Learning from Your Own Health Data with Sara Riggare and Gary Wolf

Transcript

Harlan Krumholz: Welcome to Health & Veritas. I’m Harlan Krumholz.

Howard Forman: And I’m Howie Forman. We’re physicians and professors at Yale University, and we’re trying to get closer to the truth about health and healthcare.

Harlan Krumholz: Today, we’re having a special guest, Sara Riggare, with Gary Wolf talking about the new book, The Quantified Self. But first, Howie, what’s on your mind this week?

Howard Forman: Yeah. So, this is interesting, and it’s in my field, at least for rarely this happens. This past week, the Stockholm region of Sweden announced that it will use an AI system called Lunit Insight MMG as one of the two independent readers of screening mammograms. It covers three hospitals and roughly 200,000 to 250,000 exams a year. It’s a large number. Sweden, like most of Europe, has two radiologists independently read every screening mammogram. Stockholm is replacing one of those two readers with software. Anything that gets flagged either by the AI or by the radiologist still goes to a human consensus discussion before a woman is called back.

So, this is not a machine making the call, it’s a machine casting one of the votes. What strikes me as most newsworthy is how cautious this path has been. It started with a prospective study of about 55,000 women at a single Stockholm hospital published in 2023. One radiologist plus the AI found 261 cancers. Two radiologists found 250. That hospital then used the system routinely for three years, about 200,000 exams before the region expanded it. A Swedish randomized trial of a different AI product found 44% less reading workload with no increase in cancers missed between screenings.

So, what did we learn? First, productivity. In Sweden, this frees up a lot of radiologists’ time, and one hospital reports that the wait for a diagnostic mammogram essentially disappeared. Second, there is no evidence that it performs worse than the human standard that it replaces. Detection went up, recalls did not. But third, and this is the part I want to underline, every one of these results is a surrogate. Cancers detected; recall rates; reading time. We do not yet know that women live longer or even that fewer of them present with advanced disease. There’s a real possibility we’re finding more small, slow-growing cancers that never would have caused harm. And as far as I can tell, the Stockholm study’s planned two-year analysis of cancers missed between screenings has not yet been published.

For our listeners in the United States, this changes nothing here. We don’t double read. Most American mammograms are read by one radiologist, often with some form of AI running alongside. There is no second reader to replace, so the Swedish efficiency story doesn’t carry over. What it may do is give us a template to run real trials here, randomized practices or women, and follow them long enough to measure interval cancers, stage diagnosis, and eventually, mortality. Sweden has shown this can be introduced carefully, in stages. We should take that as an invitation to generate outcome evidence, not as an excuse to skip it.

Harlan Krumholz: Yeah, I know. Interesting story, Howie. I mean, I’ve long thought that we should have two readers. I’ve told you a situation about my daughter at Yale New Haven Hospital, who was experiencing symptoms when she was a child and was given a diagnosis by our pathologist that didn’t make any sense to me, that would have consigned her to lifelong treatment and a diagnosis that was unfortunate at best. And I kept going back to the pathology suite and asking them to reread it, and they rebuffed me several times as a hysterical parent. And finally, I got someone else to read it, and they countermanded the diagnosis and said it was an error. And actually, she didn’t have that diagnosis at all, and they had just made a mistake. And I wondered if we should adopt a policy, particularly on diagnoses with important consequences. I don’t know how many diagnoses don’t have important consequences, but it seemed to me that we should be taking that on as a second reader.

I think eventually the department did do that. This is many years ago now. My daughter’s in her mid-thirties. She was in her early teens at that point, but I felt for a long time these consequence diagnoses, we should be working out how we can make errors almost impossible, and AI now may make it possible for us not to defer to it but to give us another opinion and why just two, maybe ten? There should be an opportunity for us to be able to have this kind of support when these kind of issues are raised.

I’m also chasing by the fact that our former colleague Joann Elmore had done a series of studies. She had done some on mammograms with Alvan Feinstein, a very famous clinical epidemiologist at Yale, a former mentor to so many and a colleague of mine. But she also did one in pathology, where she brought expert pathologists on a multihead microscope looking at the same tissue. And often they came up with different conclusions about melanoma.

