Tomorrow's Medicine, Ep. 3: Ben Lengerich - Machine Learning & AI in Healthcare

With Ben Lengerich, ML Researcher. Hosted by Arseniy Arsentyev. Published May 24, 2025.

https://aarsentyev.com/podcast/ep3-ben-lengerich

Machine-transcribed and lightly edited: filler words and false starts are removed, and the wording is otherwise as spoken. Where the two differ, the recording is the record. Times are minutes into the recording. Not medical advice.

[0:00] Arseniy Arsentyev: Welcome to the third episode of Tomorrow's Medicine podcast. My name is Arseniy, and today I'm talking with Ben Lengerich. Ben is an assistant professor at the University of Wisconsin–Madison. Before, he was a postdoctoral researcher at MIT and the Broad Institute, supervised by Manolis Kellis. Right now, his research focuses around context-adaptive, interpretation-friendly machine learning models that help to bridge the gap between messy medical data and actionable medical interventions. Please enjoy our fascinating discussion about computational biology, deep learning and life sciences, and much more.

[0:49] So, Ben, what is computational biology? Just tell us your way of thinking, your definition for a broader audience to understand.

[0:59] Ben Lengerich: Yeah, so computational biology is a very exciting and rapidly changing field. The idea is, when we study living organisms, when we study biology, we can collect a lot of data. And so, to organize the data into human understanding and insights, we need computational methods. So, computational biology is the process of building those methods, those AI models, that can organize data into human insights with respect to biology.

[1:33] Arseniy Arsentyev: All right. And so, what are the different trends that you're noticing right now? And most importantly, what is your advice to aspiring computational scientists, to students, to the students as I am? And how one, especially, I'm curious, how one with a technical, mathematical, computer science background can start in this field?

[1:57] Ben Lengerich: So, students come to the field of computational biology from a variety of different backgrounds. I personally come to it from the background of computer science and machine learning. So, I can speak mostly to that sort of technical side. Other people can speak better to the biological side and the understanding there. So, advice on sort of biologists' level, I'm probably not the best to give. But on the technical side of the computer science, my advice to students is to understand theory and understand how complex systems work together. It's very interesting and it's an amazing time right now as we are entering what some might call the AI era, where these large foundation models, very large institutional efforts, have shown very good predictive abilities.

[2:51] And it's kind of an open question about how the field is going to continue to change and continue to interface with these larger and larger models. But at the same time, theory always stays relevant because it tells us how do we build systems, how do we build robust systems. And so, my advice to students in general is to understand theory well and then look for applications and be adept and be adaptable when you're looking for integrations.

[3:18] Arseniy Arsentyev: All right. So, solid base and then look your way around it. Is it correct?

[3:22] Ben Lengerich: That's right. Yeah, that's right.

[3:24] Arseniy Arsentyev: I get you. As I understood from reading your research papers, you're focusing more on ML methods and deep learning. Is it correct?

[3:33] Ben Lengerich: That's right.

[3:33] Arseniy Arsentyev: Yep. And so, in one of your papers, actually, where you were a co-author, Opportunities and Obstacles for Deep Learning in Biology and Medicine. This paper was around 2,000 citations, if I'm not mistaken. So, good job on appearing. And your main conclusion was that deep learning has not, in 2018, deep learning has not yet revolutionized the field or kind of reached the inflection point where these models can significantly improve the healthcare system. What's about now? We're living in 2025. In this, let me count, in this seven years, how the field changed and what you can say about it. What is the role of deep learning in modern healthcare?

[4:18] Ben Lengerich: Firstly, that paper is an interesting experience. That might be another advice that I would give to students is if you can join a larger effort, that can be very useful and help you learn about the field. So, for instance, in this paper, I was one of many co-authors. I made only a small contribution. But I'm happy to talk about the general ideas and the trends because it's fun to talk about. So, that paper, I think, as you said, was written or published in about 2018.

[4:44] So, deep learning was really having this tremendous excitement and deservedly viewed. So, deep learning is kind of a way of function approximation. So, we have some output and we have some input. So, we're learning how to translate input into output. And with deep learning models, you don't have to specify as much the form of the function that translates the input to output. So, you can kind of learn that more from data. And if you have large data sets, your model can be flexible and learn this very interesting functions.

[5:17] And so, as a result, deep learning has, taken the machine learning world by storm. It's kind of revolutionized many approaches, many benchmarks. The standard toolkit most people's first go to is probably deep learning method and deservedly so. At the same time, as you're pointing out, it's hard to translate ML methods, AI systems, from benchmarks to practical insights and practical, even if we're talking about medicine, bedside realities. So, translating from a biomedical research lab to a clinical tool is hard enough as it is. But then add on the complexity of AI machine learning systems. These things can be even more difficult. So, overall, I'm very optimistic, right? I work in this field. I like machine learning. I like deep learning. I like AI systems. I'm optimistic about them.

[6:16] But at the same time, it's difficult to say yet that they are in the clinic revolutionizing and having their full impact.

