AI in Healthcare: Innovations & Safety

AI in Healthcare: Innovations & Safety

Audio

Listen - AI in Healthcare: Innovations & Safety

Transcript:

Generated by AI

Announcer: 0:00

Welcome to MedEvidence, where we help you navigate the truth behind medical research with unbiased, evidence-proven facts Hosted by cardiologist and top medical researcher, Dr. Michael Koren.

Dr. Michael Koren: 0:12

Hello, I'm Dr. Michael Koren and we have a really special guest today for our MedEvidence podcast. I want to introduce everybody to Dr. Ittai Dayan, who is a physician, a data scientist and somebody who has published a very important paper about the use of artificial intelligence in medicine, and Dr Dayan and I are going to have a conversation about the use of AI in medicine, and I want to learn more about this, because I'm both really excited about AI but also an incredible AI skeptic at the same time. So I want you to kind of get me clear on what I can expect and what I shouldn't expect, and also have our listeners and viewers be able to understand what's a reasonable expectation, what may not happen in any time soon and why we shouldn't be fearful of artificial intelligence in medicine. So, with that, Dr. Dayan, why don't you introduce yourself to the audience and tell us a little bit about your background?

Dr. Ittai Dayan: 1:12

Hello everyone and Dr. Koren, thank you for inviting me today and my background started from clinical medicine, worked in emergency medicine for a period of time, as well as completed research in neuroimmunology and some additional rotations.

Dr. Ittai Dayan: 1:33

I later on went to work with Boston Consulting Group's healthcare practice consulting to pharma companies, provider pay organizations, provider pay organizations mainly on topics of digital transformation and harnessing analytics into the business processes. I later on went to work for Mass General Brigham, where I led the development and deployment of AI solutions into the clinical workflow and, together with commercialization partners, through FDA approvals and the likes, also started developing model validation capabilities and launched kind of academic-based CRO for ensuring that only the best models go to market with the best corpus of evidence. In my struggles to find data and collaboration across healthcare systems, I've identified the need for a common data processing platform for medical institutions. One of the key enabling technologies of that would have been something that keeps data local and doesn't require you to transfer, consolidate, share data. The key enabling technology for that is federated learning, the topic which I published on in Including Nature Medicine in 2021, which I think you're referring to that paper most likely.

Dr. Michael Koren: 3:03

Very impressive paper.

Dr. Michael Koren: 3:04

Congratulations on that.

Dr. Ittai Dayan: 3:07

Thank you very much. And that was the kind of origin story of my company, which is taking that federated learning concept and turning that into a multi-cloud hybrid data computation platform with specialization in the healthcare and life sciences vertical, and with the attempt of bridging data silos and getting more data, more diverse data and better evidence in the development and deployment of AI and analytics products.

Dr. Michael Koren: 3:39

Beautiful, well, very impressive background and again, thanks for joining this episode of MedEvidence. So just get us familiar with just some of the terminology. I think people kind of get confused, including myself, quite frankly. So we talk about digital transformation. We've had computers and software and medicine for decades. At this point I remember I did my third year medicine clerkship at Beth Israel Hospital in Boston and it was the first hospital in the Boston area to have a computer system and at that time what the computer system did was show you what the lab values were for the patients, whereas when I did my surgical clerkship at Mass General I had to walk up to the lab and get this sheet of paper that came off of some rickety printer that had the lab values for my patients. So that was the early stages of digitalizing medicine. But that's different than what we're talking about now and it's different than some other software initiatives and I think it's different than AI. So help people understand those distinctions.

