Federated Learning with Rhino Health

Interview with Yuval Baror

This week in our spotlight on Israeli innovation, I had the privilege of catching up with Yuval Baror , Co-Founder and CTO of Rhino Federated Computing . I’ve known Yuval for over 25 years, and I can confidently say he’s someone worth listening to!

Rhino Health is tackling one of the toughest challenges in healthcare: training AI models using data from multiple contributing institutions without the data ever leaving its original location.

Sharing data with employees or partners at a global scale, feels like second nature today, thanks to tools like Google Drive or CTERA . But what if you want to compute on that shared data without actually sharing it?

If you think that sounds impossible, you’re not alone!

More specifically, how can you train AI models on data scattered across multiple sites while ensuring it never leaves the premises? That’s exactly what federated learning achieves.?

Let’s dive into Yuval’s journey, the unique challenges of working in healthcare, and how Rhino Health is turning data privacy into an enabler of scientific progress.

Aron: Yuval, it’s great to catch up. What inspired you to create the company? Was there a specific moment or challenge that sparked the idea?

Yuval: I’m a self-taught coder and serial entrepreneur with 20 years of AI experience (including work on AI-based platforms for IT, cybersecurity, marketing, and Google Duplex). I was thinking about my next startup and was looking for a meaningful problem and the right co-founder. I found both in Ittai Dayan, a physician and former head of Data Science at Mass General Brigham. Ittai had led a groundbreaking Federated Learning study with 20 hospitals, published in Nature Medicine, revealing Federated Learning's potential but also the immense difficulty of scaling it.

We co-founded Rhino Health to build a Federated Computing platform, simplifying and scaling Federated Learning across organizations while adhering to strict privacy and security standards, to unlock the potential of diverse healthcare data.

Aron: I think federated learning is fascinating, but it can feel like a technical buzzword for some people. How would you explain it in simple terms? And why do you think it’s such a vital approach for healthcare?

Yuval: Federated learning trains AI models on distributed data without sharing the data itself, only model parameters. This is crucial when data sharing is problematic due to regulations, data size, or trust issues, all common in healthcare. Despite vast amounts of daily healthcare data generated (~30% of all data generated every day worldwide!), most remains unused in silos due to privacy restrictions. Traditionally, researchers were limited to narrow, single-site data or lengthy data-sharing agreements, resulting in slow research and poorly generalized AI models. Federated Computing, including Federated Learning, offers a solution to unlock this data's potential and advance healthcare.

Aron: Healthcare is such a complex and challenging market. I imagine it’s not easy to introduce AI to very traditional research organizations that deal with sensitive and heavily regulated health data. What advice would you give to entrepreneurs trying to bring AI into this space?

Yuval: I think the most important thing for entrepreneurs is to identify a problem that is painful enough that someone is willing to pay in order to have it solved for them. This isn’t specific to healthcare of course, but healthcare has its own set of very complicated incentive structures, payment structures, and regulatory restrictions. You might think that if you have a technological solution or an AI model that can diagnose some disease better than physicians do today, that it’s a slam dunk - of course every hospital will adopt this immediately and pay handsomely - we’re talking about human lives! But unfortunately that’s simply not how it works in healthcare.?

The healthcare industry is very large and comprises several different types of organizations - payers, providers, biopharma companies, medical device companies, etc. Each has their own set of challenges, decision makers, price points, etc. You need to identify the right person to sell to, address their specific pain points, and align your pricing to the value that you are bringing them, which may be difficult or time-consuming to prove.

Aron: I’ve been thinking about how much untapped value there is in healthcare data. It seems like there’s a huge opportunity for organizations to not just extract insights but also monetize their data responsibly. From your perspective, what are the biggest challenges your clients face in unlocking the value of their data and turning it into a strategic asset?

Yuval: So we said that 30% of all of the data generated every day in the entire world is healthcare data, right? Pretty impressive. What if I told you that 97% of all of that data goes unused? Talk about untapped value…

There are 2 main challenges as I see things:

1. Data is siloed - each organization has their own data systems and cloud accounts and it’s very difficult to share data across organizational boundaries.
2. Data is heterogeneous - each organization has their own data formats, such that even if you could access data across many different organizations, it would be quite difficult to harmonize this data in order to be able to actually compare apples in Hospital A with apples in Hospital B, let alone train an apple-identifying AI model.

As mentioned, I believe that the best solution for the 1st problem is Federated Computing - just leave the data where it is and don’t require sharing it, but still allow researchers and data scientists to work with this data in a secure, privacy preserving manner.

The 2nd problem is also challenging, and there’s no silver bullet for it. There are some emerging standards in the healthcare data world such as FHIR and OMOP, but none are widely adopted and consistent enough to be a standard way to structure healthcare data.

At Rhino Health we’ve developed our “Harmonization Copilot”, which is GenAI-powered application to help sites harmonize their data to different data models, including support for human-in-the-loop expert review, and adhering to the principles of Federated Computing where all the hospital’s data stays at rest behind their firewall. We hope this will help address this challenge.

Aron: It seems like Rhino Health isn’t just solving a privacy problem: it’s enabling science that was otherwise impossible. Do you agree with that statement? Can you share some examples of customers that were able to collaborate better and improve how science is done thanks to your technology?

Yuval: Definitely - the ability to collaborate with data across multiple organizations without needing to share this data, unlocks many opportunities for collaboration. One example is a collaboration we powered between 8 hospitals around the world to train an AI model to better detect brain aneurysms. Another example is a consortium of 6 research institutions funded by the National Cancer Institute that worked with our platform to train a model for early detection of pancreatic cancer. And there are many other examples.

Aron: Looking ahead, where do you see AI going in the next 5-10 years in the healthcare space? I think the potential here is enormous, but I’d love to hear your perspective on what’s next.

Yuval: My linear mind sometimes has a hard time comprehending the exponential pace of innovation in AI, but I think that the vision of having AI that replaces all doctors is not something that will happen in the next 5-10 years. I do think that we will see AI make its way into every aspect of healthcare over the next 5-10 years.

AI will supercharge our researchers, it will come up with new drug candidates to test, it will optimize patient throughput in hospitals, staffing and workflow within hospitals, and it will reduce a lot of the data entry and manual labor that is still prevalent in healthcare (ask any physician how much time they spend entering data into the EHR today).

I think that ultimately AI will become a “copilot” for physicians that will assist them in their decision making, and be a safety net that will catch mistakes or issues that were overlooked or require another review, but this might take a bit longer to get regulatory approval and become widespread.

Aron: Lastly, for those building AI solutions in highly regulated and sensitive industries like healthcare, what’s the most important lesson you’ve learned that you’d pass on to others?

Yuval: Be naive. Coming into an industry with an outsider’s perspective and a “can do” attitude is quite powerful. For those inside the industry “the way things are done” might be so engrained that it’s difficult to consider alternatives. But work closely with someone who has intimate knowledge of the industry - you don’t want to repeat all of the same mistakes that others have made (just a few of them).

And identify the problem you want to solve first and validate the market - don’t develop a cool technology/model that you then need to go figure out what problem it might be helpful for and who will actually pay you for it.

Aron: Yuval, thank you for this insightful conversation. I think what you’re doing with Rhino Health is incredibly meaningful to our future. I’m so excited to see where you and Rhino Health take this next!

Bonus: If you're passionate about innovation in healthcare, see this: