Cloudera Evolve 2024 | Luz Erez, MDClone

Clips
News
More from Cloudera Evolve 2024

Luz Erez

CTO

MDClone

MDClone aims to solve the data issue in healthcare with Cloudera

One of the thorniest data problems is in healthcare, and MDClone Ltd. believes it has found the solution with synthetic data.Any tech company working with healthcare has to manage large amounts of unorganized data while navigating Protected Health Information or medical records. Synthetic data evades the latter issue.MDClone’s Luz Erez talks with theCUBE about the company’s partnership with Cloudera.“We had to invent or build for ourselves something called synthetic data,” said Luz Erez (pictured), chief technology officer of MDClone. “Synthetic data, what it’s doing is mixing patients so there is no one-to-one relationship between the original data and the data that you get

play_circle_outline MDClone's journey to address healthcare data complexities

play_circle_outline Use of synthetic data and generative AI in healthcare solutions

play_circle_outline Hybrid capabilities for data consolidation and sharing

play_circle_outline Overcoming Challenges in Healthcare Data Management: Importance of Quality and Models in Analytics

Info
Transcript

Luz Erez, MDClone

Luz Erez

CTO MDClone

The CTO of MDClone, Luz Erez, discussed the challenges in healthcare data management and analysis. MDClone addresses the issues of data quality and PHI compliance by providing a platform that uses synthetic data to create a secure, anonymized environment for sharing information. The data infrastructure is built on-premises with Cloudera's help, allowing for scalability and high IO levels essential for complex queries. Cloudera's hybrid capabilities are leveraged for data consolidation and movement between organizations, enabling data sharing while maintaining... Read more

explore Keep Exploring

What led to the creation of a product in the research field in healthcare? add

What role does synthetic data play in training and testing AI processes, particularly in relation to generative AI and potential challenges with hallucination? add

What are some examples of organizations using a platform to consolidate data from multiple sources into a single cloud solution? add

What is one of the biggest challenges in healthcare in relation to data quality and statistical modeling? add

bolt Powered by CUBE AI

Luz Erez, MDClone

search

Rebecca Knight

>> Hello everyone, and welcome back to theCUBE's Live coverage of Cloudera Evolve 24 here in New York City. I'm your host, Rebecca Knight, alongside my co-host and analyst, Bob Laliberte for the final stretch. This is our last interview, but I have a feeling we saved the best for last.

Bob Laliberte

>> I think so too. This is going to be a good one.

Rebecca Knight

>> This is going to be a good one. I would like to welcome Luz Erez. He is the CTO of MDClone.

Luz Erez

>> Yes, thank you.

Rebecca Knight

>> Thank you so much for coming on.

Luz Erez

>> Yes, thank you very much for pronouncing MDClone. It is really hard. I know.

Rebecca Knight

>> So for our viewers who are unfamiliar, tell us a little bit about MDClone, as well as your journey to be the chief technology officer there.

Luz Erez

>> Yeah, so I'm one of the founders, so I was born with a company. So that's how-

Rebecca Knight

>> Company was born with you, Luz.

Luz Erez

>> Yeah. Yeah. That's true as well, yeah. Both of them is true. So we are in healthcare, and I've been a lot of years in healthcare. My partner is a CEO. He's also many years in healthcare, and we've been in the interoperability field, and we saw that there is a need for a product in the research field. And we saw there are multiple problems, like there are a lot of data inside systems, healthcare systems, hundreds of thousands of tables and fields, and no one is doing nothing with the data, zilch, nothing. So why? What are the reasons? You see, you go to the doctor, he just type data, type, type, type, type. And in the end, he ask you, "What is your condition?" "You just typed so much data. Why are you asking me? Everything is in your computer." Say, "Yeah, it's too complex to read it from the computer." This is a problem.

Bob Laliberte

>> Right.

Luz Erez

>> So we looked at the problem. There's a lot of data, it's very complex, and it's not organized in the right way. There's problem with quality. And in the end, there is another small problem. There is a problem of PHI, the problem that when I'm getting data, if inside, there will be patient information, it is not legal to get this. You need an IRB in other places called Helsinki, but you need a committee to approve it to you. So you're not going to a committee because you have other things to do after your goals session this night. So that's a problem. So we went to solve this problem, and we saw that, first of all, we need big data infrastructure, but it need to be on-prem because no one is going to take the medical data and put it on the cloud, so you need it to be on-prem. And not only you need it to be on-prem, you need it to be simple enough so the healthcare provider can install it and work with it. So we built this stack, and we got to the point that we get the data, we clean it, we organize it in the right way. It's called longitudinal. This means it's based on events And then you can ask very complex questions, but there was no query mechanism that can do it. A clinical question look like this. I want to know what medication people are looking while they have a relapse in their kidney disease, and they are on a beta blocker. What is to be on a beta blocker? What is a relapse? A relapse is the frequency changes of bed events. This is a relapse. SQL does not know how to do this, so we needed to write a new database data engine that knows how to do this. So we wrote a data engine that knows how to answer those questions. And then how do you solve the problem with PHI? So this was a hard one because you can anonymize, but anonymization is not enough. If I take your data and just remove some items, it's quite easy to see who you are.