Dr. Ittai Dayan: 4:47

So health IT has already become pretty mainstream in the healthcare world, with America taking a leadership position in that for the Meaningful Use Act and additional modernization efforts, and I think that's kind of like a big part of that. As part of the health IT evolution there's more integrated systems and ability to source data in more kind of like common ways. I think analytics and algorithmic decision-making in medicine is not something new in medicine is not something new, and every person who has ever used a cell counter in a hematology lab or any IVDs have been using the kind of data products for a very long period of time. There is today a further transformation of a field or at least the beginning of a transformation of a field, with having the ability to analyze unstructured data in making more granular understanding of it and making better decisions based on that, which some of them is as easy as introducing algorithms with multiple data inputs and kind of like multivariate logistic regressions into decision making. Some of them go into neural network and now, with Gen AI and other things of that nature, a lot of the buzz in the industry has been around enabling technologies that allow that, and everybody's talking about kind of a Chat GPT, gemini, gpu, processors, platforms and all sorts of stuff of that nature.

Dr. Ittai Dayan: 6:41

I think the translational efforts of many of these technologies into the healthcare workflow are still fairly early and I think that by itself is creating kind of a lot of interest and excitement in the medical world, also a lot of kind of a Luddite approach to this in some cases. One of the things that amused me after publishing my paper on natural medicine was seeing how many papers there are that say federated learning will not cure the data diversity problems. AI is not a silver bullet and some of these technologies have barely gotten into play and already there's kind of a strong rejection of many of them I think there's a need to increase the translational research and also modernize a lot of the translational research into AI, and there's also a need to evolve regulatory frameworks, whereas today kind of a urinary catheter and software as a medical device are regulated differently but not kind of like hugely differently.

Dr. Ittai Dayan: 7:52

Many thoughts have been shared in the field and some more I believe will be implemented in the upcoming years.

Dr. Michael Koren: 8:00

Yeah, so it's interesting when you talk about the controversies. What we like to say here at MedEvidence is that we have the disease of objective I should say pathological objectivity. And we also have relentless skepticism that we use to pound out the truth. So I think that's an important part of what we do in clinical research is ask questions, be skeptical and then, if our theories hold up to that skepticism, then we can be confident that they're reflecting a truth, they're reflecting something that's reliable for taking care of our patients ultimately and for improving lives. So it's an interesting process when you get into it. So let's get into a little bit more sort of concrete examples. Is there something that comes to mind that you think is a success to date of artificial intelligence in medicine that you want to share with listeners?

Dr. Ittai Dayan: 8:56

So first of all, in the research world there's many successes and many proofs that a machine eye can interrogate things better than a human eye and many things can be predicted in kind of high degrees of accuracy. And I think you know nature and science and nature medicine and all the good papers show kind of like every month new things in terms of translating that into viable clinical products. We've seen some early successes in radiology operations and I'd call that radiology operations to a large extent because it's not always diagnosis and treatment but rather kind of triaging cases and improving workflow efficiencies. But I think we have seen some notable results there. But I think we have seen some notable results there. I think we've seen quite a lot of pharma R&D successes in which there's new biomarkers and new diagnostics, companion diagnostics, brought into studies in many kind of variable disciplines.

Dr. Ittai Dayan: 10:06

I'm seeing and we've seen a ton of automation in hospital operations in terms of reducing kind of like processing of red tape, transcriptions and things all around that nature. I don't think I've seen many cases of successful multimodal data fusion for making step change in decision making in an oncology tumor board, for example, and there's definitely some lagging behind in terms of again translating some of the findings. But I think we're trending into that direction. I think it's going to happen and I think we're seeing, in the industry side, kind of a bit of a reset in terms of expectations from new companies spinning out in regards to what we're actually going to achieve, and so I'm more focused on tooling and infrastructure enablers rather than kind of like rush ahead into diagnosing cancer. I also think there's more of a realization that the transformation of a medical practice will come from inside medicine and academia.

Dr. Michael Koren: 11:34

So how about the flip side? It's always important for us to talk about the disasters. Have there been any AI disasters that affected medicine, that were part of learning? Again, we're not here to point fingers, but just when we make mistakes, we need to learn from our mistakes, and what I like to say is that the secret of success is not making mistakes. The secret of success is recovering from your mistakes as quickly as possible. So I'm curious if there's anything that jumps to your mind from that standpoint.

