AI Today Podcast: AI and Healthcare – Interview with Vignesh Shetty SVP & GM Edison AI and Platform at GE Healthcare Digital

The AI Today podcast, produced by Cogmelidica, cuts through the hype and noise to identify what is really happening now in the world of artificial intelligence. Learn about emerging AI trends, technologies, and use cases from Cogmelidica analysts and guest experts. Hello and welcome to the AI Today podcast. I'm your host, Kathleen Mulch. I'm your host, Ronald Schmills, and we are thrilled to have you our great listeners here on the AI Today podcast. We're going now strong into our 50th year here now. Submit our fifth anniversary back in September, actually, 2021. So we're actually halfway into our fifth season now. And we are really love listening to you, our listeners. We've our audiences growing at a phenomenal rate. We saw what we were just promoted now in the top three of all podcasts on the subject of AI, top two, actually. But we know in order to stay in that position, we know we need to keep our listeners happy. If you're listening to the AI Today podcast for the first time, we spend our time understanding what is happening with AI today and learning from people who are in the industry experts, practitioners, who are actually trying to put AI projects into place and hearing about their successes, their challenges and everything in between. And of course, also we spend a lot of our time on basic education. We've heard from many of you, you know, just understanding what the various concepts are around AI and machine learning, different aspects of the market. And we've had some great series we've done in the past. If you haven't learned, listen to our AI failure series, talking about all the ways that AI projects fail, we encourage you to do so. We also have talked a lot about AI from our education. We have a training and certification in CPMAI, called the cognitive project management for AI methodology. And we've shared a lot of those insights from our training and education. And you will definitely be hearing more about that. So definitely make sure you're subscribed. But we have a great guest for you on our podcast today. And I'm thrilled to introduce and bring here into our AI Today podcast, Vignesh Shetty, who is the Senior Vice President and General Manager at Edison AI and Platform at GE Healthcare Digital. Vignesh, thank you so much for joining us on the AI today. Thanks for having me, Kathleen and Ron. Yeah, we're so excited for this podcast interview with you today. We'd like to start by having you introduce yourself to our listeners and tell them a little bit about your background and now your current role at GE Healthcare. Sure. So I leave the product and engineering for the Edison AI and Digital Health Platform at GE Healthcare. And what that really does is it focuses on enabling both AI as well as non-AI based applications that drive various clinical and operational outcomes. So and prior to this, I spent about 16 years at Cisco in various roles, including co leading, the development of both novel hardware and software, which obviously includes AI, enterprise and service provider-grade platforms. So effectively, think of it as about 20 years of both successes and misses. So knowing what to do and what not to do as you embed AI and build out a software platform. Yeah, so that's fantastic that I know you've been doing a lot of work and some of our listeners may have actually joined us on an earlier event. We had Vignesh join us at our machine learning lifecycle conference that we did at the beginning of 2021. I think it feels even longer than that. But a little over a year ago, we had that fantastic event. You were on a panel and shared some really fantastic insights. Well, definitely one of the things I know our audience is just in hearing about is like, you know, what are some of the unique opportunities and challenges of adopting artificial intelligence and heavily regulated industries such as healthcare? That's a great question, Ron. The way I see it, this is both an opportunity and a challenge today. Because the inefficiencies in healthcare lead to what about a trillion dollars of annual financial waste just in the US alone. And from a lot of the healthcare execs, the C-suite, even the chair of radiology, oncology, their report won costs as well as transparency as their number one concern. And they are clearly looking for new solutions to help solve these challenges. And where they see it and I would agree is that AI can help them both set costs, improve patient experience, increase patient volume, which obviously matters. But most importantly, increase access to care. And even the clinicians, they benefit from a lot of these AI based new technologies through the use of intuitive workflows, maybe even improving diagnostic confidence, unless rework, which allows them to spend more time with patients. Think about what happens today, for example, in heavily regulated industries like ours, clinicians rely on uro-6 and habit formation to minimize mistakes. And so what they do is they really construct workflows which are unique to them. And for a lot of these physicians, the main hurdle to AI adoption is really getting experience with the technology itself while minimizing risk and minimizing distraction. And they want to make sure that it does not interfere with those clinical routines that they've established over years. And the way the opportunity for AI is to make sure that we come out with thoughtful and targeted AI applications that are based on longitudinal patient data, because that's how you build trust. And trust leads to greater adoption, which is what will lead to AI unleashing its true potential. The second thing that the second reason I can think of is why I'm super optimistic, if you apply a simple urestic to test if AI is just a fad or is it a paradigm shift, is to run the small test. Like, for example, in my view, AI is still in its infancy, but it's used a lot by a small number of people today. That's actually good news. Folks like annotators, like radiologists, like tech. I would have been a little more concerned if it was used a little by a lot of people, because then it's likely to be something that has a sharp edge, but doesn't necessarily sustain. What this does though, this particular adoption curve, is it allows for the next evolution of the opportunity space for applying technologies like AI. And the way this would occur, I think, is by what we call native applications that are built grounds up with AI in mind. Today, and this is part of the challenge, we view the solution space through the lens of old applications or existing ways of doing things. The analogy, you may be familiar with is of customers asking for a faster horse versus an automobile. And now that's a great way to on-ramp users into AI, but it also has its limitations, because you only see what's new in terms of what it always has been. Now, I think of this as a difference between a website that in the early days of the internet, you should let you read an article, something that you could do even with an older technology like a newspaper to modern internet, where you have things that websites that allow you to both consume, create, and now even own content in real time. Think Google Maps or even NFT that introduced digital scarcity. In retrospect, these native applications, which I'm sure will emerge even in healthcare, and seem obvious in hindsight. But in the early stages, they can be very difficult to imagine. And that's one of the challenges that I see, and as well as the big opportunities for AI in healthcare. Now, in order to identify these high-value apps, what we're doing is we're working closely with two classes of people in integrating data as well as expertise. There are two worlds here. There's the world of the practitioners, think of them as clinicians, annotators, and then as a world of the developers and the startups. Both of them are passionately striving to solve the same problems, but do not necessarily talking to each other early enough. And as a result, some of these offerings do not either address the right clinical or operational lead, or are not suitably integrated into workflow, or they simply do not work. And even when it comes to deployment, sometimes there's a hurdle around how do you ensure both safety, as well as efficacy, as the AI algorithms evolve over time, because regardless of whether it's on the pitch or on the cloud, because you've got to continuously evaluate the performance of these models and assess the need for even real proofs by bodies like the FDA of these specific AI solutions. And so you need to be always designing that the benefit, the safety, and the privacy of the patient in mind. Yeah, maybe if I may ask a quick follow up question. I know that a lot of our listeners might know that part of the question about regulatory issues is that there are a lot of regulations on patient data, right? Privacy, of course, of patient data. Of course, we originally, when these laws and regulations were created, it was about information sharing. People should not get access to information. But of course, in the systems that we're talking about, both artificial intelligence, machine learning, even advanced analytics applications, which may or may not necessarily be machine learning based, we have to make use of this data. Or like we're trying to train systems for image recognition on things like medical imaging. How have you seen these regulatory challenges around access to data and data privacy? Are they, is it difficult to work within those constructs? Or is there enough range for people who might be building these kinds of applications to make use of this data? Or there may be some things that you can share, considerations that people might want to think about when dealing with patient data? Yeah, and I think that's a great question again. And in some ways, I would try that to not just data privacy, but even the varying levels of data quality. So the fidelity of data varies by, the mileage varies by application. So part of what I know a lot of us are doing to mitigate some of those concerns you highlighted is a few things. One, we increasingly leverage synthetic data, very appropriate, especially when it comes to AI model training, that both for analytic applications, work clinical, operational focus apps. And we use real world data for primarily for validation and testing, because that's also the expectation of the regulatory bodies to your point. The other thing that we do, and I'm sure other phones are the same, is we try and internalize what we call as a G, healthcare AI principles that are enshrined as we onboard new hires, new data scientists, and make sure sort of built and baked into how they work then and they are. And to scale AI, I would recommend that each of us sort of begin our AI projects with the values that are reflect both your industry and your specific company. But most of these principles have been built and enshrined after having worked closely with the regulatory bodies, with the providers, with your various stakeholders. That's another key thing that I believe is work that will definitely help mitigate concerns around data privacy, data sovereignty, as well as overall greater adoption of high quality AI applications. We also have a secret, right? And that secret really is that modern data science, I think, owes a lot of its success to harvesting what is called data exhaust, right? Data that is of little to seemingly no use to an organization, and that would typically get discarded in an environment of high storage costs. But we believe that data actually huge value in driving clinical and operational outcomes. And what we can do with that is use that to kick start low stakes experimentation without really worrying about real patient data, given that the cost of failures is bounded. The upside though is unbounded because as you iterate, as you learn more, and as you shift the wheat from the shaft, you convert data exhaust to the equivalent of data gold, right? And this then would act as a fuel for the AI fire. Because as you can know, the reason why you are at an inflection point for AI is because there's a combination of things happening in the right time. This makes it awesome. One you starting get of huge variety of data, some of which are from images, PHI, go to make sure it's compliant, GDPR and your out of there's a lot of regulations that we need to absolutely make sure we stay compliant with. But there's also data coming out of variables from sensors with remote patient monitoring, right? With the broader adoption of virtual care all over the world in part due to the pandemic, right? Also things like broader EMR adoption, the proliferation of even the internet and cheap hardware, things like cloud computing and obviously better algorithms. So all of this are really making a big difference to how you identify what data is useful, what data is actionable and then what insights can you drive? Yeah, I mean, we always love talking about data and I'm sure in healthcare we could probably have an entire podcast dedicated to this. I know you'd also talked about synthetic data and we do, we did actually have a podcast fairly recently for our listeners podcast number 241 about synthetic data. So I encourage you to check that out if you're interested in learning more about that topic. And I always do like to hear about opportunities and challenges in every industry. Different and regulated industries can sometimes throw a wrench and things, but also can provide some additional benefits or just uniqueness that other industries don't necessarily have to deal with. That's why I, we do always like to hear about what people in those industries are seeing. As it relates to GE healthcare specifically, how are you applying AI and machine learning in different application areas? Great question, Kathleen. So the way I would say it is at a macro level, we at GE healthcare are trying to meet the primary personas that we engage with the clinicians, where they are, and we are trying to empower not over the number of AI because that's been clear and resounding feedback you received. And one of the ways to do that and I sort of touched upon it earlier is to understand where they are today. Now understand they are existing workflows, their habits, and try and make sure AI is as invisible as possible and supplements does not replace the need for human interaction, and the need for unlearning and re-learning workflows. Several examples of how we do that today. We have what we call as the Edison Open AI orchestrator, which what it really does at its very essence is it simplifies the selection, the deployment, and the adoption of multi-vendor AI, not just AI built by GE, but really this broad and vibrant AI ecosystem of startups that are willing to part and solve real customer problems. So that's something that this would occur both in a departmental as well as a healthcare enterprise setting to enable an automate workflows at scale. We also have things like the critical care suite, which what it does is it analyzes x-ray images automatically for critical findings. Think of it as a collapse lung, pneumothorax, and then what it does is it produces three hours notifications so that the clinician in question can prioritize this particular patient over others, and we can make a real difference to patients at the moments that matter. Close the home on the equipment side of the house, which is where what we consider upstream AI. How do you embed AI, not outside of the equipment, but into the equipment to differentiate and provide better access to care. We have things like air recon, which is built into our MRI machine, and it's really, really making a difference. We're getting some really good feedback from our customers and the patients, because what it does is it takes uses deep learning to deliver what we call true fidelity images. This is a deep learning application. What it does, it improves the signal to noise ratio significantly. It improves image sharpness while enabling shorter scan times. And this is a big deal because usually it's this is zero sum gain. You want to have shorter scan times, you've got to compromise on image quality. Either this or that. What AI has done is it's an idea to get AI and we effectively probably couldn't have had any okay technique to do. The other thing though, this is I would say arguably even more important, is that we are building out an Edison ecosystem to solve these challenges, because no one company, not even GL care, or any of our peers can do this or not. So to be able to solve both and ease the problem of AI adoption for both providers and patients globally, you've got to be able to engage with developers to build best in class application. So a part of our job in my team's role is to provide a tool set for these developers, regardless of which part of the world they are, where they operate out of, to be able to define understand users, get access to these users and help make sure that we connect them to the customers that we engage with on a daily basis. So think about standardizing your processes of onboarding, how they build, how they validate, now they deploy, how do they monitor after deployment and get feedback from this customer, because that's one thing that GL care can do for them. So instead of the every developer needing to know everything, we build a set of capabilities that they can take access, they can get access to once, and then multiple developers can develop using that developer tool kit. We also provide a deployment vehicle, right, because it's also just as important to ensure that we help them on the back end side of the house. It's not just about building, but also ensuring that it's deployed easily. So packaging the AI embedding it either into the device, like in the case of ultrasound, or at the enterprise level for multiple devices, both G and NONG, deployed it on the edge, in the cloud, all of that is something we focus on. And last but not least, like healthcare providers, we are helping them gain access to FDA-cliored algorithms and applications by directly integrating these technologies into their existing workflows and bringing these apps from market-ready, independent software companies into our Edison platform so that those can seamlessly integrate into our customer workflow, whether those are on devices, on the cloud, or on the edge of the network. Hopefully that provides a little bit of an insight. Yeah, that really did. Actually, we got us thinking about a lot of things here because these are some real patient care applications. We've seen a lot of applications, but sometimes people like to say, well, I'm going to test AI in a proof of concept or some application that we're not actually going to, it's just for testing, right? But these things are in production, right? I mean, these are things that are starting to see in points of care. I mean, these are out there in the market right now, right? Absolutely. So we work with both market-ready vendors and especially because if you were to segment that our customers, there's folks who have very defined outcomes. that they really want to hit now. And then you've got customers who want to co-create with us, who want to potentially code developer. And so we target both ends of the spectrum using the platform. And to your point, Ron, depending on who you're talking, which is very important to understand what their key expectations are. And it sort of ties back to your earlier question, because a lot of these healthcare providers, as well as the AI companies that we engage with, are coming together to put in place robust data governance. Because then it's in production, it's a different ball game. It's not the same thing as early stage quick and dirty prototyping. You've got to have robust data governance, what have interoperability, because it's going to be multi-vendor. It's not going to be all G or Siemens or Phillips or PICC, or Friday, it's going to be a mix. You've got to be able to interoperate, have common data formats, even ensure data security, and bring clarity to consent over data sharing. Because when going back to the AI principles, we've tried ourselves on. We take our job of being a trusted steward of both data and insights, very seriously. You've got to have clear guidelines on the required level of anonymization, for instance. And even the AI research, the other end of the spectrum that I touched upon, needs to heavily emphasize explainable, even causal or ethical AI, because that's going to be a key driver for the option. Our focus collectively has to be on being as transparent as possible on delivering robust and even reproducible results, as well as guarding against either creating or enforcing any bias. Because those would be the detonals for AI adoption, might be. Yeah, it's fantastic. I think a lot of our listeners would love to achieve a lot of the outcomes that you're achieving. And actually, curing on with that question, because you were talking about data governance, she brings us into the fourth question for you anyways, which has to do with data quality. Because if you're putting these, in some cases, very much life, critical to life systems or impacting people's patient care or the quality of care, data matters, the quality of data matters. And you mentioned that a little bit earlier when we were asking about access to data. So how are you going about dealing with data at different levels of quality and just ensuring that there's some sort of consistency in sort of that whole data pipeline? Yeah, so I think it's effectively what-- just to summarize what I probably said earlier, there is a few things. First, we try and use of synthetic data helps actually ensuring consistency and reproducibility of the experience we carried out. Where it's possible, especially for training, that's something that we've also been actively engaging with bodies across the world on what's the appropriate use of synthetic data in a healthcare setting. Where is it appropriate to use? Clearly, it's not a good place for validation. But for training, it's there's some fertile ground there. So a lot of our toolsets that we're building and our data scientists is focusing on making sure that you use synthetic data that appropriate. Because that drives consistency in training, reproducibility, and even traceability about this. The other thing that I touched upon is having an active community, both with other AI companies, folks that build these algorithms, as well as providers, our customers and patients, and putting together interoperability standards, because that's, and then data format standards. Because the part of the reason we have varying levels of data qualities, I believe as an industry, we've got work to do to be more prescriptive on what good looks like, depending on the use case and question. And a step one is just making sure we put together a committee, which is of eminent folks like you guys as well, right? It's not just providers and the regulatory bodies. But to that say, hey, what does good look like? How do we ensure we maintain the appropriate level of privacy? But privacy doesn't stem me the ability to iterate and innovate to allow patients to be able to get the outcomes that they deserve. So that's a key part of our focus as an AI company that we believe we are. The last thing is, of course, making sure it is explainable. So even the algorithms that operate on the data, part of their job is to get provide feedback on how these things are performing in the field. Because a definition of what does good data look like is also a function of did the model that was trained on a set of data perform as intended when it's actually put out in production? Or did it do a fantastically good job in training? But then really operate as designed as intended as plan in production. So being able to get that feedback loop in a way that it's compliant, ensures that the appropriate privacy consideration of not just the patient, but the expectations of the regulatory body that vary by country or net is a key part of what we are trying to do. And then once you get that feedback loop, you can go back and look at, hey, the AI model did not behave as intended. What was the source of the data that was used to train the model? What have we learned? What can we go correct and sort of create that flywheel motion? Yeah, we always love to hear about different use cases, how it's actually being used, and then data issues. Like you said, at the end of the day, you really need good, clean data. Needs to how it's being used, who has access to that data, all of those data governance issues you had talked about earlier, too. As to, is it performing the way that you expected it to perform? And if not, how often are you-- what is that feedback loop look like? And who needs to be involved, who is monitoring this? All these questions, because at the end of the day, this is important and impactful to people's lives. And you want to make sure that you're getting it right. And so it's really nice to have these conversations and hear about what's going on, how things are being done, the thought and thoughtfulness that goes into this. So I mean, we have really enjoyed this presentation, but this podcast today with you and this conversation. We always like to end our podcast interviews with the same question, and we get such varied responses, no matter how many times we ask, we have never had two of the same answers to this. And that's why I love this question so much. As a final note, what do you believe the future of AI is in general, and its application to organizations and beyond? It should be no surprise, Kathleen, but I am super optimistic about AI's future. But we also got to act now to not leave that future to chance. Because I'm actually convinced that the skills for responsible leadership in what I consider this to be an AI era can be taught, and that people can build what's safe and effective systems wisely in driving progress, naturally making life better for those around them, which is why all of us, including the work that you guys are doing through this podcast, is super critical, because we have to choose to step up and share what we learn, the hard way, both through direct, as well as YK's experience. Because this is why I'm so bullish about AI, right? AI have done this, right? Has a value proposition that I think is so strong that customers would be irrational to not take advantage of its ability to both simplify clinical, as well as operational workflows if they knew the truth. What are the big opportunities in clinical AI? In my mind, there are two things that come to mind. One multi-model AI. As I mentioned earlier, the ability now to aggregate data from not just imaging CT scans or MRI scans or X-rays scans, but also combine that with genetic or genomic information, pathology information, or labs, clinical trials, and even social determinants of health are variables. It really increases the possibilities and the art of the possible. I think this will have far greater value in prescribing even therapies and predicting patient outcomes that are highly personalized because each of us is different. The second application area, which is operational AI, how do you allow our providers to do more things, give access to more patient access to care, at less cost, actually gets less spotlight than the clinical AI. Almost all the focus today is on clinical AI. But a lot of folks we talk to doctors, both in the US and radiologists and others in the world, believe that operational AI in many scenarios may likely have far greater impact on health care than clinical AI. So that's something that I think is potentially controversial, but worth spending time on. And for years, we've been also talking about the possibilities of AI in healthcare, often with the assumption that these breakthroughs will come from big tech. But we increasingly think, thanks to the R engagement on the ecosystem side, that the major innovators are likely to actually be small to medium-sized companies with deep domain expertise, right? Working closest to the action with the person or that will actually consume what they build. And then partnering with incumbents who have years of expertise in this domain, as well as relationships with the customers. And this is now possible because of cloud, the transformation of the cloud, the acceleration of virtual care, and a vibrant AI ecosystem that's enabled by a platform, right? So now as we put together our collection of voices to train these new breed of leaders that are skilled in what I think is called decision intelligence, we hope to help new generations build AI more thoughtfully and unlock the better side of technology. The same side that takes us to the stars helps us conserve resources and connects us to love one half-year-class reward. The technology, I really think, can be wonderful if you let it blossom with appropriate guardrails and I strongly believe we will. All right, well, thank you so much for that answer. I always, like I