[6:27] Arseniy Arsentyev: Okay. So, you're saying that it still hasn't reached the inflection point, but you believe it will?

[6:34] Ben Lengerich: Well, so that's a hard thing. That's a fun question because they are having a tremendous revolutionary impact. And it's just, maybe this is just the tip of the iceberg. It's just going to continue to get better and better and more and more impactful in our lives.

[6:48] Arseniy Arsentyev: Yeah. We'll definitely see. Let's proceed to the second part. It's your research and your lab. The first question is obvious. Tell us more about your lab. What are you doing there? What's kind of interesting things that you're approaching?

[7:06] Ben Lengerich: Sure. So, I'm an assistant professor at University of Wisconsin. My home base is in statistics, but I also have affiliations with computer science and biomedical informatics. So, my students have kind of a blend of interest from the foundational to the applied. I like to work from a real-world problem and then work backwards and try to find out what is the fundamental reason we haven't solved this yet. So, for instance, we might think about observational data in healthcare and hospitals are generating lots and lots of data. And it would be nice to use healthcare systems, electronic medical records as living in laboratories to tell us what's working well and what's not working well and how can we deliver better medicine. But there are so many problems, so many challenges in understanding real-world evidence.

[7:57] So, I like to start with that applied problem and then work backwards and say, can we build machine learning systems? Can we build AI systems that will help us understand real-world complexity? And then that starts bringing up a whole host of other more theoretical questions.

[8:14] Arseniy Arsentyev: Well, that's a nice overview. And you know what kind of struck me is this word, because I'm also doing some stuff in ML, but never heard about this approach. So, your research focused or focuses, I don't know, on context-adaptive models for precision medicine. Could you please explain in a way that sophomore can understand what is context-adaptive models and why especially this type of models is applicable in the healthcare space?

[8:48] Ben Lengerich: Perfect. So, this is a really fun topic from my perspective, so I'm happy to talk about this. You might think about having a simple model class, like a linear regression model. And in a linear regression, we're predicting some outcome, maybe a risk of some disease or a bad outcome like mortality in the hospital. And each input feature gets a linear coefficient associated with it. And this is a fairly straightforward model. There are always, complexities, statistical complexities in understanding. But we have one coefficient for each input feature.

[9:27] And so, we like to think about this linear model as very interpretable. And so, we can understand the data. But maybe it handcuffs us a little bit. The model itself is not very flexible, so we can't actually describe, we can't predict very complicated outcomes. And so, most machine learning people would then say, well, let's get rid of the linear model. Let's remove the linear assumption. And let's fit a more flexible model. For instance, we could fit a deep learning model where we don't have to specify exactly the form of the translation from input to output.

[9:59] A context-adaptive model, on the other hand, says, hey, let's not throw away the interpretable part of the linear model. Let's instead say that the coefficients of a linear model are allowed to adapt to the patient context. So, each person might have a different set of risk factors. They should have a slightly different weighting of the risk factors based on their personal history, their experience, their exposures. That can either be done implicitly in a deep learning model. Or it can be done explicitly by a context-adaptive model that learns to adjust the weights, the coefficients, of a linear model.

[10:40] And so, it's a really fun area to work in, to think about. You can also think about how these linear models are implicit in deep learning models. So, if your listeners are familiar with ReLU networks, or rectified linear units, you can think about ReLU networks are piecewise linear functions. And so, one way to think about ReLU networks is a piecewise linear function that is a map from some input domain to a linear function, also in that input domain. And that's implicit in the whole ReLU network. But another way to represent that might be to make that map explicit and represent it as a context-adaptive model.

[11:28] Arseniy Arsentyev: Okay, Ben. And what if I ask you, to complete the sentence, I believe that context-adaptive models possibly can. What will you say? What is the first implication that comes into your mind?

[11:44] Ben Lengerich: Possibly can predict at the accuracy of deep learning, but with the interpretability of linear models.

[11:50] Arseniy Arsentyev: Yep. And here is a key word, interpretability. Let's talk about this. So, you emphasize the importance of interpretability. Why this is crucial for medical AI, and why can't we go without it? Because before I read some stuff you wrote, I thought it's totally fine that we can have an accurate answer provided by foundational models. So, for those who do not know foundational models, let's think just of LLMs, ChatGPT, Grok, anything. And so, for me, there was only one benchmark, accuracy. But that's not true. So, let's talk about interpretability. What can you say on this topic?

[12:36] Ben Lengerich: Yeah, it's a very interesting question. And I never want to bet against foundation models. Every month, every week, every day, they're coming out with new things, new amazing things that they can do. So, I don't want to be pessimistic at all about the capabilities of foundation models. At the same time, I'm interested in interpretability because I want to understand the world around me. I want to understand biomedicine, and I want to intervene in better ways so that we have better outcomes.

