The future of AI in health care

Dr. Ellen Beckjord: Stronger communities begin with good health—for everyone. You’re listening to the “Good Health, Better World” podcast, presented by UPMC Health Plan.  

This season, we’re exploring ways to achieve good health in today’s complex world.

I’m your host, Dr. Ellen Beckjord. Let’s get started. 

We're talking with Dr. Shandong Wu in this episode to explore how technology impacts our health. Dr. Wu is director of the Center for Artificial Intelligence Innovation and Medical Imaging at the University of Pittsburgh, as well as a professor at the University of Pittsburgh and adjunct faculty at Carnegie Mellon University.

We'll discuss how technology has reshaped health care and what lies ahead with artificial intelligence.

Dr. Wu, welcome to “Good Health, Better World.”

Dr. Shandong Wu: Hi Ellen, nice to meet you.

Dr. Ellen Beckjord: Nice to meet you, too. We're glad that you're here.

So technology plays such a large role in our day-to-day life and of course, a very large role in health and health care. But what are some of the ways that AI in particular is beginning to play a bigger role in health care?

Dr. Shandong Wu: Yes. Today we're facing a lot of massive data of patients that are available in the clinical settings. I think AI is one of the new technologies that are really leveraging the availability of a large amount of data in the clinical environment due to the AI’s computational powers. We can, using AI, look at those massive data of the patient and learn new knowledge and information from those data to help us do a better diagnosis, better treatment, better prevention for disease.

So I call it this as data originated new technology, because this technology is really based on data. So, with this technology, with this computational power, it will really expand our access to new technology to improve our health care to make our life better.

Dr. Ellen Beckjord: Do you think it's fair to say—and please say if you think this is an oversimplification—it seems like the increased presence of technology in health care has been happening for some time, and a lot of the technology that's now showing up in our daily lives and in health care, when we use it, it generates data. But we now seem to be catching up with methodologies to optimize the use of that data for learning. And that seems like part of what feels like a trend or a new phase that's happening now.

Dr. Shandong Wu: That’s exactly right. So with all this data generated from health care—and we know actually health care data increases dramatically every year—most of this data, after care of patient, we don't really use it anymore or, you know, quantitatively use it or look at that information as data.

But with new technology, like a machine learning with computational powers that we have these days, we're able to look at data and look at information in the data and with massive data available with this new technology (AI, machine learning), we are able to actually now get something that we don't know before and we don't have before from the data that it can really improve our health care.

Dr. Ellen Beckjord: Can you talk about some of the ways that that is being brought into the point of care? How is it being made accessible for health care systems or health care providers to really influence the decisions that they make, or the shared decision-making process that they engage in with their patients?

Dr. Shandong Wu: Yeah. So AI in this moment, I would say, is very actively being developed and being translated into clinical settings. It's not yet, you know, massively used in the practice. But I can give you some examples [of] how this AI technology can really improve care.

For example, using AI, we can analyze a lot of imaging data, like in breast cancer screening, we can look at digital mammography images and look at the information in the data and try to predict future risk of a woman having breast cancer. This can really guide personalized breast cancer screening.

We can also use AI to make better diagnosis. Before a woman is going to be doing biopsy or other clinical procedures, we can, using AI, look at their data, and maybe we'll make a better decision. Like, you don't have to do this procedure. You don't need to do a biopsy because you're going to be fine. With all this new technology radiologists can make a better decision and more accurate decision to reduce unnecessary costs, unnecessary procedures, and obviously, you know, stress for patients.

We can also use AI to guide our treatment. For example, even though we have a lot of knowledge already about some of the treatment of disease, it's very common to see overdiagnosis or under diagnosis in current practice. With AI look[ing] at a patient portfolio data, all the data that you have in the clinical settings, we are able to use a predictive model to predict whether a patient will benefit from certain treatment or therapies.

