Kamal Menghrajani: The Limits of AI Healthcare

Hi, I'm Paul Ford. And I'm Rich Seotti. And this is The Abord Podcast, the podcast about how AI is changing the world of software and the world in general. Hello, Richard. Hello, Paul. I want us to play the theme song, and then I want us to talk to our guests because this is a really good one. Let's do it. For just a few seconds, let's explain what Abord is because people keep asking. Let's do it. You want to give me like a one sentence? One sentence? That's all you got. Give me three. Okay. All right. We ship amazing AI-powered solutions for companies. We go in, see what you need, see where there are opportunities to run your business better or your organization better. And then we deploy amazing people and amazing technology to get you there. That's pretty nice. We have about 36,000 years of combined experience building tools, and we've been riding this wave and figuring out how to actually drive real value and actually ship stuff. It's quite the wave. It's quite the wave. All that stuff you hear about 95% of AI projects not shipping, that's because 95% of people could try harder. We can do it for you. The other 5%, that's us. That's us. 5%. So, good example, go to our website, check out Make an Impact, the case study. We built a big medical dashboard that you can talk to. You can ask it questions. Making data actionable, making things work, keeping things under the rules and regulated and compliant. That's enough about us. Enough about us. Let's go to our guest. Welcome, Dr. Komal Megharajani. Hey, thank you guys so much for having me. So, I have to tell you, we've looked at your LinkedIn and there's a lot going on. Doctors are funny because they do a lot of different things. So, help us understand which of the different things you're doing. Yeah, I think right now I am a practicing oncologist who is working deeply to think about how we can improve healthcare using AI. Okay. You're teaching. I mean, give me some more. Give me a little more. Yeah, sure. So, how did I get here? So, I am a classically trained physician scientist. So, I was actually at Memorial Sloan Kettering for eight years working on genomics using computational oncology approaches to better understand how do we understand cancer? Who's going to get it? What's going to happen when they do? Are they going to respond to treatment? And it was great sort of writing software from scratch and trying to understand these big thorny questions in cancer. You're programming too. Yes. So, that's my back. That's the link. That's part of how I got into this. Okay. But to learn to program, I actually went to school at Columbia at the School of Public Health and I got a master's in statistics and they said, yeah, we'll teach you statistics, but you have to learn all of public health with it. And that got me really curious about, okay, what are the other intersections of public health and oncology? And so, from there, that interest led me to become the oncologist for the cancer moonshot. So, I moved to Washington, DC. I worked at the White House under the Biden Harris administration for a year and a half, focused both on the cancer moonshot, but also became the physician for the health team. Okay. So, I ran point for federal guidance and legislation that was coming out from FDA, CDC, CMS, and really had this 30,000 foot view of what health policy looked like in this country. So, simultaneously on the ground, lots of patients, lots of stuff going on, and then also way, way up and sort of looking at cancer at a global and national scale with an incredibly ambitious mandate. Yeah. The cancer moonshot, I mean, had these two big goals. First of how do we decrease the cancer death rate by 50%, how do we cut it in half over the next 25 years, essentially, from when it was launched. But then thinking about the softer side of how do we improve the experience for people who are facing a diagnosis, whether as a patient, a family member, or a caregiver. So, really thinking holistically. And even though a lot of our work was focused domestically, we ended up launching a global effort, which a lot of us are still continuing now that we're back as private citizens. Before we get into what you're doing today, tell us what you learned. I mean, obviously, the mandate is hovering over everyone. You're trying to get this done, but obviously, there's a lot of people involved, a lot of process, a lot of stuff. And we're going to come back to this, the culture and the people that were in the mix here. What did you learn? What did you learn with this? Okay, here's this incredibly ambitious mandate. They parachute you into DC. She's like, God, you asked me that. Such a big question. It's interesting. I think there are incredibly smart, well-intentioned people who are working in DC, at least during my time, you know, during the Biden-Harris administration, who are really thinking about how do we improve healthcare for as many people in the American public as possible? And how do we do it in a way that's guided by evidence? And it's so, you know, coming in as somebody with expertise in clinical practice and using evidence to make decisions, it was really wonderful to say, okay, I'm used to doing this on a patient-to-patient level. How do we think about this on a systems level? How do we think about this on a national level? That was one of the biggest takeaways. And I think the other was, a lot of these decisions are made by very small groups