[13:09] And if our goal is intervention and changing outcomes, maybe accuracy is not the only metric right now. So, a few reasons that come to mind when we think about this question of interpretability versus accuracy, although maybe that's a simplified trade-off. We don't want it to be a trade-off. We want both. But when we think about the question of, is interpretability important in a superhuman AI era? Let's just assume, for the sake of argument, that these AI models achieve superhuman accuracy on whatever benchmark we're talking about. And then the question is, do we care about interpretability still? I think that's an interesting way to think about it. And my answer is typically yes. We still care about interpretability for a few reasons.

[14:07] One is that accuracy is not a metric given to us by God. It's not a metric given to us externally. It is a metric that we are measuring, and maybe we are even biased to measure it in particular ways. We are incentivized to say that our model is the most accurate, and how do we actually know that we have learned something fundamental versus we have learned how to predict well on a test set? And we can go through all sorts of standards, data set splitting. All of these are good things that people do, and they make it better. But it's very difficult to know if the accuracy metric accurately reflects the underlying learning that we're trying to achieve.

[14:56] If we have some interpretability, maybe we can ground that we can perform experiments, we can move forward with it. Another reason that I think interpretability may still be important in the superhuman AI era, if we call it that, is that we want good system-level design. So there's a statistic in, I think they frequently use it in baseball, called wins above replacement. And the idea is no individual player. Their individual statistics don't matter that much. What matters is how much they contribute to the team.

[15:39] And if we are designing AI systems comprised of ML models, we don't necessarily care about the accuracy of any one ML model. They should all be good, but it's not that one ML model being epsilon more accurate is necessarily going to change the outcome of the system. So when we're trying to optimize system-level outcomes, maybe we still want interpretability, because interpretability can lead to better interactions between AI agents. It can lead to more modular components rather than just a walled garden. Everything is hidden behind this black box, and whatever happens internally, whatever reasoning happens there. If instead we have an open glass box approach where we can understand what's going on inside the system, we can have all sorts of modular components improving the performance of the system.

[16:37] Arseniy Arsentyev: So the follow-up question is, what is this wall that separates modern LLMs, modern foundational models, from being interpretable? So my question is this computational limits, or technological, or scientific limits? In other words, paraphrasing, what is the unsolved problem in the space? What is the problem that everyone is trying to solve, but no one succeeds for now at it?

[17:13] Ben Lengerich: That's an interesting question. So why aren't these foundation models already interpretable? Is that the problem?

[17:21] Arseniy Arsentyev: And why it's so hard to make the same level of accuracy, the same level of intelligence, I'm not sure it's intelligence, but the same level of quality, that's the word, and still keep it as interpretable as possible so we can see the underlines.

[18:08] Ben Lengerich: For us to be able to do interpretation on these small components, we can only experiment with a small amount at a time: we only hold one variable, we can't understand what happens if we change two experimental variables at the same time. I guess I'm more of an AI optimist in that I tend to think that there is fundamental complexity that may be beyond the human mind for reasoning. And yet, at the same time, we have, as humans, we have wonderful capabilities for breaking down complexity into modular components that we can experiment with. So my hope is that as we continue to build more and more complex AI systems, larger ML models, we build in modularity, we build in components such that these complicated reasoning engines can be broken down, can explain themselves in interpretable components.

[19:10] And you still get a very complicated summarization of complexity, but you can break it down into component level reasoning that is still interpretable. Now, I'm saying you can break down complexity into interpretable components. That might be the fundamental challenge, right? That's kind of what all statistics tries to do is to say, well, you have this data set. Now we're going to represent it in a number of interpretable statistics. But obviously, statistics is a large field. It's been studied for a while, and it's still developing. That's a big challenge.

[19:45] Arseniy Arsentyev: So a quick question. Do you believe that this attention LLM architecture right now is a more fundamental model of reasonings that we understand? Because, as you said, majority of people think it's just some statistics and some putting numbers together, a law of large numbers and making the prediction. So yes or no? Do you believe that LLMs right now possess some kind of intelligence that we do not understand?

[20:22] Ben Lengerich: Yeah, that's a hard question. So I would say that foundation models definitely are performing on accuracy metrics on benchmarks that are well above human capabilities. I don't love calling that intelligence, because I think humans are distinct in sort of a species way. But yeah, sorry, there's a few questions here. One is the transformer architecture itself versus what are these AI models doing and how does it compare to human reasoning? So with the transformer architecture with models based on that, they're not explicitly doing very much symbolic reasoning. That can be baked in after the distribution matching model is trained in an unsupervised massive way. But they're not by default learning to do reasoning.

[21:30] At the same time, the distribution matching perspective says that the simplest way or maybe the only way to get the right answer on a lot of these questions is to do some implicit reasoning. And there are very interesting studies, recent studies saying, the way in-context learning works where you train the foundation model first and then you show a couple of examples after it's already been trained and learns very quickly to adapt to those new examples. There's some interesting work now saying that this in-context learning is actually doing a gradient descent as it passes through the layers of a transformer architecture.

[22:08] And it's hard to say that a gradient descent update isn't a form of reasoning because it's approaching a correct answer via this optimization process. And with that going on, I don't want to say that there's no reasoning under the hood and then combine that with the extraordinary, amazing capabilities of these foundation models. I'm very optimistic about how much they can do. I just, I also want to say that there's a place for symbolic reasoning and that's how humans tend to approach this sort of difficulty.