Dr. Ittai Dayan: 12:07

o largely communicated catastrophes I don't believe there have been many of. I think we've seen cases where models AI, diagnostic models have gone out to market and have seen a substantial dip in performance, leading to many false positives and kind of like doctors being alerted too many times, not that dissimilar of kind of patient monitoring devices from a previous generation. I've also seen not necessarily in the clinical workflow as much, but in the payer-provider engagements, many models that could even be rule-based systems that have denied care from patients based on lack of a kind of deep understanding of the patient population or implementing model designs that may have been discriminatory or seen as discriminatory towards certain populations due to innate population characteristics. Um, so there, I think there have been some lawsuits.

Dr. Ittai Dayan: 13:12

I I think this isn't like not the best topic to go into deeply, because I'm not a big expert.

Dr. Michael Koren: 13:18

And how about this concept of AI hallucinating? We've we heard about that? Is that something that you've had to deal with? Or is that something we should be fearful of in medicine, or you think it's a small issue?

Dr. Ittai Dayan: 13:32

That's an excellent question. Clinicians themselves hallucinate and make wrong decisions. I think it's important to consider that, because we're much more accepting human error than machine error here, and that by itself is something I think it's worth a discussion. But, um, generative AI models, which you're referring to, do indeed hallucinate and say a lot of nonsense. That's how nonsense is often driven by the kind of data that was used in order to train, quote, unquote and prepare them. The level of impact this would have would be dependent on how integrated into a clinical workflow some of these solutions will be. Right now, for example, they are not deeply integrated. Our doctors around the world asking kind of generative AI questions and getting the silly answer and maybe actioning on that.

Dr. Ittai Dayan: 14:34

It might be, but I'm not familiar enough with it. I would say that I don't think that any of the kind of public broad models today are anywhere close to be a one-stop shop for clinical decision-making, and in many of these cases we need to actually package an entire product around this, add guardrails, do ongoing validation and monitoring and things that are very different from the Silicon Valley kind of way of thinking about R&D and product maintenance, and so I suspect we're still a year out from the point in which this will become a big deal. I know there's a lot of transcription and kind of processing like textual processing stuff. So it's a bit tough to say I'm not familiar enough with the performance and some of these are fairly early products.

Dr. Ittai Dayan: 15:40

Um, so it's, it's. It's a bit tough to say uh, at the end of the day, um, I don't think we're going to easily get to the point where a doctor doesn't need to at all take a glance at notes or not need to at all kind of like sign off on things. Uh, I think that will remain an issue that will be required for the time being and probably even the future. I would put on the flip side of this that today there's a lot of lack of care provided due to lack of automation, and I've seen in cases where we had like stroke networks, for example in second world countries or countries which aren't the US and others, where it's kind of like, do you prefer waiting for two weeks to get something evaluated or would you like a bot to review your imaging study and maybe make mistakes, right, right.

Dr. Michael Koren: 16:36

That's interesting.

Dr. Ittai Dayan: 17:05

So you mean one customer.

Dr. Michael Koren: 17:10

What type of customer? In other words? Is there going to be some big high-tech firm that offers an AI bot that you can own for $100 a month with your cable subscription and you can ask this bot questions and have AI evaluate? Is this something you need to go to the emergency room for or something that can wait, or maybe even suggest a home remedy?

Dr. Ittai Dayan: 17:42

I'll make the approximation of that to the exciting world of medical devices and say that I am highly skeptical about that scenario.

Dr. Michael Koren: 17:50

Okay, so be more institutional and provider-based.