of people. You know, when you sit in the room where it happens, you're looking at who else is there at the table. And it's really amazing, you know, thinking about who is there and what viewpoints are they representing. And so now that I'm outside of government thinking about, okay, well, what are the things that we can do to get things moving, to start to advocate, to start to use private industry to move forward some of these same public health principles? And how do we make sure we make a big enough splash so that the people who are sitting around that table have a sense of where, you know, we think healthcare should be going. And I think that's why working in AI right now is so exciting. You're inside this big initiative and then AI is, you can see it in the distance. What was the timing of, obviously there's a lot of conversation, well, we can finally find some cures. AI is here, right? There's a lot of that sentiment. AI is going to fix everything. They're big on that too. They love to say, I mean, cancer would be, you know, it's just a product feature that they can cure it at this point. You know, Sam Altman out there. I mean, just get a few Claude code agents going and you know, we'll solve the whole thing. We just need to get you another Mac. I mean, honestly, just need a little bit more. Get me a few GPUs and we'll just, we'll knock this out. Problem solved. I mean, look, there's a lot of people outside looking in who are very hopeful, hopeful about this convergence of technology and what we're trying to solve. But based on your mocking laughter, it's not that simple. It's not that simple. The part of the White House I worked in was called the Office of Science and Technology Policy. And literally it was us on the health team and next door on the same hallway was the tech team who were developing the AI strategy. So these two things were actually very intertwined. One of the big initiatives that came out of the cancer moonshot was a project called Cancer X. And they have multiple functions, but one of them is to serve as an accelerator program for AI specifically in oncology. And they helped some nascent startups get up off the ground and helped them create a minimal viable product so that they could actually see what are the different ways we can apply AI to solving this big thorny problem around cancer. So the use of AI, you know, even three, four years ago was something that people were thinking about. Of course, now the technology has evolved to the point that we're able to do much more than we were back then. But this is, you know, this is a train that's been chugging along for a while now. All right. So you're part of this big initiative. You're in D.C. I guess you paused your practice, I'm assuming, to go do this. You're doing your civic duty. Let's go do this big job. Did you move there? I did. Oh, wow. All the way, all the way in. Yeah. My husband is also an oncologist and he works here in New York. And so I moved to D.C. and he stayed here in New York, kept his practice going. He's a professor in a medical school. And so he kept teaching. And so when the Biden-Harris administration came to a conclusion, I just moved home and so came back to New York and decided, you know, I wanted to take my career in a different direction. Having seen the 30,000 foot view, I realized, you know, I thought AI was really going to be the next big thing in terms of thinking about how do we affect change to the healthcare system overall. And so wanted to go back into clinical practice. And now I'm up at Mass General Hospital and MGH affiliated with Harvard Medical School and practicing as an oncologist there part time. So you are practicing. I am practicing and then working in AI consulting, you know, working with different folks who are creating And then, we're hearing from doctors, and doctors are saying, I have way too much paperwork. I need just, could we start with clerical first? And I always think of this as like, the right way to bring AI into the organization is to clean up the mess the old computers made. We'll take the new computer, we'll clean up the old computer's mess. And I think, you know, our audience, it's a lot of product managers, a lot of people who work in tech. Healthcare feels almost insurmountable if you're not in it, because, and it's also when you start to work in it, people are like, they get very sort of mythological about HIPAA. A lot of this stuff is very easy to solve, but it's just, as an industry, it's got a very funny relationship with tech. But most of the challenges, so many of them are literally like, I just need to find documents more readily. I need to transcribe things more efficiently. Like, it's so many simple things that you're describing that are not related to like, cells. They're not related to, you know, internalizing and dealing with huge data sets. They're just like, can you just make it so I can go to bed? Yeah. I mean, I guess I'll pose it as a question. Adoption via, oh, I'm going to email hospital IT right now that I found a really cool piece of software that could make me more efficient, and then it just goes into a void and never comes out. And then there's, you know, what we call grassroots adoption, which is like, I can get an app on my phone that's going to listen to me talk, and then I'm going to copy-paste some stuff and route it around the organization to a large extent. I mean, how did, I'd be shocked if you tell me this adoption is the result of a lot