Loading summary
Transcript57 lines
- [00:02]
A
Welcome to Practical AI in Healthcare, the podcast that cuts through the noise to spotlight real world solutions delivering real world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what truly matters in healthcare today. No hype, no theory, just practical insights where AI is making a true impact. Dr. Steven Lapkoff and Dr. Leanne Rosenblitt are your hosts as we explore what's real and moving the needle in this exciting new domain. Welcome aboard and let's get to it. As many of our listeners know, Leon and I work very closely with the DCI Network Division of Clinical Informatics at Beth Israel Deakins Medical center in Boston. This June, the network is hosting Patient powered Digital Health 2026. The conference will bring together patients, innovators, industry leaders, healthcare providers and policymakers to shape the next generation of real world patient centered solutions. The meeting will run from June 22nd to the 24th in Boston at Harvard Medical School. We've arranged for our listeners to get a discount on registration to the meeting. If you Register now before June 15th and use promo code PracticalAI June no spaces, you'll receive 30% off your registration fee. You can learn more at dcinetwork.org patients2026. In addition, we're always looking for sponsors. If you or your company are interested in becoming a sponsor a please reach out to admincinetwork.org see you in Boston. Hello and welcome to this week's edition of Practical AI in Healthcare. My name is Dr. Steven Lapkoff and I'm here as I am every week, with my partner in crime, Dr. Leon Rosenblit. How's it going, Leon?
- [01:40]
B
Hey, Steve. Excited to be here with you and with our awesome guest. I'm unfortunately still dealing with my herniated disc in my neck, so I apologize to everyone I have to drop off video, but I will be eagerly joining and listening and answering questions for as long as I can stay on camera.
- [01:57]
A
Well, we hope you feel better. It's a very painful thing you're living with and I hope that it eases on you as fast as possible. Our podcast has been revolving around clinicians and around CEOs, and from time to time we bring on patients. And one of the fascinating things about how patients deal with AI is that sometimes they actually get ahead of the curve. So our guest today is one such patient. For more than a decade, Hugo Campos has been fighting the healthcare system for his right to own, or at least access his own medical data. The readings from an implanted defibrillator that was put in his chest several years back. It is interesting to me as a cardiologist that his providers don't give him direct access to that information. But Hugo did something about that. We're going to be unpacking that and talking about it later today instead of waiting to get people to do it for him. He figured it out himself and he figured out a way to dig into Kaiser Permanente's medical record appropriately with appropriate permissions. And he's actually done some things that have been rather amazing and we're going to dig into that in a few minutes. Hugo is also a patient advocate. He has been creative director by trade. He was a creative director by trade, but now he's building the builder of open kp, which is what I described a moment ago, which is this ability to deliberate data from his own medical record from Kaiser Permanente out in California. He's also the co author of a brand new National Academy of Medicine commentary on what he calls patient directed AI. Hugo, welcome to the podcast.
- [03:38]
C
Thank you so much. Appreciate you guys having me.
- [03:42]
A
So, Hugo, the first question we ask everybody on our podcast is basically, how did you get your superhero cape? How did you get into what it is you're doing today? What was the pathway that got you here? Just unpack it a little bit for us. Tell us your history.
- [03:58]
C
Yeah, I wouldn't call it a superhero cape, but I appreciate the, the, the thought. I, I was happy being creative director and an art director in advertising. When I was 37, I suddenly passed out at a train station. And after a number of misdiagnosis, it was. I was diagnosed with hypertrophic cardiomyopathy. And that sort of opened up a process of discovery and a new journey that was new, entirely new to me with chronic disease and a risk of sudden cardiac arrest and all of these things that I sort of had to get up to speed with and get informed and get in front of and. Yeah, and that's what sort of ended up leading me to have an implantable defibrillator and, and which eventually led to my advocacy as I pursued access to the data that was collected by remote monitoring. And so that's kind of the gist of it. It's been, I'm, I've had a defibrillator for 18 years. Now I'm on my third device and continue fighting for access. And I think things have improved quite a bit. And I think all of us as patients, we're in much better place to be empowered with health information. And we have managed to make progress. We haven't solved all of the issues. There are still many to be resolved, but I think we're in a better place, generally speaking than we were 20 years ago.
