Healthcare AI Is Failing And the Fix Isn't More Data

If you're on Netflix, if you're finding a documentary that no one knows about, right, it's on the fifth page, you complain, you shrug, you move on. But if you're in a self driving car and your car cannot detect an auto distribution event, the consequences could be quite profound. Welcome to Sasnat App Building B2B Web Applications, the podcast where we share real world stories, practical advice and tech insights for those building or thinking about starting a tech enabled business. I am your co host Aaron Marchbanks and I'm Justin Edwards. Each week we bring you the stories, strategies and insights you need to build your SaaS or tech enabled business smarter, not harder. Let's dive right in. Hello and welcome back to everyone. It is SAS that AppTime. With me as always, in his appropriate as always shirt of the Dunning Kruger Society is Justin Edwards. Justin, how are you doing man? Oh man, I don't know. SAS AppTime is fun, but SAS Naptime sounds pretty good. Oh my gosh. Yes. It's midweek, it's after lunch. Our feels like index crusted 100 degrees today so I'm now paying the piper for the heat down here. Oh my gosh. Well I will happily get to enjoy it with you some very soon. Yeah, excited to have you into town. Yeah, it's going to be exciting. I'm glad to get down there. My girls are giving me grief because they're like well we're going to be at the pool. It's like well I'm going to be at the ocean. So there my body of water is larger than yours. Your move. Also joining us today, we are excited to welcome Dr. Abhishek Jha. He is the co founder and CEO of Elucidata which is a company focused on transforming life sciences R and D through AI driven data platforms. Abhishek has over 20 years of experience at the intersection of life sciences, data science, machine learning with a career that spans deep scientific research, drug discovery and building technology to accelerate how breakthroughs reach patients. Today Elucidata is focused mainly on building high quality AI solutions that are focused on out of distribution observations and small n. They have built a data centric AI stack on their platform Poly which we will definitely get into a little bit that helps accelerate precision medicine. Dr. Shah, welcome to the show. How are you? Hey thanks Arun and Justin. Excited to be with you guys on this podcast and thanks for sharing this platform with us to bring our story to the world. And if I mention like I will deliver on the SaaS that NAP promise outstanding I got one question right out the gate which is can you tell us a little bit about elucidata and what exactly it does and who your customers are? Yeah, no, happy to. We founded this in 2015, around 10 years back now. And what we do is we build AI solutions and our key thesis is that the bang for buck is by focusing on the data side. That's an underserved market. People flex and talk a lot about models, companies, investors, media. There's a huge obsession about how many tokens you have in your parameter and that's important. That's definitely helpful. But I think an underserved need is data that goes into those models. Right. So we call ourselves a data centric AI company and we have built such solutions and that stack for a very focused vertical, which is healthcare and life sciences, mostly life science. So our customers are large organizations like you may have heard of, like the Pfizers of the world and small biotechs in Cambridge, both new and old and south sf. That's the market that we serve. So you spent obviously the early part of your career and certainly schooling being a scientist in a very broad sense. So I'm curious what got you from being a scientist, being someone who is in the lab, being someone who works for that type of a company, to being a founder for basically this industry that you're now getting into, which is very AI driven. Tell me about that job. The short answer is my naivety and my curiosity. That's what got me here. I never thought I would be on this journey that I have been on for most of my life. I wanted to be in academia and I trained for that. I spent first 30 years of my life to get there. At some point when I was at mit, I realized that's not going to happen in a time frame in a way that suits my life. There were a few other considerations, confluence of personal things and professional things which made me rethink my choices. And it was a great place to rethink because you know, you're right in Boston and a lot of high quality startups that are coming out almost on a weekly monthly basis. This is like late 2000 odds. And luckily enough, as you would guess, one of my friends who had moved to a postdoc position in Harvard, his PI was starting a company and they were looking for someone with my phenotype. I thought like, sure, you know, it's worth trying. And in hindsight I think that was a great call. Right. So that was a big transition and Agios, which was my first Job out of MIT was really foundational in my journey as an entrepreneur. My currency changed from publishing papers or going to conferences to really own a very tiny part of a much bigger picture, right, which was a patient. How can you help a patient? And Agios, again, I can talk for hours about how incredible that experience was. One of the things that I remember, and I will take it with me forever, was our CEO at Agios. He was a practicing doctor. He would see patients every so often. And he took a lot of pain for us to remember that, you know, everything that we were doing, writing code or git commits and slide decks and experiments were for a patient. And that was very