[22:51] Arseniy Arsentyev: Because, yeah, I'll double all of your arguments. And I want to just speak about my moment when I just like, yes, the LLMs have some potential. Because when this acceleration of ChatGPT started, I started using it. And it was giving me all right answers, but I can straight away recognize that this is a GPT response. That's an artificial intelligence on the first versions. And right now, any questions that I ask it conveys meaning in a way that is, for me, I'm not a native English speaker. Maybe this is the reason. And so, it conveys the meaning the same way a professor, teacher, colleague, friend would.

[23:48] So, for me, that was the kind of point when I can truly say, I believe that these LLMs are intelligent. All right.

[23:57] Ben Lengerich: It's also an interesting phenomenon where we used to talk about the Turing test, right? Can you have a machine generate language in a way that humans can't tell the difference between a human and a machine? And then that whole test kind of fell by the wayside. I don't know that I could tell you which AI model passed that test for the first time. And it just became a less and less interesting test. And now nobody really cares about that.

[24:19] Arseniy Arsentyev: And now we have, last human exam or something like this. How is it called? Like this benchmark of the toughest problems. And, okay, let's get back a little bit. So, first of all is context adaptiveness. And second one is interpretability. And we feed this into the model that these ML tools for healthcare should be personalized. So, based on each patient's history, biomarkers and stuff. But the question that pops up is, wouldn't it overfit on this patient? And in pursuit of differentiation, will we lose the generality?

[25:05] And just how you address, did you read any research on it? Just speak what comes to mind.

[25:13] Ben Lengerich: Yeah. It's a very interesting question and one of the fundamental challenges, right? If we're saying that medicine is personalized and so our statistical models driving medicine should also be personalized. Well, doesn't that go against the law of large numbers, which is the basis of all statistical estimation? That's the question. The answer so far, and one that I'm pursuing, is to say, let's go up one level of hierarchy. Instead of saying the model we are interested in estimating is the linear model, right? That I was talking about earlier that maybe there's these linear models with parameters that adapt to patient context.

[25:58] If we instead shift our focus from that linear model to the higher level of a model generator function, then we can say that that model generator function is shared across all patients. So there's some function that takes in patient context and learns to predict relevant risk factors for this patient. That generator function is at a higher layer in the hierarchy and is shared across all patients. And that allows us to overcome overfitting. We say, okay, we're going to learn this shared function. It goes across our entire data set and maybe even ties data sets together.

[26:40] And therefore, we can get generalizability. Does that make sense?

[26:46] Arseniy Arsentyev: Yeah. Yeah, it actually does. I was just thinking, trying to apply it on some cases I'm working on. So, I want to take a second to thank you all for listening to this episode. Tomorrow's Medicine Podcast is brought to you only by my desire and efforts. The best thing you can do to support myself is to become a member at my Patreon page, share this episode with your friends, like it, and leave a comment. Please, continue to listen to my episode with Ben Lengerich.

[27:27] We can transition a little bit and talk about industry versus or with academia. And I'm interested right now in computational biology, in machine learning, in medicine and biology. What is this trade-off and what can you say? What type of people go to academia? What are the opportunities open in the industry?

[27:53] Ben Lengerich: Yeah. So, industry versus academia is an interesting and difficult question. Of course, I am an assistant professor. So, I am biased to say that academia is a wonderful place and that's where I've chosen to be. And that's all true and I believe it. But at the same time, I don't know firsthand. I don't have a long experience in industry, so I can't talk about that too much. One interesting difference between industry and academia that I've seen so far is that... Actually, I'm not sure that's true.

[28:49] I don't know that I want to weigh in on that. I think there are good opportunities in both. I've seen, very successful people go towards both. And then they often switch between the two paths as well.

[29:05] Arseniy Arsentyev: Okay. So, then let's get back to academia and focus on it. And how was it opening your lab? What challenges you faced? What kind of aha moments you had? How has your life changed after opening the lab? That's what I'm saying, basically. I cannot rephrase it better.

[29:32] Ben Lengerich: Sure. So, prior to this, I was a postdoc at MIT in the wonderful lab of Manolis Kellis. And then prior to that, I did a PhD in computer science at Carnegie Mellon, where I was advised by Eric Xing. So, that was a good sort of academic training. I got to see a number of different mentoring styles and get to think about how I want to set up my advising system. So, it's been very fun here at Wisconsin setting up a group. I'm just finishing my first year.

[30:11] It's been fun. We have a couple of PhD students and several undergrads. There's a lot of enthusiasm. It's fun to be a new professor, with a lot of enthusiasm from myself and from the students as well. And we get to talk to each other a lot and push each other. And that's very, very fun. Let's think. Surprises in setting up a group. Not coming to mind. Surprises of being a professor, I can talk about that. Some of those come to mind.

