How AI and Machine Learning are Changing Healthcare and Health Research with Yuan Luo, PhD

[00:00:00] Erin Spain, MS: Welcome to Science in Translation, a podcast from NUCATS, Northwestern University Clinical and Translational Sciences Institute. I'm your host, Erin Spain. The many ways artificial intelligence and machine learning could be used to improve health care is exciting and revolutionary. But many outside of the engineering field may face barriers when it comes to applying AI methods to medical research and clinical practice. Here at Northwestern Clinical and Translational Sciences Institute, Chief AI Officer Yuan Luo is working to break down those barriers, and he joins me with details. Yuan Luo wears many hats here at Northwestern. He's an associate professor of preventive medicine, pediatrics, and at the McCormick School of Engineering, and is leading the Center for Collaborative AI and Healthcare at Northwestern's Institute for Artificial Intelligence and Medicine. Welcome to the podcast.

[00:01:05] Yuan Luo, PhD: Thank you for having me.

[00:01:06] Erin Spain, MS: Well, it's safe to say you are a leader in AI and healthcare. Tell me how you got here. Tell me about your career path and how you gained this expertise in the field.

[00:01:15] Yuan Luo, PhD: Sure. So I was actually trained at MIT in the field of medical AI back in the early days. And then I was fortunate to collaborate with many physicians in different specialties and to leverage AI and machine learning to help with the clinical decision making process, to help with understanding complex disease, and inform targeted therapies. And throughout this organic process, I gradually established the leadership and also well-rounded expertise as the hub of the AI expertise in the organization. And I continue to collaborate with many physicians across all those specialties, and it's a rewarding journey throughout the years, and I'm looking forward to continuing that journey and with many new, exciting developments these days.

[00:02:03] Erin Spain, MS: I want to follow up on that a little bit. What was it about artificial intelligence and machine learning that excited you and made you want to really dedicate your research and your career to this field?

[00:02:14] Yuan Luo, PhD: Yeah, so there are many things that is very exciting, and one of it is to greatly automate the process of, let's say, human experts looking at drug discovery or looking at making point of care recommendations, because if you think about it, human experts may have bandwidth limits in terms of synthesizing the information sources coming constantly at them, but with the help of AI and machine learning, you could synthesize and summarize those knowledge and present the most pertinent ones to those human experts and also help them catch what they might miss and overlook, right? So this is greatly automating assistance to help making robust clinical decisions at point of care.

[00:03:03] Erin Spain, MS: And how can it be used in terms of discovering disease mechanisms and making scientific discoveries?

[00:03:08] Yuan Luo, PhD: With the help of AI and machine learning, you can greatly automate the process of, let's say, protein folding or sifting through the different genetic targets for druggable targets, and then looking at potential small molecules and biological to target those drugs. And this process has been really exciting in terms of accelerating the discovery and accelerating the clinical decision making. And also we have been scaling out in terms of leveraging AI, in terms of integrating multi modal sources of the data. Think about a complex disease such as autism, right? So there you can leverage AI to integrate not only molecular information, imaging information, but also in terms of the clinical records, administrative records, together with the animal models data so that you can triangulate all the different pieces of the evidence and to look at this complex disease identifying early markers, so that you could move the screening window for the patients much earlier and administer the intervention at a much earlier stage so that those individuals can gain more functional independence when they grow up.

[00:04:24] Erin Spain, MS: You mentioned the word scaling, and that's something that you are doing in your role as Chief AI Officer of NUCATS. Tell me about that and some of your other goals and focus in this role.