Dr. Ellen Beckjord: There's been so much discussion now for, of course, a long time about personalized medicine, which of course, you're much closer to this work than I am. But from the distance that I sit from it, it has seemed like it's been about coming up with new therapies, like new pharmacotherapies that are tailored to the specifics of a person's disease, like their tumor or perhaps their immune system.

But what you're talking about also suggests that by using these technologies to learn in a much deeper and more sophisticated way from the information that we already have, it can just help us understand how to be more precise with even basic therapies that have been around for a long time, like, will this patient benefit from, say, this chemotherapy that's been used for a long time? It's not that the chemotherapy has to be new, but the understanding of how good a match it is for the disease process that any one individual has is part of what's new, it sounds like.

Dr. Shandong Wu: Yes. That's true. Because we know in the clinical setting a patient has a lot of data acquired—many images, lab tests, genomic testing. But those data we call multi-modal data. It's complex. It's very hard for a human to digest, to aggregate, to interpret, and then use it in the decision making. It’s actually harder, for example, for oncologists to look at those data and make decisions on that.

So with AI, with these deep learning new technologies, they are able to better understand this data. I think that's the beauty of the AI. And they can understand this data and identify information that are associated potentially with patient outcomes. There's information in the data that can tell us maybe a patient will benefit, will respond, or not, to a certain treatment.

And that's where, because we're using the patient-specific data, individualized data, it's actually really help[ing] us to achieve precision medicine.

Dr. Ellen Beckjord: I hadn't really put it together in my own head that AI is, of course, positioned to do lots of things. But one of the things it seems positioned to do is help us realize the full potential of all of the data that we've been gathering and generating for so long that we just haven't had the—of course, the oncologist at the point of care can't make use of that multimodal data in their own head on the spot, and we need support.

And then just the nature of integrating all the data that exist about one patient, as well as all the data that exists about patients who are a whole lot like them. Like, that's a computational problem that needs a computational solution to realize the full benefit of, like the wisdom that can generate it seems like.

There's so much to be hopeful about in that space. And I do want to talk about your work specifically. But can you talk a little bit about the processes or safeguards that are in place to make sure that the recommendations that come from AI are trustworthy, or is that even something that you in your field have concerns about? Is that something that is concerning, or is the concern more about the reliability and validity of the underlying data?

Dr. Shandong Wu: Yeah, I think all these factors are relevant. Trustworthy AI is one of the most actively researched areas today. And it also reflects a concern because, you know, people are talking about the trust and acceptance of AI.

I think part of the reason is because we all know as AI researchers or developers, AI is not yet perfect. It's evolving technology. And we know the data we're using for training AI models sometimes are not large enough, not diverse enough. Therefore, we know AI is not perfect. It may have some limitations.

I think this reflects that we need to do better work, better research to make this model more accurate so that we can trust it more. From the clinical perspective, a user of AI, for example, radiologists or physicians, they need to also understand and trust an AI product before they use it.

So, this creates a challenge, because AI is oftentimes called a black box technology. It's very hard to understand exactly details [of] how these models work. Therefore, we're trying to also help the users like a medical doctor to understand it better how AI works, try to explain some of the predictions, for example, made by AIs, so that they can understand them better, and they can trust the results better.

Making our AI more trustworthy, I think, is one of the challenges, but it is also [an] opportunity for us to advance AI to the next level so that we can really push AI and for the translation to the clinical settings.

Dr. Ellen Beckjord: Let's talk a little bit about your work. Can you please tell our listeners about the Pittsburgh Center for Artificial Intelligence Innovation in Medical Imaging?

Dr. Shandong Wu: Yeah, happy to. It's a center that we established about five years ago. At that time, we see the needs and the potential of AI in health care. So in order to really facilitate the cross-campus, cross-school, cross-subject research and collaboration, I founded this center with strong support from our institutional leadership, from our provost, senior vice chancellor for research. They provided the largest ever internal support, funding support for our center to facilitate this cross-talk of clinicians, faculties, researchers, students, to really work together advancing AI not only in Pitt and UPMC—we also work and engage faculties at CMU to work together because we know there is great talent and expertise in AI machine learning at CMU.