[31:12] For instance... This past semester was my first time teaching a course. And it was fun. I taught graphical models and probabilistic AI. That was the title of the course. I made it up. And it was very fun to be able to go from conditional independence to modern LLM architectures in an undergraduate course. So that was a lot of fun to teach. One of the surprises there was that I was actually able to talk for an hour. Before being a professor, you listen to people talk for an hour and you say, wow, they must have a lot of material prepared.

[31:55] And it was kind of interesting. It's a busy life, maybe, as a professor in your first year, getting things set up. So I didn't have as much time to prepare for lectures as maybe I would have liked. But I was still able to talk for an hour and have questions and go back and forth. And the students at Wisconsin are also great and very inquisitive. So that was fun. Yeah, I guess that was one surprise.

[32:22] Arseniy Arsentyev: Okay. So how many students was your class?

[32:26] Ben Lengerich: Oh, that class was about 40.

[32:29] Arseniy Arsentyev: 40? Wow. It's kind of tough. And you know, the toughest thing when you speak is not when you speak because I can make up like any sequence of words or sequence of discussion tracks I want. But when somebody really smart asks you a question, I was teaching Scratch for kids as part of my volunteering. And sometimes kids like ask questions about what is the loop with this variable and you're like stuck for a moment and just spending five minutes thinking about it or just Googling it to find an answer. So I think you also experienced it with your students when you like something, something they say is so new that you have to think about it.

[33:19] Ben Lengerich: Yeah, the good news is, I guess, if you're teaching a 400 level class juniors and seniors, and you position it as hey, we are talking about the state of the art, it's okay for things to be unknown, and we have to go find the answer or something like that.

[33:34] Arseniy Arsentyev: Yeah, that's great. All right, I'm thinking, where do you want to go first to the material you send me yesterday, I believe this slides? Or for like a broader discussion about healthcare as a system, not from the perspective of ML or computational biology, it's just like healthcare. So choose whatever you're more interested in.

[34:00] Ben Lengerich: I think I'm more comfortable talking about the technical foundations.

[34:07] Arseniy Arsentyev: Okay, I'll try to outline my impression and how I understood this whole concept. So you're saying that there are certain discontinuities and certain patterns in any kind of causal data in any kind of causal biomarkers that breaks this notion of causality. So we can observe them by hand if we plot several variables. I did a nice drawing. I'm a good drawer. So let me zoom in. Okay, so for the audience, here you can see the serum creatinine, if I'm not mistaken, and the mortality rate. And this I considered normal for persons. This level shows concentration. This one is unhealthy.

[35:08] And higher concentrations are also at normal risk level. So it just blew my mind, how increased concentration on something can be, can lower the risk of survival, whether it's an even higher concentration, or some biomarker can lower the risks again. So can you please expand on this concept?

[35:39] Ben Lengerich: Yeah, exactly. So you're talking about the dangers, but also opportunities of real world data. So the example you were giving there was about serum creatinine, which can be an indicator of kidney failure. So as you increase your serum creatinine, maybe your kidneys are failing. And as soon as I say that, probably listeners who have more medical background are thinking, well, that's not always the case. There are a lot of complicating factors. And maybe our measure of serum creatinine can be adjusted by different other factors.

[36:11] And there are lots of interventions around creatinine that matter. And those listeners are exactly right. And that is the problem we're seeing here, the challenge and also the opportunity. That is, if we think about creatinine as a measure of kidney failure, it doesn't actually predict in-hospital mortality. In fact, in lots of data sets, the patients who are at lowest risk of mortality are those with extremely elevated serum creatinine levels. And that is associated with dialysis and continuous renal replacement therapy, CRRT. So a purely data driven AI system that just looks at electronic medical records and tries to reduce mortality might not understand the difference between biological healthy state and low risk healthy state versus low risk because there are well known interventions.

[37:18] If there are well known interventions that are routinely given for a particular unhealthy state, then that unhealthy state now actually looks healthy to a data driven system. We need some sort of randomness to tell the difference between a truly healthy state and an unhealthy state that is routinely treated very well. Of course, in medicine, you're not going to run a randomized trial if you already know how to treat an unhealthy state. So as a result, data driven systems can learn the exact opposite conclusion. They can say, hey, this unhealthy state actually looks very healthy because these patients get good treatment.

[38:04] And unfortunately, this is kind of doubly problematic. It is not only getting the answer wrong, but it's getting the answer wrong for the patients we could help the most. So this is this is one of the classic challenges of learning from real world data. It's observational, not interventional. So we need to do deconfounding to figure out what's the impact of the treatment versus the impact of the biological causality. In some of my work we have proposed, maybe we don't have to do the deconfounding. Maybe we can just flip this problem on its head and say the strangeness in the real world data indicates opportunities for improved medicine.

[38:52] So if we were equally treating all patients in the sense that everyone got the same post treatment mortality odds. Then we would have a flat risk curve, the risk of mortality with respect to whatever biomarker would be flat. That's one notion of equal treatment. It's probably not actually our goal, but we could take it as a baseline. On the other hand, real world data goes up and down. The risk in real world data goes up and down, just like we're talking about. There's often a healthy state and then some effective treatment that actually drives the risk back towards the healthy state or even better.