[00:04:36] Yuan Luo, PhD: Sure, so one of the aspects of the scaling is to enable many, many people and researchers to access the critical data that is important to them. Taking the example of the information that's locked in the clinical notes, if you look at physicians, many of them do not have the bandwidth or the expertise to launch natural language processing pipelines to extract the structured information, let's say the medications and adverse events from the clinical nodes to conduct their own research. So what we are doing at the health system level is to bulk process all the clinical nodes in the enterprise data warehouse s o that we extract those critical concepts like medications and adverse events. Let's say if you wanted to do pharmacovigilance research and we departed those concepts extracted into a common data model table so that you can integrate data from different hospitals. You can also integrate the data with other academic medical centers so that you have this diverse coverage of the data geographically, racially, ethnically. And that has enabled the access to the clinical notes to the researchers, because now they don't have to launch NLP pipelines, they can directly query the combinator model tables that we created, the data marks that we created, and streamline this kind of access. So that has greatly expanded the scales of the access to the data. And another scaling up in terms of the data access is to create flagship data sets so that we can allow the entire research community to benchmark their state of the art models on the shared data set and to constantly improve on that.

[00:06:27] Erin Spain, MS: AI has been really successful in domains, such as image recognition and large language models, but it does seem to be lagging behind a little bit in healthcare. Can you tell me more about that?

[00:06:39] Yuan Luo, PhD: In healthcare, it always feels that we are lacking somewhat compared to the general demand. If you think about it, one of the fundamental reasons is that we don't have this large source of the data that the general domain machine learning researchers enjoy so that they can benchmark their state of the art models, such as the ImageNet and Google corpus and so on. And right now we are motivated by this gap and to overcome it. And by pulling together several academic medical centers as a starting point to create a large and diverse data set. And we all share that with the entire research community so that this can serve as a common ground sandbox for AI and healthcare researchers to constantly refine and develop their models, so that we can gain similar success as to the general domain machine learning and AI field.

[00:07:36] Erin Spain, MS: And really, you're breaking down barriers for folks. You're able to sort of bring the tools to them and make it easier for them to access AI tools. One very successful project that you've worked on to reduce barriers is the NUCATS AI for Health clinics. Tell me about these clinics and how they work.

[00:07:55] Yuan Luo, PhD: Sure. So the clinic is actually motivated by really interesting observations. And over the years we found that clinicians and AI scientists, they don't always work well with each other. And despite the good intentions, sometimes you hear AI proponents saying that AI is going to replace pathologists, AI is going to replace radiologists. On the other hand, throughout the pandemic, I think there are also a lot of suspicions and even sometimes cynicism from the clinicians to the AI in healthcare developments because many of those AI apps or models were not able to go into widespread clinical use. And so there, we think the fundamental gap is because there is lack of communication from the idea inception towards the entire development continuous improvement journey. And this is why we created this AI for Health Clinics. And the idea is to pair the clinicians and the AI scientists from the beginning of the project's inception.

[00:08:57] Erin Spain, MS: So for the thousands of physicians out there within our health system, explain what this might look like.

[00:09:04] Yuan Luo, PhD: To all the 4,000 physicians in our health system, whoever wanted to leverage AI to solve their clinical challenge can pitch their challenge to the clinics and the AI scientists and the staff will be able to help them brainstorm ideas and debate on solutions and actually implement the solutions and continuously improve the solutions after deployment. And this has created a fertile ground so that clinicians and AI scientists, they can cross pollinate each other and to avoid the trap of blindly storing the model to the data or to avoid the lack of the bandwidth of expertise of actually deploying those models. So this has worked quite well and attracted lots of attention from the clinicians and also from AI scientists across the campuses and has also generated many publications and leads to sustainable funding from the federal sponsors.

[00:09:59] Erin Spain, MS: Tell me about some of those success stories.

[00:10:02] Yuan Luo, PhD: Yeah. So, there are many, and I'll start with one of the success stories, which we partnered with cardiologist Dr. Faraz Ahmad and to use AI to enhance the identification of patients with advanced heart failure. And those patients, they transition from modest stage heart failure to advanced stage heart failure, but they are often overlooked by their primary care physicians because it's hard to detect, and this has resulted in delayed referral and resulted in higher mortality rates. So we developed machine learning models to catch patients who are undergoing such transitions by sifting through their past medical histories and current lab tests and physiological signals. And then we have nurses and physicians review and then eventually refer those patients to specialty cardiologists. And once they are referred there, they are undergoing several evaluations for advanced therapies such as heart transplantation and left ventricular assistive device and we have a few patients that has gone this route and eventually this leads to saving the lives of the patients and also improving the quality of lives after the left ventricular devices has been planted.