Dr. Ittai Dayan: 17:55

I mean you might say that either this becomes kind of a med-device-like field, when you have a lot of small companies competing with each other and none of them gaining huge scale. You might have a player who knows how to create highly customizable products, kind of like an open AI or Google Cloud kind of character, who then gets into clinical workflows and the doctors kind of like contributes to the corpus of knowledge of it and stuff like that. It seems to me highly unlikely, given the regulatory issues around that I see this more as ultimately something that doctors use like the evolution of doctors using specific tools, customizing the tools for themselves, and it could be that within the next generation, some of these tools become so sophisticated and some of these like technology capabilities become so commodity that there will just be a few big players who provide similar products in a competitive market. The difficulty of adjusting these kind of AI devices to different workflows in different locations under different guidelines and medical practice behaviors seems very tricky.

Dr. Ittai Dayan: 19:34

We tried to collaborate with European institutions because we were kind of and they were as well powerhouses of building AI algorithms and kind of a grant-funded environment, and we found that many of the problems that interested large, very large academic medical centers in America interested nobody else. Because large institutions in America have scale. The clinical triage workflow is very exciting to them, kind of like practical business problems. Many places outside have much smaller workflows, have lower degree of specialization, have lower degree of realizing operational wins from implementing AI et cetera. And it just showed to me that ultimately the market will not be conquered by kind of like the 50 leading algorithms. That will be conquered by thousands of algorithms with different optimizations for different problems and questions et cetera. That kind of moves the whole discussion to the platform level.

Dr. Michael Koren: 20:43

So opportunities for small companies right.

Dr. Ittai Dayan: 20:48

I think the world is going into small companies. Innovation is running very quickly. Large companies are incapable of identifying all the right or aren't interested in identifying all the right, kind of like small nuggets when we ourselves work with big companies. One of the reasons we work with large kind of technology partners like NVIDIA and Google Cloud and also an Amazon instance and kind of like others, is because these companies don't want to yet put on all their eggs in that basket. That's how the venture world works. That's how scale-up works.

Dr. Ittai Dayan: 21:25

Pharma companies are very, I think even maybe a bit unusual in the market in that regard of making kind of like massive bets at times on earlier assets. And even then the small biotechs and virtual biotechs and kind of like flourishing of specialized CRO's and all that are all a sign of kind of like faster innovation going to market and that needs to be de-risked in a more cost-effective and adequate way

Dr. Michael Koren: 21:56

Tell us about your company, what the vision is for your company, and then you can talk again about your great paper in Nature Medicine and how that kind of fuels what you're doing on the entrepreneurial side at this point.

Dr. Ittai Dayan: 22:09

Perfect. So in the midst of the COVID pandemic, we developed at Mass General Hospital and Harvard Medical School an algorithm that could predict, with a fairly high degree of accuracy, which patients would require mechanical ventilation, which patients would require some form of oxygen supplementation.

Dr. Michael Koren: 22:30

With COVID. These are COVID patients right?

Dr. Ittai Dayan: 22:33

Yes, yes, and we were looking for a way to disseminate this algorithm to additional medical centers. It was clear that other medical centers would not trust an algorithm that hasn't gone through some step of validation. We had the option of going to the FDA or getting an emergency use authorization, which seemed impractical, and we decided to take a federated learning approach in order to retrain and test the algorithm on a large number of medical institutions.

Dr. Michael Koren: 23:03

So explain to people that terminology federated learning.

Dr. Ittai Dayan: 23:07

So federated learning is a way to train algorithms on disparate data sets without needing to exchange data. The technology itself is based on running training tasks of AI models, of different versions of the AI model in the different locations, merging of the parameters of the model in a central orchestrator and then updating the models that sit within the data sets in order to ultimately generate a globally optimized model.

Dr. Michael Koren: 23:45

So you have a learning model and you add another institution and then how does that get into the learning model and how does it train the model? Maybe that might help people understand a little bit better train the model.

Dr. Ittai Dayan: 24:03

If I give a really kind of like intuitive and a bit simplified version. I'd say that if you have a linear curve that says Y equals A times X plus B, then instead of centralizing all the Xs and all the Ys and then putting one scatterplot and kind of running the line all the Xs and all the Ys and then putting one scatter plot and kind of running a line, you only show the A's and the B's of the function and use them in order to deduce how an optimal curve would look like.

