This AI Urine Test Detects 50 Diseases and Reads What Genetic Tests Miss | #194

We started talking and saying, look, you know, the dog is smelling the urine and the nose is detecting the things that evaporate from the sample. How can we replicate this, this result with the hodge?

They built a fully AI hospital. No humans. They treated tens of thousands of patients. You walk in, it's a fully AI-equipped hospital. It's absolutely incredible.

I want to optimize the most precious resource we have with the technology that I have in my hands. And the most precious resource we have is time. It's doctors and nurses hours in the day that they can see patients and actually do medicine and not fill out forms. I don't know of any physician who can keep up with, with new technologies coming to market. Why not use technology to help you with that?

And I think with the way that the new way of how medicine is done, it absolutely needs AI because you're actually uncovering layers and layers of data that sometimes are subclinical manifestations of conditions. Hello, welcome back to the Dr. Joy Kong Podcast. You know, my mission here is to make sure that you live a longer, happier, and healthier life. So today, I'm so honored to have a guest with us to teach us about technology in medicine and what we can do to detect diseases earlier. So I'm really excited about our conversation. And Dayan, so happy to have you here.

Thank you very much. It's a pleasure to be invited.

Yeah, so we were just talking about Bulgaria. You actually, you're from Bulgaria, but now you're in the US. So I'm going to introduce you just a little bit so people just get a little bit more background. So you are a technology entrepreneur with over 30 years of experience building companies at the intersection of software, and healthcare. So Dayan is a co-founder of Luventix, is a diagnostic startup using AI to detect over 50 diseases from a single urine sample. So that's an idea that came from him watching a study about dogs detecting cancer by smell. He has navigated FDA processes and clinical trials for nearly a decade. And is also the founder of Panaton. Did I say it right? Panaton. Correct. Is a health IT firm specializing in connecting different healthcare systems in hospitals and labs and clinics all across the US. So thank you so much for being here, Deyan.

It's a pleasure. Thank you very much for having me, Dr. Kong.

Yeah. So I'm so curious about what you have found out about this diagnostic tool. We all know that animals are very intelligent. They've used animals to detect seizures and in this case cancer. So what was the study? What popped into your sphere?

It's a funny story, actually. One of my two co-founders and I were sitting in his home after a day of skiing, watching television, and we saw this TV program of dogs smelling urine samples and sitting down when they detected cancer. And, you know, both of our backgrounds are technology. He's a biophysicist and spent most of his professional life doing drug discovery using machine learning methods and AI technology long before it was sort of a household word like it has been now for 4 years or so. And we started talking and saying, look, you know, the dog is smelling the urine and the nose is detecting the volatiles, the things that evaporate from the sample. So what is that? How can we replicate this result with technology? Well, the nose of the dog is essentially a gas chromatograph. You know, it's quantifying the presence of certain molecules in, in, in the vapor or the, the, the volatiles. And I said, well, what happens then? Well, then the brain's dog is doing classification. It is determining whether, uh, the molecules that the nose detected are indicating the presence of a condition or the absence of a condition. So how can we do that? Well, you know, we have now artificial intelligence technology, or more properly, it's a branch called machine learning. Which allows us to train models to do classification, basically saying, give us data from a population of patients that have a condition, give us data from population of patients that don't have the condition, and more than likely, in most cases, not all cases, we can build a little bit of software that will then detect a non-classified, an unknown sample in one of the two categories. Uh, it's not quite as simple as that, obviously, but that is the fundamental theory, which is very straightforward. Yeah.

So what, what disease did you go after first?

Well, uh, Martin, my co-founder, did his initial two studies as a proof of concept with bladder cancer and prostate cancer. And they were very small studies, but, uh, they demonstrated that the technology was viable, that we were actually being able to detect the signal and do classification. It took us a few years to start Luventix and then fund our first clinical study in which we did 4 different GI diseases, colorectal cancer, Crohn's disease, SIBO, which is a small intestine bacterial overgrowth. And so we are in the final stages of closing that study, and you can find it on clinicaltrials.gov. It's IRB approved with about 1,250 patients, and the results are actually very, very encouraging. We are able to build models that do classify the, or correctly identified the presence of the conditions.