[00:11:18] Erin Spain, MS: Well, that's really exciting. For folks who are listening who might be interested in pursuing one of these clinics, how do they do it? How do they go about getting involved?

[00:11:27] Yuan Luo, PhD: Well, so it's fairly easy. And we will regularly share a signup sheet, and anyone who has interest in terms of leveraging AI to help their clinical challenges are welcome to sign up, and then you just need to present the issue to the clinic. And then we'll have a brainstorming session, followed up by regular meetings if there is enough sustainable interest, and in many cases we're able to continue the process and leads to either publishable outcomes or sustain federal funding to help scale up the project.

[00:12:02] Erin Spain, MS: As I mentioned at the beginning of the show, you also have an appointment at McCormick School of Engineering in Evanston. You kind of have one foot in medicine, one foot in engineering, and these clinics do bring together engineers and clinicians. Tell me about bridging that gap and how powerful that is.

[00:12:19] Yuan Luo, PhD: Right. So that's a great question. And in terms of the gap, we actually have a physical gap between the Evanston campus and Chicago campus. It's about 40 minutes drive. We also have a lot of enthusiasm and also a keen interest from both campuses to partner up to address some of the hardest clinical challenges and to tap into that great enthusiasm we have many events. So for example, the AI for Health Clinics is also open to participation from the Evanston School and many engineers and AI scientists from there can work with the physicians at the hospital and the health system to collectively embark on a journey to address the clinical challenges. And recently we have also launched the Healthcare AI Forum that's open to the entire Northwestern University and actually open to the research community in the greater Chicago area. There, we were organizing the presentations from the state of the art developments of AI in healthcare in really short and biteable snippets and this kind of presentations can come from students and trainees from both the Chicago campus, the medical school campus, and the engineering schools. And these pieces, they serve as the intuitive introduction to people on how these recent developments can help address some real clinical situations, the challenges. For this to happen, we organize the Healthcare AI Forum on a regular basis, bi weekly basis, so that people from both campuses can collaborate. And over the past couple of months, we have also organized a data submit that work to attract all the leaders from both campus and from the health systems, and to discuss what are the top priorities in terms of standing up the AI governance and establishing some of the lighthouse projects that can demonstrate early ways for AI in healthcare to address some of the clinical challenges that are the real pinpoints of the health system across the country. And so this has also received wide participation and enthusiasm from both campuses from the health system as well.

[00:14:42] Erin Spain, MS: This is just an incredible amount of resources for people. And you mentioned the frequency, these are biweekly meetings, these forums. But in your spare time, you're also doing research. Tell me a little bit about your research. What are you working on right now?

[00:14:56] Yuan Luo, PhD: Sure, so one of the focus of our research is to integrate multi model healthcare data so that to power AI and machine learning algorithms to really sift through patients histories and present lab measurements and imaging and molecular data to understand the disease mechanism and to inform targeted therapies. And another focus is to do it in an iterative fashion to be able to adapt to the quickly changing patient's evolution. And for the first one, we have been developing different kinds of multi model AI machine learning algorithms. One of those is to leverage on the currently very trendy large language models in order to generate a coherent representation from all those different sources. And think about it, it's actually quite analogous to how humans were reasoning. You are looking at a table of the measurements, and then you somehow match that into an intuition that can be expressed in language. You're looking at a patient's genetic mutations and internally in your brain you're generating some insights that's expressible with human language. And similar to medical imaging findings and pathologists and also radiologists are constantly doing that, writing down their interpretations in those reports. And the bottom line is that through all those transformer models, advancements in those models, we can leverage them to map all those different sources of information, whether it's from imaging, molecular data, structured data, clinical notes, to this coherent language representation and use language as a common inference vehicle for reasoning about the patient disease progress and reasoning about the best point of care recommendations. And this is something we are keenly interested in and in terms of developing models to achieve that.

