[0:00] AI at Weill CornellHello, The guest today is Faye Wong. He's a professor at WOW Cornell, a leader in health informatics. So hi Faye, what is your role at WOW? And hello everyone, it's my great pleasure to speak here. I am currently a professor and division chief of Health Informatics and AI in Department of Population Health Sciences at Wow. I'm also the associate team for AI and data science here, and I also LED a institute called AI for Digital Health and wow. That's incredible. What is the role of AI within the Wild Cornell ecosystem? Yeah, yeah. We are currently talking more and more about AI and you probably saw that we just have our new strategic plan called care. So C stands for A clinical care, A stands for AI, R stands for research and E stands for education. A clinical care, research and education are our three at while and AI is at the center cross cutting all our missions. It is very important to help us to boost across. All and then like are you within a specific division within while Cornell or is it like overarching? Role here, I'm in a particular division in the population Health Sciences department and at the same time, so my institute is sort of like, although assists within the PHS department, but I interact broadly with pretty much all the clinical departments here to help the clinical investigators to resolve their problems using AI tools and also my associate Dean as a AM data science role interacts, I mean helps me interact more within the while institution, but also across campus. We also work closely with Cornell Tech, where I also have an appointment as a senior a faculty fellow clinical AI, as well as Cornell Ithaca I assistant recently recently established AI Console. Super. What's a specific example of something where AI is used at well? Cornell yes. So we can talk across, as I stated before, like 3 different missions. Like in research, you know, the wire investigators here has been

[2:21] specific examples of AIapplying AI broadly in different kind of research problems from basic science discovery like you know, molecule design or target discovery for novel treatment to using AI to help with the clinical trials, including both, you know, matching the patient to the right trials as well as using AI agents to help with the clinical trial design or building multi agent system to help with the clinical trial execution as well as on the clinical care side. So our investigators have been building like risk prediction tools to help forecasting the risk like you know, in Women's Health like postpartum depression as well as the onset of preeclampsia. That's because Women's Health is one of our major initiative. But beyond that, we have also been building all kinds of models for like subtyping complex diseases like Alzheimer's, Parkinson's, a covet, long covet and so on and so forth. So and you know we have been you know obtaining like external supports from federal agencies, industries as well as working closely with our hospital partner New York President here and to have some of these models implemented in the workflow to really a clinical workflow to really benefit the provider. So that's on the reason. On the research side and on the education side, so, you know, under the leadership of our senior Associate of Education, Doctor Joseph. So we have a working group and we have a group of people actively talking about how to incorporate AI contents into the, you know, medical doctor training curriculum. At the same time also how to leverage some of these AI tools to, you know, enhance the learning experience for the medical doctor training program. One of the example is a tool that actually built by our talented students in computer science at Isaac and faculty called Medsim AI and where they use AI models to kind of simulate the patient scenarios so that the medical trainees can recognize the situation better and come up with the right decision. Very talented students, yeah. And that's on the education side and on the clinical care side.

[4:55] clinical care improves with AII think we are, as I said, we are working closely with our hospital to incorporate some of the AI technologies like the ambience scribe and other things to help the clinicians to hopefully offload their burden like documentation and some other like repetitive jobs to make their work more streamlined and and also more efficient so that they can have more time to interact with patients. And actually on that line, we are also kind of like not not just incorporating some of the third party vendor tools like the scribe tools like that, but also our investigators are building some of their own tools like the postpartum risk depression model that is built by one of the talented faculty in my division. Her name is Yazan. So hopefully these homegrown decision support tools can also help our practitioners to do their practice better. So that would be something that came out of research then goes yes. So what's the like? So for something like that, like the postpartum one, how does the how does the research get started? Does a professor in one

