AI in Healthcare: Hype vs. Hope with Mark P. Mills on Future Innovations

[00:00] Bryan Hyde: Welcome to the Health Policy Podcast. I'm Brian Hyde. Today, I am joined by Mark P. Mills, founder and executive director of the National Center for Energy Analytics. Mark, welcome to the show. Would you take a moment and just tell us a little bit about who you are and what you do? [00:15] Mark Mills: Well, thank you for having me, Brian. And it may seem discordant to your listeners to have someone from the National Center for Energy Analytics talking about healthcare. Setting aside the fact that we all care about healthcare, the reason we're talking about it is because I write a lot about and study and have been working in the technology fields of sensors, imaging, of how energy causes things to work, or how we produce energy, how we use energy. And this would include things like X-rays and magnetic resonance imaging. I'm a physicist; all these things are connected. Right. So, in the world I've worked in, both semiconductors and data centers and the oil and gas and the nuclear industries, they all intersect, not surprisingly, if you think about the underlying physics and engineering, they all pivot around healthcare domains. So, I'm very interested in them, I've written a lot about them, studied them, in my book, The Cloud Revolution, it's a few years old now, a major part of it's about AI, And a significant section is on healthcare and the future of medicine and drug discovery. So it's part of the reality that I have studied and work in. It's the intersection of the 3 domains: technology, which includes energy and compute; finance, what we can afford to pay for, how we get the money, who makes money; and politics and policy, because Permissions always matter in the real world we live in. There's already regulations exist, and whether we like it or not, they could be added to that famous line from, you know, Benjamin Franklin, you know, about death and taxes. I guess I would add regulation. [01:59] Bryan Hyde: Well, you know, AI holds some very interesting promise, and yet there's a lot of hype out there. And frankly, Mark, I'm not sure who to believe. On the one hand, I hear AI is going to find a cure for cancer, and on the other hand, I hear that it's going to wipe out the human race. Maybe you could help us have some context about, as it applies to healthcare, what are some of the promising areas you see where AI could actually be beneficial? [02:23] Mark Mills: Yeah, I think this is really important because the misunderstandings and the hype are epic. I mean, as you pointed out, we're giving a framing that essentially it's magical and going to cure and solve every problem and therefore take away every job, so the jobs apocalypse, or it's going to kill all of humanity. Maybe, you know, it's just possible that it's indeed a remarkable technology and will do remarkable things. And like all technologies, there are side effects and negative effects we have to be sensible about. You know, this is an old adage, but put this way, when new technologies come along, they always bring risks with them. This is not new. We had to figure out how to manage them. Without getting hyperbolic, to put it simplistically, the invention of the car was also the invention of the car accident. [03:11] Bryan Hyde: True. [03:12] Mark Mills: That's why. So, here's what's fascinating to me. Arguably, the single most promising near-term benefit from AI is in the healthcare industry. And, it's not to cure cancer, although I think it is arguable, and I think a lot of scientists have already believe this and agree with it. I used to be a scientist, but not in healthcare, but in the underlying systems that relate to it, you know, the physics of systems. But, it is true that AI, as a research assistant, is extraordinarily powerful, arguably as big a step from computing to AI as it was from adding machines to computing. So, if someone had talked to me in the 1950s and '60s about why computers are important, and what they would change. Most people would not have imagined, in the 1950s and '60s, the world we live in today. They would not have imagined Amazon and Uber. They would not have imagined online shopping. By the way, a couple of people came close in those times, and they were science fiction writers like Arthur C. Clarke, but nobody made those kinds of forecasts in the business world. The transformation that we have underway is comparable, which means it is just as big an opportunity and disruptive in many ways as this last great revolution we had. In healthcare, let me frame it this way, and this is my bias on how I think about healthcare industry, the healthcare desire. I first think about it not as a costly economy, but a business that everybody wants. This is a framing I steal from Steve Forbes, because he always talks about how politicians talk about the You know, the burdensome cost of healthcare. Everybody wants more healthcare. That means it's a business. What we want is more effective healthcare, and we want our costs lowered. We want better outcomes for a lot less expense. That doesn't— that's normal, right, in any product or business, but especially something so personal. So, when it comes to technology and money, and thinking about how AI applies, I divide healthcare into just 3 buckets. And they're simplistic, but they sort of makes sense. The first bucket is what I'll call the science bucket. We want to understand how biology works, how bodies work. This is the science of disease discovery. It's the science of discovering how physiology works, how neural systems work, how diseases propagate. That's science. And the science of discovering new cures, that's science. AI is going to be an incredibly— it already is— powerful research assistant to accelerate discovery. And that is very promising, but it will take time. It's not going to do it overnight. The tool is very powerful, but it's still not powerful enough. It needs to be, frankly, 1,000 times more powerful than it is today to be truly useful. The good news is, it's going to be 1,000 times more powerful within a decade. That's the track we're on. It's 1,000 times more powerful now than it was 5 or 10 years ago. So that's an astonishing rate of improvement in an incredibly powerful research tool. But only 5% of all money in the whole healthcare ecosystem is spent on science. 