Yeah. So in all conditions, I'm just curious and interrupt you. So the urine is, I guess, detecting a whole group of compounds. It would be interesting, you know, from intestinal diseases that is showing up in urine. So is that like a cluster of compounds?

Well, the fundamental sciences or biology is that when an organism develops a certain condition or just changes, that's reflected in their, in the metabolism of the organism. And your metabolic profile is expressed in your urine. So when the cells of your body start behaving in a different fashion, the metabolic behavior changes, the, what is being produced in your urine changes. And so we're essentially correlating the changes in, in the organism to its metabolic profile. There's about, I would say, close to 10,000 different metabolites expressed in urine. And so that is an enormously strong dataset that can provide a signal for many, many different conditions. You know, theoretically, if a disease or a condition, not necessarily a disease, but just a condition, you know, you know, presents itself and it reflects in a metabolic change, then we should theoretically be able to detect it as a change in the metabolic profile expressed in analytes in urine.

Right. So I'm curious, let's just pick one example, you know, bladder cancer or sebum. Let's just say sebum. How many of the markers are going to manifest changes among this 10,000 different metabolites?

We're getting into the interesting part.

I'm just curious about the scope.

We do not know. And that is one actually of the innovations that we're kind of very excited about and one of the things that we claim in our provisional patent. In that we, unlike the traditional methodology where you first have to identify the biomarker and then develop technology to detect the biomarker in a sample, we operate on the raw data, meaning that we don't necessarily have to identify the specific biomarker. We look at the data pattern across all present analytes. And that is the power of the, sort of the AI technology that we could holistically— Interesting.

I'm sorry, it just dawned on me. This is like, it's reminding me of Chinese medicine, right? Right. They don't know exactly what compounds are the problem, but on a system level, they know there's something that has a particular tendency. Right. It's almost like this merging of, you know, the system approach in Eastern medicine and then, you know, the, the cool scientific investigation of the Western science. So I just, it just kind of dawned on me, like, where if we're really going in that direction, then we are bridging the gap almost.

It's very interesting you, you brought that up because, uh, again, my, my chief scientist and co-founder and close friend Martin talks about network biology and how the organism is completely interconnected. And while, you know, historically and traditionally we've been looking at diseases and isolated biological system in the organism, the reality is a lot more complex. And so when we are able to look at the total picture expressing itself in a, um, digital twin of the organism, which represent— is represented by its metabolic state, we have a much better opportunity to take into account, uh, what we call, uh, the, the, the full dataset without dimensionality reduction. So, which essentially in fancy Western science words says what you just said, which is, you know, we are looking at the, uh, holistically at the, at the health of the organism.

Yeah. So that's interesting. So how did— because you mentioned initially, you mentioned there are things in diseased urine that may have, you know, there's something markers may show up and then there's, you know, the healthy urine. So that sounds like a traditional approach, right? So that's how you entered. How did you come to the point where you don't know How many markers are there?

Again, we, we don't start by trying to identify which markers correlate to the condition. We look at every— a gas chromatogram basically is, uh, in our case, 38,250 data points that say quantitatively how much of a particular molecular weight we're seeing in the sample. We don't know which molecule is expressed at any point in the chromatogram, but we're looking at all those 38,250 data points and from 200 patients that have a condition. And the machine learning, uh, models that we build are capable of selecting the, uh, data correlation or the, the the signal, if you would, that is associated with the presence or the absence of the condition. The technology actually has additional applications which are also very interesting because you mentioned clinical trials. You obviously, you have gone through that and lived it yourself. But in clinical trials, it is very, very, very important to select your cohort, you know, the exclusion inclusion criteria of the people that are going to participate in the trial in such a way that you maximize the likelihood that you're going to get a response to a treatment. Well, with our approach, we can build a model that can predict whether or not an individual is going to respond to a treatment or not. Which, you know, it's, in pharmaceuticals, you know, they call this companion diagnostics. Once a, you know, a treatment is brought to market. But during the clinical trial, if you want to reach your endpoint, you're much better off finding people that are going to respond to your treatment to bring, you know, to optimize your cohort. And so technology works the same way. If you give us a group of people that have particular properties of their organism, we are likely to be able to build a model to recognize them by their metabolic profiles.