[6:14] how collaboration happens at Cornelldepartment reach out to the professor in another department? How does the people work together? Amazing question. So I think across departmental or cross disciplinary collaboration is actually the spirit of the medical AI or medical informatics. So essentially for that particular project, but also broadly applied to other projects as well, it's really we have this a clinical problem that is important that either came to us. I mean the people in my division who are typically technically talented, but I mean I also have MD's in my division as well. So either from the community or the literature or comes from the first line of practitioners themselves. And then so you know, we kind of like discuss and see if there is any like technical hurdles like the data, like the model and so to see if it is feasible for us to train some model that can help with resolving the problem to a certain extent. And then so that's the pretty typical like retrospective study and research. And then so we kind of like pull our data and a lot of the times we also need to validate on some independent data to make sure the model is really generalizable. So it is not like something we call overfitting to a particular scenario. And then typically around them, we're going to have some good collaborative manuscript published. And then after that, so we're going to talk to our hospital colleagues to see what would be the pathway if we really want to implement this model in the hospital. And then there you have to also talk to like, like in this case would be the clinical partners at Obgan to make sure we have appropriate clinical championship and understand because after you build a model, if you think about integrating that into the clinician's workflow, it is actually a whole different world. It's not about model itself, but also about like how if the model got implemented, how the clinician going to use it and what do they like to see and what they're going to consider to be useful and what they're going to consider to be even additional burden. We don't want that, right? Yeah. And we certainly want to understand how that will benefit patient, but rather than do additional harm to the patient. And we're gonna also have a whole bunch of plans like how are you going to, let's say, after the model was implemented, how are you going to continuously monitor the model performance and let's say if you observe some abnormalities of the model. So what are you going to do with that? So all of these are kinds of like the processes, you know, we have to work with, you know, the clinical department, the hospital, you know, all together as well as you know, the EPIC, you know, the informatics folks. So that our model can be really integrated into like the electronic health record system like Epic. So it is really a team or into involving into people with interdisciplinary expert expertise so that we can all make sure that this model really leads to benefit. So many stakeholders yes to desire user data providers, software people. Yes, absolutely. It's a it's a absolutely a complicated process. I think that also is one thing that, you know, uniquely positioned while but also in general why Cornell is, you know, like, like Tony always

[9:52] why Cornell is the best for AI health researchsay, the best place for conducting health AI research because university. Well, it's not just at while, but we have pretty strong technical expertise at Cornell Tech, but also Cornell, Ithaca, computer science, data science, statistics, operations research and so on and so forth. We have very technical, strong technical experts to help us to develop the model or validate the model and develop the theories at the same time, while we have fantastic clinical investigators to ensure that we really have the necessary supervision and advice. You know, while we are developing these models, this should be aligned closely with the actual or clinical workflow, not just the imagination like how these practices should be like. We really want to make sure it lies with the real world practice. And at the same time surrounding us is not just the New York Presbyterian, but we have also other partners like Hospital for Special Surgery, Memorial Sloan Catering. So we have a pretty, you know, how to say Rich, uh, clinical ecosystem here, including both, you know, the medical school, but also different type of top notch. They are, they're all top notch hospitals in their respective specialty. So that we have really abundant resource to do lots of health AI research and show that our research can benefit reorganization. I make sense. And then for something like the the postpartum, I'll go back to that example like how long does it take? What's like the typical timeline and like framework thinking about how the research is progressing?

[11:39] integrating AI with clinical workflowYes, yes. So for that, that is more of like trailblazing type of efforts that takes a quite a while because that was for that particular posted part of depression prediction project. That is a, you know, sort of like one of the very first several efforts that we really go through from the very fundamental a problem identification model development all the way to clinical implementation. With that, that actually gives us a lot of precious experience and we have of course, again working together with our NYP colleagues, we have fairly established A pipeline if you want to do some research with the hope of your research will ultimately integrate it into clinical workflow. So how are you going to design the research from the very beginning? How what are you going to expect? So like how are you going to evaluate your model? What type of evidence is expected once you have your model you proposed to, we have a Tri institute, the governance committee when you go through the committee. So what are you expecting, expecting them to see and all of these things? I think, you know, I wouldn't say the process is a completely smooth yet, but at least we have a pathway that we can follow and we do have the different stakeholders and decision makers at a different stage to help us ensure that what we build are really ethical and also beneficial once it is implemented. Or I think on that front we are still progressing and there are already quite some progress over the year. So once you have one success, you have more success. Yes, that's the precious experience that we can get from those successful existing examples. That makes sense. How much of it is like