5%. The next part of the healthcare industry, if you like, of business, is what I would call diagnostics, right? Those are the physicians and the machines that try to figure out what's wrong with you. I mean, and we all live that. So diagnostic, diagnosticians, and diagnostics are incredibly important. If you— we all know this personally— if you get the diagnosis wrong, you're going to get the cure wrong. Like, this is not a scientific observation. It's a no-duh. That is incredibly important, and that's where, again, AI is powerful. Remember, we were told 3 years ago that AI would eliminate all pathologists because it could read, you know, X-rays better than pathologists. Turns out, in just 3 years, we've discovered that's not true. What it's done is made increased the demand for pathologists, which has, in fact, happened, but it's made pathologists and pathology reading better, faster, more accurate. That's beginning to echo into the system. That's what AI will do for physicians and diagnosticians, for X-ray machines. It'll make the machines more effective, faster, cheaper, ultimately, in terms of per image, per reading of a blood sample. That's already beginning— Excuse me. That's also a little slower, than we'd like because these things are all regulated, for good reason, right? We want them to be safe. So, that so far is 20% of healthcare spending. Where's the other 80%? It's in the deployment side of all the cures and diagnostics. That's the healthcare system of hospitals, clinics, doctor's office, all the infrastructure of deployment of how we administer healthcare. Maybe this shouldn't be surprising. It's 80% of the total outlays. spending. Okay, so what would you target to cut costs first in the deployment? And here's the key fact that's exciting about why AI is important. 25 to 35% of all healthcare spending is administrative overhead. [08:16] Bryan Hyde: Wow. [08:17] Mark Mills: AI is really good at that and fast at getting better at that fast. So let's just do spitballing. Let's assume AI is good enough to cut administrative overhead costs in half. That would take almost 20% of the costs out of all healthcare overall, dollar spending. This would be tectonic. That's huge. Nobody has any path to take 20% out. Now, does this mean some people will lose their jobs? Yes, some specific jobs in healthcare administrative back office will disappear. But overall, it's not going to diminish the workforce. It'll just slow the rate of growth of that workforce. The largest growth in the workforce in healthcare has been in the administrative overhead, not in doctors, not in clinicians, not in scientists, but in administrative overhead. Who wouldn't want that to grow slower and get better? Precisely what AI tools are profoundly good at and already good at. We don't have to wait a decade for 1,000 times better Compute. [09:25] Bryan Hyde: No, that makes sense. I mean, I think of how many people I know who have gone into healthcare, not so much as a care provider, but there to, you know, to document or log the, you know, the various insurance codes and so forth. So yeah, I could see AI being a real game changer there. [09:45] Mark Mills: Yeah. [09:45] Bryan Hyde: Talk to me a little bit about What are some of the biggest challenges in medicine beyond the bureaucracy? Where else do you see AI, you know, upping the game of, for instance, the doctor-patient relationship? [10:00] Mark Mills: Yeah, well, again, this is what's fascinating about the healthcare industry. Everybody has an opinion because we're all subjected to it. And we all have had this increasing experience in recent years and decades over the You could see the administrative burden for the nurses, and the clinicians, and doctors when they're with you, because they have to do record keeping. And for good reason! You want them to have good records. You know, they see a lot of people. You'd like them to remember correctly what you told them. Who wouldn't, right? So, they're very careful about recording these things. They're writing them down, put it in the right forms and formats. This is holy grail land for AI— listening to conversations, reading documents— and literally, Listening to a doctor or a nurse talking to a patient and transcribing and recording and putting into the right formats and framing and security systems exactly what was said, exactly what was heard, how the patient responded, how you and I talked back to the doctor, what questions we asked, all that transcribing has already— there are already AI tools One of them is called a scribe. It's called AI scribes that can listen in real time. It's secure. It's as secure as your current health records and do all that work for the doctor or the nurse and the clinician. That alone, first, it's cheaper, faster, better. But just think about the experience now. You're now talking to the doctor instead of— or the nurse instead of having them stare at a screen and typing away. So, it makes it better. It makes it more accurate. And it makes the experience that you have with the