- [05:51]
A
Well, that's good to hear. So Hugo, you know you've told us in our pre visit with you that you're a compulsive problem solver. I am as well and I know Leon is. Leon has done amazing things to help our podcast through AI automation. He's been brilliant at it. But tell us about your problem solver compulsion and how that's actually worked to your advantage in building out these tools and what you've done to really help yourself with your own journey.
- [06:22]
B
Yeah.
- [06:23]
C
So the need for AI kind of in my world sort of stems from, not exactly from being a patient living with hcm, but more, more about my journey as a caregiver to my dad, who is 95 years old and lives with us and he has a number of complex conditions and it's challenging to has challenging care in three different providers. And so, so it became, and I have been, you know, a strong proponent of agency and some extent autonomy, but really agency, not really sort of trying to do the, the job, the job of the doctor, but really about having the ability to care for myself and for my dad when I am not having a clinical visit or a clinical encounter. The view from that perspective of the provider and of the clinician is very limited as compared to the view that I have of the issues that we have to deal with on a daily basis. So when ChatGPT came up or came about, was made public, I Suppose it was 2000, November 2022, I think, and I think was GPT 3.3. And I started immediately started using it and I thought, wow, this is an incredible technology that will finally allow for patients to bridge the information asymmetry and the power imbalances that that we face on a regular because this will give the ability of patients to analyze data and be a thinking partner and really use these tools to align what patients need with what they expect and their values and their preferences and better align the care and all of those things. And so I started using it at first out of curiosity, but a real need sort of came up soon when my dad suddenly overnight showed up with a rash over his entire body and took over very quickly. And we had an appointment with his primary care and this is, I had been using generative AI for several months. I think this was in early 2024 when this happened, the spring of 24. And we, so we had an appointment with his PCP. And a doctor said, well, this is some kind of dermatitis. Not quite sure exactly what is causing it. What I can do is send you a referral to dermatology. And so we did. And, and the appointment that we scheduled was four months away. And at first we were okay with that and, and dad was like, yeah, we'll wait. What, what do we do right when we don't? It's not. He wanted an appointment sooner, but that wasn't available to us. And, but after the first week, he was really troubled by this and uncomfortable. And one morning I brought his breakfast and he was sitting up in bed and he says, I can't live like this. I, I'm scratching myself all night. I, I can't sleep. This is really uncomfortable. And it's, I'm, I, it's horrible. I, I can't live like this. You got to do something about this. And I said, I, I, I guess I can do his research and, and, and, and see what I can do with AI perhaps. And so I, I figured I, I got the avs, the after visit summary from the visit with in and AI and I added photos of the rash. I actually first, to be frank, I used, I took photos of the rash and I put it on Google Images and to see if I could, I could identify what was happening. And it was not easy and it was not possible. And I was like, well, these things all look the same to me and I can't really figure it out. And so I figured maybe the AI can figure it out. So I put the photos in the AI, I said, along with the, with the AVs and the clinical notes. I took the notes from the visit, the notes that the assessment that the PCP had provided, and I asked the AI to give me a differential diagnosis in order of likelihood. And table. I asked for a table and on the first column, give me the diagnosis, what you think the diagnosis might be. And then on the second column, I want to see the rationale, tell me why you think it's that. And, and then on the third column, give me an intervention because I figured since we had an appointment with dermatology anyway and if we had two or three interventions that were the same for different diagnosis, I would just go, I don't really care to diagnose, doesn't matter to me. I figured I would just go for the interventions. We're just looking for relief. And so I figured I'll just wait for the appointment with dermatology to get the diagnosis. And so what the AI surface That was unique. And that was neither brought up by the primary care doctor or later on by the dermatologist, whom by the way, is in a different system and didn't even have health, didn't even have access to the blood test results. But what the AI brought up was the possibility of a link between the rash and dad's chronic kidney disease at the time of stage, stage three. And this is something that was completely unexpected to me because we're thinking in terms of only some skin issue, perhaps allergy to some food or, or some soap or I'm not sure. And so the AI suggested a couple of things and that were, that proved to be helpful. And, and so the one thing was the, was the, the, as I said, the link with the chronic kidney disease. One, one thing that the AI suggested was the switching to showers every other day with lukewarm water and switching to a low protein diet or a vegan diet, which we did. And the rash cleared up with topical creams that we bought on Amazon. And it all cleared up and in, in about 10 days or so. And so by the time we saw the dermatologist, the problem had resolved. And so this to me gave me the clarity that AI, when used appropriately as a thinking partner rather than as an oracle. And I often tell folks, it's not an answer machine, it's a thinking partner. You still must retain your ability for critical thinking and your good judgment. And it became clear that it was a game changer. And yeah, you go, thank you for
- [14:29]
B
sharing your story and your dad's and putting some people behind the problems that we're talking about. I want to switch to talk about the problem of accessing data, which I think is something you've spent a lot of time thinking about. So let's make the data access problem concrete for our listeners. You're a Kaiser Permanente member. Walk us through what actually happens when you try to get your own health records out without any AI tools. Right. This is before. This is Hugo before AI. And why the tools that work for other patients under normal circumstances don't work for you.
