← Practical AI in Healthcare
Loading summary
Transcript83 lines
- [00:04]
A
Welcome to Practical AI in Healthcare, the podcast that cuts through the noise to spotlight real world solutions delivering real world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what's truly moving the needle in healthcare. No hype, no theory, just practical insights where AI is making a true impact. Welcome aboard and let's get to it. Foreign. Hello, and welcome to this week's edition of Practical AI in Healthcare. I'm Dr. Stephen Lapkoff.
- [00:40]
B
And I'm Dr. Leon Rosenblatt.
- [00:41]
A
You know, before we kick into this week's episode, Leon, I thought we might talk a bit about last week's episode and reflect back on what we heard from Adam Bloom.
- [00:50]
B
Yeah, he was super cool, man. I like him. I mean, we know Adam from dci, right? And seen the awesome work he's doing on the patient facing bot. But I thought the discussion we had was just really in depth and got me super excited about the work that they're doing. What did you think? I know you're really close to the clinical trial recruitment space because you've worked on both sides of the aisle, right? Yes, I have.
- [01:14]
A
I've been a Republican and a Democrat in that space. And, you know, but to that end, you know, I recognize from a pharmaceutical perspective that a lot of clinical trials fail because they can't recruit patients. And when you're a patient and you're looking for a trial to get on for a cancer, you're desperate and there's a gap there. You'd think it'd be a Venn diagram where that overlap would be enough to get patients really enrolled in trials. Well, but it turns out that there's a big gap there that's just not working well in the real world. And Adam has come up with a solution that is helping to bridge that gap and make that Venn diagram overlap a lot better.
- [01:49]
B
Yeah. His approach involves some very thoughtful applications of modern AI technologies. And it also reveals an architectural pattern that I think everybody in healthcare should understand. It also highlights something you and I have talked about a lot, which is lack of standards. In particular, lack of any data standards for computable data in protocol representation. So lots of fun stuff there. But let's turn to today's guest, who do we have queued up?
- [02:20]
A
So we have Shashi Shakar and he is with a company called Novelia, and they focus in on helping to get patients their own data and helping to harmonize it. We're going to hear a whole lot about that in just a few minutes. So right now, though, let's welcome him to the show. Sashi, how do you do?
- [02:40]
C
Thank you, Steven. Thank you, Leon. I'm doing great. I'm doing great. The weather's a balmy 40 degrees here in Manhattan, and we're surviving, I think, one of the most intense winters that we've had, but all going well here. Yeah, thanks for asking.
- [02:52]
A
We had 18 inches of snow here in Stanford. I'm a little just up the 95 from you, and I spent most of Monday digging out, so I'm just mad
- [03:01]
B
at the weather at this point. I want a refund.
- [03:04]
A
So, Sashi, why don't we dig in? You know, our first question on our. On our show is generally, you know, how did you get your superpowers? How did you get into this role? And tell us a little about your background.
- [03:17]
C
Yeah, I actually never wanted to be an entrepreneur. I had zero interest and zero curiosity in it. I sort of wanted a very stable job. And I think I mentioned as a joke before we started recording my parents, I would have been very happy if I'd gotten an md but it wasn't in the cards for me. And so I wound up working in the biotechnology industry. I joined Genentech at a university in San Francisco and spent nearly a decade at Genentech and then ultimately Roched, a parent company out in Switzerland. And I loved it. It was a super fulfilling career. My coworkers were incredibly bright. It was very supportive, and you got to work on some of the most transformative therapeutics on market, and it was a wonderful environment. But then two things happened, I think, that led me to say there's no other choice other than starting a company. The first was from a professional standpoint. We bought troves and troves of data. And I know, Leon, you have an informatics background, and so you're probably very familiar and have your own PTSD with Real World.
- [04:13]
B
So does Steve. We're both sort of thinking informatics nerds. Yeah.
- [04:18]
C
Yeah. And so I think you both probably relate to this, but we would spend inordinate amounts of money on data from organizations I think everyone on your listener base is probably familiar with. But the data were always lacking, and we thought it was sort of the best we could do. There were always gaps. You couldn't actually follow patients through their treatment journey, especially when it comes to something like oncology. You really need to understand sort of the totality of someone's care. So, one, I was buying tons of data and constantly frustrated by it and shocked that we spent so much. But the second thing that really Nudged me to leave and to start a company was on the personal side. So totally unrelated. My grandfather, who I was really close with, we called him Fatah, he was very surprisingly and shockingly diagnosed with late stage gastroesophageal cancer and he passed away in about four months. And that journey I think was eye opening because I'd only really thought about cancer from the Genentech Roche standpoint, working on therapeutics. But to see it up close and personal, to see him changing doctors every one to two weeks, to see the CTs and the medications and the lab results, and having no idea what it all meant, I thought, you know, there's gotta be a better way that consolidates records for patients. And then man, imagine the insights and sort of the discoveries you could do on the therapeutic land front as well. So that's sort of what ultimately convinced me to go leave and try to start a company.
- [05:36]
A
You know, it's interesting, unfortunately I share something in common with you in that regard. My grandfather likewise died of the exact same cancer and that left a profound effect on me. I was a medical student when he died and I watched that journey and it was a tough one.
- [05:52]
C
But it's more to hear that.
- [05:54]
A
Thank you. And I'm sorry for your loss as well. But it looks like that was a seminal event in your life that really gave you the direction to really take this on as a company. Um, so from that side of things, what, what happened next?