doctor better. But also, this is the subtle thing that people miss, I think, and it's obvious when you state it. If you, as a professional, are trying to figure out a complicated problem, and all diagnosis is complicated, it's a lot of gut instinct in it, but you're also distracted to have to multitask and do a pedestrian thing, type this down, put this in the right form, make sure you've said things the right way, all these things, you are diverting your, you know, the wetware of our brains, its neural capacity, to focus on the real mission, the real task, which is paying attention to the patient, listening to the subtle signals. You know, really good doctors are good at reading signals and understanding, you know, what you're saying, how you're saying it, and their abilities to focus and do their task that they're much better than computers at, is compromised by multitasking. So, it reduces that, too. So, anyway, that's, you know, one small example. The other one, to your point, is that once a diagnosis has been made, and once you are put into the system to have something done, some medical procedure, you might have to have an operation, there are literally dozens of different kinds of steps, forms in that ecosystem supply chain to coordinate with. That involves the complicated sort of labyrinth of forms and coordination of time and who's available when. AI is really good at that and getting it right, getting it fast. So, again, that will not only make the experience better— you don't have to wait as long to figure out what the path and how long you have to wait, where you have to go— but it'll be done by AI much faster than people can do it and with far fewer errors. [13:20] Bryan Hyde: I guess I never really appreciated until you just laid it out here, the incredible amount of data that has to be a part of the healthcare process and making sure that all of that communication and directions, prescriptions, and doctor's orders, you know, are not only logged, but are correctly and accurately logged. [13:39] Mark Mills: Well, add to this the data that are associated with diagnostics. So almost all of us have had an X-ray. Well, X-rays are all digital now. So digital imaging, digitally stored. MRIs, data stored digitally, digital imaging, digitally stored. All the new treatment tools are digitally imaged, stored, all this massive quantities of data, which AI can manage, first, not just to store it, but to mine it, to look for patterns that doctors might miss. In a very sort of simplistic way, I'll give an example in a totally different domain. Some of us may remember when the world was treated to the first image of a black hole. I don't know if you remember this a few years ago, because everybody likes their science fiction movies, at least a lot of us do. No one had ever taken a picture of a black hole because you really can't, because it's a black hole. But it is visible at all kinds of different wavelengths that our eyes can't see: X-ray wavelengths, gamma-ray wavelengths, cosmic rays. infrared. And by collecting all of those signals and blending them through, literally, an artificial intelligence tool, these are massive amounts of data, and processing it, the computer system that the astronomers designed was able to produce a photograph, an image of a black hole. This was treated as a revolutionary step in astronomy, and it was. That took a massive amount of compute power. Here we are, it's only a few years later, we have 100 times better compute power, about 100 times more compute power, and apply that logic to all the multimodal imaging you could do of a human body and get pictures that could never have been taken before, literally never been imaged before to show a physician or a clinician. That's what AI can do for X-rays and MRI by combining all of the different imaging modalities and produce a comprehensive image. The holy grail in medical diagnostics is that I could take an image of your body and your internals with such high resolution and precision that I'd be able to figure things out. In fact, doctors already use 3-dimensional imaging with X-rays for complicated surgery to map out and plan difficult surgeries in a virtual environment before they do it on you. This will become much more common, become much more precise. All of it's data-centric. So, when you look at the quantity of data involved, and you do the arithmetic on it, this brings us straight back into my world, which is why I study the imaging and data world. You need big data centers with massive compute capacity, and you're going to need them, never mind any other application, you know, making making fake videos of your least favorite or most favorite politician, all that fun stuff, or making movies, or CGI for movies, or cyber hacking, all the things that are talked about, the biggest single demand vector for data and processing it is, in fact, in the healthcare domain. And imagine if, instead of calling these AI data centers, we gave them different names like Medical Research Center, Medical Imaging Analysis Center, or you pick your name to describe what it's in fact doing. And we might not have as many protests of what we're trying to build. [17:08] Bryan Hyde: Well, and that's, you know, this is kind of the last question I wanted to ask you is how do we move it away from being such a political hot potato? Because right now data centers, They're a touchy subject to a lot of people. I'm not saying people should blindly trust them, but how can we help them see that there is an upside? [17:27] Mark Mills: Well, what I described in healthcare is the upside. In fact, let me give you one other example before I answer the question about tamping down the irrational data center resistance, which is different than the rational objections. to building big industrial facilities near people's houses. 