And it just shows me that sometimes there’s not even agreement, and we rarely are willing to talk about that kind of uncertainty, let alone the disagreement that occurs. So, I think this is an exciting moment. I’m looking forward to seeing how this plays out, but I don’t think in medicine we should be afraid of these other opinions, but the question is how best to deploy them in order to get the best outcomes.

Howard Forman: It’s also worth pointing out that our peer nations who have nationalized healthcare services seem to me to be more inclined to do these types of randomized studies, these types of clinical trials that need to get done. In our country, so much is motivated by the profit motive that it doesn’t seem to get done, and there are real benefits from us doing more here. So, I hope we can move in this direction.

Harlan Krumholz: Absolutely, absolutely. All right. Let’s pivot to our interview. We’ve got two great folks coming, a book coming out and I think people are going to enjoy this interview. Most of us have tried to change something about our health, our sleep, our exercise, some part of our daily routine. And then we’ve wondered, did that really help? And how would I know how to answer those questions?

Gary Wolf is a journalist, a former contributing editor at Wired, and the co-founder of a group called the Quantified Self. His new book, The Quantified Self: Learning to Observe, explores how we use scientific methods to learn about our own lives. We’re inundated with data these days, but how do we use it to answer questions? How do we use it to improve our health? This book comes out October 6th. I’ve read it. It’s great.

Today, so we’re joined by Gary, Gary Wolf and my good friend Sara Riggare, a patient researcher at Uppsala University in Sweden, and just someone who has inspired me, whose work, both coping with her own condition, but also advancing our thinking about how we can take control of our lives, how we can own our own destiny, how we can understand our illness, how we can take responsibility for improving our own health. She’s lived with Parkinson’s disease for decades, and her research brings together scientific methods and really the knowledge that comes with living with a chronic condition but also the scientist within her who’s continuing to push these frontiers.

What interests me most about this work is, it’s not just about measuring more. And like I said, we’re being inundated with data these days. It’s about deciding what matters, how we ask better questions, and recognizing both what our observations can teach us and what they can’t. Gary, Sara, welcome to Health & Veritas. It’s great to have you on.

Gary Wolf: Thank you. It’s great to be on.

Sara Riggare: Thank you. Happy to be here.

Harlan Krumholz: Good. So, let me just start off. I’m going to toss a question to you, Gary, and then I’ll have one for Sara. And then Howie and I just tend to go back and forth a little bit. So, Gary, you came to this as a journalist. Was there a particular person or discovery that made you think there’s something important here? Writing a book’s a big endeavor, so you’ve got to invest a big part of your life in it. So, you’re thinking, this isn’t just about collecting numbers, but we can get people thinking as a different way of learning about themselves. So, how did you get to the point where you wanted to spend the time to do this?

Gary Wolf: Well, really it was a three-step process. It began, as most of my projects have, as a journalistic exploration, some research, and as part of the research, my colleague Kevin Kelly and I convened a meeting and we just invited people who were using empirical methods, self-tracking. This was early days, first consumer GPS devices that got location.

Harlan Krumholz: What year was this?

Gary Wolf: This was 2008, so 18 years ago. So, we got to know some people who were doing some advanced things with what we called “very personal computing.” And we convened a meeting and very informal, and somebody came in late and Kevin said, “You came in late. You introduce yourself first.” And instead of doing a regular introduction, he just opened his computer and showed this detailed time diary of everything he’d done in the last year in 15-minute increments and had it in a graphical form, and it was amazing to look at.

And then we said, “Okay, anybody else want to show something?” And people just showed their work. And spontaneously, this format emerged, which we called the show and tell. And we just started having meetings where people brought their self-research work and shared it and answered three questions. What did I do? How did I do it? And what did I learn?

And the second stage was that, just spread all over the world. I mean, at one point there were something like 100-plus meetings going on in 30 countries, and I became the archivist and maintainer of the community resources. So, we have thousands of talks that people gave.

And then the third step was really connecting with Sara and some other researchers, some academic and some non-academic, to ask the question, “Okay, now that we’ve seen so many projects, have we learned anything about what makes it possible for people to really get meaningful results from their self-tracking, since it’s not always easy?” And that led eventually to the book The Quantified Self: Learning to Observe, which the title itself gives the essence of what we discovered together, which is that it’s the observational process and really the details of how you set up that process, that’s the missing element in self-tracking and that ends up really delivering the discoveries.