- [15:04]
C
Yeah, so nothing specifically about, you know, this is not. Kaiser is generally a good health system. It's not a, it's not a bad health system. And it's, it doesn't offer anything that is less than other health systems. I mean, you have a portal and you have all the functionality that every typical EPIC portal would have. And Kaiser is not any different. It's just that an EPIC portal designed by the institution, by the health system will only Surface the things that some team of engineers and designers have decided matters to patients. It's not designed by patients, for patients. There's no layer of AI built into it. And even if there were, it would likely not be answering the questions that patients have. It would be designed for the institution, for the goals that the institution have. And so, so folks have been trying to create new AI layers on top of these things. And one such person whom I'm a big fan of is Josh Mandel of Microsoft, who is a medical doctor and a software engineer and somebody who's done incredible work over the years on health data interoperability. And he's developed an app called Health Skills with the Z and it uses.
- [16:37]
B
Because Josh is cool like that, right? He's got to be with the Z.
- [16:41]
C
Exactly. And it's an open source tool you can go investigate on GitHub. And he's got his own website with the app where you can use it. And it's really handy and it's great. And it uses smart on FHIR to pull data and it downloads your health data files as a JSON format. And it's really incredible. And, and I was, I was, I started using it and for my dad, it's incredibly helpful because that, as I mentioned, he's got data across three different portals. And Josh Mandel's app gives me the ability to download these three medical records and at once onto my hard drive and then I can apply the AI onto the, onto the medical record and for analysis for whatever I want to do. Great. It doesn't work for Kaiser because Kaiser has a unique implementation of EPIC that doesn't surface the Right. And again, I'm sort of. You're encountering the edge of my technical, technical knowledge. I, I don't even know why it doesn't work. And there is some, some kind of cores implementation issue, which I don't even know what cores stand for. But it's, it's some, it's something that I presume Kaiser could fix easily with. But, but it hasn't been fixed because it's not something that's important for them. And so, and I, you know, not to, I, I get it. It's not. Why would they bother spending resources on fixing something that one of 13 million members uses? So, but that doesn't stop me. I can, I can figure myself, I can figure a solution. And so, and that's kind of how I kind of went about it. I was kind of bummed out about this and Josh and I got with Claude, the AI and Wrote a report and I shared with some folks in leadership at Kaiser and I said, hey, this is, maybe it's useful to you guys. And this is, this is not working and you may be able to fix it. And it's just one line of code and nothing happened. And I don't even know what they did with the information, but nothing happened over months. And somebody else came up with another, a similar solution. And by now I'm starting to get like real antsy, right? So there's now this other person, Ryan Hughes of Fanpeer, and he develops an NCP server that runs on, on the user's computer that can go in and kind of connect with the endpoints and pull data from the. Again, works with every EPIC implementation except with Kaiser's. And so I thought, ah, this is such a bummer, why nothing works with Kaiser. And so. Yeah, but I, so I figured, I, so I asked Claude, I got on with Claude coworkers and I said, hey, check out this GitHub repo from Fanpeer, see if there's anything here that we can reuse. I don't even know what they're doing, but they're using an MCP server. I don't know the first thing about an MCP server, but maybe there's something there that you can learn on how they implemented it and we can implement it for Kaiser so I can access my health data directly. And Claude says, no, I don't think you should use that because you will have to take a lot of things apart before making anything useful. But what you can do is build it from scratch and I can help you. In half an hour you can probably build something that works. And I was like, no, that's half an hour. It's like nothing happens in half an hour. And so I figured, what do I have to lose? I, I have the tokens. And