- [06:13]
C
Yeah, so I had started to do a lot of research at Roche. So I'd been promoted into this role, digital health and in the global product strategy. And my job was to understand what kind of investments or projects Roche could pursue around digital health technology. And so the joke I used to make, and I'll still make it at my own apparel, is I was so deep in pharma and knew, knew so little about technology that I kept coming across APIs in all the research I doing. And for anybody that works in pharma, I think that's active pharmaceutical ingredient. And it took me a long time to realize what it actually meant in the world of technology. So I was really, really very new to the world of tech. But I came across all this research that there had been landmark legislation passed In America, the 21st Century Cures act and the regulatory unlocks and interoperability. And my mind was blown because in pharma we, we, I think many people still believe the only way to obtain real world data is, is buying it from clearing houses and sort of same companies that have been around for 30, 40 years. And all of a sudden, I'm reading this research paper that says patients themselves now have an unfettered right to go collect and consolidate their data. And so when I left the company, I thought, if we can figure out a way to harness this technology and make it really simple for patients, it's a wonderful unlock for them. And there's incredible granularity of data that biopharma researchers could benefit from. So I left, and the very first thing I was in San Francisco at the time was I started to get coffee with any venture capitalist I could find. And luckily, the joke is you could throw up a nickel in San Francisco and run into 15 different VCs. And so I would get these endless coffees, and I was so unaware of how fundraising worked that I go to get coffee and I'd leave the chat, and I think, oh, such a wonderful chat with Stephen or with Leon. And then on my way to my Uber, my way back home, I'd get an email and I'd say, you know, great idea. We're going to bow out. What the heck are these guys bowing out from? And I didn't realized that those folks actually thought those were fundraising pitches. And so that's how little I knew about how to raise money.
- [08:09]
A
Wow.
- [08:10]
C
And it was a pretty brutal journey, but eventually, you know, you figure out how to tell the story in a way that resonates with customers, investors, potential employees. But, yeah, it was a long journey to raising our first prey, and we raised that at the end of 2022.
- [08:24]
A
Oh, interesting. Jeez.
- [08:27]
B
That was a tough, tough raise environment, too. But that's really funny. What I want to know about venture capital is are they going to run and pick up that nickel if you throw it?
- [08:37]
C
Depends on the boom and bust, I think.
- [08:39]
A
That's right.
- [08:39]
C
Which side, which cycle. So you're going to write a thesis about it first, though.
- [08:45]
B
I love that you're framing here. Right. So what you're discovering in your personal journey and your professional one is the data landscape's broken on both sides. Patients can't navigate their own records on the one hand.
- [08:56]
C
Right.
- [08:56]
B
Because your grandfather had to switch doctors. And it's incredibly difficult with a complex cancer case. And then at the same time, pharma is paying tens of millions of dollars for data that can't answer basic questions. So let's unpack about what's actually broken. So let's pretend we're a pharma company that buys a data set from my former colleagues at iqv, who are wonderful. And I Want to wave at them here or Flatiron. Right. Which is a company I know Roche acquired because they want in part because of the data. What are they actually getting and what's missing?
- [09:31]
C
Yeah, as you said, both extraordinary companies and I think certainly have a place. And at Novellia we always talk about the real world data stack. So sort of all of the tools in your armamentarium that you can use. So typically what's happening when you're purchasing a data set from iqb? And I also, you know, the company might have shifted and they might be working on things behind the scenes. This is just my, my best guess as an outside observer now, not buying their data anymore. But from my recollection and I think from my experience, iqvia has extraordinary amounts of claims data. It's very, very valuable because when I was in commercial I spent time in meta fairs and commercial and then launch preparedness. So when you're in commercial it's wonderful data because you can actually look at the incidence and population level trends. But the challenge with claims data, there are really two issues. One is it's frustratingly broad. So if everyone's evergroup tried to unpack an ICD10, even with oncology, okay, great, you can see that such and such event happened for a patient. But what about the phenotype of that patient, the tumor type and what's the progression and what's the IHC or the FISH score and what were the other therapeutics that they cycled on and off of? You have no understanding of what's going on there. And so I think the biggest challenge with claims data, it's very, very broad that enables you to have population level understanding because of the incidence, but it's very broad. The second challenge I think with claims data is it's very, very, there's a huge lag window. We were just on a call this morning with, with another customer and their biggest question was, is it a six or nine month lag with Nolia? I said there's, there's, that's unconscionable. I mean you guys have quarterly priorities, sometimes monthly priorities. So one thing we try to do at no is to really address that lag. So it's an issue with I think claims databases. If you look at a company like Flatiron, which I have tremendous respect for, and there are a lot of folks at the value from Flatiron 2, we've learned a heck of a lot about, about what they've done. Really, really well. The biggest challenge is you got to look at the data Source for Flatiron, and it's Onco emr again, you know, wonderful product but heavily prevalent in community oncology settings. And the reason that matters is if you think about fictional woman living with breast cancer, let's call her patient Patty. Well, patient Patty probably goes to her community OB GYN or a GP very first moment, and then she eventually gets referred to an academic specialist who then diagnoses her. Then she'll probably go get some biomarker tests. Those biomarker tests have to be interpreted. And then she goes back to the academic specialist, gets a diagnosis and then a treatment plan, and then she gets that treatment plan back at her community site. She's probably going to her ob GYN again, probably going to the PCP and urgent care. So if you think about Onca emr, it's got great depth at that community site, great depth. But what about all of the other sites of care that patient Patty is touching? What about her journey before she was ever diagnosed with cancer? All of that data is lost. And so at Novali, we always talk about longitudinal data versus single site emr, which I think Flatiron is phenomenal at. And again, our belief is those are great tools, you should keep using them. But Novelli is a really nice supplement to put into that data stack that helps shed light sort of across those different sites of care and episodic treatments as well, if that makes sense.