Harlan Krumholz: The book is built for consumers, but I really felt like research scientists should read it, because it really gets you thinking about how we should be observing. Sara, I want to turn to you. You and I met years ago, and I thought of you as like an optimizer. You were someone who was a keen observer of your own experience and almost taking N-of-1 experiments to the extreme by saying, “I’m going to be observing my own body, my own capabilities, my own experience and try different things and be able to be doing my own tests, which was going to be very different than the experience I was going to have when I go to a doctor episodically, even just annually for a checkup where they really don’t have any view about what my real life is like or what’s working for me.” I wonder if you just tell your story just a little bit. Give us a sense of yourself.

Sara Riggare: Yeah. So, I had my first symptoms of what much later would turn out to be Parkinson’s disease when I was 13 years old, which is unusual and very rare but not entirely unique. I know others who have been diagnosed at that age, but it’s unusual for sure to get it that early. But I didn’t know that until I was 32 and was diagnosed with Parkinson’s disease.

But when I had been diagnosed, it was a pretty dark time. I had a little baby at home that I was taking care of, and it felt like I fell down into a black hole. But thanks to the baby, I had to get myself out of that black hole, of course, and I had to start doing something, start building my life back together. And I did. And then a few years later, I decided that I wanted to combine my patient experiences with my engineering skills and try to improve things for myself and others with chronic diseases.

And I then decided that I wanted to become a babel fish, if that reference talks to you from Hitchhiker’s Guide to the Galaxy. I wanted to become a babel fish between patients and the researchers. And I thought initially in Parkinson’s disease mainly. And so, I realized I had to learn more about medicine. And I realized also that you had to be taken seriously by researchers, you had to be a researcher yourself.

And then I was sad because I told my husband, “Well, and I’m not.” But he said, “But no, you’re not now, but maybe you can become one.” I was like, “No, I can’t do that.” But then I thought, “Maybe I can.” So, I then started studying at the master’s program in health informatics at Karolinska Institutet in Stockholm with the aim of doing a PhD. And so, 2022, I actually did defend my PhD thesis and became a doctor, and the way forward is still open.

Howard Forman: I want to follow that up because it’s part of the story you talk about—

Harlan Krumholz: It’s an incredible story, right, Howie? I mean, this is just incredible, right?

Howard Forman: But I think a lot of the detail that comes out in the book comes out really in your story that you tell in the book is something I think our listeners would be fascinated with. And two things that I would love you to just comment on. One is this, the finger tapping as a metric for how you’re managing your disease and how you use finger tapping as a proxy for how much your medication is working for you and when it’s fading off. And then secondly, you have this great quote about how one hour with the neurologist versus I think 8,765 hours. I may have it wrong.

Sara Riggare: Correct. No, that’s correct. Yeah.

Howard Forman: Yeah, 8,765 hours alone. How do you manage that so that you maximize the time with the neurologist but also use all the information you gather?

Sara Riggare: Yeah. So, that image with one red dot and 8,765 blue dots is my second biggest masterpiece, after the 23-year-old. And it’s something that I’ve become well known for all over the world actually. That image has been at conferences in countries where I haven’t even been. So, many people use it as an illustration of not only Parkinson’s disease but chronic disease care overall. We spend very little time with our clinicians, even in very, very complex and complicated diseases.

So, the one hour I spend with my neurologist is then the red dot, and it’s only during that time that he can observe my symptoms and we can discuss my treatment directly. So, it’s only during that one hour that I’m in direct contact with neurological specialty healthcare with its clinical guidelines and practices. But it’s during all those blue hours, all the 8,765 blue dots, that I am putting that treatment into action and taking care of myself as best I can.

Howard Forman: Can you speak to the measurement, just so people understand what I’m even talking about because I think it’s useful.

Sara Riggare: Sure. So, I wanted to understand how my medication effects varied across the day. And I had come across a test that was used in the clinic or they used finger tapping as a measure of Parkinson’s progression and all sorts of things where you tap your fingers together as fast as you can and then they observe and see how it changes over time. But they had used a test with a touch screen, but I thought maybe I can use that in my everyday life as well.