transformational. It not happened in one meeting or one pep talk, but it was very visceral, right? Like, for example, when you walked into Agios, there were pictures of people who were lost to cancer. That was a very visceral reminder of what is at stake. Every once here, when we had our off site, we would have a patient who would come and talk to us. And that finally led to Agios finding four drugs. All four of them were first in class that we brought to the clinic. All four of them may have been approved by the fda. So that really changed my currency to think more about patients and how can I help in my own way, how tiny that is, right? To make a meaningful contribution there. And seven years later at Agios, the company was in a different place. I was in a different place. I thought like, hey, look, from very basic first principles, right? I don't have the tools and services that I need to do my job better. I'm not alone. There are many people like me. And the third point was, it'll only get worse with time. So that was a very basic first principle analysis that I thought, like, okay, let's get on this path of entrepreneurship. And it has been an incredible privilege to be doing this for last 10 years. I would attribute that to my naivety. I had no idea what I was signing up for. I was very curious. I thought, like, let's try it out. My wife was very supportive. Ajios was very supportive. I was very lucky to find a good, very good, outstanding co founder. And here we are. That's fantastic. I'm trying to think back, Justin, but I. I think Dr. Cha is the first medical professional or someone even in the medical field that we've had on the show. And in my mind, that puts you in a very, very small group, obviously, but also an extraordinarily important one when compared to other organizations and certainly other pieces of software that are out there. But it's interesting to me because the approach is very similar. You know, you had a passion, you had a curiosity about something. It was missing a set of tools or a solution or a utility, and you took it upon yourself to go fill that gap. It's a very common arc that we see, you know, across all of these different organizations. And I, I doubly love the fact that the people who find it from being in that type of work. You are a medical professional, you are a scientist, and, you know, that is what led you to the technology rather than the other approach of I'm in technology and I'm looking for a problem to solve, which is very cool. So talk to us a little bit about elucidata itself. What did you learn getting that off the ground? What was the initial venture there? You saw a problem. Tell me about that. Yeah, and before I go into that, I just want to comment. My medical professional friends will not defy me. As such, when you get on a flight and you have to identify yourself as a doctor, I do make a qualification that I'm a usage doctor. I cannot help patients. I'm too far removed from that. Right? No. But as I said, the genesis of Felicita was based greatly on my experience at Agios, which was a very exciting front row seat to see my colleagues at my company work on a problem over years and bring a meaningful product for patients. And that plus some basic first principle analysis. Right. Of a need. At that point, when we're starting, we had a choice. One was to raise capital and build a product. Right. The second was to just dive into the market and build what the market would actually pay for. Both of them have their own pros and cons. For whatever reasons, we chose the second path, we bootstrapped the company. And of course, you know, there were some favorable things that happened. Like, for example, Agios signed us up right away for some support. Right. On day one. So we had some cash. We were cash flow positive on day one. Right. That allowed us to bootstrap the company, be fiscally disciplined, be very thoughtful about whatever we build is addressing a real need, which the best proxy for that is someone's paying you for it. And that became, in hindsight, a very critical part of our D day. And we have raised capital three times since from investors, some of the world's best investors. But we have been very disciplined with deploying capital. Right. To understand where it leads to growth. Right. So we had a thesis as bullish and as hopeful we were about the thesis, we were very clear about not, you know, drinking our Kool Aid and like be not romantic about it. And we just tested out in the market, right? What does the customer think? And that has served us really well. Right? That's how we started. Real quick. This episode is brought to you by Delta Systems, which is what Aaron and I do when we're not talking through microphones to you, the people of the Internet. We've got a really, really great software team here and we love to work with cool people on cool projects. So that sounds like you and you've got a problem or you're in some kind of a jam, go to deltasystems.com grab a time with us, we can beat up on your problem together and if there's a fit there, amazing, we'll help you out. So Deltasystems.com grab an appointment and hey, maybe we can work together. So you talked about being data forward or being less focused on models and more focused on data. So I assume that your customers, you need to go to customers that have maybe not a lot of data, maybe unique data or data that they're not currently using. I know that you mentioned that you work specifically with small n and out of distribution data points. Can you tell us a little bit more about working with maybe smaller sets of data and unique sets of data or how