- [12:34]
B
Yeah, no, that's a great explanation. And so another place some people might be tempted to turn for high quality data are hies health information exchanges. The data looks beautiful on paper, but you mentioned that it falls short in practice. What did you learn about the gap between the promise of health information exchanges and what you actually get from them?
- [12:57]
C
So I'm really bullish on where HIES can go, and I think the country, especially over the past, you know, 18, 24 months, there's so many initiatives right now with the administration, but there have been tremendous strides made when it comes to things like IES and tefa, and we're seeing a lot of wonderful commitments. But the challenge with HIES is it is a subset of a subset of a subset. So if you think about patient Patty, again, how does her data go from being generated at that site of care when she goes to see maybe Dr. Leon or Dr. Steven, to winding up in an HIE? Well, that process, first of all, is really opaque. It's very, very opaque. The second part is it's, it's a. Many times it's voluntary contribution of patient Patty's data onto an hie, the access to the HIE is predominantly limited to treatment use case or it's fundamentally limited to treatment use. And so if I am a researcher and I don't have treatment use, even if I'm able to get patient Patty's data, well, it's probably extremely thin and it's not the totality of her care. Again, because I'm pulling it from somebody who's voluntarily contributing data. Up to an hie, our belief is you got to do a lot more leg work and use a lot more elbow grease to get the raw source of the data. You have to partner with patients, which is why we say our entire company is founded around the patient of the center. If you work with patients and you're able to put together that story from the source of care. HIEs are so distant from what's actually going on, and I think they sound great in practice, as you said. But anybody who's actually worked with HIE data, there's, there's probably a reason you don't see a lot of publications at leading conferences or scientific meetings that say source of the data is hie. Right. I actually don't know any off the top of my head that I've ever said HI data were so good that we were able to publish this data and it changed the course of treatment.
- [14:41]
A
I haven't posted.
- [14:42]
B
I've seen a poster. I've seen a poster. No, but you're right. Yeah, yeah, no, I, I, I, your I take your point is really well taken. I was just, I was like. But I've literally seen one or two posters out of HIE data, but it's not dense, not to be a really rich source. So I like that observation a lot. Now, your solution at Bavelia, right, is to develop a model where patients control their own data and get to curate it over time. That's a really interesting model, but it sounds an awful lot like something we used to call phrs Personal Health Records. I am not surprised that a lot of the VCs were giving you the Heisman after the conversation, which, because the graveyard of PHR companies is full to the brim, right? We've had Microsoft Health Vault, we had Google Health, we had, you know, a bunch of others, much smaller ones, I, I, and at every meeting I go to where people ping me for advice, somebody comes up with a, you know, Personal Health bank type vision. Right? It's a really good vision. But what was, you know, but tell me about the reaction you got and how were you able to overcome it? And why it would be different this time.
- [15:56]
C
So at the time we were raising funding, it's changed a little bit now. I think investors were allergic to the very notion of any consumer product, especially when it related to healthcare. It was, it was wild. And you've totally seen that shift. But at the time, you know, in 2022, if you mentioned consumers or you mentioned patients, everybody's focused on enterprise. Nobody really believed in anything consumer first. And that was, I think, a difficult learning. Um, but the interesting thing about PHRs, and again, we are the beneficiaries of all the work that those organizations, so many smaller ones that I don't think anybody, you know, takes the time to applaud or recognize. But we've learned from a lot of those. And I think the challenge really came down to two things. With all historic PHRs, and I think we've done a pretty, a pretty effective job at trying to circumvent those. So the first part was it's just the timing of when those products were coming out. We talked about the regulatory unlocked, we talked about the interoperability unlocked. But we also talk about, at Novellia, a lot, patient awareness and desire to go collect their data. In fact, a lot of the feedback we get from our patients is around, hey, I am super thankful that this tool exists. I'm generating so much data at scale. It's really helpful to have one place to put it. So I think, one, the timing of when those PHRs came out, they didn't really have the benefit of, first of all, interoperability. Second of all, generative AI or any of the LLMs that can go clean and consolidate that data at scale. But the second challenge with many PHRs that we've really studiously avoided is they tried to build everything for everyone. That's a huge challenge because everyone's health is very different. So if you think about sort of a bell curve of health, and on one end of the skinny part of the bell curve, you've got the VCs who are into biohacking. They're super healthy. The middle, you've got normal folks. And then at the very skinny part of that long tail, you've got folks living with serious, complex chronic conditions. Well, that's exactly who we've been building Ovelia for. And it just so turns out that those folks really, really want a solution like a phr and they're willing to co create with you. And so I think our change was we benefited from the timing, but we also really focused on who are we designing for how do we bring them in as co design partners and try to create this product together?
- [18:04]
B
Yeah, I find that strategic decision really illuminating and worth noting for our audience. So the scoping, this, making the scope force have a force function to look at the skinny tail of the health curve and look at the most complex cases and where the demand is greatest and the need is sharpest is really an interesting strategic choice. One thing I wanted to follow up on though is a third source of highly curated patient data. And it's something Steve and I worked area. Steve and I worked in a lot and written a lot about, which is patient registries. So typically, if you ask me, where do I get the highest quality data about a particular disease on that far end of the curve, I would say, well, let's find a patient registry run by a patient advocacy group or medical society. What is your positioning or Novellia's positioning personally relative to patient registrants?