So, I started experimenting with my first smartphone and started tapping them for 30 seconds at a time with my right hand and left hand a number of times every day and then made the graph and saw how it varied. And then I also put my medication timings in the same graph. And I found things I couldn’t have realized without having done that because I found that I had a dip in function around lunch. So, I moved the medication timing 30 minutes, and I had a more even effect across the day.

Howard Forman: That’s fascinating. And that’s the type of learning that I think is so... what we all hope that we could get with whatever illnesses we’re facing. And so, I turn to you, Gary, because it’s maybe a similar question in a way. And that is the ultimate goal of all of this, that more people will be able to do these things and gain from them individually? Or is the goal that individuals, particularly people like Sara and others who are really dedicated to this, can help us learn more about diseases that then can be exported to a much larger population? What do you imagine? You’ve written this book. It’s going to come out next week. We’re very excited—next Tuesday. What do you hope is the outcome of this?

Gary Wolf: Well, that’s a great question. Sara and I just talked about this this morning because I think we agree that the uncharted territory, the most benefit that people can get right now is in the individual help they get coping in day-to-day life. There’s a lot of techniques of thinking, reasoning, discovering things like Sara has discovered and the stories in the book and really thousands of other examples that even just 10 years ago would have been a lot harder to learn.

And now they’re easier to learn, not just because we can make observations using sensors but even just keeping the records of the observations and displaying those records in a time series graph or even just in a table where you order them from the most to the least and see the dates where things were especially bad. All of these maneuvers which are at the bottom of scientific practice—I mean in a good way at the bottom, foundational to scientific practice—they’re more accessible to people. And so, I think Sara and I agree that skill development, sharing methods, telling stories, this really could help people a lot.

There is another part, which is the part you also mentioned and you used the word “outcome,” and we were just talking about this because outcome is really a healthcare word. It’s not really a personal learning word. And outcomes are also important. And I think we share an intuition that outcomes in that healthcare sense could also be improved.

But there you get into some really tricky areas. I mean, if we had hours, I would ask both of you to help us think about outcomes because we want those outcomes, those healthcare outcomes for one reason mainly, which is that we want to make a claim on social resources. We want to say, “Hey, you should help people more learning from their own data and support this with your clinicians and your allied health.”

But in order to make those claims, you need to show evidence that there’s going to be generalizable and systematic benefit. I think we’re really far from that still. And a phrase we use is the epistemic gap. There’s a gap in the methods. They look the same. There’s graphs and there’s numbers, but really how you think about them is very different, and that’s also uncharted territory.

Harlan Krumholz: Sara, you were really a pioneer in the era of patients taking charge of their own health. I mean, there’s always been some of it, but the way that you did it was really distinctive and very science-based. And honestly, you got this PhD, but you were always acting as a research scientist. I mean, your mind was always thinking in this way. And yet I know that when you encountered the healthcare system, it wasn’t always encouraging or welcoming for this kind of thing.

And there has been some evolution in many countries. I think in the U.S., you’re seeing more and more, especially with people’s access to AI and everything, but I still think many people lack the skills that you acquired that enabled you to systematically evaluate how you were doing and to be able to come up with inferences that had some validity to them as opposed to a hunch that you had.

And by the way, there’s nothing wrong with hunches, but the question is, do you have an opportunity to test the hunch? And what are you thinking about what it’s going to take so that people just don’t think like, “Well, gosh, I’ve got so many things, and I’m feeling overwhelmed, and I just can’t do what Sara Riggare did. I mean, that’s amazing what she did, but that’s superhuman, and it’s beyond my ability.”

Part of this book, I felt, is a way to try to hold people’s hands and try to show them they can do this. But what are your thoughts now about what’s possible for people, and what would a healthcare system look like if we really did encourage patients to be more active in this way?

Sara Riggare: Yeah. So, you call it “research scientist.” I call it having an engineering mindset. But I also am fortunate enough to really, really enjoy learning things. I think that’s to be curious, to want to understand and learn more and try and test and evaluate. And that’s something that can be learned. I mean, it’s not something I was born with. It was something I probably found very early on that I’ve learned along the way.

And I often hear this, “But Sara, not everybody can do everything you do.” And that’s of course true, but also, I wasn’t this active 10, 15 years ago. I’ve learned along the way. So, knowledge is not a constant. So, I mean, with the right support, everybody can do more than they do today. And I want to talk about as well that I did this tapping test a long, long time ago now. I think it was 2012, even. And I rarely use it today because by doing that then I’ve learned to recognize things in my body and feel things more.

And so, the burden of tracking, even though it can sound like 30 seconds with each finger for each hand for a few times a day isn’t that much of a time demand, but it actually is a burden to do it. So, I don’t do that these days. I use other things to learn and to evaluate. So, it’s about finding your own methods. And as Gary said before, it’s about, in personal science, what’s generalizable is the methods more than the results and the outcomes of the research or the study experiment. So, we talk about, we don’t share data. We share methods in personal science.

Harlan Krumholz: One thing I wanted to ask you, Gary, was one of the challenges I think in taking this on, like I said, the data’s everywhere. Now you started before wearables have become ubiquitous, and there’s lots of different sources of information. It’s interesting about the tapping test, by the way, because now in Parkinson’s, people use that on apps to help track symptoms and progress or lack of progress in Parkinson’s. Again, Sara was a pioneer in helping to figure out how to do this.

But Gary, one of the things, it’s hard to hold one thing constant. So, if somebody wants to look at one change, usually one change though is still surrounded by a lot of other changes in your life. And it’s hard for you to say, “This week, my stress level, my challenges, everything else was, and all I changed was what I had for dinner.” But there are other things going on. And one of the things that you mentioned in there was the power of the time series. I wonder if you wanted to explain that because when you have that data over time, even though there are things changing over time, you can begin to isolate.

So, there can be noise, oscillations in your measurements because of lots of things changing. But when you change one thing, you can begin to look and see was there an inflection or so forth. You want to explain a little bit, because that’s a powerful concept you talk about in the book, the time series. What is time series? And for people who aren’t familiar with it.

Gary Wolf: Sure. One of the things that we learned in watching, supporting, and doing so many personal science projects is that there really are just—even though there’s infinitely many questions and an extremely wide range of the ways people make observations. When it comes to looking at the records of your observations, they really do clump together in just a few basic forms. And at the bottom of these forms is the time series.

And what is a time series? A time series is a series of observations over time. And you can think of it really straightforwardly as just looking like a table, like any other little spreadsheet or table that you might see. In the left-hand column is a set of times. And that might be all you have. For instance, if you’re recording incidents, how many headaches do I have? You might just record the time. Each time in the table represents a headache. You don’t need something complicated to be like, how valid is this measure of a headache? It’s just your headaches. So, all you need is the times.

Often though, there’s a second column, which is a count or degree or an intensity. Say you’re tracking body temperature. I just did a project with a family member who was in and out of the hospital, and we tracked her body temperature. And we were given a threshold where, okay, if it goes above this threshold, maybe there should be a readmittance. But it was very useful for us to have a little table which showed the times and showed the body temperatures.

On the one hand, it was reassuring. On the other hand, as we crossed that threshold, we could see, okay, maybe we’ll skip urgent care and go right back to the hospital because we could see that it was heading towards that threshold and balancing around it and finally it crossed it. We’re not going to go spend an hour in urgent care and then have them send us to the hospital.

So, that’s an example of using time series data, very simple data. Nobody’s going to get a Nobel Prize for getting two columns and saying, “Okay, I crossed the threshold. We’re going to the emergency room”—but so useful. And we’ve seen that time and time again, that this simple arrangement of your data with the time and the intensity just yields treasures.

Howard Forman: Okay. As we get closer to the end, I wanted to give each of you one additional question, but also offer that you could add anything on you want to make sure that public—our listeners—get to know more about the book. But for me, when I listened to this at first, I thought initially, “This is biohacking, this is that bridge between wellness and social media and an entire enterprise.”

