So, when I was asked to contribute, the idea was that maybe having three decades of experience in telehealth adoption might have some lessons that we could possibly apply to what's going on with AI today. So, that's sort of the background with which I approach you today. And it should be fun to take a little bit of a trip down memory lane. And I'll try not to be too sappy or play the old timer too, too much. But I think there are some lessons learned, to be sure. These are disclosures. They're all companies that have AI-based products, but I won't mention any of them in my talk today. Right. So, here's a picture of yours, Truly, in the mid-1990s. I tell people that I walked out of the Mass General one day, took a wrong turn, and wound up an accidental tourist in this world of digital medicine, we called it. It was mostly telemedicine at the time. But more importantly here, the pictures on the right are worth reflecting on for just a minute. Now, this isn't a talk about me walking uphill to school both ways. But look at what we had to deal with in the mid-1990s. Now, we had this vision of time and place independent care. It was very clear that we thought that was not only our future, but our present. So, off we were to the races, and we dealt with things like, so I'll just point out a few of them. The top left-hand picture is a video conferencing unit. It cost about $60,000. You had to string three separate special telephone lines to connect them. It was a big deal. And the one on the top right, that's a vintage Macintosh computer. You'll see the two slots, those were two 1.4 megabyte floppy disks. So the biggest thing we could store on one of those was 1.4 megabytes. The images that we started taking were three megabytes, so we had some storage challenges from day one. There was this new thing called a hard disk. We could get one with 130 megabytes, so we were pretty excited about that. But then again, how do we move them? We don't have any network, no Netscape navigator yet. The final one that I'll just take you down memory lane is the one in the bottom middle, which is the Kodak DCS-420. That's the camera that we started working with. Cost $10,000. It was about one megapixel resolution. Now, again, this isn't too, too much about reminiscing, but really more about my point is that, and we've heard so much this morning that I think confirms this. You work with the technology you're given at the time you're innovating, and you have to take, my wish for all of you that are working day to day in AI is to take solace in this, because think where we are now, and what you will be in 20 years with what we're doing. So that's really the reason. Now, the challenge, because you look at that and you say, well, that's great, because I got what I got now. Am I supposed to ignore that? No, that's not my point. And indeed, it is a challenge to be able to look ahead accurately. So I love this particular picture. This is from a film many of you may have seen. I don't know. It's a cult classic now. It's called Blade Runner. And what I love about Blade Runner was it came out in 1982, and it predicted the future almost 40 years hence, which turns out to be 2019. So you can watch this film today and sort of get a fun glimpse at what people thought 40 years hence of 1982 was going to be. And this picture is particularly relevant to those of us in the world of digital medicine. So on the left, you have this blurry image of Harrison Ford. He's the protagonist. He rather fancied the woman that you see pictured on the right, and he was having a drink at a bar and lonely, so he wanted to see if she would come join him. So he makes a video call. That's the innovation that got right. In 1982, we had not heard of a video call, but he goes to a phone booth to do it. That's the part they didn't get right. So this is what it's like when you're looking into the future and trying to figure out how it's going to work. Because what? Well, we didn't know this was going to happen. Again, I showed you pictures of this big clunky thing. All of a sudden, in 2007, we have the network, the computer, the camera, all in one little tiny thing, and it takes beautiful images, and that's changed our world. These are things that are hard to predict. So what's the AI equivalent of the iPhone going to be? And I don't know the answer. I'm just suggesting that that's part of how we think about innovating in this space. Now the other thing, and I gather that we're in this room, it's a bit of a safe space here, I think. Most people here are enthusiasts, but not everyone outside of the room is, I can guarantee you that. So this is some examples of things that we encountered early on. One of the first things we did in teledermatology in my group was to put our work onto a web browser format, so you could actually upload images. This sounds quaint now, but at the time it was because the web was brochures at the time. So you could upload images, and then someone could put in their diagnosis, care plan from afar, et cetera, through time and place independent care. And I was excited to present that at a dermatology meeting in New York City in 1998 or so. And two things I remember about the audience response, you're cheapening our specialty. This guy came up. He was very animated too. And then the other one that I love was, you're going to turn us into cutaneous radiologists. So this is the kind of crap that you deal with, and you just have to realize that that's what it is. And there's plenty of it out there. Now, we also did some work in the early days in remote patient monitoring before any of these systems existed. And remember, I met the Mass General across town, it was the Brigham and Women's Hospital. At the time, it was the, without a doubt, the epicenter of cardiology in the world because of Eugene Bromwell mainly. And there was a woman there who was the heart failure expert, perhaps world best. And this is what she said to me about remote patient monitoring. You can't do that. It's not going to work. So this is the kind of stuff you get. And again, it's easy in the retrospective scope to prove who was right. Some people said things like that to me, and they were right too. And of course, my favorite, favorite, favorite, which you'll always get is, my patients will never go for that. That's a sign that someone feels really threatened when they say stuff like that, because they really probably haven't checked with their patients whether they'll go for it or not. All right, so telehealth adoption went through, I think, three main phases. So this might be relevant to those of us thinking about AI going forward. The first was about proving out that it actually was quality care. It's a little different in a sense because telemedicine is a care model intervention. AI is a software intervention. But most, or I would say most, but many of the themes this morning were very much about this idea of proving out quality. And I think we need to invest time, energy, and resources into that for sure with current models and as they improve. So that seems relevant. The next two I would put up interchangeably because in different scenarios, these would take one before the other. One of them is about integration into workflow. Doctors would say to me, I'm not going to do this unless you put it in the electronic record. That was particularly a theme after 2008 and the HITECH Act. And then some said, I got to be paid. Thank you very much. Which, of course, we want to be