[00:16:56] Erin Spain, MS: Another focus of your research is about making this process really adaptive and iterative. Tell me about that.

[00:17:02] Yuan Luo, PhD: The healthcare landscape is really changing fast. And fundamentally, we see a gap in terms of the conventional machine learning models to adapt to this really fast changing involving landscape because the steps of the conventional machine learning, they react to the human expert input, and this is inadequate because you cannot flexibly address the data shift and buyers issues. Let's say if your data is developed on this particular portion of the population it may not generalize to another portion of the population in let's say another continent or from another race, right? And I was advocating for a notion of proactive AI in machine learning, which features the model to data feedback loop that enables the automated targeted data collection and automated augmentation of the data. And the idea is that once you have this machine learning model up and running, the model itself can identify weakness of the data where it has the largest uncertainty, let's say the wide confidence intervals, when it is making its recommendations. And that can be used to feed back to the model to guide it to automatically suggest where you should gather more data or where you should really leverage external, let's say, knowledge bases and some external sources of information to complement with the current uncertainty or lack of the data set. And this can happen on an automated basis because we have these tools such as reinforcement learning with human feedback can help you do that And I think people should be constantly leveraging on these opportunities. And especially with the dynamic healthcare landscape. And this not only applies to clinical decision making process, it also applies to things like drug discovery, right? So, nowadays, you even have this notion of adaptive platform clinical trials where you can adaptively randomize patients on different arms of the clinical trials based on the evidence that you gathered so far, which seems to be more effective towards certain types of the patients. And as you collect all that information, you can adapt to randomize the patients instead of sticking to the randomization that's set in the beginning, which can quickly fall out-of- date. And then people are doing that, this kind of clinical trials in practice. And I think AI can also help that to more systematically, comprehensively gather those information and make more robust recommendations such as leveraging some of the reinforcement learning algorithms that I mentioned. And so it's not only in terms of the clinical trials, but also in the entire value chains of the drug discovery, such as the research and also medical affairs, and gathering real world evidences, and even in terms of targeting patient population who would benefit from the therapies the most, you can be constantly evolving this kind of evidence to data, to model, feedback loop, and then continuously refine the recommendations the model is generating. So this is very similar ideas across the entire value chain for the pharmaceutical industry as well.

[00:20:24] Erin Spain, MS: There may be folks listening to this podcast right now who are investigators who are interested in using AI. What message do you want to give to those listening who are interested in this world, but they have yet to delve in?

[00:20:38] Yuan Luo, PhD: Well, so I think one of the messages that I wanted to give out is to encourage everyone to start trying. Nothing will happen if you don't try at all, right? So, it's very similar to, I think, the time where we see the invention of the automobile, right? So I think our grandfathers or great grandfathers, they might be used to riding horses and then they might be afraid or held back in trying those automobiles. But eventually, for those that did, they were able to constantly evolve the designs and the efficiencies and also evolving the safety rules. And I think we're at a similar, although much faster paced era, and with all those developments in generative AI and large language models. I think the notion that AI is going to replace certain professionals in certain industries probably isn't going to come, but, people who use AI are going to replace those people who don't use AI. I think that's probably something that more and more people started to gravitate towards that notion. I think this early adoption, early exploration is really important to figure out the frontiers, figure out the landscape, and also making machine learning AI as a proactive versus a reactive process. We really need to engage the entire organization as the talent pool for this kind of activities. And that brings the question of how we can really disseminate and democratize AI literacy resources and tooling across the organizations.

[00:22:22] Erin Spain, MS: Well, you are certainly taking those steps here at Northwestern and it's really impressive what you've been able to build here in just a short amount of time. So, thank you so much for sharing what's happening at NUCATS and within your other roles, and I hope that those listening will reach out to you.

[00:22:37] Yuan Luo, PhD: I'm looking forward to that and again, thank you so much for having me.

[00:22:41] Erin Spain, MS: Subscribe to Science in Translation wherever you listen to your podcasts. To find out more about NUCATS, check out our website, nucats.northwestern.edu.