So as a team, we really work together to advance research, education, and translation and commercialization of AI, trying to leverage the strong expertise in the institutions and also the clinical expertise in the Pittsburgh area.

Dr. Ellen Beckjord: That's fantastic. Are you aware or does the center engage in any work that is having an impact on medical education, whether it's medical school or residency or fellowship? Is there already new coursework or curricula that are developed that integrate that AI into the education that new medical providers are getting?

Dr. Shandong Wu: Yes, definitely. At Pitt, UPMC Health Sciences, we are actually developing different levels of educational programs for different audiences.

For example, for our medical students that are early in their medical education, we have been developing, we call it a mini course, short lecture courses for them, to teach them about AI, to help them understand AI's potential, AI's limitation so to help them understand AI in their early stage of training.

So actually, one of our associate deans in our medical school is actively developing several different pieces of educational courses for our medical students right now. And I think we have something already in shape that we can actually share with the community, with the students.

Other than that, and we also develop different research pipeline educational programs for medical residents, for example, and fellows. Those are the young and next-generation medical professionals. So we're helping them to leverage their medical training, but also teach them data science, AI, so that they can really become the leaders of the field for AI research and AI uses in the clinical environments.

We're also working to develop specific degree programs, which is more advanced education for students, for example, trained in the computational background with a STEM background. And many of those students are actually interested to work in the health care or medical domains. And then we're trying to teach them medical knowledge, health care knowledge, so that they can really learn how to fit and use their technology to address real-world biomedical research questions.

Dr. Ellen Beckjord: That's very exciting. Would you share some of your own work with us? As I understand it, you are well-funded right now to do research in the area of the application of artificial intelligence to treatment, decision making or to help inform treatment protocols specifically for women with breast cancer. Do I have that right?

Dr. Shandong Wu: Yes.

Dr. Ellen Beckjord: OK. Tell us about your work.

Dr. Shandong Wu: Sure. I can share a few examples. We're developing an AI model that is capable of predicting future risks of having breast cancer for women under screening. This is a work that we can really look at longitudinal mammography imaging women have in the screening settings and using those imaging data to predict their 5-year, 10-year risk of having breast cancer so that it can customize and guide the additional screening and supplemental screening for those patients.

We are also developing tools and AI models to predict recurrence of risk of breast cancer. For example, women after treatment, they may be having cancer coming back. But we're using, again, multimodal data—imaging data, pathology data, lab tests, demographic data, and genomic data, if available—to put them into one model to have AI to learn all the information about this patient history using this data, this model, to predict their outcome.

So these are the models that can really help a patient to understand whether they will benefit from a certain treatment. Therefore, better to adhere to a certain treatment or switch to different options.

Dr. Ellen Beckjord: I am so excited to hear that. A long time ago, when I worked in the field of psychosocial oncology and a lot of my research was focused on post-treatment cancer survivors, one of the most, if not the most common, psychosocial concerns is fear of recurrence.

And so the idea that there is work happening that will perhaps not completely eliminate, but possibly minimize the uncertainty [and] arm people with really actionable information about what course of treatment or maintenance treatment or, you know, maintenance screening, etc. just the surveillance in the post-treatment period, how to optimize that against their own personalized cancer risk, I can imagine will bring just a lot of comfort and confidence to so many people.

Dr. Shandong Wu: Yeah.

Dr. Ellen Beckjord: That’s very exciting.

Dr. Shandong Wu: I can give you another example. For example, we have a recent work looking at women under endocrine therapy. We know this kind of treatment helps women reduce recurrence risk but women need to take this kind of treatment for 5 to 10 years long and we know there's a lot of side effects associated with the treatment. So in reality, a lot of patients are not able to adhere to that. So, we're trying to look at, how can we develop some AI models (or we call them imaging biomarkers or response biomarkers) that can help evaluate the responses of women for this treatment.