Bob Laliberte

>> Right.

Luz Erez

>> In order to do this, we have to invent or build for ourselves something called synthetic data. Synthetic data, what it's doing is mixing patients so there is no one-to-one relationship between the original data and the data that you get. In the end, it is PHI-free, but any statistics that you'll do in it will be almost always correct. The only thing will not be correct is outliers because outliers are special people, and you threw away outliers because you don't want to know about special people or unique people. Everyone is special. Sorry. I know it's not correct to say not special. But in the end, we've built this, and now we are provisioning this platform. You don't need an IRB. You don't need to know programming. You ask a question, quite natural. You get data. It's synthetic so you can bring it to someone else. Think of one of our clients is the VA. So their typical client, patient is not a typical person.

Bob Laliberte

>> Right.

Luz Erez

>> It's unique. The normal protocols in the literature do not feed them. How do they change the protocol? They need someone that work on this data, but they can't give the data outside because of PHI. So they write the query, they get the data in synthetic form, and then they can work with outside, like universities, research organization, and so on. And you can go to the top tier Chicago . They need to be cooperative with other organizations. Now, really the top tier. How do you do this without synthetic data? You can't take your data outside.

Bob Laliberte

>> Right.

Luz Erez

>> So this really enabler for collaboration, for building models, and so on. And so this is the main thing and the reason why we are here today, so this advance.

Bob Laliberte

>> Yeah, no, it's super important, and that ability to share that data universally across the healthcare space is something that's really needed. Because we are here today at the Cloudera event, maybe you could talk a little bit about the role that Cloudera plays in enabling you to be able to share this data, create the data, share the data, and get insights out of it.

Luz Erez

>> This is the condition. Those are the conditions. We need to work on-prem because most organizations wouldn't just give to the cloud the data, first thing. Second thing, there's a lot of data, and you need very high level of IO.

Bob Laliberte

>> Yeah.

Luz Erez

>> You can't use indexes because the questions are too complex. So you need scalability. You need a stack that will work for you, for search, for queues, for everything that you do. So you can take open source for each one of them. But then if you have a problem and you have clients like we have, then you are stuck. You have to solve this. If you have any type of issue, it can be a problem in the source, but it can be a security breach in one of the component, you need to have someone to pick up the phone and call to. So just an example, we had Log4J security breach. If we weren't using Cloudera and we've built our own stack from scratch, we will have to go each one of those components and make sure, and then we have to reinstall for all our clients. It is a local installation, very heavy. We pick up to Cloudera. It's very good because we have a time gap between United States and Israel. And then the next morning, we have installation package that we can deploy for our clients. It was one that was approved by Cloudera. So for a client like, we have DOD, we have VA, those clients, they want to make sure, be sure that someone, that you have a back. And this is very important for small company that deploys something very unique and have enterprise level customers. If you don't have enterprise level, if you are selling TikTok, there's no problem. You would complain. If you have VA, if you have Chicago, those organization really needs... Yes, it should work fine, but you need the backing of someone that will able to provision solution near real time to real problems, enterprise grade. And without Cloudera, this is something that couldn't be. In the beginning, we have Hortonworks and Cloudera. We selected Cloudera, but in the end, it's not really important but they converged to one entity, so it was fine with us.

Rebecca Knight

>> What role does AI and other LLMs play in helping to generate this synthetic data?

Luz Erez

>> Okay, so our traditional product use AI, but not generative AI in order to make the journey. Now, we enter generative AI into it, and we got new capabilities. The new capabilities that we got now is the ability to create synthetic data in order to train and in order to test-

Bob Laliberte

>> Okay....

Luz Erez

>> AI processes. So now, we don't only supply it as a solution for PHI. Now, we have a new magic tool and this magic tool enabled to validate that processes backed up by generative AI doing the correct job. So one of the biggest problem with generative AI is hallucination, and this is the biggest problem for the clients that want to work with it. Say, very nice as a Copilot, it summarized Gmail wonderful. Will it open the valve on this mechanism correct? And you need really to test it. 95% is not good enough. 99 is not good enough. You need to make sure with 9.99999%. You need data. And how you can get data, synthetic data is the solution, is currently is recognized as a solution to get data for a specific job. Because without it, how do you get data for a specific scenario?