[39:35] If we are in the case where the unhealthy state, the unhealthy treated state actually has a lower risk than the healthy state. Then we have inverted causality. We have some sort of paradox in our real world system. And we know that there could be better interventions, better targeted interventions. We could spread the outcomes better among the population. It doesn't tell us exactly what the right answer is. How do we actually find the right answer? We need to run trials. But it tells us that there is a counter causal paradox in our real world data.

[40:11] And if there is a counter causal paradox in our real world data, then we know there is a group of patients that are not receiving as good of care as they could be otherwise. Does that make sense?

[40:26] Arseniy Arsentyev: Yeah. And so is the solution to create kind of a third model? So imagine here is our some ML health engine. Here is the data. And do you think that we need something in between that will be kind of a catcher of these counter-causalities?

[40:45] Ben Lengerich: Yeah, exactly. So I was talking earlier about interpretability, about how humans break down complexity into modular components. We can imagine a very similar system that is automated instead of having a human do the data analysis and a human do the statistical fitting. We can take our statistical models, break down the complexity into modular components and then ask some foundation AI to analyze these components for surprises. And this is kind of what we're getting at just now is that when there are surprises in real world data, that often means there is opportunities for improvement.

[41:29] Arseniy Arsentyev: So have you tried doing it? Do you have some statistics, something interesting?

[41:34] Ben Lengerich: Yeah, yeah. So we've done this on a couple of different data sets. Some of them are private data sets, but some of them are more public, like MIMIC is a standard benchmarking data set in EHR. And we found, once you automate this process of looking for surprises, there are surprises everywhere. So there's this classic results of saying, hey, patients who come to the hospital with pneumonia, if they also have asthma, they have a lower risk of death in the hospital. That's kind of a classic result that says, hey, if you have this comorbidity of asthma, then maybe you take your pneumonia more seriously and you come to the hospital in an earlier stage or with a less severe infection. And so data driven risks look like asthma is good. It is protective against pneumonia.

[42:26] So that's like one classic example. But when we automate the system of looking for these surprises, of looking for these counter causal paradoxes, we find them everywhere. So it's not just asthma that that does this, but all sorts of histories like a history of chest pain. If you have a history of chest pain, then data driven risk says that actually your likelihood of dying with pneumonia is lower, probably because you've taken your chest pain seriously and you've gone to the hospital and you understand that things can be more serious, more severe, more quickly.

[42:59] Interestingly, we've benchmarked against doctors looking for surprises, experts doing data analysis. And we found that on the models that we've tried so far, although the models are always changing, the foundation models are always changing. On the models we've tried so far, the doctors are better at individual anomaly detection. So the anomalies they find tend to be more grounded in scientific knowledge and more grounded in the data. However, the doctors conserve their efforts and they only look for surprises in features that they kind of already think are going to be surprising.

[43:46] The LLM, on the other hand, has no notion of effort, right? It'll do whatever you tell it to do. And so it will look equally at every feature. And so it actually finds almost twice as many surprises in real world data as do very good expert data analysts. And I expect as the foundation models get even better, this gap of the quality will reduce and maybe the foundation model even becomes superhuman in this data analysis test.

[44:12] Arseniy Arsentyev: And are you doing it multidimensionally? Because I can imagine doctor or even myself making such graphics about two or three variables, maybe like somehow have intuition for four. But so how many dimensions your AI models look at?

[44:32] Ben Lengerich: Right. So in this experiment, we made the graphs for the doctors and we made the graphs for the LLM to look at so that they would be on equal footing. In this data set, we looked at 46 different features. So they had 46 different graphs, just like you drew there to look at. And you can imagine, of course, going forward, that now there's going to be interactions between these 46 features, which humans are quickly not going to be able to verify.

[45:03] Arseniy Arsentyev: But that's interesting. Imagine if we would find this kind of surprise on dimension 46 and somehow outline it for a person. So that's that's another level. Have you seen something like this?

[45:24] Ben Lengerich: So we're getting there. Nothing is currently in the clinic to say, oh, we don't pay enough attention to this feature. Another interesting finding, by the way, from this pattern is that humans tend to conserve their effort in not looking at or not analyzing deeply features that came back with zero coefficient. So if your statistical model estimates that a feature has no impact on the output, humans tend not to think about that very deeply. They skip that analysis. But LLMs will think, hey, I had a prior expectation of this feature actually positively contributing to risk and it's coming back at zero. That is surprising. That's something surprising. So that's kind of an interesting pattern that we see repeatedly is that on the zero effects, the LLM really looks and the humans tend to skip it.

[46:16] Arseniy Arsentyev: So, okay, building a complete picture: context-informed personalized models that have some kind of backdoor for interpretability and for understanding what's inside it. So, also that have certain anomaly detection. We got this picture. How do you believe this would fit into the hospital that we have right now? So would it be kind of an assistant also chatbot for a doctor? Would it be some online tool? Yeah, more broadly in the terms of healthcare system. How do you envision it?