Dr. Michael Koren: 24:31

So that's kind of the mathematics of it, simplified. But you're talking about this concept of data sharing and the guardrails around that and the sensibilities around that, so tell us how that fits into this process.

Dr. Ittai Dayan: 24:47

Absolutely so. In a federated learning collaboration, different parties can contribute their data for specific tasks without having to actually move the data outside their firewall, outside their ownership, and so, in the case of the exam study which ultimately was published in Nature Medicine, we were able to set up 20 medical institutions worldwide within less than two months, because we eliminated the need to share data, and the federated computation engine we provided was a common platform for processing data in each one of these institutions.

Dr. Michael Koren: 25:30

I understand.

Dr. Michael Koren: 25:32

Yeah, interesting.

Dr. Ittai Dayan: 25:33

We kind of solved an IT problem and solved the privacy and data sharing problem at the same time by allowing the different partners to approve a specific use case of their data and not having to diminish the control over the data.

Dr. Michael Koren: 25:51

Got it.

Dr. Michael Koren: 25:53

Yeah, so that's super interesting. So let's get into some of the clinical practicality of it, specific to the COVID-19 pandemic and some of the concerns early on in the pandemic and the learning that we needed to have happen in the beginning of the pandemic. So obviously, when we first were dealing with COVID, the prognosis, particularly for older people, was horrible and we had to make decisions about who was going to go into the ICU, who was going to be put on a ventilator. I remember I was in the hospital in active during these times and there was one theory out there that maybe putting people on a ventilator early and quote, resting the lungs was the best strategy to improve prognosis, but that turned out to be really, really wrong. So tell us a little bit about your computer model and how it helped us make these practical decisions about who goes to the ICU, who gets this type of treatment or another.

Dr. Ittai Dayan: 26:50

So the algorithm used.

Dr. Ittai Dayan: 26:54

To begin with it predicted a score, a continuous score, so it wasn't a yes no answer, but by itself testing its calibration and understanding, being able to compare cases side by side and define specific cutoffs for decision-making, but without the need of saying, oh, the bot said yes or the bot said no, but by itself itself was already helpful.

Dr. Ittai Dayan: 27:20

Second of all, the algorithm used data that was fairly objective, like lab data, vital signs, age, things that the clinician can discern using some laboratory tests, some physiological tests and simple medical record examination, and not require clinical impression. Many of the algorithms at that time had a lot of clinical impressions in them but made the models very not generalizable because they also were very kind of curator-dependent. In addition to that, we leveraged multimodal data and added an imaging study of a chest X-ray. So additional kick to the model. And in addition, when you included a good number of different inputs, you created a more robust algorithm that was able to kind of like sift through cases that otherwise you may have made the mistake, with a high degree of variance and inaccuracy in it, cases that otherwise you may have made the mistake with a high degree of variance and inaccuracy in it.

Dr. Michael Koren: 28:21

Was the output something like likelihood of mortality or likelihood of needing the ICU?

Dr. Michael Koren: 28:35

Give us a little bit more flavor in terms of the clinical output that the clinicians were able to see based on the learning

Dr. Ittai Dayan: 28:38

We optimized. So it was a regression model. It provided a score from 0 to 1 or 1 to 100, depends on how you set it out and we applied the different cutoffs and different normalization to an entire score in order to discern patients who are mostly likely to remain on room air or be discharged, patients that would require low flow oxygen, patients that require high flow oxygen, mechanical ventilation and death. Okay, so these were kind of like cutoffs, and the scoring we had to give to the algorithm itself was based on a weighted averaged area under the curve for making each one of these discriminations.

Dr. Michael Koren: 29:21

Got it, and this could happen, obviously, relatively early in the course of presentation, and I would imagine that's when it would have the most predictive value for a clinician.