of large IT purchases from hospitals. I think there are. So, you know, it's interesting when you think about where is AI seeing the most utility in healthcare today? So one is what we were talking about in terms of, can you speed up the documentation? One way that AI is doing that, you've referred to, is sort of this ambient dictation software. I have a clinic visit. Somebody comes in to see me. I bring in my phone. I turn on the app. It's listening to the whole conversation, and it turns it into a note at the end. Now, I don't have to do that typing. And so that can be really useful and a real time saver. The other place where there's a lot of AI adoption is what's called clinical decision support. So I'm taking care of a patient. You know, I do oncology, but this person, there's something going on with their kidneys. I'm not entirely sure. And so I'm going to put some stuff into this AI, and it's going to tell me, oh, you should be thinking about these five potential problems. Here are the next steps to work it up. Why don't you order these labs, order these tests, and then you'll be able to figure out what's going on with their kidneys without necessarily having to go and call a specialist right away. And it gives, these clinical decision support tools will also give you evidence. So they'll give you, you know, here's the literature. Here's the New England Journal paper. Here's the guideline that is supporting this, you know, systematic way of trying to work up what's going on with the patient. Wow. That's a big deal. That's not just note taking. It's not just note taking. It's really helping physicians think through problems in a way that's evidence-based and up to date. I think there's an important thing that I've noticed is, like, I didn't realize this about the craft until we had twins, and it was a high-risk pregnancy, and I did very little. And but I didn't, as we got further and further into the pregnancy, I realized there was a weekly meeting about our pregnancy, along with all the other things going on, right? And that doctors talk, and I don't think when you're a patient, you realize that, because when you're saying this, like, that information comes in, but it doesn't just stay with you. Like, you're going to talk about what, you're not just going to be like, okay, the AI said it. It's good. Which is kind of what's happening elsewhere in the world. You're going to take this information kind of into the culture that you're in. You're going to share it out. You're going to be like, I think this. I think that. And you kind of don't, like, it just always struck me, because we think of doctors, I think, a lot of times, as just kind of these brains that operate. And, you know, like… Yeah, input, output. Brain in a jar somewhere, you know, floating around. We don't think of you, like, with a PowerPoint presentation in a room somewhere in the hospital with four other doctors being like, man, I don't know. Right. What do you think? Is this… I don't know. Like, the idea that you could all be somewhere shrugging is alien to patients, right? But I'm sure it has to happen all the time. Yeah. Yeah. People don't read textbooks. They present the way they present. And, you know, the study in the New England Journal or whatever other, you know, wherever you're finding it, it's done in a specific population in a very controlled way. And most of the patients who present don't 100% fit. And so that leaves a lot of white space in the practice of medicine. Sure. And so there's a lot of room. Judgment. Exactly. A lot of room for experience, you know, people who have tried something again and again and again and have seen it work. And so we do our best to practice evidence-based medicine. And I think AI can be really helpful in terms of, here's what the evidence shows. Here's the, you know, the ideal way of facing this issue. But there also needs to be more space where clinicians can discuss difficult cases, cases that don't fit the regular paradigm, and bring in that clinical expertise to layer on top. I think AI is not there yet. No, I don't know if it can be. What do you think? I think it's a great question. I think this could be a real differentiator for where AI is headed. I think if we can find a way to not only integrate, here's the literature, here's the guidelines, but here's clinical expertise on top of it, allow for more of those physician-to-physician conversations to happen, not just in the same room with a PowerPoint, but going on across the country or potentially across the world, then I think we'll really be able to take these tools and turn it into improved health outcomes in a way that, you know, wasn't possible before. There's another thing I'm thinking a lot about, which is LLMs as a technology, it's very hard. They're not inherently reliable. They're trying to make them reliable, but they're not inherently reliable. Meanwhile, the history of medicine has an unbelievable number of things around knowledge graphs and sort of data-driven methods and machine learning that are far more reliable, right? And I think we're just at the very beginning of making a loop between those things. Like the LLM is great for querying. It can figure out what you're asking. It can kind of, you can get a lot of intent and then you can go consult really large corporate databases, like vast sets of PDFs, but also vast sets of data. And I think that loop is going to be really exciting where you're like, hey, help me get into this