that kind of comes to work for you? That's a big part of our thesis, that that's how we have built our stack and our people and processes. The argument is quite simple, right? The kind of data that tech companies and traditional AI at large has built on, right? Think about search, advertising, product recommendations on a marketplace like Amazon, right? Insane amount of data which is just not a possibility for healthcare and life sciences. So even think about the best, richest clinical trial that has ever been run. It's been tens of thousands of patients, not tens of millions of patients. So you have a small end to begin with. Compare that to, you know, Facebook has, you know what I think half of the world, 3.5 billion people unique logins on a daily basis, right? And all their clicks, journeys on the platform, right? The scale is just off by orders of magnitude, not even close, right? And if you compare that to the best, richest clinical trial, that's tens of thousands of patients. And then there's a huge number of rare diseases, small cancers, which that's where the median is. So you inherently have a ceiling of how much data, how many ends you can have to begin with. And that's a common objection that we hear about, is AI going to be even meaningful for us or not? That's one part of reality where I think traditional AI just does not translate very well. The second one is that output distribution is a key tenet, very well defined mathematically, a key assumption of any supervised learning model where if you are trying to predict on data that it's outside of the training distribution, you will be compromised. Your model will see proofs. That's a very established bleeding edge of AI, period. Right? Is this just an academic point or is this has some real consequences? And that's something that we talk a lot about, right? So if you're on Netflix, if you're finding a documentary that no one knows about, right? It's on the fifth page, you complain, you shrug, you move on. But if you're in a self driving car and your car cannot detect an auto distribution event, the consequences could be quite profound, right? Similarly in healthcare and life sciences, right, if there is a group of patients that are not responding to a drug which was doing really well in pre clinical settings or phase one, right. My worldview is that most likely it's a distribution that is different, that biology, that label has shifted, that it's the same disease, maybe the same tissue, but those variables are not going to completely define your distribution and you're in a. So both of those things are key aspects of our vertical that we're focusing on. If you buy that, the third point we make is that traditional AI will just not serve because it over relies on scale. So what do you do, right? How do you drive value from that? And our this is data centric AI. We are not the first ones to talk about it. Most famously Andrew Ng Coursera Founder, Stanford he's been talking about it. We have learned a lot from his blog posts and interviews and such papers. But that's what we bring to focus in healthcare and life sciences. Again, these are very simple arguments. If you buy that, which we do, then the stack that you build is very different. How do you find data that matches your use case? Is that multimodal? How do you clean it up? How do you link it to each other? Because even if you have small N, the advantage that you have in healthcare lives, you have lot of data, you can have lot of different types of data for that small lid. And if you know how to link it up, this you can reinforce the signal from a very noisy background, right? How do you choose a training data set? How do you prepare for model to be generalized? How do you test it? How do you add trials to that? So There's a whole stack that we have built which are very data centric, which we have showed we have just submitted in Europe last week to consistently improve model's performance without changing the model at all across five different architecture types. That's the stack, that's our focus. Does that mean models are not helpful? Most certainly not. They are very helpful. But the argument that we are making is there's a whole broad community of academics and companies that are publishing models. We want to shine the spotlight on the data part which again becomes very critical for classic ML and also for large scale models as well. So for the people that you serve, I'm going to guess behind the scenes, but I'm not going to tell you what my guess is. Are you kind of an outreach sort of an approach? We do this thing and we think it might be useful to you. What do you think? Or are you all more closely aligned with like a referral based model or a reference model where, you know, people are like, hey, I have used their service and it's really great. You should too. Is it a combination of those two things or is it one more heavily than the other? It's at this stage of our journey, it's a combination of both dominantly customer references, right? Like, you know, I think it's very noisy out there, right? So if I put myself in the shoes of our customers, right? Everyone is talking about data and AI, right? Like it's kind of exhausting. I get exhausted by that. So any social evidence, prior experience or a friend who has worked with us, right? Anything that allows us to cut through that noise helps us a lot, right? That could be a paper in Europe, which is a probably good paper, right? Then it helps you cut through the noise that hey, these guys have done something real, right? So to answer your question, it's both, but dominantly it's a small group of people