Dr. Ittai Dayan: 29:35

The model was trained on data and tested on data that was captured in the emergency department, largely with patients that showed up with respiratory symptoms, meaning somebody who felt like they needed to be evaluated for respiratory illness. So there was some selection in that regard.

Dr. Michael Koren: 29:58

Okay, yeah, terrific. And can you share with us how the institutions involved were able to deploy the data? Give us some examples of how it made a difference for these institutions?

Dr. Ittai Dayan: 30:10

The model itself by the time we completed the study, and so it was implemented. It was used for population-level studies, and many institutions have not tracked each and every one of them, but some used it in kind of a prospective evaluation of the model with the intent of implementing it. The model itself was an excellent test case of AI and an excellent test case of federated learning. Training also required a lot of multimodal data, as I said, and it was not a trivial one to deploy. We ultimately did deploy that at Mass General Hospital,

Dr. Michael Koren: 30:53

And any feedback from ER staff or anything that's shareable at this point?

Dr. Ittai Dayan: 31:01

I think that's a bit later than the publication and me starting Rhino Health. I'd let the ER speak for themselves.

Dr. Michael Koren: 31:12

All right, We'll let them speak for themselves. Sounds great. So you mentioned Rhino Health, so tell us a little bit about that. I'm fascinated by that. So tell me where'd you get the name and also what your hopes are with the company.

Dr. Ittai Dayan: 31:27

So the study itself showed that the federated technology is very powerful and it moves the needle in terms of setting up collaborations, executing collaborations, bringing in more data, bringing in more diverse data and getting better algorithms. That was great. The sustainability of these kind of efforts was very questionable in the sense that you needed a lot of IT collaboration. You needed a Rhino Health of IT collaboration. You needed a lot of kind of like local collaborations, a lot of skill set in each institution. While the computation method of federated learning was already reasonably established and still evolved quite a bit, the need for an edge compute-based platform that would support a collaboration like that also became very clear.

Dr. Ittai Dayan: 32:14

Edge compute, in kind of like simple terms, is the computation happens at the data source and doesn't require kind of like centralizing data. Edge compute in this case allows you to create multiple computation tasks, not just federated learning. It can be federated learning, it can be execution of a model, it can be data analysis and it can be many other things. Rhino Health is an edge compute-based platform that allows the usage of different applications on data sets, including federated learning. And a lot of the blueprint for defining the platform was influenced by the experience of publishing findings in a highly rigorous paper such as Nature Medicine, which asked us a lot of questions on the underlying data identifying sources of bias, being able to sustain ongoing access to that data in order to test this with external independent reviewers, the ability of expanding this collaboration into additional sites, and on and on and on, which are not covered by kind of like federated learning per se, and that influenced a lot of our roadmap and how to build something that would be relevant to the medical world that's at large.

Dr. Michael Koren: 33:37

Yeah. So for some of our listeners who may not be familiar with Nature Medicine, it's a very high impact, highly respected international medical journal and if you get something published in that you know the author had to go through a painstaking process of dealing with a lot of questions from the referees and the editors and I personally have gone through that process so I can attest to the fact that getting that paper published was really, really impressive. So again, congratulations. Now getting back to the Rhino Health thing, tell me a little bit more about how you started the commercialization process and again, I'm really curious about the name Rhino.

Dr. Michael Koren: 34:26

Where does that come from?

Dr. Ittai Dayan: 34:27

We started out by working with academic centers as a design partners and enabled pushing several academic collaborations that were stuck due to data sharing issues, including with the National Cancer Institute and different NIH-sponsored efforts, a European consortia and others. We evolved from there into providing more commercial use cases, including with the biopharma industry, professional societies such as the American College of Ideology, continuing deepening our relations with different big tech companies, taking us even broader into a more global scale. Today we have around 50 healthcare organization nodes worldwide and we're growing quite quickly and I expect us to get beyond 100 by the end of this year. The name Rhino Health came from the need to find a name for the company that isn't already trademarked and the fact that we liked the idea of the rhino as a symbolism for breaking through obstacles, so we saw kind of like data sharing and data silos as a serious obstacle.