Yeah, I absolutely love it. I mean, you're really leading the future of medicine, right? This is what's so exciting about how things are going to be very different.

I, you know, I am a victim of the Silicon Valley way of saying that we've invented the great next thing, you know, hot water, the wheel, that kind of thing. But the reality is that all of these things are Lego blocks in understanding, you know, biology and medicine. We take a metabolic approach. If we marry that with a, you know, the genetics approach that people have been pursuing very actively in the last few decades and have heavily invested with, we have an even better understanding of the organism because, you know, genetics tell us what we are possibly programmed to do. Metabolics tell us what our organism is actually doing. It's like the difference between our—

Metabolics.

Yeah.

So you're referring to metabolomics, right? So, so like, like what, what exactly is that like? Because that's a big word.

Metabolomics is simply the study of the metabolism in organism, how our cells produce energy essentially or behave. So again, you know, the other word that we use a lot these days is multi-omics. That takes all the omics and we combine all the knowledge that we know, the information that we gather into a bigger, better and better picture of an organism. The, again, coming back sort of down to earth in reality, we talk a lot about personalized medicine and the technology leaps that we're making in, you know, recently do enable us to get closer and closer to very narrowly personalized medicine, which is a good thing and make it accessible to, you know, a broader broader population.

So you guys already have the technology that you can predict certain responses from drugs?

I mean, is that just for a few diseases or a few drugs? Well, we have technology that can do patient stratification depending on prob— you know, metabolic profiles of the organism, whether it's detecting disease or detecting likelihood of response or a any negative response to a treatment, you know, like a toxicity of a proposed new medicine. All these things are possible. Now, like all other things in medicine, until you've done the studies and gathered the data and demonstrated it for each particular condition, you know, we cannot claim success across the board. We're trying to be very, very disciplined in our science and make it all reproducible and not be in a position to have to backpedal making claims that are not substantiated.

Hello, my loyal listeners. I want to interrupt this podcast for just a second because I want to let you know that if you're looking to help solve a major health issue, Or you're looking for health optimization and just for longevity purposes, you can find us at Chara Health. We are considered the leader in the regenerative medicine and longevity space, and we utilize modalities like stem cell therapy, peptide therapy, plasmapheresis, ozone treatments, laser therapy, all to create a beautiful symphony to get you to where you want to be. You can find us at www.charahealth.com. Chara-health.com. That's C-H-A-R-A.

Yeah.

Well, this 50+ diseases, uh, which are they and how did you pick them?

No, no, we're— 50 was a, a number that we kind of picked out of, uh, uh, the air as we said, you know, what would define success if we can build a single urine test that essentially, you know, you pee in the cup once, you run it through the, uh, laboratory instrument once, you get a data file. And then if we have 50 models for 50 conditions, we can give you a panel of 50 diseases. If you have 200 models, we can do 200. But 50 was kind of the, you know, wouldn't it be wonderful if we can do 50? Because in reality, you know, 50 conditions cover a lot of ground, especially because technology is not only limited to oncology. You know, we, you know, we, we picked our study, you know, uh, Crohn's, celiac, SIBO, uh, colorectal cancer specifically so we can demonstrate that technology works across different types of conditions, not just oncology. Um, and so that, that was part of the reason we built our first study that way.

Yeah. So which other conditions are you able to detect?