who know us, have drive value from us. They take us to their new jobs, they take us to their friends and colleagues, right? That has served us really well. But we also have something more scalable that we are reaching out to. But it's very noisy out there, right? Very. Yeah, yeah. Well, related to that then. Talk to us a little bit about your pricing model, how you put this out there. I mean, is this a licensing type of a situation? Is it a one time use sort of a situation? Is it an X number of uses or per study? Something along those lines maybe? The use cases are so varied, right. We have embraced a managed services model is some combination of subscription plus Access to our experts that is again driven by more of a first principles approach. That's what our market understands, that's what they see value in. So there's the projects which end in three months, there's some engagements that have gone for six years. And I think that is the reality for AI solutions companies. I think the classic traditional SaaS model per seat based thing just does not translate very well. We have certainly lived it firsthand. We have experienced that we have changed our pricing models over the years from a subscription, annual subscription of these many seats and whatnot. That just was not meaningful. We're trying to force fit a model that was not natural to our offering in our market. There's a very interesting blog that came out from A16Z a few years back which argued that AI companies are struggling with SaaS like models because it's very, you know, there's some purpose built configuration that you have to do for almost every use case. Right? So the combination of the offering that we have and the market that we serve, it's a managed services model that we are using to deploy, using to deliver, and using to price it as well. Got it. I'm also very curious because you are in a field that touches a lot of data that is very sensitive. I'm a magic, so. And I realize that a lot of times by the time it gets to data modelers and things like that, it's stripped of a lot of identifying types of data. But I'm curious how you all do handle your data being as sensitive as it is within the studies that are going on across studies that you might be leveraging things along those lines. So a couple of comments here, right? There's a huge amount of data in public domain that is extremely valuable and we always draw the attention to that. Right? There have been very successful large enterprises that have been built on public data load, most famously Google. That is perhaps not to that magnitude, but that's definitely true in biology as well. Nih, the biggest funder of R and D activities, publishes so many good quality papers every year. Right. There's a huge wealth of data there, right? Number one. Then there's a class of restricted access data, right, where you have to write an application as to why you need that data. NIH again supports that. Other countries do that too. I think what you're talking about is propriety of our customers which are very sensitive. We handle that too. And there are at least two ways to do that, right? One is contractually we commit to not using the data to improve our model or anyone's model, etc. Etc. And technology, we have all the certifications like HIPAA, SoC2 and others. Right. And any personal data, very well defined, they're stripped. Or even if it's not personal data, like there's a lot of sensitivity around compound structured data that companies are using to find new drugs. Right. There's a lot of IP associated with that. So both contractually and technically we are capable of handling and addressing that. But that is a very legitimate concern, very specific to our industry, I think, not unique to our industry. I'm sure financial industry must have similar burden, but we certainly do and we have very actionable responses to that as to what kind of certification we have, what kind of contractual obligations we'll commit to, to respect that. And finally, it's most important for our business, right? We'll be, you know, our business if there's one such glitch, as important as it is to our customers. Ip, Right. It's quite existential to us as well. Are you typically using models that your customers already have and then you're kind of enhancing, improving them with today and your approach, or are you also using other generally available models like, whether in the context of like cloud or OpenAI? Is it mostly in house stuff that you enhance or is it mostly their models? And then the second piece of can you speak to the velocity of new discoveries and growth and how the models are performing specifically? In the last year or so, it seems like everything out here on the consumer side has really gone crazy. It has gone crazy, yeah. So we do just in all of that. In fact, if I were to make that picture more complicated, we also have models coming from academia on a weekly basis. Right. Which captures the imagination for two weeks and then there's another model and people move on. So let me contextualize this for the audience at large. So if you put one class of frontier models right, like clocks or ChatGPT or Gemini and whatnot. Right. Or Llama. I think there's a huge gap between the frontier models to delivering value for our scientist at Pfizer or Moderna. Pick your favorite. Right? And that's where I think our managed services business model has been very helpful. How do you make Claude understand your proprietary data, the context of clinical data or single cell rnc? All those things are very hard to get out of the box. So what I say is you will get a response and will