Bob Laliberte

>> Right.

Luz Erez

>> Give me this special scenario. I don't have this patient. Okay, but I need to test all possible solutions. I need to... What will happen if you have a five-year-old kid with Alzheimer's. Oh, we never had it in the last years. What will happen if you have it next year?

Bob Laliberte

>> Right.

Luz Erez

>> Okay, can you create? Yes, I can create synthetic person like this. And let's see if it handle right as a pipeline. So synthetic data become very important for generative AI, and I think this is something that is really will be important, much more important that what we are doing today and the future.

Bob Laliberte

>> Got it. Yeah, no, I think that makes a lot of sense. I think one of the things I wanted to ask about was you mentioned a lot of your data is on-premises, but so much is stored in the cloud today as well. There's lots of data that's up there. How do you leverage Cloudera and its hybrid capabilities to help your clients? I'm sure if you're sharing data that can't be on-prem. That's going to be somewhere else. So I wonder if you could touch upon the hybrid nature of-

Luz Erez

>> Sometimes there are such needs, especially, again, if you work with multiple organizations. I give you two examples. In the COVID time, NIH used our platform in order to get data into the cloud from 60 different organization, synthetic data into a single cloud solution. Today, we are doing for the FDA devices, data consolidation. Those tools needs to do not only processing data, but moving backward and forward processes as well, like federated learning.

Bob Laliberte

>> Right.

Luz Erez

>> You don't want only to use the run data, because sometimes to get the data consolidated, this is very complex, sometimes you only need to move the knowledge. So we call it weight matrix. Weight matrix are the holder of information and consolidated in one place. So we give these capabilities using hybrid. Usually when we speak about hybrid, this means having the data. It can be on private cloud, public cloud, but usually it'll be on real prem, on real service, and moving it to a hybrid centralized solution that will gather the information from multiple sources. And we have a very good product today that does this, FDA, NIH. There are multiple uses for this because there are so many use cases.

Bob Laliberte

>> Right.

Luz Erez

>> And many organizations are not reluctant to share data. We're all saying, you're going... Sharing is very nice, but when you really get to the low level. So again, our capability of collecting synthetic data and to conceal the source as well is very important. Because if you ask a healthcare provider, I don't want to give names, but almost anyone, they wouldn't like to give the information that they lost.

Rebecca Knight

>> Well, because privacy. This is sensitive information and-

Luz Erez

>> It's not only privacy. They don't want to know that their system is not the best.

Rebecca Knight

>> That is another reason. I see.

Luz Erez

>> Being frank, in frank, it's not only PHI. It's more than PHI.

Rebecca Knight

>> I want to ask you though, because you talking about this reluctance to share data too, is that the biggest challenge that you're seeing when it comes to healthcare data analysis and management? If it's not, what is it? And what are the solutions that MDClone, with the help of Cloudera, are working on?

Luz Erez

>> Yeah, so again, Cloudera is our infrastructure, so everything I tell you about MDClone technically is done over Cloudera. So I can't speak about MDClone without Cloudera. Without it, we really have no capabilities at all. So yes, reluctance is issue, but I can't solve it, and really, Cloudera can solve it. So what capabilities that we give is to conceal information about patients and origin of the organization. The biggest problem with healthcare is the quality of the data. The quality is very poor, and there are two ways to do. One is to clean the data. This is very hard because second law of thermodynamics means something is bad.

Rebecca Knight

>> Right.

Luz Erez

>> Fix it very, very hard. But second way to do it is to do a lot of statistics and build statistical models over it. And usually, this is... So you need to collect a lot of data. If you don't have enough data, you can't build proper statistical models, and they will be very poor. So one of the biggest challenges is to get enough data, the right data, and in the end, enable to build the models that those models will make sense. I see all the times, three times a week, people that build models, write articles. I wouldn't like to be their patients. Totally

Rebecca Knight

>> I think you said it well, enough data, the right data, and models that make sense. Luz Erez, thank you so much for coming on theCUBE.

Luz Erez

>> Thank you for having me. Thank you.

Rebecca Knight

>> A really fun and interesting conversation.

Luz Erez

>> Thank you.

Rebecca Knight

>> I'm Rebecca Knight for Bob Laliberte. Stay tuned for more of theCUBE's live coverage of Cloudera Evolve 24. We will have our CUBE insights coming up or right after a quick break. You're watching theCUBE, the leader in enterprise tech news and analysis.