Well, we've done, we've done not with our data, lung cancer. Again, I mentioned bladder and prostate. We believe we can do neurological decision, neurological conditions as well. We, you know, we've done studies again, not with our data. So we, you know, we can't claim scientific rigor there, but we've modeled diseases with publicly available datasets. So, so far. We have not performed an experiment that disproves the hypothesis. In other words, we haven't failed yet, so we believe that our technology does work as we expect it to.

Yeah. So can people use it already?

No, no, no. It is not available. You know, in diagnostics specifically, there's two paths to bringing a new test to market. One is through the FDA process of pre-market approval. The other one is through a process called the LDT, a lab-developed test. We will probably do a lab-developed test first because it's a less complex regulatory path. But, you know, as you know, you have to dot the i's and cross the t's and not get on the bad side. But then it's also just ethically, you need to be comfortable with bringing something to market that actually works. And so we'll probably build an LDT path, probably with some GI diseases or a panel of them. And then we'll continue developing additional conditions, probably in groups of specialties. We may do kidney and urinary tract things next, or neurological, depending on, depending on which, what partners we find first. Bottom line, yeah.

This pathway is not under the FDA? Is it a parallel pathway?

The, so the lab-developed test It is still under FDA supervision, but in order to bring an LDT test, you're limited only to the lab. The lab is responsible for validating the science and complying with the regulatory requirements to bring an LDT test to market. There's actually probably more LDTs today being used in independent labs in the US than FDA-approved diagnostics. And I could be wrong about that, but the numbers are very high. And that is not a new practice. That regulation, I think, dates back to the '70s. And it's, you know, it's very—

You said in the '70s?

Yes.

So it's outdated.

Well, the LDT regulation is old. I don't think it's outdated because it allows science to be done with less regulatory burden and still have strict controls about its validity and, you know, be, you know, not, not allow people to market and sell snake oil, but actual real diagnostics.

So has the FDA been pretty supportive of this arena as far as AI-powered diagnostics?

Oh, very much so. It was actually a couple of years ago when the first FDA materials went out for comments on how to regulate, you know, AI-supported diagnostics. It's a very interesting field. One of the interesting things about it is that, you know, the FDA really likes to lock things down and when they approve something to be sure that they can control the quality of what comes out, that the drug is exactly the same regardless of which factory makes it, that the test is exactly the same regardless of which lab, you know, whether it's a Quest lab or a LabCorp lab or some independent lab performs that test, it's exactly the same. The results are exactly the same as it should be. But with AI-developed products, we're in a different world because literally every 6 months we can improve the model. Just like our colleagues in OpenAI and Anthropic and Google, you know, now every 2, 3 months there's a new version, you know, there's ChatGPT 4.0, 5.4 and 5.5 and, you know, Anthropic 4.8 and Stable Diffusion 5. Every 2 months there's, you know, there's better and better models. The very much the same applies to AI-driven diagnostics. We can continuously improve. So the dichotomy there is how does the FDA, how the regulatory body define the rules of making sure that the next version of the diagnostic model performs at least as well as the previous version without having to go through another clinical trial and 2-year approval cycle, et cetera, et cetera. So, so how do, how do we build the regulation in such a way that it accomplishes its goal without killing innovation? So that, that's the real challenge. But, uh, I think, uh, the folks that run the agencies are fully aware of this and highly competent and very much engaged in making these things, uh, practical. So, I have faith that things are going to work out.

That's lovely to hear optimism. Yeah. That sounds good. I'm sure, you know, people definitely, you know, we need great tools like this, you know, the diagnostic tests. And, what's the main— I mean, a lot of people are going for genetic tests, of course. So, Like for you, um, you decided not to go into the genetic direction, right? Like how?

Well, uh, different branches of scientific inquiry. Uh, but you know, genetics are a little bit like an architectural drawing. Uh, what the house looks like when it's actually built, or it looks like after you've lived in it for 20 years, is very different from what the original architectural drawing looks like. And so your genetic Genetics are very much the same way. They provide the template, but what actually gets expressed in your body is very much influenced by many other factors. And so, you know, think of the, you know, we think of genetics as the predisposition to certain things. And metabolics is telling you what the organism is like right now today.