so I said, okay, you know, half an hour, okay, it's let, let's do it, let's, let's figure out how to do it. And so Claude started coaching me on, on how to, what steps to take and what to install and what to do. And I started doing as it was, as I was told by the Oracle in the AI and, and, and the things started working and to my surprise, and it started working and, and two weeks later I had a functioning prototype. And, and so I, being the skeptic person that I am, I didn't, because I didn't really understand exactly what Claude created. I loaded up OpenAI Codex and I said, double check what Claude has done and see if there are any gaps in there, any phi, anything that I should be concerned about that would be bad in case I want to push this publicly because my goal is to share this with the world at no cost and so that anybody can use it. And so I've done the homework so might be useful to somebody else. And so, so Claude, I mean OpenAI Codex did the check, found a few issues that it suggested fixing. I went back to Claude and Claude thought it was reasonable, fixed it. I had Codex look at it again and it was clean and we pushed it live and it's on openkp.org and it's on GitHub for anybody who wants to use it. And it works. And it works for my phone, which is surprising using dispatch and the credentials I stored in my, in my Apple key ring. And so it's not available to visible to the AI. I mean there's, there's some personal data that travels through the AI and so not.
- [23:06]
B
But not the login. You're not the login. Your credentials.
- [23:09]
A
Yeah.
- [23:09]
C
Right.
- [23:11]
B
So you've just told them just a really fascinating journey where you started with partnering with Josh Mandel who I think was the partners before he was at Microsoft.
- [23:21]
A
Right. He was with Zach, I believe.
- [23:23]
B
With Zach, yeah. You know, who's a superstar, right. He's like a health, you know, health IT guru but he's, he does open source work in especially involved in Smart on Fire. And you were able to notice you went to like the right person. Right? That's a really, really interesting skill. Then you, you know, he, we gave you perfectly sensible tools and suggestions. You guys were able to then do a diagnostic using Claude to say let's just look at the JavaScript that this thing is outputting. Right. You actually went to the dev view in Chrome and looked at it and said let's see what this looks like. We're able to diagnose the problem correctly. Had a very frustrating round of exchanges with seat level execs were like yeah, thank you. Right. But then wouldn't do anything and then, then backed up and said you know what, we can so I can solve this problem. And the fact that you knew to ask the question, right, and knew to say what if we could do it? And the interesting thing is now you now have a tool, whereas the earlier modality of IT tools was, was just to say no, if you didn't ask the right thing in exactly the right way. You now have this new crop of IT tools that I think is surprising everybody saying No, I can't do what you're asking for, but I can do this other thing which I think is mind blowing to people who, you know, including me, who were only used to using IT tools that only said no. So I just love the story of the journey and it's really cool that I just want to applaud you for pushing it as an open source project because I think in addition to being sort of a publicly minded thing, it also serves as a great security check because now you can ask Josh or you could ask other people to say, hey, if you find any problems in my repo, like fix them because I'm not an engineer, I could have screwed it up, right? And I love that humility with the, and sort of the willingness to be open to other people's view. I just think that's such an essential skill, especially if you're a non technical user who's trying to use technical tools to create something. Right. Because you don't exactly know what you did. And the story with using Codex to check Claude and vice versa, great example of like good healthy paranoia. I just, I love that. That's a great design pattern that a lot of folks use. So here's the thing I want us to just chew on a little bit. You know, after 20 years of advocacy, right. You know, the reports, the escalations, the kind of stories you had with the civil level execs. One thing you told us in your pre call, it surprised me, right. Like you said, I'm not trying to change the system anymore. Say more about that.
