Why We Started Rhino Health
AI can only truly transform healthcare if it’s built, trained and maintained using diverse data
By Ittai Dayan MD, co-founder and CEO of Rhino Health.
Starting a company during a global pandemic makes sense — if your mission is to change the way the healthcare market works today, and improve the outcomes of a broad patient population. Ours is.
The Federated Learning (FL) powered Rhino Health Platform makes it easy for hospitals to accelerate the creation and adoption of AI solutions, in order to improve patient care. This can be done independently, or in collaboration with commercial developers that can scale solutions to the entire market. As patient privacy is of the utmost importance, patient data is always protected — data never leaves the system of origin. This point is essential to successfully facilitating access to the volumes and diversity of data required to build the next generation of health AI models. Used in real-world clinical settings, among today’s increasingly diverse patient population, these models will deliver consistent and accurate insights.
The Data Diversity Problem
Clinicians can tell you that there are still too many decisions made based on heuristics — broad generationalizations and quick approximations — rather than in a truly data-driven way. When I treated patients struggling with serious mental illnesses, I was exposed to a lot of available but untapped information — from neuro-imaging studies to electrophysiology exams to clinical notes — that were not processed and analyzed, at scale, for decision-making. I was excited about new developments in Machine Learning that could consume all this data to create predictions that could lead to improved medication selection, risk stratification and diagnosis. It was clear that AI had the potential to help physicians make better, data-driven decisions.
So I started my journey into AI, first as a clinical researcher, and later leading the Center for Clinical Data Science (CCDS) at Mass General Brigham. As our group grew, I oversaw the development of 25+ clinical AI applications and came to understand firsthand the underlying challenges associated with accessing and activating the data necessary to develop and train superior AI models. I launched an ‘AI CRO’ (Clinical Research Org) that tested algorithms from multiple commercial developers. Despite the many advancements in healthcare AI, the road from development to product in market is long and often bumpy. Solutions crash (or just disappoint) once in the clinic and supporting the full AI lifecycle is nearly impossible without ongoing access to a broad swath of clinical data. The challenge has shifted from creating models, to refining, validating and maintaining them.
The problem is that data is still very siloed. This means that AI models are generally trained using a small subset of data, without broader real-world context. When the model is then used to serve a larger, more diverse patient population over time, the results fall short of expectations.
The consequences for healthcare AI are significant:
- If patients receiving AI-supported care don’t experience better outcomes, clinicians lose confidence in the solution and it will never gain traction.
- Inaccuracies can lead to missed or incorrect diagnoses, blossoming into loss of life, lawsuits and cascading inefficiencies.
- AI developers are at risk of having non-sustainable unit economics, which will impact their ability to grow and scale their AI portfolio.
- Regulators will find it difficult to approve efficient solutions for model maintenance (ie, “learning models”) if they can’t tell whether the models actually improve or just “over fits” to data that doesn’t represent the targeted patient group.
Bottom-line, lack of ongoing access to a large dataset from a diverse patient population collected in different places and grown over time is holding back the positive impact of AI-based healthcare solutions.
The Federated Learning Solution
We’re building the Rhino Health Platform, that leverages Federated Learning (FL) technology, to converge AI models created on disparate datasets. Thus, we are making it possible for healthcare AI developers to leverage data across hospitals and health systems without moving data, transferring ownership, or risking patient privacy. This allows hospitals and other data holders to maximize the value of their existing IT investments, since there is no need to create redundant copies of the data (such as required by most cloud-based solutions today).
Our approach will provide the quantity, quality and diversity of data necessary to train AI models for clinical success, broadly speaking. It also accelerates the process — eliminating the need to strike and manage one-off relationships between dozens of care providers and technology developers. Further, it eliminates the complexity, expense and risk of moving and managing huge volumes of data.
The EXAM Study I previously co-led with Mona Flores, MD, global head of Medical AI and the NVIDIA Applied Research team, brought the promise of Federated Learning to the forefront. We focused on developing an AI model to determine whether a person showing up in the ER with COVID-19 symptoms and receiving a chest X-Ray would require supplemental oxygen even days after an initial exam. We used the powerful NVIDIA Clara Train SDK, and built on Dr. Quanzheng Li’s work, combining medical imaging and health records to help clinicians more effectively manage hospitalizations during a time when many hospitals were seeing a significant influx of patients.
We collaborated with 20 hospitals around the world — making this the largest, real world, most diverse Federated Learning initiative to-date. Researchers at individual hospitals were able to use a chest X-ray, patient vitals and lab values to train a local model — and then share only a subset of model weights back with the global model. This maximized the security of the federated training and ensured the preservation of patient privacy, while rapidly improving model results.
The distributed data came from a multinational, diverse dataset of patients across North and South America, Canada, Europe and Asia. The impact was striking. In a matter of weeks, we achieved a top-notch algorithm that has the potential to impact clinical care. We demonstrated the rapid and robust development that FL can enable. It was clear to me that there is so much potential to build on this type of success, and also that there was more work to be done in order to apply the FL framework to real-world settings.
During in-depth discussions with my co-founder, Yuval Baror, it became apparent to us that healthcare AI, with its own unique challenges, can leverage many advancements from other ‘AI-enabled’ industries. In Yuval’s prior role at Google, he has scaled AI solutions used around the world. Many of these learnings can be applied to healthcare. We decided to create a platform company that could benefit all the AI developers in this market, and ultimately all patients.
Busting silos by keeping data local? Automating maintenance of AI solutions in a regulated market? It was a moonshot (and still is), but everything about this made complete sense to us. In any case, if it’s not hard, is it really worth doing? To us it’s clear that the time is now.
After the recent seed-round investment, we are now scaling the team to accelerate our development efforts. We are continuing to evolve our Platform to incorporate imaging data, pathology data, structured clinical data, clinical notes and genomic data — these constitute the most clinically relevant data for impactful AI-based solutions at this time. We are collaborating with leading academic and industry creators of medical devices and software applications focused on diagnostic imaging and digital pathology. And we are helping hospital-based researchers translate new AI models into clinical impact.
Watch this space for more to come. Only the Rhino Health Platform offers a turnkey Federated Learning solution that allows researchers and developers to use FL to support the full lifecycle of AI solutions. Interested in exploring what that might look like in your area of focus? Reach out at firstname.lastname@example.org.