2 classes of protests going on, so to speak. Scientists have been working on the idea of what's called a virtual digital human, digital twin, for a long time. Digital twins exist for machines, computers— not computers, but computers make a digital twin of a motor, in fact, a whole aircraft, and you can You can simulate flying in silico with digital twins. I don't just mean flight simulator, but everything down to the engine components. The holy grail in healthcare is I have a personal digital twin— me, you— and what it does is constantly harvest, securely, all of my medical data, not just when I go to a doctor, not just when I get an image. It keeps that, too, but also, you know, my blood rate, because I'm wearing a smartwatch or smart ring, and glucose, and what I eat. All these things can be harvested and create a virtual digital twin. And, it's your digital twin that allows doctors to make more rapid and accurate diagnoses about you and your health and what you do. Who wouldn't want that? I mean, the amount of data it would take for each digital twin It's astronomical. I mean, if everybody had their own personal digital twin for their own health purposes, it would increase the amount of data required globally millions and millions of fold, billions of fold, staggering amounts of data, which would be a great thing because it can't be used to control you, and it can't be— it doesn't violate your privacy for you to have a digital twin. The only thing that would be violated is if you have information about your health, which you often have, have, which we don't want somebody else to know. Well, that's already the case. We already have these things to protect us, and we'll continue to do that. So, that's the trajectory we're on, those kinds of opportunities across all kinds of domains. But, it is true that there are negatives to the digital domain, social media being the obvious one, the misuse of social media. Social media, it wasn't created by AI. I mean, AI makes it easier to do search. It makes it easier to do fake videos on social media. Social media, problem identification, and these are real problems that we're trying to solve culturally, but they have been bundled into this broad inchoate data center resistance. People who hate big tech or hate the oligarchs, to use Bernie Sanders' language, or they worry about cybersecurity or cyberattacks, all these things bundle, then we don't have enough power, we don't have enough water. We go through all these issues and separate them into 2 buckets: things that are not AI-related but are genuinely concerns, your properties of social media. Okay? Data centers that make social media are also the same buildings where you get navigation, mapping, shopping, collect your videos, and get Netflix streaming. Same buildings. It's just different businesses in the same building. It's like a shopping mall with different stores in it. when you're protesting that maybe you don't like a shopping mall built next to your home. Okay, fair enough. Then that's a zoning issue that has nothing to do with the data center per se. I think this is going to sort out. I think the big tech guys have finally figured out they have to start talking, not condescending, to citizens and communities. It's already beginning. They have to begin to change their— I won't call it their posture or political posture, but that's probably the right phrase. Rather than being partisan about who they like or don't like in broader, non-technical political issues, they should stay out of those lanes. Politics matter, but it's not— it's fine if you're a tech CEO and you have political opinions, stay in your lane, but talking about your business, you should stay out of the politics. So people, we don't bifurcate this issue as a Republican-Democrat issue. It's neither. It's a human issue about, well, we care about zoning and big buildings. The big warehouses are like big data centers and big factories. They're big buildings, and you build them, they can be ugly. Maybe you want them in different locations. The last thing I'll say is, in terms of the 2 principal animating tropes that are surviving, is that they use too much water and use too much power. We can do the power part. This is not— They use a lot more power than people realize, but enough to trivialize it. It's not a problem. It's not a problem of, can we make the power? It's a problem, again, of permissions, where we want to actually do it. The water issue has been genuinely, profoundly hyped, overhyped. These systems can be what's called closed-loop cooling. It's like a car engine. You put water in the radiator once, and you just circulate it. for the life of the car to cool it. Air comes across your radiator and takes the heat away. That's a common mode of cooling that's possible and increasingly used by data centers and many power plants. Whether one uses water, what's called a once-through cooling or closed-loop, just depends if there's enough water around and water permits it. All right, you do a closed-loop. Data centers use far less water than golf courses. In fact, the big water user obviously is irrigation. It's the biggest single one in the whole country by factors of, you know, 10 to 100-fold. And that's where we want the water to be used, frankly, to grow our food. [23:25] Bryan Hyde: Again, we've been talking with Mark P. Mills, founder and executive director of the National Center for Energy Analytics. This has been a very worthwhile discussion. Mark, thank you so much for joining us on the Health Policy Podcast. [23:39] Mark Mills: Thanks for having me.