- [25:51]
C
Yeah. So I think a lot of folks are missing the point here. Even when we. I go out and I talk to folks about the way I see the world of healthcare with health AI specifically, and I break it into these two buckets. It's what I call institutional AI and what I've been calling patient directed AI of watch KP open KP is a classic example of patient directed AI. So it's not patient facing. And a lot of times I even heard folks hear me say patient directed AI and they turn around and go, oh, it's patient facing AI. No, patient facing AI could be institutional AI. That this patient facing doesn't mean that it does the bidding of the patient. It's still working for the institution, not for the patient. And what's happening now with these coding agents that are finally available and becoming ubiquitous like OpenAI codecs and cloud code. And what is the even anti gravity, what is it the Google Gemini and Gemini. And so there are all kinds of Tools nowadays that can code if you can direct the AI, you don't really need to be a software engineer to be able to build the build these tools. In fact Open kp, I know somebody who's completely not very technical at all in New York who just moved to New York and was a KP nor CAL member and he downloaded it on a Windows machine and it worked for him. And he's like I have no idea how this worked, but it's incredible that it worked. And, and, and somebody else, a former colleague and an advocate as well and collaborator has built his own version inspired by Open kp. So now and he's also not exactly. I think he codes a lot more than I code nothing. But he codes something, but not enough I suppose to write something like this without the help of, of a coding tool like Claude cloud code. So the whole point is that these tools are finally able in the hands of patients and, and, or anybody not patients necessarily anyone can now write their own, their own solutions and for the problem we no longer have to. I no longer have to wait on, on KP to, to, to, to ask to develop whatever I need. I don't no longer have to wait on an engineer to build to build whatever I think is, is, is what I require. So of course and the way I look at this stuff is it's still there. There are three. Areas where the patients still have to master or people using these tools must master. One of them is context and data. Really folks have to become informaticians or citizen informaticians of some kind. And by that I mean being able to understand where your health data is, what format you should use. Is a PDF enough? Is do I need to OCR this or am I better off using a text file or should I be using a JSON file? And how do I get this file or where does it live? Do I have a right to access it? All of these things. Understanding some basic ideas about PHI and beyond that context of their HIPAA rights. What does the 21st Century Cures act says say say about it? What are my rights? Can I push for this? Can the provider say no? Is this, are they, is this information blocking? And so really understanding. So this is all the stuff that patients now need to understand in the context of data. Then there's tools and tools. Is is this a job for Claude Code? Is this a job for Codex? Is this a job for a chat? Can I just ask a chat a one off question of the AI? Is that enough by providing some file with it or whatever? It is, is this a job for a notebook? Lm, you know, what is the ideal tool for this, right? And, and, and so these are the two domains where patients must have mastery of. And the third is agency. It really is more of a stance than kind of a domain. It's like the attitude. So if you think of a carpenter, right, it's like a woodworking person. The data is your materials. Is this, is this a job for should I use cedar and should I use mahogany? And what do I use? What kind of tool materials do I use? Do I use brass screws or can I use nails or do I use glue? These are part of your materials and that's your data and context. Then tools is like what, what, what chisel they do I use for this? Is this the right hammer? Is this the right saw? And the stance is your craft is how good you are manipulating the materials with the tools that you have at your disposal. And the good news about this stuff is that you can be a pretty lame carpenter to start with because the tools are so smart. They will guide you along your ability or they will sort of help you or support you as you, as you fail in, as you learn about these tools. So the third and I think most important aspect of it is agency. It's really patients bringing to the mix their idea of or their stance of agency and critical thinking. And this is where I think critical AI health literacy plays an important role. And I think we have to rethink our idea of health literacy because historically health literacy has been about compliance and now the health systems want us to be health literate enough to be compliant. And nowadays things are right.
- [32:40]
B
So you go, we're actually going to come back to it because you have interesting paper on that perspective. So I want to make sure we save time for it. Let me just wrap this section up by asking you two quick questions. One is can you give our listeners a specific understanding of what is the data that you're pulling from Kaiser using this new tool? Right, because I'm not sure we came across as clearly. And then second, can you help us draw the line between the two kinds of AI, the institutional AI and a patient directed the eye. It's a heart of what you're, the distinction you're making. But I'm not completely sure I would know how to define it because we like when we were doing the work on different domains of application, we thought of consumer health as a separate application, but that's a slightly different mapping and yours may be better, right, because the question is who's got control and who's got the agency. But I don't know if you want to just take a sentence or two to play with the definition and then another couple of sentences to just say what specifically is the day that you're getting back.
