession that allows people to connect with their personal desires to help people as opposed to do paperwork? Can we create that situation that will motivate people? And that is the most important problem for us as technologists to work on. Yes, it’ll be great for us to solve genomics with AI, to solve cancer with AI, to have better radiological imaging techniques with AI. All of that is great. But at the end of the day, if the one thing that we can accomplish is to have AI make a dent in this kind of workforce shortage and then day-to-day worker satisfaction in healthcare, we’ll have really done the world a great service.
MOLLY WOOD: Healthcare is obviously such a unique industry and it presents its own set of challenges. But you can imagine that these are also lessons that extend into other industries. I wonder, in your learnings, what is your message about the way that leaders across industries should implement AI in this way to bring more time and potentially more empathy?
PETER LEE: This is going to sound funny, but the way I explain it is that generative AI, that a large language model, is not a computer. You could substitute any type of information worker for this, but let’s imagine you’re a nurse. Your mental model of a computer is a computer is a machine that does perfect calculation and has perfect memory recall. So, if you ask a computer to come up—let’s say you do a web search, it’s going to come up with precise answers. If you ask a computer to do some calculations, it’s going to come up with a precise answer. The thing that’s odd about a large language model is it’s similar to the human brain in being very faulty with memory and very faulty with calculation. And so, it will make mistakes. If you ask it to do a big pile of arithmetic, it’ll get it wrong in ways very similar to the way a human being would get it wrong. The thing that is so important for people to realize is that this is now a new type of machine, a new type of tool, that doesn’t have that perfect calculation or perfect memory capability. There’s a professor at the Wharton School at University of Pennsylvania, Ethan Mollick, who really puts it nicely. He says it’s better to think of a large language model as an eager and tireless intern, and so if you are a doctor, it can be dangerous to use the large language model as though it’s a computer. It’s much better to treat it like an intern. And the answers you get from it, you have to assess and you have to think about in the same way as you would from your intern. And it’s high stakes, particularly in the world of medicine. If you don’t understand this, you can end up hurting someone. And so, as I’ve gone around to healthcare organizations around the world over the past year, I always start with that lesson.
MOLLY WOOD: Yeah, that is a very different mindset. And actually seems like an important one for using these tools in any industry. So what’s your general advice to leaders for how to use AI in a way that really taps into those strengths?
PETER LEE: The way to start, of course, is to be very hands-on with these systems. And the easiest way for a human being to be hands-on is to do it through a chat interface. And you can just talk to it. There’s another stage where, if you have a whole bunch of data, you can ask the system, Can you figure out how best to structure this data and prepare it for analysis and machine learning? That’s another thing that’s rising in tremendous importance. A great project in Microsoft Research involves clinical trials matching. So, right now, when there are potential new therapies and new drugs, new diagnostic techniques that are proposed by medical researchers, they have to go through a validation process. Part of the validation process involves standing up what’s called a clinical trial to kind of test under circumstances, whether let’s say some new therapy is both safe and works well. A sad thing is that over half of clinical trials that are stood up fail to recruit enough participants. And this holds back the advancement of medical science by huge amounts. It’s really a sad thing. And part of the problem is that when you look at clinical trials documents, they’re incredibly complicated things to read. And they’re highly unstructured text documents. What we’re learning is that a large language model like GPT-4 can read all those clinical trials documents and put them in a structured database that allows tools to better match up patients with those trials. It just opens up the possibilities that we’ll be able to accelerate the advancement of medical science by doing that. And so each one of these stages, you know, where you just start with the raw large language model, then you give the large language model access to tools, and then you use the large language model to make sense of all that data out in the world. Those three stages, I think, is a natural progression.
MOLLY WOOD: And again, we should say those stages are applicable to almost any industry. It’s really sort of that mindset of thinking about it and sort of understanding what you should adopt for and what you shouldn’t.
PETER LEE: Oh, yeah, absolutely. I mean, transportation, retail, manufacturing, law, finance, you name it. These same ideas apply across the board.
MOLLY WOOD: When you hear reluctance to engage with some of these tools, what’s your sort of go-to response?
PETER LEE: I just try to show empathy. You know, when folks first showed what we now called GPT-4 to me and explained to me what it could do, I was super skeptical. Like, give me a break. And then I passed from skepticism to annoyance because I saw some of my Microsoft Research colleagues getting what I felt was duped by this stuff. And then I got sort of upset because it became clear that my boss, Kevin Scott, and his boss, Satya Nadella, were going to make a big bet on this technology. So I thought, what? This is crazy. And then, with my own personal investigations, I got into the phase of amazement. Because it was true. These things that OpenAI was claiming about this thing were actually true. They were happening. That led to a period of intensity where you try to figure out, okay, so what is this going to mean? How can we use it? Then you get into a period of concern because you start to encounter problems like hallucination, issues with bias, transparency, and so on. And then you realize this is a real technology that’s going to change everything. And so I share my own journey because I’ve seen so many other people go through the same journey. And I’ve seen whole organizations and businesses step through these things. And so what I tell people is, you need to have patience. Everyone needs to go through this. And you need to understand this is a process that people have to go through because it’s just very challenging to believe that this technology can even exist.
MOLLY WOOD: And then finally, in the medical field in particular, is there something, is there a moonshot that you think you really want this technology to take on?
PETER LEE: You know, when I think about what is the most important thing to accomplish, there is a concept in medicine called real-world evidence, RWE. The dream there is, what if every healthcare experience that every patient has could feed directly into the advancement of medical knowledge and science. And so here’s my favorite example from the pandemic. In the first year of the pandemic, some doctors around the world were randomly discovering that if they had a very sick COVID patient in respiratory distress that they could sometimes avoid having to intubate that patient by having the patient stay prone for 12 hours, stay on their stomachs for 12 hours, and they would start to share that knowledge actually on social media. And so other doctors started to do the same thing, but it was very random and ad hoc. A few months later, a network of medical research institutions around the world banded together and formed a clinical trial, a clinical study, to study this. And a year and a half later, they determined that, yes, for some patients in severe respiratory distress that this worked. That year-and-a-half gap is something that, first off, leads to thousands of patients being intubated when maybe they didn’t need to be and some of those patients dying needlessly. What if we had systems that could observe every single experience in every single medical encounter that patients had? And that feeds in directly into the storehouse of medical knowledge. That’s the dream of real-world evidence. And when I see what AI is becoming today, I cannot escape the feeling that some aspects of that dream of RWE are actually within our grasp. And that’s where I’d like to see the world lead to.
MOLLY WOOD: Peter Lee is President of Microsoft Research. Thank you so much for the time. This is phenomenal.
PETER LEE: Thank you, Molly. It’s been great to chat.
MOLLY WOOD: If you’ve got a question or a comment, please drop us an email at worklab@microsoft.com. And check out Microsoft’s Work Trend Indexes and the WorkLab digital publication, where you’ll find all of our episodes along with thoughtful stories that explore how business leaders are thriving in today’s new world of work. You can find all of it at microsoft.com/WorkLab. As for this podcast, please rate us, review us, and follow us wherever you listen. It helps us out a ton. The WorkLab podcast is a place for experts to share their insights and opinions. As students of the future of work, Microsoft values inputs from a diverse set of voices. That said, the opinions and findings of our guests are their own, and they may not necessarily reflect Microsoft’s own research or positions. WorkLab is produced by Microsoft with Godfrey Dadich Partners and Reasonable Volume. I’m your host, Molly Wood. Sharon Kallander and Matthew Duncan produced this podcast. Jessica Voelker is the WorkLab editor.