[47:05] Ben Lengerich: Yeah, it's a difficult question. I think there are maybe two ways to think about this. One is at the population level and the second is at the patient level. So the population level we are fairly comfortable with as academics, right? We have some data set, we analyze it, and we can write some some takeaway in a paper. And I think that's very promising. That's a good direction that we should continue going down, to say, we have these large models, good representational capacity. We can take context, we can build complicated models of the predictor, but we can also interpret them. Well, we can find surprises in data at the population level, and then we can give recommendations to the system about how to improve or what tests to run to actually measure improvement.

[48:06] The second, the personalized medicine way, is to say, based on everything we've analyzed at the population level, now you specifically have these factors that may contribute to or take away from risk, and they lead us to recommend you taking these steps. And I think that takes being deeply integrated in the healthcare system, right? So maybe this is built into electronic medical records, maybe it's built into insurance systems that do kind of follow ups or preemptive communications with patients. All of these are very important steps that I think are being done, but it's a little bit harder to do those from the academic seat.

[48:54] Arseniy Arsentyev: I guess so you think more kind of people closer to administrative tasks are the ones who will figure this out how to put it into work, and how to make money work.

[49:08] Ben Lengerich: Yeah, I think for the personalized action items, that really has to be deeply integrated in the system. The EMRs need to be accessing the AI system. And there are efforts towards this end. But healthcare is private for very good reasons. As a result, you need to build this from inside the system, pretty much.

[49:41] Arseniy Arsentyev: Okay, and speaking again about money, I forgot whose saying this is, like, chase the money and you'll find the truth. And who's right now paying for the research? So how do you how do you feel? All the things that you do, context-based personalized medicine: is there more corporate or private sector sponsoring, or is it, as it should be, more kind of governmental initiatives that fund this all?

[50:13] Ben Lengerich: Okay, well, yeah, interesting as an academic, I guess I'll say if anybody listening wants to fund me, please reach out. We're not going to turn away, interested donors. So who is funding this sort of research? There are a good number of efforts in corporations, either insurance companies, health innovation systems, healthcare partnerships, venture capitalists that are investing in sort of an AI vision of medicine, and they are kind of building the architecture. And that's a good number of efforts in the healthcare system.

[50:55] On the academic side, right, I think maybe there's more towards the population level, understanding the risks and improving care, looking for equity, all these sorts of concerns. And in the US, this is mostly driven by US federal funding, like the NIH. There are also support from corporations, but maybe there's less for healthcare than for computer science based methodological innovation.

[51:22] Arseniy Arsentyev: So the balance is shifted towards the government for the research.

[51:28] Ben Lengerich: I don't know that it was ever any different. I think for a lot of academic research, it's a lot of it is supported by federal funding. And then for kind of integration with the system, a lot of that is more corporate funding.

[51:45] Arseniy Arsentyev: Because it's so interesting to see how different entities, different governmental organizations change their proportions and amazing changes in the academia and in the industry happen. And speaking about funding and stuff, I literally, in the place where I'm at, so I literally know a computational biology lab that is fundraising $2,000 and cannot fundraise it for a month. They literally started a month ago fundraising campaign, and they cannot fundraise $2,000. And like, literally, I believe hundreds of investors, like look at it, and they're like, probably better invest in oil or something like this.

[52:38] Ben Lengerich: Yeah, I think you're outside the United States. And I think the funding mechanisms are probably country specific. I know there are certain countries that have, kind of a baseline, they all have different grant mechanisms for academic institutions, some are more generous, some are less. And yeah, I don't know that it's easy to generalize across the world.

[53:02] Arseniy Arsentyev: And taking my chance to ask you how this works, do you have certain goals that you have to achieve to keep the funding? Or even if all the research fails, it's not like credit score, your research failed, you're on a blacklist. Is it comparable? Or am I wrong?

[53:28] Ben Lengerich: No, so in the United States, they are fairly protective, I guess, with the tenure system, which tends to say, the idea is, hey, if you've shown you can do good research, then maybe if you have some slow years, you're probably doing creative research, rather than just not being able to do research. And alongside that, funding can be hard to get. And there are all sorts of ways to get funding. Different countries have different approaches. I talked to some friends in Switzerland who just have like a baseline funding from the government, whereas in the United States, it's more of per project that you want to get done, you write a proposal and they fund you.

[54:14] And it's public money, right? It's taxpayer money. So you have a moral obligation and a legal obligation to do the research that you have promised to do. Beyond that, I don't know if there's any system for counting outcomes.

[54:34] Arseniy Arsentyev: Okay, and speaking about territorial resources or like location based, is it true that US for now is the center of computational biology and machine learning? It's the first. And a little deeper, can you name some universities or some research centers that are like, if you want cutting edge, you go there? I know you're biased, of course, to say it's UW–Madison, but yeah.