you get a high quality response or not? They are two different things and that's a gap that we see is quite obviously There number two is, I think there's a much broader range of what I would call task specific models, which are very good at one task. Right. Most famously, alphafold, it does one task of predicting protein structure. That's one of the like really hundreds of tasks that you need to do, maybe thousands of tasks you need to do before you make a product and bring it to the market. Right. So that world is less mature, more fragmented. Right. Very good models coming out on a weekly basis from labs, from companies, but they all perform one task really well. Right. And what we have talked about is that any task as simple as, let's say scheduling a meeting, right. Has its own spectrum of complexity. For example, let's say we want to schedule a meeting. It's easy enough, I share a cardly link, you click it, find it, get it done. Right. But if there are six of us in different time zones traveling, it becomes a nightmare. Best AI solutions will not generalize to that scale of complexity. So what we have tried to do is we have taken tasks through the entire value chain of drug discovery, very narrowly defined tasks which are actually being performed by scientists at Pfizer and Devartus and whatnot. And within that task, we have defined a scale of complexity, what is a simple version of that task and what is a more complex version of the task. And we have built diagnosis methods as to evaluate different models that are coming out. That's our benchmarking activity with a very specific lens how generalizable these models are. That goes back to our out of distribution observation. Right. And we see a drop in performance as the paper might be very compelling and promising, but the moment you try it on your data, it collapses. So we have put a benchmark to diagnose, right? Just to say that, hey, look, this is where you stand. And then the next step is prescription, what kind of priority you can add without changing the model. Of course you can change the model if you want to, but just to make it apples to apples, we show that if you add data priors, it increases the context of your training corpus. Your models are becoming a little more generalizable than before. That is what we bring to your models, public models. We don't do this to close models because we cannot. But yeah, that's the idea. And in terms of performance of models, speed of discovery, are you seeing crazy speedups on your internal stuff like we're seeing on the consumer side? No, I think the speedup has not been the bottleneck to, at least my viewpoint is the performance. The performance is quite poor. Right now, I guess I was asking if you guys are also experiencing the improvements in outcomes that the rest of us are seeing or if it's been a more steady improvement. It has been more steady improvement and not impressive enough for many tasks. Right. There's some very good tasks where the models have done really well. Alphafold again is a very good, celebrated example, won the Nobel Prize, has received a lot of attention, has seen a lot of adoption. Digital pathology is another one. But there are a lot of tasks where the performance has not been accepted enough of the models. So any optimization of cost or speed is kind of moot. So you know, when we talk about the complexity of tasks, the other axis is performance. It should pass a minimal threshold of performance. Right. That's how you evaluate people too, Right. Like can they perform tasks credibly enough on a range of complexity of a task? Right. We are putting the seamless for AI models, but I think a huge majority of activity in the space just do not pass that performance threshold. And you see poor adoption in the space for that. Right. It's not universally true, but it's majority of the models do not pass that threshold. I'm wondering how much of that is just due by virtue. Just admit, in our space, so much of it is driven by large language models. It's built by a certain degree of control and everybody seems to be a little bit more loose on what they're okay throwing at it to train the models versus in the medical field, I would say that probably people are a little bit more skittish and rightly so, to just start throwing a whole lot of data at it just to get those models. Because everybody wants to take a lot of care up front, recognizing the security that's needed. I'm wondering if it's just we're dealing with a volume imbalance right now. So there's definitely a school of thought that argues for that. Right. But irrespective, like if you talk about volume or the quality, right. You're talking about data to the model. There are initiatives like a more famous announcement that Lily made about tuned labs, which clearly puts money where the mouth is, where they're partnering with small biotechs, where they put data, they get access to the model. Right. And what I can read between the lines is that they're hoping that more data or different types of data leads to this. There's definitely another example of federated technologies which have received a lot of traction. You have very large compelling biobanks from Finland, from uk, from us, India, Middle East, Asian Countries and the data cannot leave the country. Right. That zip code. So there are federated solutions which allow you to access all the data without the data moving. Right. Again, solving both the volume and the quality issues around data that feeds into your model, but especially in healthcare life sciences, I think it's still very early and large language I think is more mature in terms of