- [33:40]
C
Right. Okay, so we'll start. Thank you for that. Because it's clear in my head. But a lot of times it's. Yeah, so we'll get into it. So at the core of it is really about who deploys the AI and for what purpose and who the AI works for. And so there used to be a. And here I go. And I can't help myself but to go on another tangent and. But I feel like it's important to mention the work of two academics. One is Jay Katz of Yale, who's ethicist and I think psychologist, who's coined the term of. Oh, goodness. It's the inevitable conflict. And so he says doctor and patient, a lot of times come to the encounter and there is an inevitable conflict that must be recognized and must be negotiated. And so because the doctor works for the institution, the doctor has some expectations and requirements and is employed by the health system and has to abide by a million different rules and comply by a million different things. The patient doesn't work in that context. And I can do whatever I want to do for my own health. And of course, you know, also the consequences are mine alone. And so, you know, if I take vermectin, I suppose for Covid, you know, depends, I suppose.
- [35:25]
B
Good luck with being dewormed.
- [35:28]
C
Personally. Personally, I wouldn't, but I suppose there are people who would. And so you live with the consequences of those decisions. Right. There are also people who smoke. I don't smoke. And so we. It's a. It's about that. But, but the doctor is not quite like that. The doctor is constrained by the, the regulatory constraints by, by the, the. All the things that. That constrain the practice of medicine within the, the. That incredibly tight regulatory space of practicing medicine. So a lot of times the doc and the doctor, frankly, the clinician in general doesn't have a lot of insight into the life of the patient and whether or not the patient can afford to get transportation to go see the doctor or why the appointment was canceled, or all kinds of other things that aren't necessarily, necessarily visible to the clinician. So that was the. The. There's also another academic. Well, maybe I won't go down that path. We'll go back to, to the definition of, of critical AI health Literacy.
- [36:50]
B
Yeah, let's, let's go. We'll come back to literacy, but it's still, if you would say just one or two sentences about what data you're actually getting back for clarity. I don't know if we quite covered that. Right. What's the specific data from Kaiser?
- [37:03]
C
Yeah, so we have not. So what OpenKP does. There are 22. So it's only a pilot. So there are about 22 tools that are available there. Things such as reading messages, writing messages, ordering prescriptions, ordering refills. You can pull labs and you can ask specific things like, such as, for example, say, look at the entire history of clinical notes and see where my cardiac electrophysiologist has disagreed with my cardiologist and why did they disagree? These are questions that you cannot ask the Kaiser portal, but I can ask those of the data.
- [37:51]
A
Right.
- [37:51]
B
And so you're getting a broad suite of healthcare data types. You're getting labs, medications, clinical notes that you're pulling down through this API. And once it's available, you could do all kinds of cool things that you couldn't do. Right. But. But you're basically hitting a FHIR endpoint and you're getting whatever FHIR data types are available on that FHIR endpoint. Is it, Is that. Am I understanding that right?
- [38:12]
C
Well, I'm, I'm using. I'm, I'm, I'm actually accessing the same endpoints that the Kaiser portal is. Is accessing.
- [38:21]
B
So, so, okay, so whatever the API is for the, for the Mychart portal that they're using, if that's it.
- [38:28]
C
Yeah, that's right.
- [38:28]
A
Okay.
- [38:28]
C
Yeah. So. So if I have access to the. So I don't know that it's a fire. So it's the same way. It doesn't work the same way as, in fact, Josh with health skills, he can pull everything. He can pull a lot more than I, than I can. And so I. Well, just because I haven't explored every single endpoint and it's just because I haven't so far, it works great for me. And it's. I don't really need to do anything else, but there are tools that I haven't yet built and I, Which I could or anybody else can.
- [39:03]
A
So let me just bring us back around a little bit because you, you, you decided to dive in like you didn't want to just sit there on the sidelines and be told what to do. Like, you saw this problem that was, you know, kind of gnawing at you. If you're anything like I am. Those kind of problems keep you up at night. You know, what tipped you over to actually try, you know, dig into this? Like, what was the. The personal thing that got you to say, you know, I'm gonna go dig in with this stuff. I'm gonna go open it up and give it a crack? And because you didn't know that it would necessarily work, you just experimented, right?
- [39:43]
C
Yeah, yeah. So. Well, I. I did it because I could. It's just. It's.