And the more I read, the more I listen to you, the more I listen to your TED Talks, other things that are out there, you are not a biohacker. You are somebody who really wants individuals and the public to be more informed, but not because you think that somehow we’re going to reach that 125 years of age and still be cycling every day. Can you say a couple of words about how you differentiate yourself from what is an enormous and lucrative industry and how you didn’t get sucked into that because it is a lucrative side of the business?

Gary Wolf: Sara, can I jump in for a second? I’m very curious what you’re going to say too, but we really did have a lot of intersection with the biohacking community, especially at the beginning, because the biohackers were also interested in discovery and they were also using measurement. But it was confusing because that really wasn’t the ethic of the Quantified Self community, and we had to think about what’s the difference. I think the difference, the most clear difference is that the biohacking community is really focused on interventions. It’s just intervention after intervention after intervention.

Now it’s peptides, but there have been many other interventions that the biohacking community has gotten attached to. Whereas the Quantified Self community is a much broader set of questions. And often the interventions come in very late or often even very early like, okay, I have this therapy that I’m trying or that I’ve been given by my clinician, but I’m concerned about a side effect. Is it really a side effect? Can I tell whether it’s a side effect or not? You can see that questions like that, they’re not really about what is the next intervention to try.

Sara Riggare: So, I also met a lot of biohackers at the early Quantified Self conferences, and there was overlap, but the difference is, I mean they want to optimize things to not become unwell, but I already live with a chronic condition. I want to stay as well as I can with this condition, and that requires different kind of judgments and different kind of weighing things differently, and it should also be worth my time.

Could I stay well by applying lots of interventions? Probably, but it’s not worth my time to try it if I’m well enough. So, I track to be able to live my life as well as I can. And when I then found something that works reasonably well, I don’t want to keep tracking. I want to live my life.

Harlan Krumholz: I just want to jump in with actually two quick questions here at the end. One is, so what’s the difference between this and N-of-1 trials? And do you think that we should be better equipped to do N-of-1 trials? And for people listening, these are just, usually we think about randomized trials as involving large numbers of individuals to come up with an average result about whether A is better than B. An N-of-1 trial is just taking a single individual and actually in some ways manipulating exposure. For example, you could be using a pill, you’re asking whether a pill works or it’s a side effect that if we could get manufacturers to give us the opportunity to say, “Take these two pills. We’re not telling you which bottle is real and which one’s not. Take it on/off for a week, several weeks, and see what it does for you.” And then at the end we’ll unblind and let you know, tell us which bottle you thought was effective.

And what often happens in that kind of experiment is people are convinced that a pill is effective. It turns out that they can’t tell the difference between the placebo bottle and the active ingredient bottle and that they learned something from that. But that turns out to be very hard to do today because people can’t get that configured for themselves. Is this adjacent to that? If we’re going to push the Quantified Self, should we be pushing it even further and saying people should be allowed to be able to be doing these kinds of trials on themselves?

Sara Riggare: So, there is a lot of overlap between N-of-1 trials and personal science to some degree. There’s also a lot of overlap between N-of-1 trials and a really good clinician-and-patient collaboration. And the main difference in my mind is that the main objective of personal science is to learn for yourself. A classic N-of-1 trial is the learning comes to the clinician or the person running the trial. So, that’s probably two ends of the spectrum and then with the clinician/patient collaboration is somewhere in the middle.

Harlan Krumholz: No, that’s helpful. Gary, let me just throw this one to you, just so I got the second part in, was, so where does AI fit in all this? Because some people are going to say, “All I need to do is take all of my data and just dump it into ChatGPT or Claude or Grok or Gemini, whatever I want to use and let it just churn the data.” And I’m teeing you up for this because I think it’s about whether you actually compose your question as opposed to just hoping AI is going to come up with some insight that you haven’t even asked. Where does AI fit into this?

Gary Wolf: It is a great question and so timely. I think AI is changing the landscape of personal science a lot, but at the same time, not in the way people think it is. So, this idea that you’re just going to throw all the data in and you’re going to learn what to do, this has been around in the Quantified Self community for years. We’ve seen many projects like that because people have a high level of skill where they’re just like, “I don’t really have a question, but I just want to collect all the data I can and I’m going to build a model, a second version of myself, a mirror me.”