[13:40] creating a culture of AI and healthinfrastructural, like how wild Cornell was organized and how much of it is cultural? Like how do people think about it? Yeah, yeah, yeah. I think that's a great question. Culture wise, I joined a while in 2016, so actually May, so next month, actually it will be my yeah, 10 year anniversary during these ten years. So I really appreciate the collaborative spirit and environment that according to my experience for here. So it's really people are very enthusiastic especially when it comes to the topic of AI and data science about collaborations, because this is in nature, this is a collaboration topic and we work closely with the clinical partners. Let's say if we have some ideas, it is fairly, I wouldn't say that straightforward, but not difficult at all to find the right clinical collaborators. When you reach out to people, people are always friendly. And even if they are not the right people, they're going to connect you to, you know, to the right people to start the collaboration. I mean the same thing vice versa. So let's say if clinicians have some interesting problem there, they need help from technical people like my team. So we are also very happy to help them to design the right technologies to solve their problems. And there are several examples that we have built with clinicians with different specialties through that way. So I think I am feeling pretty fortunate to be in such a collaborative environment or I mean, I think what I stated is mostly talking about collaboration with clinical departments, but the same thing for let's say, Cornell Tech. So I work closely with the folks there like Tanzim, Mark, Deborah, you know, we all work on, you know, different collaborative projects based on the need and also the expertise to match. The collaboration has been always fruitful and successful. That makes sense. And like what is like the resource allocation like a funding sources like for across collaboration research, like is there usually one big grant or is there a different brand? So as you can imagine, that is mixed up, right? So there, there could be a lot a large center type of grant and, and certainly this the nature of

[16:13] funding grants, and cross-campus collaborationthis type of work we do requires a lot of collaboration and it fits in the, you know, a lot of times what those center grants are calling for, right? So, but at the same time, certainly those grants are highly competitive and they cannot be always guaranteed they can get them. So at the same time we also pursue those more regular research grants with, you know a different investigative partnerships, those clinical and unclinical. But even beyond the federal agencies like for example Cornell has that cross campus collaboration, the seed grant mechanism and I did I think participated in I think at least three of them which ultimately got, you know successfully founded. And one interesting mechanism of that is it requires that you at least have one investigator per campus. So that is intentionally encouraging cross campus collaboration. Seed grant. Yes, I remember there are different tiers, but I think the tier I was involved in mostly like 75,000, which is a big deal. But if you think about getting things started, that's very helpful. And the hope is after that, let's say over a year, you can get large grants, which a lot of my collaborations leads to like RO ones and these type of things. And I think that's again the spirit of this university. And in addition to those, there are also like certainly industry, you know, founded projects and sometimes it also involves like both investigators from clinical, basic science and technical perspectives. And they are also like foundation grants like I have actually, I started my research journey here with the support from a foundation called Michael J Fox. So actually Michael J Fox says if you know, it is a particular foundation focusing on supporting Parkinson's disease research. And through that I got collaboration with our neurology department. They have a Parkinson's movement disorder Institute and I am still collaborating with the neurologists there. And through the support of Michael J Fox, we kind of like successfully secured some follow up NIH, I mean much larger NIH for us, so you can see. So it's all especially in the current environment. So this versus diversified portfolio is becoming more and more important. And I think this collaborative nature of Cornell helped us to really be able to getting, you know, diverse support from different sources. Yeah. And then like, do you see like the funding sources, are they being more excited about AI as well?