- [19:03]
C
Incredible tools. I mean, every single phone conversation meeting we have with customers today, the word registry comes up, if not within the first five minutes, certainly five or six times for the rest of the conversation. And the reason is registries, I think, have a reputation as really, really rich sort of continuously maturing data sets. Right. Because they're actually prospectively, many times are prospectively following patients. The challenge with registries though is can they get beyond one modality, can they get beyond one sort of data set or can they get beyond one site of care for a patient? And so I think for us, we have built a lot of registries with some of our customers. We try to approach it as a slightly different registry, which is more technology first and more multi site and more multimodal. But I think registries. The reason a lot of pharmaceutical companies invest so much time in developing proprietary registries is it's really difficult to get that level in granularity of care from other data sources, such as, you know, an IQV or a flatiron, but highly valuable.
- [20:08]
A
So, you know, as we're talking about registries, Leon and I published what we think is one of the landmark paper up at the DCI network.
- [20:17]
B
Well, we think so, yeah,
- [20:20]
C
it is, you know, what, 100%.
- [20:23]
A
We've been decided quite a number of times. And the paper is really about, you know, the complexities of building registries because a lot of folks who dive into this space really dive into it, not appreciating the complexity of what it takes to actually do it. And when you're getting into next generation registries which are big, these big multimodal Genomic registries with transcronomics and pros and emr. The complexity of bringing all that data together is really challenging and the fragmented nature of the healthcare system in the US doesn't make it any easier. And that's where I want to go next with you in terms of that fragmented data. You know, Leon and I both worked on a project at the Multiple Myeloma Research foundation called the CURE Cloud. And one of the challenges was bringing data in from EHRs all over the country and then merging that with other data sources. And back in, when we did this, it was 2018 to 2021. There was no 21st century cure, there was no smart on fire, these things. Well, there was smart on fire, but it wasn't ubiquitous. I don't know if it's ubiquitous yet, but it wasn't nearly as prevalent as it is today. So you're working in a space where the data is literally all over the place. How you, what's the secret sauce or the secret program that you're doing to bring it together in a cogent way?
- [21:34]
C
First of all, you have to, you have to send me the landmark paper that sounds, sounds amazing. I'd love to read it and I'd love to, love to share with our team. Yeah, the answer, it's so, it's so cliche. But, but we live and breathe it every day. Patience. We believe the entire answer to almost every frustrating, vexing question, Healthcare is the patient. And so for us, our belief, which was so radical, and I know, Leon, you said, well, what, what did, why did VCs give you the Heisman? This was a critical part of it was we were adamant in saying we're only going to work with patients, we're only going to work with patients. We're not going to go buy data, we're not going to go lease data, we're not going to have, you know, smoke filled room handshakes with data brokers to go somehow acquire massive data sets and completely cut the patient out. So we wanted to begin with patients. And Stephen, to your point, I heard you mentioned, smart on fire. And so our belief was, man, if you can partner with patients, because patients so badly need this at that, you know, the long end of that skinny curve. Folks with serious conditions, they so badly want a way to consolidate their records. Now it doesn't mean that they're going to look at their medical records five times a day, but gosh, when they need them at the time of care, it's critical that they have as accurate a historic representation of their data as possible. And so one of the first things we did was we spent a lot of time talking to patients and we realized that a lot of these folks not only remember where they received care, they actually have all of the credentials. It's a little unusual. Right. We're right by Bryant park in Manhattan and I, I promise if I went to Bryant Park, I'll Admit I asked 10 people, do you remember all your patient portal logins? They'll have no idea what I'm talking about. Right. So I have to acknowledge that. But that's sort of the middle part of the bell curve for the folks that we were designing with and for a lot of those folks know exactly where they're receiving care and they have a lot of that data already. So by using patient portal logins and partnering with patients, at least from what we've seen from user activity, those folks connect at least three, four, sometimes five different data sources going back up to 20 years, and we're able to consolidate all of that data. But hopefully that answers your question, Stephen, in terms of what is a secret sauce? I think it begins with patience.
- [23:33]
A
It does, but I'm going to, I'm going to push on that button a little bit because, you know, I, as we talked in our pre interview, I did a big project@pfizer 20 years ago looking at phrs and the difficulty of using them. And I'm wondering, you know, timing is everything, right? So the idea of a PHR was always a good idea. However, the complexity of all this data stuff we're talking about and the lack of a unifying tool, was my chart kind of a catalytic event that, because it's so ubiquitous today, was that one of the things that really made what you're doing more possible because the patients have always been engaged, they've always been interested in their own care, but all of a sudden they now had a tool that was usable and provided this access. So was that a catalytic event for your company's existence?
- [24:25]
C
Without, without question. I mean, without question. If you even mentioned the word patient portal, the number one thing I'm sure if both of us went and asked five of our friends that aren't in this space, patient portal probably say MyChart. And so I think Mychart, you have to give EPIC their well deserved flowers. They totally changed the way that patients interact with access and understand their care. And what we said was they've done amazing work. Well, how do we take that a step further? Which is not just the epics in the my charts of the world. What about Cerner and Athena and Nextgen? What about lab data? Right. Multiple modalities of care. And how do you instantly put that into a clean timeline so that patients understand all of the things that happened? But in a word, Stephen. Yes. I think that my chart completely changed the landscape in a way that's very beneficial for patients being able to understand and access their data.
- [25:11]
A
So with that in mind, and now that we have access to data, because it didn't exist really until not that many years ago, or at least easily, tell me about what you're actually pulling out. Like you're able to get, you know, SOAP notes or you're able to tell me the pieces. I don't want to steal your. To your story.