[55:04] Ben Lengerich: Yeah, right. So I am American. I was born in the US and I've worked my whole life in the US. So that's what I know best. I am very happy at Wisconsin. I did postdoc at MIT. That's a great place. I did PhD at CMU. That's a wonderful place, too. I have friends all over the country; there's a lot of very good universities. My European friends would probably want me to list some of their universities. I actually collaborate with folks all over the world.

[55:41] So some people that I work with are based in Egypt or, yeah, lots of places around the world, we can just, with a remote connection, we can just pick up and work. And that's very, very nice.

[56:02] Arseniy Arsentyev: Okay, so, because when I'm speaking about Europe, only places that come up to my head, some like, UK-based, maybe like Cambridge, Oxford, and ETH.

[56:17] Ben Lengerich: Yeah, those are great universities. I don't know that I'm a ranking expert or anything like that.

[56:25] Arseniy Arsentyev: Yeah. Again, I was at the MIT as like, visitor, touring it, and had... When I was going there, I thought I would think that I'm in the place like where Albert Einstein, every second is Albert Einstein, but what I've seen is completely a different picture. And there is like different thought kind of strings. So, for example, I haven't seen it on campuses of other universities. At MIT, there is literally this perception that people from completely different subjects work together. And that they kind of always work on intersection of something. Is it true? And even if it's not, can you tell us more about your time at MIT, being a postdoc in the Broad Institute and CSAIL with Manolis Kellis?

[57:42] Ben Lengerich: I'm happy to talk about MIT. MIT is a great place, special place. I think the people are great and the resources are amazing. That being said, people are people everywhere, right? So that's true. And research is research everywhere. So if you want to do great research, you can do it at MIT. If you want to do great research, you can do it elsewhere. What I found really special about MIT was being surrounded by people that kind of held themselves to a very high level of expectation of excellence. And maybe that bleeds into your own thought processes. That being said, it doesn't change, the research, the work that you're doing. That same work can be done elsewhere and is done all over the world.

[58:34] It's just when you can be surrounded by friends that have kind of an expectation of excellence, that can be helpful.

[58:43] Arseniy Arsentyev: Yeah, friends and family. That's actually very true. And I kind of faced this time when I really made the priorities clear to myself where I want to be.

[59:01] Ben Lengerich: I just want to be encouraging that MIT is a great place, but there are lots of great universities all over the world. And even, people doing independent research can do, amazing things these days. So I wouldn't say we should just hold out hope that you get to a institution, then you do it right. It's you do the work and then maybe the institution is the right place for you. And maybe it's not.

[59:28] Arseniy Arsentyev: All right. We're approaching one hour landmark. And I want you to just put your imagination, put your kind of heart, soul and tell me what is the healthcare that you wish to see in 10 years, not like in 1000 years. But what are, in your opinion, possible changes? And like how you wanted it to be just any anywhere ranging from the hospitals, how you get to your doctor, how's the treatment process? Anything you would like to discuss on this topic.

[1:00:09] Ben Lengerich: Wow. Yeah, that's a fun question. And I am not a healthcare administration expert or a legal expert. So this is totally imaginary.

[1:00:18] Arseniy Arsentyev: I can go first. It's always simple. I told my friend Oscar, I want this sci-fi machine that you just lie into, it's not in 10 years, in 100 probably. And it will tell you like, eat less cookies, exercise more, you'll be fine. You'll live to 150. That's like my future in 10 years. I wanted that. I have my only wish that all of my data stored in one place or somehow like put into standardization because right now medical data is a mess.

[1:00:57] That's my only wish for the next 10 years.

[1:01:02] Ben Lengerich: Yeah, that's very, very interesting. So maybe I would say, just thinking about this for the first time, but maybe I would say that I hope in 10 years, we understand how to return medicine to being a form of an art. Maybe this is a strange thing to say, but a lot of people, a lot of medical doctors that I've talked to have said things like, oh, well, medicine, is an art. And it is you know, it's a very difficult subject with a lot of technical knowledge, but it is also knowing how to apply that technical knowledge in the real world with real people and connecting to real people.

[1:01:50] And I am optimistic that as we build better and better AI tools, as we build better and better machine learning models to predict outcomes and explain them and build systems that talk to each other. And I don't know that we will be replacing doctors, or nurses or anybody in the healthcare system. But maybe we can build a system where the complex diagnoses, the complex calculations that humans are not great at can be offloaded such that medicine can return to being a form of an art and a connection and being personal and holistic. And right now, if you go to the doctor in America, they spend most of the visit typing on the EMR and nobody likes that system.

[1:02:38] So I think there's a lot of work towards returning the doctor patient interaction to being face to face. And yeah, I'm optimistic that as our computational methods are getting better and better, as our assays are getting better and better, this is all going to work together and empower, empower the medical professionals.

[1:03:00] Arseniy Arsentyev: All right. It was such a great conversation, discussion that truly provoked my kind of aspirations and wake me up to reality a little bit. So thank you very much, Ben, for attending this podcast. Everyone, support me on my Patreon page, subscribe to Tomorrow's Medicine podcast. And let's just accelerate from here. Have a great guests and great discussion. Thanks, Ben.