said, love to hear everybody's answers. They're all unique and really get to, you know, bring your own perspective to things with what you see in both your personal and private life. So thank you so much for joining us on this podcast today. Yeah, and go ahead, Vigneche, I'm sorry. Thanks, sir. Thank you. And thanks for all the fantastic work you folks are doing, educating all of us during this video. I'm a big fan, not the best. Yeah, so if you're just as enthralled as we are and listening to Vignesh and hearing what he had to share with us here on the AIT today podcast, you can actually hear more from Vignesh because he is speaking at our May Enterprise Data and AI event in May. And, you know, in that event, which will provide a link to on our show notes, he's basically going to share more insights into what is happening in this space, what is happening in healthcare, and this will be an opportunity for you to interact with Vignesh and ask some questions. We always have a highly interactive audience. We don't always share our upcoming events. We do events both on the government side. We have our AI in government events series as well as our enterprise focused enterprise data and AI event series. But our listeners, these are free events to attend. And if you like some of the speakers, and sometimes we share what our speakers are sharing after the events are happening. But now we're looking to share it with you before the events have even started. So you can just go to events.cognolidica.com. That's events, eveanties.com.click on the sidebar for data for AI. You will see the event coming up. If you're listening to this podcast before May, you will basically see this event coming up very soon here. I'm actually just getting some of the details and the exact timing of that event. Maybe Kathleen might have that. But it's going to be May 5th. So it's the first Thursday of the month in May. So it'll be May 5th from 1130 to 1pm Eastern time. We always love our audience interaction. If any of our listeners have attended, you know that about half of the time is for the presentation. And then about half of the time is for Q&A from the audience because we have a very engaged audience. And you know, I really do enjoy if you've listened to this podcast, please listen to this beforehand. And then listen to the presentation and then we can dig a little bit deeper into what was discussed. That's why the Q&A is always so wonderful. While the podcast is great because we're asking the questions we want to ask, I always love to hear the audience Q&A because it's questions that you want to ask and have, you know, additional insights from some of the additional content that's presented in in the event, which I know is also nice too because sometimes it's visual content where podcasts are just audio content. So we can dig a little bit deeper into some of those visuals as well. But like Ron said, it's May 5th. Thursday, May 5th from 1130 to 1pm Eastern time. Definitely pre-register. It's free to attend. All you have to do is just click on the link. You can register and you can also watch it on replay as well. So in case you're, you know, listening to this after the podcast, after the, you know, event has happened, you absolutely can watch it on replay. So we're looking forward to that. Hopefully this was a little preview of what we're going to discuss and some of the topics we'll dig into. So on that note, thank you so much for listening. And again, a big thank you to our fantastic guest, Vignesh. Thank you. All right, listeners. So thank you so much for listening to this podcast. And we hope that you have subscribed so that you can get notified of all of our additional episodes. We have some podcasts coming up with speakers. Like we said, we'll be previewing some events or have some of the speakers who were previously on for either enterprise data and AI or AI and government event series on the podcast to talk, you know, a little bit in more detail about some of the topics they've presented on. So definitely make sure to subscribe to AI today so you can get notified of all of our future episodes. And also if you like this, please make sure to rate us on iTunes, Google, Spotify, or your favorite podcast platform. We always like to listen to feedback from our listeners and, you know, hear what you like, hear what you'd like us to continue to dig deeper on. So please do make sure to rate us on iTunes or your favorite podcast platform. I'll make sure to link to the some of the things we discussed in the show notes, including a link to our May enterprise data and AI event, our podcast on synthetic data and also our cognitive project management for AI or CPM AI methodology and certification. You know that Ron had brought that up early in the podcast. So thanks so much for listening and we'll catch you at the next episode. And that's a wrap for today. To download this episode, find additional episodes and transcripts, subscribe to our newsletter and more, please visit our website at cogmalidica.com. Join the discussion in between podcasts on the AI Today Facebook group and make sure to join the cogmalidica Facebook page for updates on this and future podcasts. Also subscribe to our podcast in iTunes, Google Play, and elsewhere to get notified of future episodes. Want to support this podcast and get your message out to our listeners, then become a sponsor. We offer significant benefits for AI Today sponsors, including promotion in the podcast and landing page, an opportunity to be a guest on the AI Today show. For more information on sponsorship, visit the cogmalidica website and click on the podcast link. This sound recording and its contents is copyright by cogmalidica. All rights reserved. Music by Matsu Grabas. As always, thanks for listening to AI Today and we'll catch you at the next podcast.