both volume data, how to represent that the tasks are better defined. Not entirely true for all tasks in healthcare and life sciences. Well, a couple of big picture questions that I've got and then I'll adjust and have another swing. So just looking at, I'm going to call it biotech in general, but certainly anything that's high stakes data, you've done this now, you've been successful at it as a co founder, certainly as someone with a history and a background in the field that you're working in. But what is something that other people or you might mentor or encourage or help other people who are trying to get into this space and trying to do it in a trustworthy manner and also in a way that moves the needle, you know, moves things forward. What would you tell people who are entrepreneurially minded but trying to get in this space and maintain that level of trust and also solve problems at the same time? Yeah. So one of the lesson that I keep relearning even after 10 years is that as scientists and engineers we over index on technology and not on people. Right. I think that seems to be the harder problem. I'll bring it close to home where AI that has again captured the imagination as very few technologies have in the last 20 years. Cloud did not have a aha moment like ChatGPT had. Like, you know, everyone was on it right out the next week. The kind of growth that we have seen, right. Leadership in companies large and small have co opted this language. Right. But I think there's definitely a people problem when the leadership does not really understand what that means. And the guidance is they do something about AI. The board says that. Right. And the CEOs talks to the VPs and VPs don't know what that looks like. So I think anyone who's getting in the space, I, you know, congratulate them and invite them and welcome them. But a big component of all of this is people. How do you educate? How do you tell them about the problem, how best to think about it? Right. Because it's very polarized otherwise if you follow your Twitter at LinkedIn, there's an apocalypse coming waiting to happen or it's all a Hoax. The reality is something in between. Right. And that nuance is a very people driven. It's not tech limited. Often some things are tech limited, but a lot of things are not. And as scientists, I tend to over index on, oh, I have this better solution, people will come. Does not really happen that way. Distribution trumps product. So very cliched. But these are things that you learn after years and years. Right. And there's a reason why they are cliched. There's a lot of truth in that, definitely. All right, Justin. All right, I'll round us out here. What is something we should have asked you about but didn't? Or what is something you have to give that we didn't touch on? I think, you know, as I've lived this for last 10 years. Right. The romance of an entrepreneurship journey is definitely not the entire truth. Right. There's a lot of collateral damage starting from your own body health. You deprive that your family, your wife, your friends, your kids. It consumes you. Right. Like, I think no amount of. Amount of Excel modeling of how much will you make and when will you make justifies this. Right. It's a gut call. If you feel very compelled to do this, you do it. Otherwise maybe you should not. That is something that I think should be part of the conversation as well. For entrepreneurship at large, it takes a toll. Right. It's also a privilege, but we cannot ignore one or the other way. And that's often overlooked. As I talk just like that, I realized that I've been doing this for 10 years. I've learned a thing or two. And when I meet founders who are preparing for YC or whatnot. Right. Like, you see, the energy is really inspiring and it's really exciting, but maybe at some point they will have kids and they will marry and they'll lose kids along the way. Right. All of that will happen. So that's something that we should make more part of the discourse. Yeah. It's that whole notion of 1% inspiration, 99% perspiration thing that's very common across every industry. Well then. And that you're not an island unto yourself. You're dragging everyone around you into that situation. They cannot sign up for that. Right. Like, you get the highest, they get the lowest. Yes, that's the truth. Yeah, absolutely. Well, that kind of brings us to a close. For today's episode, we have been chatting with Dr. Abhishek Jha, co founder and CEO of Elucidata. Dr. J. A. How can people reach you if they are so inclined. I'm quite accessible on my LinkedIn. I don't have a Twitter handle. I do have one, but I don't use it a lot. LinkedIn is the best way to reach me. You can always email me. This is a great platform to meet new people, share ideas, learn from them. So yeah, LinkedIn is the best place to find me. Fantastic. And thanks as always to my co partner here, Mr. Justin Edwards. Thank you, sir. Do I get to turn my AC back on now? You can turn your AC back on now. And special thanks as always to our listeners and subscribers that are out there. If you like what you're hearing, do hit that subscribe button, send us some notes, what you like, what you don't, what you want to see later, and we'll get back to you and maybe get it on the pod. So thank you all again and that will wrap us for this episode of Sass that App. Thanks for cruising along with us on Sass that App. We hope you grabbed some insights that were inspiring, actionable, or at least entertaining. If you enjoyed the show, don't forget to subscribe and leave a review until next time. Keep building, keep growing and keep those apps sassy.