So.

Putting the two together is very powerful. Now, it's also, I wanna be, you know, careful here. Uh, when we talk about genetics, some of the most modern genetics-based diagnostics actually identify the genetic code of tumors or viruses. They don't, they don't rely on the genetic code of the person themselves, but they, they find what is actually expressed, which is actually, again, real-time identification of the presence of biomarkers. So we need to be careful when we talk about genetics, what approach we're talking about.

I see.

Having said that, metabolomics does express the state of the organism at the moment of time much better than your food genome can do.

Yeah. Yeah. So how close is Lucentix when it comes to actually having technology people can use at home?

We're a couple of years away. And, you know, I have to be perfectly honest, you know, we are entrepreneurs. We are not, none of us have previously taken a product, you know, a diagnostic product to market. So we're looking, you know, our, we've done a one round of financing mostly with friends and family and people that have invested in our previous endeavors and have been willing to invest with us again. But our next step is going to be probably with a much bigger partner. An organization that has the reach and depth to really develop the technology. So in reality, that can be as little as 18 months or as long as a decade and somewhere in between. But, you know, we have demonstrated what we set out to do, which is that the artificial intelligence, machine learning-based approach of metabolomics based on urine samples is a viable technology for diagnostics and for companion diagnostics as well.

Yeah. And I'm just curious about, since you're using AI technology, are there some major misconceptions when it comes to AI in healthcare? Are you encountering anything that you want to—

That could take probably several conversations rather than one. I spend a lot of time actually in my day job working on artificial intelligence for personalized medical records analysis and summarization. Fundamentally, you know, healthcare cost is growing faster than GDP. Across the world, and that's unsustainable. I personally believe that artificial intelligence technology is one very powerful tool to allow us to bring that cost into, you know, to control that cost. And so, you know, as I'm, you know, I'm an engineer and I say, how do I solve the problem? I want to optimize the most precious resource we have with the technology that I have in my hands. And the most precious resource we have is time. It's clinical time. It's clinicians' time. It's doctors' and nurses' hours in the day that they can see patients and actually do medicine and not fill out forms and type data into Epic, right? So the AI is actually very good at this. And you see the deployment of ambient listening technologies and note-taking that automates that. On the front end, what we—

Writing notes.

Yeah. On the front end, what we are doing is we're saying, let's have the AI prepare the full summary of the medical history of the patient before they walk into the door. So you can save, you know, the first 5 to 15 minutes in interviewing the patient and doing the anamnesis, but be able to jump, you know, to have a very summarized view of what the current complaint is versus the full medical history and be able to hit the ground running. See the summaries of previous diagnosis, current medications, do the medical reconciliation automatically, see any, you know, chronic conditions, surgeries, et cetera, et cetera. And so that technology is available today. And so, uh, we see it as a very powerful sort of a tool. Which can be incredibly useful in clinical decision support for both patients and both clinicians and patients.

Oh yeah, human doctors can miss things for sure. I mean, because we have limited brain capacity.

Yeah, I, you know, I'm incredibly fortunate in that I'm quite healthy, but you know, with 30, 40 years of medical records, when you drop them and you see, you know, you have hundreds and hundreds of pages of stuff, which part of it actually could be related to what I'm complaining from today? You know, so being able to see the full picture and not miss something, be able to see a summary of possibly relevant pieces of information, be able to have a menu of possible diagnosis so you don't miss something and possible treatments, which quite frankly, I don't know of any physician who can keep up with new technologies coming to market. They seem, you know, and why not use technology to help you with that?

Yeah. And I think with the way the new way of how medicine is done, it absolutely needs AI because I can see how, you know, functional medicine, I think, is taking the forefront of where the medicine's going because you're actually uncovering layers and layers of data of, you know, the sometimes are subclinical manifestations of conditions. So, nutritional, microbiome, right, hormonal, all these complex data are coming together. And as a clinician trying to analyze everything, it just become— it becomes more and more and more unmanageable. And so at some point, you've got to borrow bigger computation powers than your own brain. So I definitely think the way medicine is going, we need AI to help us.