- [25:28]
C
Yeah, absolutely. I mean, you mentioned a few of them. The, the one, the one caveat I'll put out there is we just had a data meeting this morning and truthfully, we continue to be surprised by the data that's being acquired and ingested and pulled in as well. So we're continuing to kind of explore almost this frontier of this landscape of all of the data that's coming in as well. So from what we've seen and from what we've understood about the data, I think you can really break it down into three different categories. One category, the structured elements that we talked about before and those are highly valuable to patients. Sort of the first thing that folks want to understand their lives. Labs, vitals, allergens, medications, doctor's appointments, so on and so forth. The second thing that you mentioned is an incredible trove of unstructured data. And the unstructured data many times isn't accessible to patients within their patient portal. I think my chart was a decent job sharing your doctor's notes, but a lot of other, a lot of other patient portals don't readily surface it, but it is available. And so what winds up happening is when patients join Novella and they connect multiple patient portals, we're able to pull in all of that unstructured data. It's in SOAP notes. There's also XML files, there's HTML files, RTFs. And so in some instances, you're getting an incredible amount of data that's unstructured that eclipses the structured data as well. And the reason that's valuable for us is our data team sits down and they reconcile both accounts. Now, sometimes you might have discrepancies, right? And sometimes you might have disagreements in the data. And what we do is we share that back with the patient and that gets into the third Part which I think is the most important kind of data is the data that the patients themselves are telling you and sharing. What makes Novella so unique, I think as a data company, is we allow patients to interact with the data that we surface. They can edit records, they can hide them, they can correct them, they can upload new information, they can snap picture of a different medical record. We'll digitize it. And the reason that third category is so important is when you talk to patients many times, not many times, but sometimes they'll find aberrations in their data that might not reflect their real treatment care. And so they're able to kind of correct the record or support for them. But for pharmaceutical companies as well, they really want to understand what's happening to patients outside of the clinic. And so that third bucket of patient generated health data, I think, has been really valuable as well.
- [27:39]
A
So I'm going to just probe on one other thing real quick before I hand it off to Leon. When you give the rights to patients to affect their data and change things, you know, this isn't the single source of truth for the patient. It becomes a single source of truth. But my concern about this is a little bit around. If you ever watched House, they had a philosophy that patients always lie. And I've lived that. I'm a physician, you know, and patients lie. And if you give patients the right to change their records, some, not all, some will try to change their narrative, try to change their story. How do you protect for that? What do you do about, you know, people that will delete important pieces of information, Perhaps illnesses, perhaps medications they've been exposed to? Because that's a problem. And I've seen it firsthand.
- [28:30]
C
Yeah. And I, I'm, I'm glad you share that. Our, our belief is we almost, in an overly zealous way, believe and I think center everything around what's best for patients. And our belief is that you might have some instances where patients are perhaps changing data that might not reflect accurately what's happening to their care. But our belief is patients are the most incented for the best course of care possible. And if they're changing something in their record, well, it doesn't change the actual source data. Right. So the data that's within that data set, it's left unchanged. But there is a, there is a record there that the patient has commented or, or corrected something about it. But our belief, which I think is, is a bit radical, is we really want patients to have a data set that they feel reflects the accuracy of their Data that doesn't mean that when somebody changes a dosage, for example, that the novelli account of it fundamentally and irrevocably changes. It means what's in the patient's portal with the novelia reflects what they believe is more accurate. We haven't seen instances of patients sort of dramatically correcting or changing what happened. Usually it's sort of oh, that medication was prescribed in September versus April or I was in a different dosage or it was sub q instead of IVs, those sorts of things which are really valuable. I'm not doubting that, that it does happen, that patients could do that. But we haven't, we haven't seen it. And I think we're perhaps overly, as Alice, as you said, about protecting the rights of patients to manipulate and sort of control their own data as well. At least in our perspective. Everybody else has had control over patient data. Insurers, physicians, health systems, except for patients. And so our goal is to really aggressively flip the script and put that power back in patients hands as well.
- [30:19]
A
It should be, it should be right there. Yeah. Leon, you want to take a. Yeah.
- [30:24]
B
I mean I'm, I'm probably not as cynical as you, Steve, about patients deliberately lying, but there is very, you know, heartbreaking research about data quality that shows that data across EHRs, PROs and clinical trials is inconsistent and there is no single source of truth. So you wind up, you know, which made me kind of cry for a while. But then you get over it and you say, well, all right, you know, there is no single source of truth. So you have to converge and triangulate. Right? That's really the answer. Or you basically develop a data enrichment and curation process that improves the data given all the sources you have, and tries to move it up that quality ladder. But let me kind of get into the AI piece here, right? And the processing of the data. So you already got this rich data coming in, you're grabbing it, you're grabbing the XML directly from the FHIR endpoint on the HR and you're getting some of the data patient didn't even see in a portal. And then there's a normalization deduplication pipeline at a high level. What does the process look like? Are you, you're normalizing to OMA for foreigner customers. What else are you guys doing?
- [31:34]
C
Yeah, normalizing.
- [31:36]
B
This is the nerdy, this is the nerdy part of the interview. This is, you know, this is where the non informaticians just go, what are these people talking about?