database of prior outcomes and figure out what happened over the last 20 years. And I think it can be an amazing interface for that, but I think that world is just starting where we sort of bring those two together. Yeah, I think there are some enterprise AI solutions for healthcare that I imagine are trying to do this. And it brings up a few interesting questions. Number one, how much of the information that is actually used for patient decision-making is structured in a way that an LLM has access to it? Because yes, you do get the labs and you do get the radiology read, so on and so forth, but so much happens in conversation and that's not captured anywhere. So is the data that's even available to the AI adequate to make appropriate decisions? Is that data even making its way back to the LLMs, though? Even the structured stuff? Probably not. No. Structured notes are sort of a different thing. There's, you know, there are efforts in different EHRs to try to structure the data better. anonymize it, get it out there so we can see patterns and such. So when we think about how do we actually use patient data for research, which is sort of what you're alluding to, how do we find patterns? Yeah. There's a few things involved. Number one, it has to be anonymized. Number two, we have to make sure we're asking patients for their consent in order to use their data for research. And then number three, there's a risk when you look retrospectively that you are going to find a statistical fluke. It's much easier to accidentally find something that turns out to be incorrect when you're looking back. Yeah. And so that can be a real issue with sort of real-world evidence. So making sure that we have appropriate statistical frameworks, for use of a wonky term, to make… Oh, you're in a safe space. No one who listens to this podcast… …no one's like, this lady, this doctor with the statistical framework. They haven't been funny in a while. Yeah, yeah. No, I mean, you're good. You're good. You're very safe here. Yeah, yeah, yeah. Because, you know, if the data is prospectively collected and appropriately controlled, that's when we sort of of interacting with AI. And so as a person who wants to do their best taking care of another person, you're going to do everything you can to try to make sure you're making the right decisions. And so whether you're using the information from AI or not, you're probably still going to go to your peers. You're probably going to go to senior attendings and say, am I making the right decision? Am I thinking about this the right way? So we may sort of see, you know, sort of a delayed period of training as people ask more questions and learn. But once people have to develop that independence, I think they'll realize the importance of being able to critically think through things and not just relying on the AI. I gotta tell you, I get a little sense that Dr. Majorani's students would never. It's not, it's just not happening. Don't go there. No, I know. You get the sense, it's all very, very nice, but you also get the sense that there is a steely, steely glare that can happen. And it's, it's not to be trifled with. We do our best to make sure that those who are in training are well-prepared to take exceptionally good care of their patients when they're done. And so, you know, that requires the bar is high. They know it's high. The expectations are set and they meet them. Well, this has been amazing. Thank you very much. We've learned a lot. Have you learned a lot? Yes. That's good. It's good to see something get in there. Great conversation. Really interesting time. I feel like if we chatted in a year, we'll have a whole new set of topics to talk about. It's been so fast. Absolutely. Liability and regulation and who's paying for this and AI and drug development. And yeah, there's a lot more to talk about. All of that. But let me summarize it this way. Don't count on AI to replace your doctor. Use it to learn about public health. Yes! There we go. More people should learn about public health. Go ask Claude about good smoking cessation programs globally and sort of how that affected cancer rates. Let it draw you a graph, right? That's a thing we should all be doing. You know, doing the things that your mom told you were good for you and making sure you're staying consistent with those things. Getting outside, getting lots of sunlight. I think Lex's mom handed him a pack of cigarettes every day. My mom told me to do that today. Oh, that's true. Cigarettes, yes. Oh, yeah, absolutely. Sleeping well, eating properly, getting your vegetables, you know, a social interaction. These are the things that build health and wellness over a lifetime. And so whether or not AI is where you get that information from, make sure it's a trusted source, you know, not trying to sell you supplements or something, but actually trying to help you live the healthiest, best life that you can. Great ending. I mean, it turns out mom was right. Yes, absolutely. It turns out mom was right. Reach out. Hello at aboard.com. Show topics. Need help. Check us out at aboard.com. Dr. Makrijani, if anybody wants to sort of get in touch with you or sort of learn about you, where should they go? You can find me. I have a page. It's called Health Insights LLC. And so you can just Google that domain and you'll find me there. Great. We'll put it in the show notes. Thanks. Thank you so much for coming on. Thank you so much for having me. This was a blast.