Dr. Michael Koren: 35:33

Interesting Okay yeah, is this crazy statistics that rhinos are actually responsible for a lot of deaths in certain parts of the world because they can be fierce and unpredictable, from what I understand. So hopefully you're a fierce company and hopefully you're taking something that's unpredictable and making it more predictable.

Dr. Ittai Dayan: 35:55

I think that's a very fine way of positioning that.

Dr. Michael Koren: 36:00

So tell us what your aspirations are for Rhino Health and who you envision your customers are, and what problems will your technology solve for these customers.

Dr. Ittai Dayan: 36:13

Today we've expanded from doing federated learning into doing a full stack of data activation, from harmonizing data with a Gen AI powered co-pilot, by the way, one of the low risk, high impact users of Gen AI into data analytics exposition and creating collaborative workflows between different institutions and as such, in fact, we've kind of gone beyond AI and into analytics, reporting, quality and et cetera, with excellent partners in doing all of these, and we've always been kind of a platform enabler of existing medical needs. As we progress onward, we want to see ourselves as activating the data of more and more hospitals worldwide and enabling a new collaboration method, one that doesn't care about federated learning, doesn't care about edge compute, doesn't care about any of these , but needs to create a private, secure and responsible sharing of data insights in order to promote medicine and drive a transformation of practice.

Dr. Michael Koren: 37:22

So give us an example of a customer, what type of customer you have or, if you can, a specific customer and the kind of question or problem they want you to address specific customer and the kind of question or problem they want you to address.

Dr. Ittai Dayan: 37:33

We have customers that need to develop biomarkers for clinical studies based on biometric data, which is a very big problem for secondary leverage. We have customers who are worried about the safety of AI and want to test it at scale in multiple institutions but cannot centralize kind of like entire archives of each institution just to test a product on it and don't want to integrate each AI product one by one just in order to test it. That's a very prominent emerging need of kind of like a safety net for AI users. We have customers using us for longitudinal patient analyses, with data that remains tethered to the medical record, so you don't have a need of kind of getting to a level of redaction in order to centralize it to a centralized cloud. We have customers working on outcome research in a way that allows different hospitals to understand aggregated outcomes with a certainty that there's no kind of discovery of the data in a way that we disagree with, so the ability to enforce a contract in that regard.

Dr. Ittai Dayan: 38:44

We have customers who use us for bringing the GNI to the data, since most hospitals' medical records and data warehouses are still on-prem or on private cloud and they are less interested in exposing it to public cloud APIs. We have customers looking to harmonize their data at scale into FireR4, FireIL and OMAP 5.4 standards without just being kind of like a year-long process by using the GPTs and different generic capabilities. So, as you can see, it's multipurpose and it allows you to do whatever kind of computation, analytics and pipelines you do on centralized data without the need for centralizing data.

Dr. Michael Koren: 39:34

But it's still one of the things I'm gleaning. It's still mostly in the data analytic realm. It hasn't really transferred to the patient question realm quite yet or the clinical process dilemma question that a lot of institutions are dealing with.

Dr. Ittai Dayan: 39:52

Much of it is. We're also being used for hospitals to a certain quality of their actual workflows and optimize for workflows as part of data analytics. We're not a clinical workflow solution in the sense that you'd punch a number and Rhino will tell you what to do. That is not at all what we do today. We are enabling researchers who have developed some of these products to optimize them and make them ready for clinical deployment. We may, in the future, also support that, but frankly, that's a bit of a more well-trodden space and not something that we are actively doing right now.