- [31:42]
C
But please, so funny. I know, my CTO co founder and our data science folks are probably banging down the doors to jump into the conversation and share their thoughts. So OMOP for pharma. Yeah, it's typically the standard, but it sometimes depends on what the pharmaceutical partner is looking for. And you'd be surprised at how variant or maybe neither of you would be surprised, probably know exactly what this looks like. Right. So there's such heterogeneity in terms of how they want the data transposed and shared back with them. But the process, I think, Leon, if we zoom way out for a second, answer your number one question about how this work. So patients join Nevelia and they authorize Nevelia to then go pull in multiple forms of data. Smart on fhir. Obviously it's a multiple patient portal, sometimes it's lab data. We also have some CMS payer data. The reason I mention all of that is the ability, I think technically to pull in not just multiple different formats of EMR data. Steven, I know you're a practicing physician. Even one emr different instances across a site of care could actually be quite variant too. So to be able to pull in all of that data from an EMR perspective and then you've got lab data and sometimes you have claims data too. We take all of that data in and the very first thing is the deduplication of it. And many times actually now with HIEs and interoperability, sometimes you'll have duplicate records within multiple forms of EMRs. Right. So Leon maybe had, maybe was diagnosed with something, or he talked about something with his PCP and that record is pulled into his rheumatologist EMR as well. Right. But it didn't happen twice. It happened once. It happened at the PCP side of care. So the very first thing is understanding what's probably realistic and rational and deduplicating those records. The second thing is in normalizing that data, I always use the example of dosage. And so if you look at dosage data, you can have fractions, decimals, written words, phrases. And so putting that all into a normalized context, a single field for us, usually decimals, is extraordinarily valuable. It's sort of the grunt work that's not sexy, not exciting. And then the last part, which is heavily leveraging LLMs, is the harmonization of that data. So what category does this fit into? What do the data actually mean at scale and sort of what's the missing context behind each of these data elements? The final thing is that data is shared back with patients almost immediately, but to transfer it to a research partner to share it back with them. There's an extraordinary amount of work that goes into the data modeling. So usually we'll sit down with the partner. What's your inclusion exclusion criteria now that we have the ie, what are the table shells you're looking for? What's the power? Talk to me about the statistical analyses you want to run. And all of that eventually goes into a data model that's relevant and specific only to that biopharma partner that we then run all of the data through and then it's transferred back in a way that's anonymized, de identified, fully safe for the patient as well.
- [34:30]
A
Yeah.
- [34:30]
B
So what I think is really clever about your model is that you've inserted a data curation group that is patient driven into the post, into the post collection process. So it benefits the patient, but it also benefits the downstream data use. So it's not just engagement for engagement's sake. Right. It's like patient engagement, what's actually making the data better. And it benefits both the patient and a downstream data user.
- [34:57]
C
So that's it, you share. Absolutely it.
- [34:59]
B
Yeah. So thanks. I would just want to bring that up to the, to the audience because I think architecturally that's a really good pattern for people to understand. You shared a specific.
- [35:10]
C
Because you. I'm going to steal the way you talked about that because it was so articulate. I think a lot of it. If you go back to the investor part about the Heisman for consumer, there's a fixation on engagement. There's a fixation on it. Right. Because their whole thing is, oh, it's a consumer app. What's the engagement, what's your download ratio? And talk to me about the different funnel metrics and et cetera, et cetera. Well, I think what folks fail to understand is when it comes to health data, engagement means something different at the most fundamental level, which means the completeness of that data and the value of that data both to patients and the pharma. And you nailed it when you're talking about engagement, which is how do you partner with the patient and engage with them to develop the most complete single source of truth that's maybe still going to have some gaps, but it's as accurate as possible, both for patients and for pharma and not engagement for vanity metrics sake. And I think the way you said that was, was really, was really sharp.
- [36:04]
B
Oh, well, thanks so much. I mean, you know, you can you, if you get, get into it you can look at the slides from my course on data quality improvement that you know and how those processes usually work. So. But you shared an example of where, when in our earlier conversation where lab values in EMR was off by digit from the raw XML right in the phys likely mistyped it. That's a great example that I think will illustrate how the data curation can actually work. Can you describe how those discrepancies are detected and how they're resolved just to give it. Just to solidify our understanding of that curation process?
- [36:42]
C
Yeah, absolutely. So from sort of a technology perspective, when, when we're trying to automate the harmonization of data and you've got multiple different data sources, you inevitably, with real world data will come across discrepancies or disagreements about the same exact data element from two different sources. And so essentially what happens at that point is in this example I shared with you, we were currently doing a research project for one of our partners and lab data is incredibly valuable. Right? And so if you looked at just the lab data within the structured elements, and that's what the physician would type into the emr, place into the emr, it had one value, we thought, huh, that's kind of interesting. It doesn't contextually make sense with what else this patient's going through, but okay. And what's interesting is using artificial intelligence and machine learning, our systems were able to consult the unstructured data. I believe this is in an XML file. And so looking at the xml, you're actually able to see the raw result of that lab. Right. So not what the physician saw and jotted down, but what actually happened with the lab. And this is very valuable if you think about cancer care. So not the pathologist's interpretation of the lab, but the lab in and of itself. So we looked at the lab in and of itself and the lab value and we saw that it was off by a single digit. From what the physician and what the lab value said in the unstructured field was far more contextually accurate than what was in the structured field. Now, we will never make a judgment call and say the physician mistyped this or go tell a biopharma partner, hey, we, we know the answer, it's the unstructured. But at least what we can do that I think other real world data companies might struggle with is we could present both to the biopharma partner and we can say, look, there's a discrepancy here. Here's one value, here's the other value, and this is Sort of the context behind both, but it's definitely the power of LLMs in machine learning today is you can automate that entire process. It'll detect the discrepancy and it'll tell you which one might be more valuable because of the context of that patient without any human intervention. I think. Which is, which is fascinating.
- [38:41]
B
Yeah, I think you are right to pick up on the technology shift that is making this kind of effort much easier. Kind of goes along with the theme that Steve and I have been picking up throughout the podcast, which is modern AI and improvements in IT automation are making some mundane things easier. And this is a typical kind of data cleansing task that takes 80% of any data scientist time and annoys the hell out of everybody.
- [39:10]
A
Right.