paid for the work that we do. So those two things have happened in telehealth and now there's almost every EHR has an integrated module. You, of course, can accept images. You can do video calls, get them right into the record, et cetera. And as you know, there are diagnostic codes to support this activity. We don't have really much of those latter two. And I would argue we still have to work on the first one in AI. So what about some mistakes that we made? Well, we worried, I said this earlier, but we worried too much about current technologies. And so while I'm not suggesting that we throw out things like hallucinations at all, I'm suggesting that just take a deep breath and look beyond that when you're thinking about how you're going to build out your vision because that can weigh you down. The fact that we had to worry about a three megabyte image seems so quaint now, but at the time that's something we focused on. Or thinking about how we were going to justify having anybody purchase a $12,000 camera to do this work. Sold it as efficiency. This I still see a lot as people say, well, or the third one, it's the right thing, right? Patients demand this. It's the right thing to do. Those are tough business models. The right thing to do model in case anyone, at least in the US, nobody really pays for that. I'm so sorry to say it, but it's just the truth. Whereas the other one, you can get some traction around efficiency if you're careful about it, but it's much better to bring in something as a revenue enhancer if you can. So can we make the same case, what can we learn from telehealth and AI? Well, there are some similarities. Wait a minute, sorry, back up. There are some similarities. There's this, again, fear of loss of control you see in the general population, fear of something new, and it's disruptive to current practices. That's all stuff that is about change management, and we definitely can learn from history on that. And not just history of telehealth, but of adoption of innovations in general. But there are differences, and I alluded to this a moment ago, AI is a software tool. So it has different things about it in terms of how we manage that adoption. And whereas telehealth is a care model innovation, it is front and center in care delivery. Hard to hide that. Much easier to blend in AI in certain circumstances. And we, I think, heard from the audience about some of those earlier, about problems that you would solve. And we said, I want something that will help me diagnose skin cancer. So I would call these our first generation challenges, and I'm not an AI expert. I don't even play one on TV. So this is just my idea from the sort of reading the tea leaves, if you will. And it was mentioned before, but I'll just say it again. This is scary stuff, because not only is it wrong, but it's wrong with confidence, right? And we know of a certain politician that got elected president because of that. So more than one, perhaps. But anyway, the idea that this software is telling you straight out something that's completely wrong, that's kind of scary. We got to fix that. And so therefore, we need human in the loop. How long do we need human in the loop? Will that, will we get beyond that? That's a drag on the future business case for this tool set, I think, just. And how important is explainability? Do we really need, at some point, will we trust the tools to do the work that they do without having to explain everything? All questions that I think are first generation challenges for this type of tool set. So in terms of accelerating adoption, I have one set of categories that are about business or technical things that I might suggest, and then one that are personal. So here we go. Go back office. This came up again earlier today a couple times. Let's solve some problems that everyone wants to solve that are non-threatening to our intellect, like prior authorization, things like that. Stay under the radar at first. Didn't get a chance to do that with telehealth. It's out there. You can't really hide it. Tackle something non-threatening, something that's a revenue answer. That would be great. Ensure privacy and security. There's a little bit different spin on that with AI, for sure, than with telehealth. So that's a difference. And then double down on the evidence base. I said that earlier, but I can't emphasize that enough. People, doctors respond to, number one, evidence. Number two, can they actually make a living doing something? And number three, does it make their life easier or less easy in the clinic? It's really important. So the evidence base, so important. Personal things. Well, make it about the greater good and not you. So we all struggle with this as innovators, because people are coming in and telling you, you're full of shit. You don't know what you're doing. It's hard at first. But if you have this greater good mentality, you can sell just about anything. So it's really well worth thinking it that way. Giving others credit. I forget who said this, but there's no limit to what you can achieve if you give other people credit. And that is so true. Just, again, don't make it about you. Make it about your team. Make it about everyone else, and you'll do much better. Don't pick fights with Luddites. They're out there. They're always going to be out there. Sometimes they retire. Sometimes they die. But it's not worth it. You kind of want to do that. It's not worth your time. It's just not. Just let them sort of go about their business. And sooner or later, they won't be there. And surround yourself with people smarter than you. That's just good advice in general, I think. So we won't belabor that one. All right. The last couple of slides we're getting there, AI reimbursement conundrum. So I spend some time helping the AMA on CPT codes for digital medicine, and we are grappling with this currently. It's because it's a software tool. So without spending time on how CPT works and how your reimbursement works, because I don't have the time, but there's this component of every time you get a reimbursement from a payer, there's a component for practice expense. And software, particularly your EHR, falls under practice expense. So most of people that are innovating in the business of AI are really worried that that's where they're going to fall, because there isn't much room to move there. It's not like you can build on more. So that's an enormous challenge. But right now, because it's software, that's how a lot of payers are thinking about this. The current best practice is to have a human in the loop. We talked about that. That person should be paid for what they're doing. How do we work that out, at least temporarily? Again, that's an unsolved problem right now. And then finally, at scale, AI should make clinicians more efficient, i.e., take work off of our plate, not at it. We haven't got there yet. And once we do, since most of the coding is about doctor's work, how do we pay you the right amount for taking work away? So these are challenges, and these are going to be challenges for some time with this software and getting into the mainstream. Last thing is just for me to plug my colleagues at ATA, our annual meeting coming up in May in Phoenix, and listen to our podcast, Health Virtually Uncensored. Thanks for the time. These are my contact information, and I want to make sure I give time to Justin. Thanks.

telehealth adoption

AI trends

challenges faced

quality care

AI adoption issues