Dr. Michael Koren: 40:28

And my final question to you is part of our mission in MedEvidence is to have physicians talk to one another and have people who are listening to the conversation glean some truths. So I'm curious to see how you perceive your background as a physician as something that's helped you in your current space. You mentioned to me off-air that you're not currently practicing, but you've certainly practiced extensively in the past and tell us a little bit about how those clinical sensibilities, that training being a physician, how that's impacted your ability to be successful in your current role.

Dr. Ittai Dayan: 41:07

There's a huge kind of a big phenomena of a solution looking for a problem in tech and I have been able to bridge better the solutions to an actual problem that a clinician decision maker would be making, rather than something that kind of sounds good would be making, rather than something that kind of sounds good, and I think I've been better or well positioned to analyze the kind of roadblocks and hurdles and the way of implementing that technology in a regulated healthcare setting.

Dr. Michael Koren: 41:44

Wow, I love that answer. Thank you, Ittai. This was a great conversation. Thank you for educating me. Sometimes it's kind of confusing hearing people talk about AI and how computers are going to revolutionize medicine, and you know there is skepticism. You alluded to that and not all computerizations have been necessarily physician-friendly. But I have a better sense for what you're doing. Keep up the good work and thank you for being a guest on MedEvidence.

Dr. Ittai Dayan: 42:15

Thank you very much for having me and for the stimulating conversation. Speak soon.

Dr. Michael Koren: 42:20

Okay, take care now.

Announcer: 42:21

Thanks for joining the MedEvidence podcast. To learn more, head over to MedEvidence. com or subscribe to our podcast on your favorite podcast platform.

Video

Watch - AI in Healthcare: Innovations & Safety

Ever wondered how artificial intelligence could revolutionize the healthcare industry? Join us on MedEvidence as we sit down with Dr. Ittai Dayan, a distinguished physician and data scientist who shares his expertise in emergency medicine, neuroimmunology, and AI development at Mass General Brigham. You'll gain an understanding of federated learning, a groundbreaking method that allows data processing without transferring data, fostering collaboration among healthcare systems. We also differentiate between digital transformation, traditional health IT, and AI, addressing the enthusiasm and skepticism surrounding AI's role in medicine.

In this episode, we delve into AI hallucination in medical contexts, contrasting human and machine error while discussing the limitations of generative AI models in clinical setups. Dr. Dayan highlights the critical importance of product safety measures, ongoing validation, and vigilant monitoring. Discover the creation of a sophisticated multimodal algorithm by Rhino Health aimed at predicting patient outcomes in emergency departments, and learn about the hurdles of data sharing, regulatory challenges, and commercialization in the ever-evolving field of medical AI.

Talking Topics:

  • Artificial Intelligence in Medicine Explained
  • Federated Learning in Healthcare
  • Advancing Healthcare Technology With Rhino Health

Learn more about Dr. Ittai Dayan:
As co-founder and CEO of Rhino Health, Dr. Ittai Dayan transforms how healthcare AI solutions are created, adopted and measured. The Rhino Health Platform provides access to a large, distributed dataset from a diverse group of patients, powering models that deliver consistent results and, ultimately, improve health outcomes for large populations of patients. Drawing on his background as a clinician and researcher, Ittai is passionate about creating equitable access to advanced AI-based diagnostics and treatment pathways - across increasingly diverse patient populations. He led the EXAM study, published in Nature Medicine, the world’s largest and most prominent study to-date utilizing federated learning (FL) to train a healthcare AI solution on diverse data across institutions.

Ittai graduated from the Johns Hopkins Bloomberg School of Public Health. He earned his MD and his Bachelors of Science from The Hebrew University of Jerusalem. He serves on the Editorial Board of Nature Digital Medicine, a leading publisher of peer-reviewed scientific studies. His own research has been published in journals including those from Nature group, IEEE’s Journal of Biomedical and Health Informatics, and at the MICCAI conference.

Follow Ittai Dayan

Follow Rhino Health
 

Be a part of advancing science by participating in clinical research.

Recording Date: May 31, 2024
Music: Storyblocks - Corporate Inspired