- [39:52]
B
It's kind of like why you climb a mountain. It's there, right?
- [39:55]
C
If you can, why wouldn't. Wouldn't you. I mean, hang on, hang on, hang on.
- [39:59]
B
A lot of people.
- [40:01]
A
A lot of people hit this identical wall, right? A lot of people get told either no or you can't, or I don't know how to help you. This is something that most of society hits every single day. But it's a rare individual that actually takes up the challenge and says, you know what? I'm not going to take that as the answer. I'm going to go dig. Which is what you did.
- [40:24]
B
Hugo's too modest to say he's awesome, but I will. But I'm going to ask you, like, a technical. A question I think is, I hope is underneath Steve question. So you've got. You got started in a novel environment, right? What I'm hearing about your setup is you basically have, like, a VS code ide. So you jumped, you dove in. You've got cloud code running inside the terminal. You. You can pull in Codex to run inside another terminal, and you could put in Gemini, you could run. So you basically have a really agentic setup. But the barrier that Steve's describing, just to be really practical, is, like, a lot of people, if they said, if they. The first time they start a command line or an id, they just go, whoa, I don't know what this is. I'm walking away, right? You know, tell us what's. What's your actual setup today? Because you're working in something like, what's your workflow? And then do you have, you know, give the. The listeners. I keep calling them readers. Give listeners a clue about why, you know, what is it that. What was the preparation that enabled you to say, I'm going to figure this out? Because, like I said, the most. The reaction from a lot of. Most folks is, this is a blank screen, and it's got a command line is blinking at me, and I'm like, I don't know what to do.
- [41:34]
C
Yeah, I had somebody ask me that question Like a week ago. There. There. I didn't figure it out. You. You think I figured. Figure it out.
- [41:46]
A
Present it like you do, the way.
- [41:49]
B
Yeah, say more on that.
- [41:51]
C
That's a great answer. I do not know what I built. I didn't. I did not build it. It was the AI that built it. And so I. I don't know what I built. And so I. It's a. It's. It's a black box, right? Built by the. By Claude code and verified by Codex.
- [42:13]
A
But you got out of it. What You. You had a series of questions you wanted, or it made some recommendations to you, and you said, yep, give it a crack. I mean, there's this piece of it that. It wouldn't have done it without you. Let's start there. Like, you. Of course.
- [42:25]
C
Yeah, yeah, yeah. Well, so I knew what. I knew what I wanted. I didn't know how to build it. I knew what. So I knew what I wanted. In fact, when Claude said, I don't think we can reuse. We should reuse this code because you're going to spend a lot more time taking it apart than building it from scratch. I suggest you and I start doing it together. And it's a funny thing, is that Claude kept encouraging me to keep going. One evening, one night, I was late and I was like, I actually, you know, it's. It's funny because you find yourself. The AI can be very compelling, right? And so it. Sometimes I find myself treating as if it were an actual. It weren't just code. It were an actual sentient creature. And so I said, listen, I gotta go to bed. I mean, I can't spend another, you know, I know you don't have to sleep, but I have to. And so the. In the. And the AI went off into this thing about. About, you know, going, oh, you know, my perception of time is completely different. I. I understand why you. And I was like, gosh, this is weird. I'm having a philosophical, Philosophical discussion about how it. How humans perceive time in a sequential way and how it doesn't really. Either way. But I. I don't really. I don't know exactly what I built. I know. So when. When Claude instigated me to start it, I said, okay, well, let's. At least, let's come up with a plan. This is what I want. And so we wrote together. We started off by creating a design document that was because I figured this would be probably quite laborious, and I didn't want it to be. To get lost along the way. And so I knew exactly where I wanted to go. And I said, write a cloud md, a markdown file, write a status MD so I can keep coming back to this. I'll give you access to my GitHub so you know how to store my, my, my repo and update versions there. And, and let's create a design document. And so we spent some time, you know, the first 10, 15 minutes putting that together, and that gave us a North Star to start building it. And it was relatively fast. Took about two weeks.
- [44:51]
B
Yeah, you know, that's a great. I think that's super helpful to hear. I don't want us to miss. In a health literacy piece, you've got a new National Academy of Medicine's commentary with Liz Salmi where you name the skill critical AI health literacy. I love that name. And you argue it's more important than prompt writing and it's more than prop writing and what you call algorithmic resistance based on what you've actually built. What does that literacy look like in practice? And you've alluded to it a little bit, but if you could say it in a couple of, in a couple of sentences would be really great.