But none of that has really paid off because actually, there’s infinitely many questions. There’s all kinds of possibilities in the world and if you don’t have a goal or you don’t have a question on your mind, there’s no machine in the world that’s going to give you an answer to a question that you haven’t posed. But on the other hand, many of the barriers that keep people from answering or exploring the questions they have, AI is really good at solving.

And those barriers, they’re almost so boring you can’t even talk about them, but they’re very real. They’re like, okay, I have a bunch of data, but some of the measurements are in Celsius and some are in Fahrenheit and now I’ve got to go through and figure out, okay, maybe it wasn’t properly tagged in the start or when I downloaded them, they got all messed up or I want to look at this data, but I want to take the weekends out. I only want to look at the weekdays. Or I’ve downloaded all my data from Google Health, but I really only want to know my sleep onset time. I’m just concerned what time did I actually fall asleep every day.

There’s hundreds and hundreds of questions that you have the data that you could use to explore them, but they’re so fiddly and it’s so painful to get it into the shape that you want. And AI is proving terrifically useful for that. And I think it’s going to open the doors of discovery to many more people.

Howard Forman: We just both want to thank you so much for, first of all, coming on the podcast for the book, which comes out on Tuesday, October 6th. The book is The Quantified Self: Learning to Observe. It’s available in all bookstores and Amazon October 6th. We’re wishing you both the best. I think there’s a lot more to learn from both of you, and we look forward to following both of you.

Harlan Krumholz: Yeah, thanks so much.

Gary Wolf: Thank you so much. And I just want to say Sara and I both follow your work and really learn a lot, so thank you for that.

Harlan Krumholz: Thank you.

Sara Riggare: Yeah. Thank you for having us.

Harlan Krumholz: Well, that was terrific. They’re good.

Howard Forman: That was great.

Harlan Krumholz: That was really good.

Howard Forman: Thanks for bringing them on, Harlan. That was great.

Harlan Krumholz: Yeah, my pleasure. Yeah.

Howard Forman: So, tell me a little about what’s on your mind today, Harlan.

Harlan Krumholz: We’re not used to taking each other’s roles.

Howard Forman: I know, I know, I know.

Harlan Krumholz: Howie, today I want to talk to you just a little bit about an application of artificial intelligence that’s a little less exciting than discovering a new drug or making a difficult diagnosis, but I think raises a consequence question about the future, particularly of U.S. healthcare. And it also illuminates the perversity of our healthcare that this is the place where we think AI can be making such a big difference, probably that some of the most successful companies in AI and healthcare are in this area. It’s called revenue cycle management, or RCM. It’s about the machinery of getting paid, documenting care, submitting claims, following up on payments, and appealing denials.

Many have been describing how these new AI agents can take on much of that work and do it much more effectively. The question is, are they going to simplify this machinery and make it easier to do or actually make it more powerful and in some ways disrupt the entire healthcare system because it all depends on how this works and the friction that’s created and the kind of perverse incentives to code.

So, I was struck by two pieces that I read recently. The first is an article from McKinsey, the consulting firm, that appeared last January in which it talked about a touchless revenue cycle. It estimated that AI could reduce administrative cost of collecting payment by 30% to 60%. Not the total cost of healthcare but the cost of getting paid.

Howie, it is crazy how much money is being spent just preparing the billing and then pursuing the payments. It’s insane. And the idea is that this AI could be helping that. And there’s nothing inherently wrong with that. Hospitals should get paid accurately for the care that they provide.

But then here comes the second piece. A lot of these AI platforms are configured to ensure that each patient is getting the maximal amount of coding. And an analysis from Blue Cross Blue Shield Association found that the share of inpatient cases classified as medically complex rose 37% from the beginning of 2023 and 40% by the end of 2025. The association estimated that the increased coding intensity added almost a billion dollars to the members’ plans. This was just over two years. And this is all for the same cases. I mean, nobody thinks that the patients got more complex, but the machinery of being able to assign these codes led to their labeling is more complex and led to the payment now of more than a billion more dollars to these plans.

So, the ideas that the association which was generating this themselves, so you could say it needs to be validated from external sources, but that the adoption of these AI systems are going to be raising insurance rates because they’re built on an old system that had certain assumptions about how you could tell the complexity of the patients that people were handling.

