Washington University in St. Louis has created a university-wide research institute focused on applying AI to solve the most important health problems using data-driven approaches. Leading the charge is Chenyang Lu, Ph.D., an IEEE Fellow and founding director of the AI for Health Institute at Washington University, St. Louis, who recently spoke with Healthcare Innovation about the value of taking a multidisciplinary approach to uniting AI researchers and health professionals.

Lu is the Fullgraf Professor of Computer Science & Engineering at Washington University in St. Louis, with joint appointments in Anesthesiology and Medicine. A Fellow of the ACM and IEEE, Lu received the 2022 Outstanding Technical Achievement and Leadership Award from the IEEE Technical Community on Real-Time Systems. He also serves as the editor-in-chief of ACM Transactions on Cyber-Physical Systems.

Healthcare Innovation: Could you start by talking about the type of projects taking place through the AI for Health Institute there at Washington University?

Lu: Washington University in St. Louis has created a university-wide research institute focused on applying AI to solve the most important health problems using data-driven approaches. This came about three years ago when we realize AI was going to be the new frontier of medicine and public health. There are just so many opportunities where AI is advancing at this explosive rate. While data is becoming abundantly available throughout healthcare and public health with electronic health records and imaging and text, we recognized that we needed to combine the expertise of AI experts and clinicians and public health experts, so that we could work together to bring the best AI approaches to solve the most important problems. So we got together across the engineering school, medicine, public health. In fact, we now have over 120 faculty members from all eight schools across Washington University. That includes the law school and business school, because there are so many important legal and business issues. We also include the school of art and design, because there are a lot of design issues as well. It’s a lot of fun and we’re solving important problems.

HCI: Washington University has a big medical campus that involves other health systems. Is there an opportunity for working together across those health systems as well?

Lu: Very much so. Barnes Jewish Health is the healthcare system associated with Washington University School of Medicine. There are over 20 hospitals associated with the system across St. Louis and Kansas City and other places. Through a partnership we get access t to the BJC data, and we can implement and pilot our solutions in collaboration with BJC.

HCI: I understand that your research focuses on developing machine learning models to predict health outcomes using multimodal data. Can you describe that work?  How the data is gathered and analyzed, and what how those predictions are used?

Lu: Well, it’s a wide range of data that we use and we solve a pretty broad range of problems as well. For example, we do a tremendous amount of work on surgery, which is one of the highest-risk procedures in medicine. In one example, we  look at longitudinal electronic health records, basically looking at diagnostic codes and labs to predict this condition called CSM [Cervical Spondylotic Myelopathy], a very common sort of spine degradation problem that in some cases leads to surgery, and it's notoriously difficult to detect. Oftentimes the diagnosis was delayed by months and years. Basically, we address this problem by looking at the structured medical codes in the longitudinal electronic health records. This is essentially similar to large language models, where you read a certain amount of text, and you guess what the next word is. In our case, we read a whole bunch of codes, and we are predicting that CSM will come in a few months. The patients suffer for a very long time before they finally get a diagnosis and intervention, and procedures are more effective if you do it earlier.

HCI: I read that your institute recently awarded $300,000 in a seed grant program to support six interdisciplinary teams. Can you talk about that?

Lu: We want to encourage people to work across domains, across schools. One of the challenges of doing AI for health is that it is hard to get started because you have AI experts on one side of the campus, you have medicine and public health on the other side of the campus, and they don't know each other, they speak different languages, and they have no prior track record of working together and finding successful solutions. So this seed funding program is used to get it started, with AI experts teamed up with health experts to write joint proposals or new ideas, and then we select the most promising ones to fund.

HCI: What about the impact AI is having in medical school itself? How are medical schools trying to figure out how to train the incoming cohorts of physicians to use AI in a way that's helpful, but not de-skilling themselves by becoming too reliant on it?

Lu: This is an extremely important, timely question. WashU Medicine has a curriculum committee that's taking up this problem of how to incorporate AI into our medicine curriculum. There is an AI literacy course, for example, to get everyone started. I think it should go beyond just one course, obviously. We need to train future physicians to understand what AI is saying, and understand the risk, understand the uncertainty, and be able to critically evaluate what AI is telling you. You want to take advantage of the benefits of AI to make you more efficient and at the same time, be very sensitive and mindful of potential errors in the AI outputs.

HCI: What are some issues that health systems are finding in implementing new AI tools? Are there governance issues or algorithm monitoring issues to make sure they don't drift or there isn't discrimination built into the models?

Lu: These are all very important issues. Certainly you need to have governance now. Many medical schools and hospital systems are setting up AI committees to make sure to vet these models and tools before they get deployed. Monitoring is an important issue. I call these spatial/temporal challenges of AI models. Spatial in the sense that you're trying a model at WashU Medicine, and then you try to deploy it at Massachusetts General and it may not work as well. It basically means if you take a model that's developed by a vendor or at a different hospital system, before you deploy it, you’ve got to test and verify it and you might have to adapt it. We had a recent paper that showed that even these large foundation models with hundreds of millions of parameters, sometimes they do not really transfer very well across different hospital systems.

The temporal challenge means that initially the model might work pretty well in your hospital system. Over time, the population changes, the hospital procedures change, the society at large changes, and the model performance degrades silently. Physicians may have noticed the model doesn't seem to be as accurate as before, but in reality now we realize this is a systematic problem. Everyone has to monitor their model performance and detect degradation, and then take action when it happens.

HCI: We are hearing a lot now about agentic AI on the administrative side, for coding, prior authorization, and revenue cycle management. Do you think that AI is having more of an impact early on in that area than it is on the clinical side? And do you think the potential on the clinical side is much higher overall?

Lu: That’s a very good question. Naturally, for the hospital systems it's an easier decision to adopt the efficiency tools, because they are not as safety-critical, so there are fewer liability and safety concerns. That is why documentation tasks such as generating referral letters and generating discharge notes, handling patient messages are taking off first. But I do think the other side eventually will happen at a very large scale as well. I think we certainly have a lot of initial evidence that AI can do a really good job in things like differential diagnosis and identifying patients at risk. Personalized treatment strategies are potentially just tremendous. We just need to work out all the workflow issues and safety issues to make it work. 

HCI: Is it already having a big impact on clinical decision support tools within the EHR or have those not really been replaced yet by AI versions?

Lu: I think this is starting to happen. There are more and more AI tools being deployed within the electronic health record platform, so that physicians are actually seeing them all the time. Of course, in this area, the most mature area has been radiology. I think radiologists are now very used to having the AI putting on markers of suspicious areas and contours of cancer areas. I think the other areas of clinical medicine are catching up, but it's really happening because you have all these big vendors deploying tools now.

HCI: Do you have ideas about regulation of AI in healthcare? Do you think it's better to use a consensus-building approach and best practices and transparency, or are we going to see more heavy-handed regulation from government?

Lu: There's potential harm or loss both ways. If you over-regulate, the loss of opportunity to improve care can be tremendous. But on the other hand, of course, if you become too loose about it, then you risk harming patients. That is why there are a lot of legal experts involved in this process and trying to decide what's the right balance or what's the right legal framework around all this.