[19:10] funding sources for AIYes, I think it's pretty apparent, especially during my 10 years of journey here. So because now AI is not just like a medical school or university or priority, but it is also like national, you know, priority because you know, the reason is a pretty apparently like really the unprecedented power and the capability of these AM models. I think we are still trying to understand what actually they can do. Where is the limit? You know, these founding agency are like through these years, you can clearly see their enthusiasm on supporting, you know, AI related research projects. But we can also see especially over the recent years, the focus is also evolving. Let's say five years ago, maybe the focus is really supporting really innovative ideas using incorporating AI to. Generate, let's say a novel discoveries or whatever. I mean, which is still very important now, but because for more people are using AI to generate all kinds of discoveries now, as you can see, more and more those costs are shifting towards like, say evaluation and validation. Like whatever you discovered is something true, not something really like what we already call false positives, right? So make sure your discoveries is reproducible and can be validated. And you know, actually we have a fantastic biostats division in my department and the Chief of Biostats division, Doctor Judy Zong, she actually recently got a pretty major center grant focusing on like the reproducibility and the validation of autism related research. And we're going to see more and more of this kind of call following up. I think not only that, we talked about like implementation of these models into real world clinical workflows. So people want to see all these funding agencies want to see your model is not just looking good on paper, but they can benefits, let's say real world patient. So we're going to see like it's if you are familiar with the Gutner's innovation curve. So initially we are kind of a climbing up the heel and the people recognize that is very powerful. And now people becoming more and more realistic. And then we want to see like what actually these tools can help us with, I'm sure. I think very soon we're gonna see what do they call the plateau of establishment, which is sort of like you're gonna see more and more actual results that you can see AI benefits in the real world. So, so the funding is, do you typically like work with like OBGYN for example, you work with them to apply for a grant together whereas they separately funded and then they collaborate afterwards? Yeah. Yeah. That's great question. So, I mean, not, not all began for myself yet, but I'm actively exploring collaborations with all began. Actually, I I just delivered a grand run not too long ago at all began. Hopefully we can we can build some collaborations there. But definitely like I think let's say in urology, in critical care and autolaryngology, you know, you know, and also oncology, a lot of these other, you know, clinical departments, I've been working with their investigators and we have successfully secured a number of federal grants either initiated by myself or initiated by our clinical partners. So we can definitely show like 1 + 1 is larger than two. In this case we form a synergistic group that can, you know, boost the success rate or grant application but also do better size. So like the clinical partners will be more likely to win their grants on. That front actually I have, you know, very proudly, I have successfully helped sync 1314 junior clinical investigators who are MD's successfully getting their training grants like K because you know more and more of these whatever grants you apply, people appreciate you have some AI components. So I was sort of like, as I said, our investigator clinical investigators are top notch award class. So I'm fortunate to work together with them to work out like if appropriate particular plans on the AI and you know, machine learning part to further boost the competitiveness of their, you know, application. And also pretty much all of them are interested in getting to know about AI and machine learning themselves. So through training grants, so a lot of them I got to meet with them regularly. So it's really a pleasure to see they kind of like getting more knowledge about this new, I wouldn't say new field, but to clinicians or to medicine, these AI is really sort of like new capabilities that they get to more knowledge about them. Some of them even get the curiosity to try some of the AI exercise themselves and then really augment their expertise and capabilities to do their clinical practice better. That was a really satisfying process. That's incredible. And how do you organize your

[24:56] research group organizationlike a research review? Is it like 1 researcher per per practice area or do you people work across multiple? Ones that's that's a that's a great question. First, I am a flexible in my managing style. I don't like to like strictly ask, ask somebody, let's say you can only work on this or you can only work on that technique, these kind of things. So collaboration is very important in my own group as well. But at the same time, you know, like especially like for postdocs and PhDs. So we need grant support for their salaries or whatever expenses. So that means that each of them will be mentally supported by 1 or maybe 2 projects. So which means that typically each of the people in my group, they do have one, at least a one major project for them. But a lot of the times, one person per project is not enough because you do need this collaboration. A person with one major project doesn't mean he or she gonna only work on that project. So they're gonna still collaborate across projects based on the need for their expertise and they're gonna also benefit a lot through this collaboration process. So it's really sort of like everybody have their own like you can think of like let projects while at the same participating in other projects to facilitate collaboration. And in my group, we do have like more senior researchers, research associates, some of them become faculties. We also have postdoc researchers and we also have PhD students and actually proudly my PC students are from all over the place within Cornell. Like not just to the graduate programs at a while like PBSP or Tri I, but also, you know, from Ithaca like computational biology, computer science, ECE and so on and so forth. And, you know, there are also masters students for some of the projects that they can contribute. And even I think we have also some successful experience engaging some high school students. Wow, actually I, you know, probably, I think we have a mentor sort of like 3 or 4 high school students. So that's really kind of like the experience. Like how I'm thinking that researchers in my group should work with each other and you can see the latter like the different levels people working. So, so I just met the two most incredible assistant professors in the in your lab. They just joined. How do you introduce them to like the entire Wild Corner ecosystem? There's like so many different moving parts. How do you help them get their first up our projects off the ground and going?

[28:00] starting as an assistant professorThat's an amazing question. I can tell you like for the two assistant professors you just met, so they have the advantage in a sense before they become assistant professor, they worked here for post docs for a while. So through that process, they already get to know the environment and already got their clinical partnerships for both of actually both of them are successful. So Chang actually again already this is his third year, but he has already secured 1R1 and there is another one very promising. And for Chen XI just got a JIT for one of his R1. So essentially both of them for the starting R1 projects are extensions of what they have done during their postdoc research. And based on the connections they have built and they quickly can become independent. They're independent for those for a Full disclosure, those our successful RO one project, they pretty much work on their own. That's fantastic. And now they're building their own labs. But in general, like because I'm also the division chief of my division, we also have other faculties. I think for me, I mean my emphasis is, is always like, you know, feel free to let me know what your interest in exploring and if there is any clinical partners that you want to reach out and you want me to introduce, you know, I'm always happy to do so. In addition to that, I'm also proactively involved especially junior investigators in some larger team discussions, either it is institutional initiative related or grant related so that they can get themselves exposed to you know the larger environment involving you know investigators from different clinical departments and some of them may lead to like follow up collaborations and joint research. Again, we are very collaborative and I am doing everything I can to help with especially clinical investigators. And then, uh, like as, as AI goes more and more into like healthcare, like there's like industry, there's academia, like where do you see like the balance between where AI should be within academics and industry?

[30:22] balance of industry and academiaThat's a great question. And especially like holidays as I think everybody can see most of the dominant innovations from AI are from like industries. So it's not surprising because they have the resource resource, I mean like data, human, especially compute to build those really large models. And people, some academic people are getting panic about South. What is the role of academia? So I am thinking of, you know, now it's really like we we should really seriously think about academic industry collaboration. Like, I know this is always a topic, but now it's just becoming more and more critical in a sense, like something that industry can really be good at, like what we just talked about, like large length model training. But some is really academic are good at and we need academic to do that. For example, like a theoretical breakthroughs of model, like a large language model. You know, they perform pretty well. Is that all will be the case? You know, the model is almost at their extremes in terms of scale and we have used up all data. So what are we going to do to further improve the performance? And we call this artificial intelligence with sort of like the intention of helping these model to be able to mimic human brains. But this is not like how human brain work. We don't we don't need like 10s of trillions of samples to recognize an apple. So is there any other like a better architecture, let's say that is that kind of leads transformer the next generation of like AI? Who knows, right. So, but in academics, we have a lot of talented theoretical computer scientists that can do this kind of study. Industry may not have that appetite or they're just not there really men job. And at the same time, let's think about like especially healthcare AI, right? So this is the intersection between AI and healthcare or medicine. Like in academic Medical Center, like wow. So we have the clinicians that are taking care of patients everyday. So when you think about healthcare AI, so you should really say about like how it can be integrated into these day-to-day clinical practice business instead of the other way around, like using healthcare data as a benchmark, let's say to further improve your model. Sometimes that's a feedback loop is also important. But the ultimate goal should still be you build a better model to benefit real world healthcare practice, right? So if you think about that, even for a large or prestigious tech companies who develop these state-of-the-art models, they still need to collaborate with us to show their model can benefit because clinical care, they, they don't have clinicians, yes. So they can show a good retrospective performance of their model, but that's on paper. So if you want to see the real world impact, they have to collaborate in that sense. So we have like more deeper understanding about the clinical care process and we know sort of like what is needed and we can provide necessary feedback. And our investigators can also sort of like either adapt or fine tune or whatever word you want to say to make whatever these industry develop the model more appropriate or a fit into the healthcare workflow. So make the AI better to particular this vertical area and certainly there are other area that's we are good at, let's say agriculture, right? So Cornell is really the top notch university in terms, in terms of agricultural research. So those vertical intersections, AI is more of like a horizontal tool, but this vertical intersections. And it's not like you just can blindly plug in those AM models into these different verticals. And basically it can work deep domain expertise in the understanding about the application workflow. And that is where those, you know, a more applied schools like medicine free culture hotel, you know, can provide. So that's how we're gonna work together with industry. It's like everybody gonna compete to build another, yet another.

[34:59] how industry can work with Weill CornellPosition ourselves to be the most attractive for industry and like how does industry get started? In your last, you know, a podcast you interviewed the Lauren right, who was, you know, one of the leaders in our enterprise innovation team. So we do have a pretty good enterprise innovation team at while led by Lisa Blakanika, whom I work closely and also Donna Ron. So our enterprise innovation team really I would say like interface that talks with both industry and wild side to facilitate like the match and also the successful collaboration. They want to work with Cornell, but they may not know and this work with same the same thing vice versa. So I may have a great idea. I don't know which company might be interested. Not only that, I mean, we also have talented researchers have great ideas that can be translated to commercial values. And let's say, you know, do IP licensing, you know, patenting all of these things. And Lisa's team have been critical, you know, to, you know, have all of our investigators have that, you know, mindset like if you have something that really have great commercial values of what you need to do. And they have all kinds of programs to help our investigators to realize that. And, and the same thing for Cornell Tech. So, you know, because entrepreneurship is one of the spirit when they set up Cornell Tech and they have, let's say, the wonderful like wrong wrong way a post a program and Doctor Tanzim Chaudhary, who was recently appointed as the chief health innovation and coronal tech, we work closely and we are thinking of like, you know, how to do something jointly to further boost of the entrepreneurship and commercialization activities. So I'm thinking of with the help of all of these people, enterprise innovation and faculties like Tanzim. So really help us to bridge the industry side as and you know the academic side. So it is not a challenge but really a booster to. Us so you're 10 years like what? Well, like 10 years ago. Like what excited you about the the AI part of medicine? Yeah, yeah. I mean for me it is easy because

[37:26] Fei Wang's path to AI and HealthI am trained from the AI parts. Actually I got my PhD from Tsinghua, that's one of the top notch universities in China. And I study AI and hardcore AI and machine learning for my PhD and very algorithm focused, not particular to any specific application area. And after I came to US, so I joined after two years post I joined IBM, that was a year when IBM went the Jeopardy challenge. What's the machine you want the Jeopardy challenge And they're thinking of like, you know, the business plans after wasn't when the job the Jeopardy challenge. So one of them is healthcare. So they were setting up within the IBM teacher Watson Research Center. They were setting up this healthcare analytics research group. I am one of the earliest members group. So that then I started to focus on, you know, the healthcare world. And I do think that's a critical application area because that is related to everybody's daily living. And certainly I had a great time when I was at IBM getting to know a lot of friends in a different circle, because in the past, my circle is mostly in the machine learning world, but with my research at IBM, so I got to know like people in medicine, but also medical informatics circles. And later on, I found this direction is #1 promise in #2 really interesting. And I would appreciate flexibility because, you know, in industry there could be constraints, like you cannot really freely explore any ideas that you want to explore. But yeah, in Healthcare is just so broad. There are just too many upper, you know, possibilities and potential. So I think, you know, going to academia, we all know like, the difference between industry and academia. So like freedom is 1 and the signature of academia that you can pursue more of whatever that you're interested in. So that draws me how the idea of moving to like academia. And Cornell, of course, is one of the top notch schools and they have that opportunity back to 10 years ago. So I joined and over the years, I still maintain pretty good relationship with IBM. So that kind of like give me as one of the advantage, default advantage. If you talk about industry relationships and I think that was a great decision and certainly over the 10 years we have, things have changed a lot globally and locally and we certainly have achieved a lot as well. So that was. So for an amazing question, I always ask the guests, what's the kindest thing anyone's ever

[40:23] closing questiondone for you? I. Think that's that's too much so for me. So there are a lot of difficulties as you can imagine. For example, I don't have education background in the US. As I said, I got my PhD from China and after I came, I came to USI have to change my research area and. Moved. To academia from, you know, industry and I'm currently a faculty at medical school that I have never imagined I I can be end up with before when I did my PhD. So there are lots of changes. I achieved a lot and you can't imagine. So it is not something I can accomplish with myself. They are all my friends, my mentors, my students and my family support, you know, my wife that's willing to go with me moving from different places. Certainly, you know, we have two beautiful girls and my bigger 1 is now looking at, you know, college very soon. This is just too much. Too many people did too much like you said kind things to me. Otherwise, I cannot imagine I can survive like as a as a person do not have US degree and change research field and in a top notch medical school of whatever I can achieve today. So I really appreciate all of the help from everybody, and I hope I can turn this into my gratitude to help other people as well.