- [39:10]
B
But yeah, because you have to write Python scripts to detect that kind of stuff and then create the output and it's become much, much easier. And I think that's really interesting change. So let's shift a little bit and talk about the business model, whatever. At the end of the day, right. Your business is going to be selling structured data sets to pharma today is. It's, you know, just let's getting specific is going to be like a CSV or JSON object or the OMOC data set. What's an adoption looking like and what are your customers looking like today?
- [39:40]
C
Great. We started out primarily with oncology companies and the reason for that is, I mean, you mentioned Flatiron earlier. Right. I think the interesting thing about oncology data is there's so much data generated for folks that are living with any form of oncology. There's, you know, there's pre diagnosis diagnosis, there's cycling through multiple lines of care and then there's survivorship and there's an enormous amount of data that's being generated. But the irony with oncology data sets before Novellio was it was all episodic or single side of care. So we said this is a therapeutic area where patients have extraordinary high unmet need. They need a place to consolidate their records and researchers badly need to understand the totality of care. So we started out in oncology and we had a single partner at that time. And today we've grown largely across the top 10. Although excited, there's a number of smaller emerging biotechs that we work with as well. And now it's not just oncology. There's oncology, there's rare disease, there's cardio mets. Increasingly we're looking at chronic, chronic conditions and neurology. We're not yet active in but that's starting to emerge quickly too. And so what you've seen is year one to year three and change at this point. Both the expansion across a number of customers, but also therapeutic areas that we're servicing today has grown quite a bit.
- [40:56]
A
So you, you really made some incredible. The story is very, it touches both Leon and I, our hearts very, very much. We both worked in this space on many different levels over our careers. You've built like this technology first platform. You've got patients helping to curate it. But you've also hinted in our pre call there's something else down the line. Are you able to give us a hint of what that might be?
- [41:24]
B
Yeah.
- [41:25]
C
There are two things to think about on the patient side because I know that that's. And I can see the passion. And Steven, your own story about your grandfather. Right. It's touches all of us on the patient side. There are a couple of features we're working on right now that are, I think are going to be extraordinarily valuable to patients. And without getting too into the details, I'll share more about pharma too in a second. But I think right now we live in this world. If there are so many headlines that come out where so many different technology companies enter into healthcare. Right. And you can think about all of the conversational chatbots that are now playing in the field of personalized health recommendations and PHRs as well as our belief is a lot of those organizations are approaching this with the same, I think, mistaken belief that a lot of traditional PHR companies did, which is we're going to build everything for everyone all at once. And I think some of those tools are great. If I get an mri. And I want to say, hey, did you think before I go see my orthopedist, is this a sprain or a tear? Can I have a couple beers with my buddy tonight if I've been prescribed this medication? Those sort of like comical point of care questions I think are definitely great and valuable. But for us, we're always thinking about how does AI fit into the fabric of a patient's overall journey and meet them at the moments that matter. So, okay, here's my overall treatment plan. What about ASCO and CCN guidelines? A patient might not know what those are, but tell me about how my treatment care aligns overall over the past five, six years with what the national evidence says. There are emerging therapeutics every single week, it seems almost a month, and cancer and a lot of other therapeutic areas. Well, what does that mean for me? What should I talk to my physician about? Are there trials that I could latch into? And I believe Novelli is uniquely positioned to automate all of that from a patient engagement and education standpoint, too. So there's a lot of exciting stuff coming out on the patient side.
- [43:12]
A
Go ahead.
- [43:13]
C
No, no, please, Stephen. Yeah, no, it's a conversation I would love to.
- [43:16]
B
Yeah,
- [43:18]
A
I was sort of thinking about this, you know, going back to the patient tail and the most engaged patients. I mean, what's really, really different about your approach, which others simply don't do, is they're trying to be, you know, a generic data tool. And the fact that you're looking at patients who are that, that long skinny tail and the bell curve, those are indeed the most engaged patients. And they're the ones who are most likely to be, who have stickiness with this because they're going through a journey, they're scared, they need something to anchor to. And giving them the ability to be able to get all that data together in a unified way, it may well give you the best kind of the best quality data that might be out there, aside from those that may cheat. And I will grant you that probably is a low number, but I will stand by my statement. I know many people, unfortunately, who do cheat. But when it gets to the question of how this is going to be used by the pharma companies, how this is going to be used by h EOR departments or biostatistic organizations, how do they unpack this in this way? Because if you're going to be at such a the small end of the tail, those use cases may have actually some bias in terms of what the population looks like. So how do you sell that? Or how do you navigate the potential bias that could be intrinsic just by nature of how you collect the information and who you collect it from?
- [44:43]
C
I would. And I'll circle back to the other question you had, which is, well, what's also on the horizon for pharma? Because I think it gets to this question too, is I would say if folks look at the existing data sets that are out there, many times they're extraordinarily biased as well. And I think the benefit of a tool like Novelli, again, it's a complement into someone's real world data stack where what we're able to do is go really, really, really deep within different patient cohorts, whereas folks like Ikevia do a, a really wonderful job going very broad across multiple different cohorts. I think bias is something we think about a lot, but because of the granularity of the data, what we're able to share back with Biopharma partners is a level of specificity about what that cohort actually looks like without having to say that we're making assumptions or guesses about the representation or the generalizability of that data set. But, Stephen, I think bias is something that's so critical to understand. I don't know if we have the answer to it yet. I think that we're still exploring it, and I don't want to get ahead of my skis in terms of where the company is with that, but totally recognize that that's critical. But on the pharma side, in terms of what's on the horizon that we're thinking through, I mentioned it on our call. Stephen, I know you did a lot of this work at Pfizer before, too. Imagine a world where when you're purchasing or working with the data partner, you as the director of Evidence Iteration, or hgr, within that tool, you have the power of five or six junior data scientists. And you don't need to share this data set, the CSV or JSON, with a team of data scientists or an outside consultant or an analytics vendor that that tool in and of itself has agentic capabilities where you can ask any question. And because of the value of a tool like Novelia and the depth of that data, it can give you incredibly specific analyses. In plain English, I think that's really the future is how do you make it much easier for folks at biopharma companies to get the answers they want instantly? So right now we're able to offer really, really rich and valuable data. But imagine a world in which that data is already analyzed or easily understandable by folks that don't have to rely on data scientists. Of course they'll always have a critical role, but on huge teams and have to ship that data out to a very expensive consultant or vendor. We get all of that information all at once. So I think we're building that foundation of proprietary, rich, valuable data, which means the AI tools you build on top, some of those agentic platforms, I think will be extraordinarily powerful as well.