By using some imaging data before the treatment, after treatment, and other clinical data, we actually can really develop a model predicting the future risk of recurrence for those women. Therefore, it has a greater potential to help women understand, even after maybe a few months in treatment, they maybe have a good sense of whether they will benefit from a certain treatment. AI models leveraging a patient’s individual data can really help achieve these kind of goals.

Dr. Ellen Beckjord: I do have a couple of more questions, but just to pause on that, I mean, I could talk with you about this for hours because to me it is a specific example of a more general challenge. As a behavioral scientist who has focused on health and health care, this is the foundational problem. Most health behaviors—you're talking about this specific example of adhering to endocrine therapy after primary treatment for cancer is over, but you could substitute almost any health behavior, you know, eating well, maintaining appropriate levels of physical activity, getting enough sleep—all of these health behaviors are done in an effort to either preserve your current health state or prevent negative future health state. And a lot of those behaviors are hard to adhere to because, in the case of endocrine therapy, there are difficult side effects to manage.

And so that's a real central problem in behavioral medicine of how can we help people be intrinsically motivated to engage in health behaviors today that could benefit them down the road? And so this kind of work that can help people have confidence in reliable predictions about their future health states could be incredibly helpful in motivating behaviors today that could help to prevent the potential for predicted negative outcomes later. So I think that's so interesting.

What are some things that you would like listeners to know about the role of these technologies in health care in particular? What are the things that you may be most excited about or hopeful about?

Dr. Shandong Wu: Yeah. So I think, fundamentally, it is very clear now: AI is going to transform medicine and health care. I think that there's no doubt about that. From what I can see in the clinical setting, in research settings, educational settings, people are excited about this technology.

Several years ago when I went to some AI conferences or a clinical conference and I can see physicians and medical doctors are still kind of skeptical about this technology, whether they're going to be really helpful or not. But after, you know, years and years of time, I feel now it’s very different, you know, they are convinced. They know this technology can make a difference, can provide new information, can be helpful. So, again, it's very clear AI will transform medicine and health care. From the operational perspective to make things easier, to make procedures faster, save our time and also can save cost, for example, right. It can also improve diagnosis or provide predictions that are not currently available. And I think that those are the potential benefits we will be seeing from AI.

Other than that, you know, you were talking about trustworthy AI and I think, there's another important topic [that] is about the access of AI, potentially, because while we're building a lot of AI models, we have to think about whether everybody can have access to AI surveys or AI-based models or products because we know there are some limitations with AI. For example, some AI may be trained with a certain amount of data or a certain type of data. It may not be generalized very well. So this creates a potential limitation. This AI may not be widely available or accessible to a broader range of patients. So therefore, we need to be working on this, but also be aware of AI’s potential limit in this regard.

And also in actually using AI, it is very important to understand that AI may make mistakes. AI is not perfect yet. So how do we ensure safety of patients while embracing AI into the practice to help us? It's also a challenge and also a topic that is under research, that we need a lot of effort to try to move forward this technology to be able to use it safely to patients. Yeah.

Dr. Ellen Beckjord: Yeah that's, that's very interesting. And certainly when it comes to patient safety, there is such a high bar that needs to be set. But again, as a behavioral scientist, one of the things that I find the most interesting about the increased presence of AI in our lives is, what kind of expectations might we have of AI and what kind of grace might we extend to the limits of AI today? And how might that affect the expectations that we have of one another, and the grace that we're willing to extend to our fellow human beings?

There's so much about AI that I think is almost like a mirror, that as we interact with artificial intelligence, how might it, I guess, represent the way that we interact with one another or change the way that we do that?

Dr. Shandong Wu: Yeah. You bring up a very important topic about humanity, you know, human-AI collaboration and human-AI synergies. Actually, thinking about AI as a tool, AI is going to transform our practice, but AI is not going to really replace medical doctors. We see them as a tool to augment our work, to help humans, to help radiologists, for example, to make better decisions.

We can really collaborate with AI, we can leverage AI strengths, but we can also leverage human strengths and try to work together as a team so that we can deliver the best care to the patients.

Dr. Ellen Beckjord: I think that is perhaps the most exciting thing of all. And again, in navigating that and gaining success in that, I hope that it helps us become more successful in our just human to human collaboration.

Dr. Shandong Wu: Sure.

Dr. Ellen Beckjord: Is there anything we haven't talked about that you'd like to share with our listeners?

Dr. Shandong Wu: Yeah, I do have something else that I think we can share [which] is about the recent large language models and foundation models and I think these will become another topic that are really going to show benefit in health care.

As you may be aware, those large language models, they can provide a lot of information. For example, you can ask a lot of questions and they give you answers. And I think, in the clinical setting people have been trying to adopt these kind of tools to help the patient communication with the physicians. For example, to help you better understand those terminologies, indications, of the reports of the findings of a patient. So, using a large language model, I think, is going to be another important tool in the clinical practice.

Dr. Ellen Beckjord: It's another example, it sounds like, of how a large language model tool, for example, could augment or a human could collaborate with that tool to equip themselves with understanding and knowledge and even questions to ask. If I'm someone that's been newly diagnosed with a disease and I don't have expertise in it, how might I partner with an AI tool to not answer all my questions so I don't have to see my doctor, but help me understand what questions to ask when I see my doctor to get the most out of that interaction and equip me—empower me, really—to engage in the kind of shared decision making that so often we think of as central to patient-centered care?

Dr. Shandong Wu: Yes. So again, I have to emphasize that those large language models are not positioned to make a diagnosis for a patient, but they do can serve as [an] additional source of information a patient and can use to ask questions, find out more information about a disease, and better frame their questions with their doctors. We see this kind of large language model tool as additional resources and a new means that were not available before, right? So, they can serve as a general knowledge base for us to ask questions to learn more from it.

Dr. Ellen Beckjord: Yes. A long time ago, I worked with a group at the National Cancer Institute in the Health Communication and Informatics Research branch, which at the time was led by Dr. Brad Hesse, who's also a guest on our podcast this season. And we were using a lot of national survey data to look at how people used the internet when it came to looking for information about their health. And that, of course, still happens. And at that time and still today, it can be an experience that is fraught with challenge, because it's difficult to sometimes know the reliability of the information or the credibility of the information. Or it might be hard to understand the information that you encounter.

And so it just strikes me that that still happens today. Many people go to the internet to look for answers to their questions about health or all kinds of things, but augmented by or in collaboration with a reliable AI tool, could really assist someone and help to mitigate some of what can be challenging about going out to the internet and looking for information without any assistance. Again not replacing the need for professional input and diagnosis and treatment planning, but just to help people make use of the information that is available to them in a way that I think, when done appropriately, can be really empowering.

Dr. Shandong Wu: I agree with that. Yes.

Dr. Ellen Beckjord: Well, Dr. Wu, it has been wonderful to talk with you today on “Good Health, Better World.” I'm so excited to learn about your work and it's exciting to know that the Pittsburgh Center for Artificial Intelligence Innovation and Medical Imaging is doing so well and doing such great work. So thank you for the work that you're doing. And thanks for taking the time to talk with us today.

Dr. Shandong Wu: Thank you for having me. It's a great pleasure to be here. Thank you.

Dr. Ellen Beckjord: We hope you enjoyed this episode of “Good Health, Better World.”

Be sure to tune in next time, and visit upmchealthplan.com/goodhealth for resources and show notes.  

This podcast is for informational and educational purposes. It is not medical care or advice. Individuals in need of medical care should consult their care provider. 

Views and opinions expressed by the host and guests are solely their own and do not necessarily reflect those of UPMC Health Plan and its employees.