No, absolutely. I can speak to that from personal experience. My again, very close friend's daughter for several years developed a brain infection condition that was very, very rare, maybe only a few cases in the country. And literally the way it was diagnosed by trying different treatments until something was found that worked for her. And very complex case, you know, pre-AI. But I would like to believe that if we had the AI tools that we have today, physicians would've seen additional possible treatments early on rather than having to navigate specialist to specialist and, you know, 10 states until, uh, you know, everybody's suggesting, well, maybe it's this, maybe it's the other thing until they hit the right treatments that, aha, that's what it is.

Exactly. Exactly. Yeah. Uh, are you, you've heard about that AI hospital in, in China actually is, in the university that I grew up in, Tsinghua University. They built a fully AI hospital, no humans.

I have not. I'm ignorant of that. That's very interesting.

They treated tens of thousands of patients. One thing that China is not short of is people. So they have a lot of patients. And doctors are incredible because they've seen, by the time they finish medical school, they've seen more doctors, I mean, more patients than a doctor in the US who I've seen in 10 years, right? So, that's a power of experience, but they're using, yeah, AI, everything. You walk in, it's a fully AI-equipped hospital. You don't need human. It's absolutely incredible. So, that'll be a place to visit.

You know, until you're, for now, uh, until you have to get some surgery, you still probably want a human surgeon supervising at least. Uh, so, uh, you know, speaking as somebody who has, uh, two total hip replacements, uh, and my legs are still the same length, and I can still— I can ski again, for example— I, I do appreciate my human, uh, orthopedic surgeon very much, even though I know he's looking using robotic technology to do the procedures.

Right, right. Yeah, yeah, this is wonderful. So when you look at the future, so in the next 5 years, what does success look like for Luventix?

Diagnostics available in the market. Ability to either give a urine sample at the point of care, you know, when you're seeing your physician, or to get a kit mailed to you at home and mail it back to a lab and get a panel result back. You know, even at the level of triaging conditions to put you on the right diagnostic pathway, that would be a success. Obviously, we are much more ambitious than that. You know, again, at the macro scale, if we can bring in technology which costs only a few hundred bucks but can do panels of dozens or theoretically hundreds of diseases, that puts, um, population-level screening and early diagnostics, uh, into an economically feasible position. And I think Everybody has acknowledged the fact that the earliest treatment is the least expensive and the most successful one. And so, you know, for us, the sort of the ultimate success would be, you know, to have a diagnostic panel based on our technology from a urine sample that is, that costs so little that it's accessible to essentially anyone around the world.

Yeah. Something like a pregnancy test. You can just buy it and then put your pee on it and then, uh, you get the results. That's right.

That's right. I mean, again, I mean, we go to the doctor now when we get sick. Um, but we can catch conditions before we feel bad enough to go to a doctor. Anywhere between 3 to 18 months in advance if we had the screening tools at a cost-effective level. So fundamentally, you know, I personally believe that, you know, that annual exam should cover dozens, if not hundreds of conditions, not just, you know, the 4 or 5 killer diseases.

Right, right. Amazing. Well, thank you so much for your, your courage and your drive to make something new in the world and bring something that's super helpful and hopefully saving a lot of lives. Yeah. So thank you for your work. And where can people find you and follow your work and, and follow your progress?

Our site is levantix.com and, you know, The moment we have tests available, we will be marketing them. Maybe we even have another, another conversation with you and tell you about the products in the market at that point.

That would be so fascinating. Yeah, good luck to you.

Thank you very much.

Thank you so much for sharing your, your story and your your knowledge. Thank you for watching and listening. If you enjoyed this kind of content, please subscribe and turn on your notification bell so you don't miss any of my future content. And if you want to learn more about regenerative medicine or other modalities in anti-aging medicine, please go to my YouTube channel, Joy Kong, MD, and there are all kinds of videos to help you keep learning.