- [47:13]
A
Okay, so I want to just drive back to another, another piece which is, How does this fail? Is there, is there a failure mode in here? Is there something that, that, that unwinds this in a way that you haven't anticipated? If it does, like, what is it that's going to break?
- [47:38]
C
I'll pick a bit of a surprising one, which I think is external. That is is out of our control. And then I'll talk about internally what could break it and evaluate. So my biggest fear right now is externally, you have a lot of folks that are beginning to experiment in healthcare, particularly personal healthcare. And I think folks are well aware of these organizations. And my, my main concern is, is the level of trust and sanctity and care that those organizations have for patients and patient data. Now, I think OpenAI incredible organization, and they've developed extraordinary technology, but they are not a healthcare company. They're a technology company that's extremely innovative, but they're also focused on so many different verticals. Lawtech, fintech, as far as I understand, a lot of other verticals as well that have virtually nothing to do with healthcare. So you have to think about, is healthcare an experiment? Is it a fascination that's ephemeral? Or do you actually really understand the sanctity and the sacredness of patient health data and what that means to individuals? And I know anthropic is also involved there. Right. And my biggest concern is when you have organizations that spend $40 million on funny but dueling super bowl ads making fun of one another and how sycophantic their tools are, and the fact that some of these organizations are now talking about dismantling safety protocols so that they can win larger DoD contracts, are those really the organizations? And maybe they are, but are those the organizations we should trust with the most sacred data of all, which is our health data? So my biggest concern right now is these organizations are so aggressively pushing into healthcare. My concern is it might not be for the benefit of. It could be for the benefit of patients. But if that trust is broken, if that data are shared without consent, if something happens to patients we've seen as sycophantic, some of these tools are, if they give wrong advice, who's liable? Who's liable for that information? Who's liable for data that's mishandled, that could erode trust across the entire ecosystem and be highly detrimental to the true innovators that are actually working hand in hand with patients. So my biggest concern is the trust risk. Right now, with all of the focus on some of those larger players internally, I think my biggest concern is always speed. So how do you balance speed with the rigor and the value of a product? We believe that it's paramount that we're developing and putting out features for patients that are best in class and highly innovative and doing the same for pharma. But it's an extraordinary amount of work, I think, to Balance. I said this when we chat about an enterprise company that commercialize this data with a consumer based company too. So I would say those are the two, those are the two areas of risk.
- [50:16]
B
Yeah, those are really good observations and I think we, we also worry that large tech companies of all sorts that get into healthcare wind up breaking their teeth on the complexity because they simply don't understand.
- [50:29]
C
Right.
- [50:30]
B
It's a, the domain has a really different field both ethically and procedurally than others. And I think they need domain expertise not to break and not to move fast and break stuff they don't mean to. Yeah. So I mean we try to get some field notes at the end of the conversation to give our audience some take home messages. One thing I do want to mention before we go though is what you talked about is really aligns well with some of our close collaborators in the Society of Participatory Medicine. And I just want to encourage you at Novella guys with Novelia to reach out to them. So particularly in Price, Danny Sands and Dave debrankart, I think that'd be really good contacts to align patient interest with what Novelia is making possible. So we hope we can facilitate the conversation to happen. But imagine now you got 30 seconds in an elevator with a pharma chief medical officer in an elevator or you know, some other venue where you're stuck with them. And what's the one thing that you want them to understand about patient consented data that probably don't understand right now?
- [51:42]
C
Yeah. Patient authorized. Real world data parwd as we call it at Nevelia is the single most comprehensive and granular record of what's happening to a patient. It doesn't mean it's 100% complete. But imagine a data set that follows patients. It's dynamic, it's always updated, it's multi site, it's multimodal, it's multi physician. You're able to understand the true journey of a patient's care. That's exactly what Novelia does. And I think we've, we've done a lot of really good pioneering work in that space and we're excited about that. Biopharma partners we've worked with and a lot of the publications you've gotten out of that work.
- [52:16]
B
Where can people find out more if they want to look up a Novellia how and or you. How should they reach out?
- [52:23]
C
Yeah. Go to novelia.com I know that this is not a video just yet, but novelia.com so that's n O V E L L I A dot com and click on the Life Sciences tab and super easy to get in touch with us and we'd love to hear from you. We're always looking to see how we can support organizations both large and small across a variety of therapeutic areas.
- [52:44]
B
Shashi thanks for sharing not just the business story but the personal motivation behind it. It was such a pleasure to chat with you. We wish you the best of luck and look forward to hearing more about your future trajectory.
- [52:57]
C
Thank you so much Steven Leon it was a wonderful conversation and thanks for giving me some time to share my
- [53:02]
B
story and I just want to thank our audience and invite all of you to join us next week on an exciting episode of Practical AI in Healthcare.
- [53:16]
A
Thank you for joining us this week on Practical AI in Healthcare. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay practice.