Now, some of this may be correcting underpayment in the past and that may be good, but there’s going to be many sides and many debates about this. And in a way, this is an arms race dynamic where the payers are getting AI systems to deny, the hospitals are getting systems to appeal, and there’s a great incentive to just assign all these codes to make everybody seem like they’re as complex as possible.

Let me just end by just saying that these two articles that I’m talking about weren’t really testing the same thing. One’s talking about whether you can make this administratively easier and drop the cost of that machinery. The other is saying, “But they’re also adding features.” The reason hospitals are buying some of these is because they can increase revenues. And that second article from Blue Cross Blue Shield is talking about the consequences of that.

Now, whether or not the AI systems were fully responsible for this is still something that’s unknown, but it seems very likely—but we need more evidence and we need greater validation. That’s the second thing. And I do think we shouldn’t be calling the whole enterprise wasteful because we could actually improve efficiencies within the system. But if a lot of this is about figuring out where can we milk more revenue out of a system that was really built—and now you’re taking advantage of it by trying to build in these, more codes, then in the end we could be just all creating more burden for the consumers who are going to be paying higher prices. I know what you think about this, Howie, you’ve been tracking this a bit.

Howard Forman: Yeah. Look, revenue cycle management and revenue maximization is just business. It’s what we’ve done in medicine for decades. It’s what I’ve been involved in for decades. Nothing wrong with trying to get what you deserve. The introduction of Epic, our electronic medical record at Yale and other places, one of the most important goals from a system point of view is always to optimize the appropriate payment for the patients that you’ve seen. Nothing wrong with that inherently, but we’re really right now in this—

Harlan Krumholz: What does that word mean, “appropriate,” by the way?

Howard Forman: Well, you should be paid for what you do. You should be accurately paid for what you do. And if a primary care practice, for instance, is not properly noting that their patients are more complex than someone else’s patients and they’re being underpaid for it, that’s not good. You don’t want someone to feel like, “Oh, my god, why can’t I do as well as my neighbor? Is it my fault that my patients are more likely to be diabetic or more likely to have glaucoma or something else?” People should be paid for the work they’re doing, and they should be able to capture that.

But what we’re seeing now is this arms race, as you described it, but I’m going to put it in a different framing, between the insurance companies and the hospitals, all the other providers, and eventually the patients to try to figure out how each one can get on top of the other. And ultimately that is not going to make things more optimal. That’s just going to be a game and watching it play out, introducing yet more layers that are going to frustrate people.

Harlan Krumholz: Yeah. We make the joke, that’s AI talking to AI when we’re on these appeals. It’s hard to know where it’s going to go.

Howard Forman: Right. It’s what it’s going to be.

Harlan Krumholz: Yup. Thanks, Howie. You’ve been listening to Health & Veritas with Harlan Krumholz and Howie Forman.

Howard Forman: So, how did we do? To give us your feedback or to keep the conversation going, email us at health.veritas@yale.edu or follow us on any of social media.

Harlan Krumholz: And give us feedback. We love to hear from you. We love for you to comment on us. We’d love you to recommend us to your friends. Any and all of the above. Very welcome.

Howard Forman: Health & Veritas is produced with the Yale School of Management and the Yale School of Public Health. To learn about Yale SOM’s MBA for Executives program, visit som.yale.edu/emba. And to learn about the Yale School of Public Health’s Executive Master of Public Health program, visit sph.yale.edu/emph.

Harlan Krumholz: And my favorite part, I love to give gratitude and to talk about our superstar folks. We’ve got our superstar undergrad today, Gloria Baek. Howie, do you know how great Gloria Baek is? She’s amazing.

Howard Forman: I continue to be amazed by everything that she does, and she did an amazing job helping us this week.

Harlan Krumholz: She did, and Donovan Brown, and we have a fabulous producer, Miranda Shafer, and I get to work with the best in the business, Howie Forman. Thanks, Howie.

Howard Forman: Well, I can’t thank you enough, Harlan.

Harlan Krumholz: Talk to you soon.

Howard Forman: Thanks, Harlan. Talk to you soon.