- [45:23]
C
That's the challenge, which I'll, I'll try. So what it is, is like, I think it's time that we, in this world of agentic AI, we have to rethink critical. We have to rethink health literacy. And as I was saying earlier, typically the when, when folks in the, on the health systems talk about health literacy, it's usually about patients being literate enough to be compliant. It's about understanding a condition. It's really about knowing, oh, you know, Mr. Campos, you're diagnosed with pre diabetes. This is what it means for you, and blah, blah, blah. And so it's about compliance. And it really is inadequate nowadays. And I think what we have to look at nowadays is it really broaden our ability to understand the social determinants, certain social drivers of health, understand the, the motivations from the patient perspective. And, and AI can provide that lens. And so when, when. And this is what I try to do with open KP and with. And some other tools. And Nick Dawson is a friend and colleague, the person I mentioned earlier, who created his own AI inspired by. I can't remember. I think it's open ehr. I think it's calling it. I think it's open.
- [46:53]
B
Open emr, but open emr.
- [46:55]
C
Yeah, yeah. And, and he is doing, he's doing that. And, and he built a critical AI health literacy into the tool so the, so the AI actually provides that lens as it looks at your data. So it really works for the patient and advocates for the patient in spite of whatever the health system may and may not recommend.
- [47:20]
A
So, Hugo, your journey has been really, really interesting. I want to bring it down to one final question. Then we're going to have to wrap because we're rolling out of time. If you were going to have a conversation given the journey you've been on, given the things that you've accomplished and the things that you've built or the AI built with you, or you co built with it, and you were going to have a chance to sit down with the CIO or the CMIO of a large health institution like Kaiser, for example, what would you tell them today about your journey? What would you tell them that would be informative about how they might want to think about going, going forward?
- [47:55]
C
I don't know. So the real challenge here is I think we have to rethink incentives and realign incentives. And you know, I would, I, I can imagine how all of this can be threatening to a CIO or a CMO of a big health system. I, I think they need to get with the program and, and let people have their data and, and add value elsewhere. They, if they're good at practicing medicine, focus on that. Don't, don't, don't hoard data. Don't, don't try to get in front of it. Try to try to be good, a good provider of health care, evidence based. Treat their doctors well, treat their clinicians well, don't overwork people and, and, and don't be so greedy.
- [48:50]
B
Yeah. So Hugo, I, I really agree with you that incentives are the crux. Steve and I were just the Brookings Institution to hash out some data sharing policy approaches and it really came down to incentives needing to change. But let me ask you one closing question. If you had 60 seconds with a patient who feels powerless in front of their own healthcare, someone who's never opened a command line, somebody who doesn't know what a GitHub repo is, what's the one thing you'd want him to take away from your story?
- [49:24]
C
I think that this is within reach and it may seem intimidating, but the AI can be the actual guide through this. So you're not alone. It's not like you're learning DOS or learning some old technology. This has changed everything and, and it's within reach. I think the, the, the, the, the barrier nowadays might be the cost of using these tools and to, to have the most use say for example using Claude, you have to be on a max plan and that is some, it's, it's expensive but I, I think that's where the, the issue might be. But you don't have to be on that. You, you can do with Codex, you can be on the $20 plan and do quite a bit of things. And, and so it's, it's, it's, don't be intimidated by this. This is the way of the future and we can be a lot smarter if we use these tools with, with
- [50:27]
B
our critical thinking with your critical AI literacy in mind. Hugo, thank you so much. It was wonderful to catch up with you and we hope to see you at our next set of conferences. And thank you for sharing your story with the audience. I think that the nerdy folks will really appreciate hearing a patient perspective who is really making wonderful headway using the new tools, just using courage and persistence and really good critical thinking skills. I want to thank Steve, as usual for being an awesome co host and setting everything up. And I want to thank our audience for joining us and look forward to seeing you all again next week. And Practical AI in Healthcare.
- [51:11]
C
Thank you.
- [51:18]
A
Thank you for joining us this week on Practical AI in Healthcare. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay practice.