Stanford University professor Steven Lin tells AIMed what inspired him to create the Healthcare AI Applied Research Team (HEA3RT), why it’s not a must for clinicians to pick up coding and computer science, and how to facilitate the deployment of AI in actual clinical settings.
What initially sparked your interest in medicine?
I have always liked the puzzle-solving nature of medicine, merged with the opportunity to build long-lasting relationships with individuals. When I started college, health disparity was a big passion of mine. I began working with different nonprofit organizations to address some of the social determinants of health (SDoH) that drive racial and ethnic health disparities across communities, and that got me interested in healthcare.
I decided to enter medical school and focus on primary care and population health. I went on to complete my residency program in family medicine and joined the faculty at Stanford, teaching students the same thing.
When did you become involved in AI?
That was more of a happy accident, rather than an intentional career move. I started in operational leadership, running one of our academic family medicine clinics at Stanford. That was when I realized there are many different areas where technology-enabled solutions can make patients’ and clinicians’ lives better, decrease healthcare costs, and improve health equity.
One of the first projects I got involved in was looking at the issue of documentation burden for healthcare providers. Documentation is a huge challenge – especially in the US where healthcare providers are spending, on average, two hours in front of a computer doing clerical work for every hour seeing patients. So, we created a voice-enabled digital documentation system, or digital scribe, and that project opened the door for me to the world of medical AI.
As I walked down that path, I discovered so many exciting things that were happening in the basic science space. There are many models and algorithms that are being developed, getting increasingly sophisticated and impressive, but at the same time, as a frontline practitioner, I didn’t see any of that technology materializing in the real world. So, I hypothesized that there’s a translational problem in healthcare AI and we need more investment to get better tools from code to bedside. That’s why I started the Stanford Healthcare AI Applied Research Team (HEA3RT).
Since you established HEA3RT, and especially with the ongoing pandemic, do you think translation is getting smoother – as in, we are seeing more AI solutions channeling to the bedside?
I see progress on the front of translation, but I wouldn’t say that the floodgates have opened and suddenly we’re in this brand-new world of applied AI. Even though I work at an academic medical center where world class AI scientists are sitting next door, I still don’t see many AI-driven tools at work. But I do feel there’s an increased awareness of the need for out-of-the-box thinking and new models of care delivery that doesn’t reside in physical clinics and in-person visits only.
So, the expansion of telemedicine – virtual visits, digital health in general – I think has opened up opportunities for AI to augment that type of virtual care delivery. I think there’s greater awareness, excitement, and enthusiasm. To some extent, I also see some patients coming around to the idea of virtual care delivery, and healthcare providers at the same time coming around to the idea of it. But there’s still a huge gap between all the stuff that’s happening on the basic science front and what we are seeing on the frontline.
Why the gap?
There are many barriers that prevent this translation from happening. One is that, unlike pharmaceutical drugs or traditional medical devices, AI is unique in the sense that it’s almost completely dependent on the data that goes into building the models and feeding the algorithms. An AI model might work well at one institution, built using that institution’s data, but it does not always translate well to a different environment. We see from time to time that a model with 90% predictive capabilities at one institution performs like a coin toss at another. We must validate the models, anytime when we are taking it into a different environment, against that environment’s data. That’s far more difficult than, say, a pharmaceutical drug or medical device that you can pretty much plug in anywhere and expect it to work.
So, the nature of the technology is one problem that leads to another problem which is data infrastructure. We have very poor data interoperability. We have silos of data and datasets that are often very messy. We don’t have a lot of high-quality data to show how models perform in the real-world, outside of these highly controlled experimental settings.
An example I can think of is that there are many computer vision and deep neural networks that take photographs of skin lesions and generate potential diagnoses. Their performance can be quite impressive compared to board-certified dermatologists, and many have a high predictive value. However, what we don’t know is whether putting that kind of technology into the real world – into the hands of patients and healthcare providers – is going to be helpful to patients in the way that we hope, in the sense that they can get quick answers to their dermatologic concerns. Or is it going to cause undue anxiety for patients who don’t understand how to interpret the findings? Does the increased patient anxiety result in unnecessary utilization of the healthcare system when access is already an issue? We also don’t know how primary care providers will use this type of technology. Is it really going to help them make better decisions?
So, does the tool make meaningful differences in management and patient outcomes, or does it add complications and confusion to the system that doesn’t benefit patients and providers at the end of the day? We don’t know the answers to any of these questions yet. We need to go beyond the studies that are happening to validate model performance, and we need to see how they work in real-life because we don’t know what the unintended consequences might be.
What have we done well and not so well in medical AI?
I think we have done well in the following:
There’s no shortage of innovations that are happening. On the basic science front, we are taking a vast amount of data, applying cutting edge machine learning methodologies to the data, and discovering incredible new ways of predicting clinical deterioration in both inpatient and outpatient settings, incorporating algorithms to traditional devices like stethoscopes to increase our ability to detect murmurs and arrhythmias.
There are so many promising things that are happening that I think, theoretically, could truly improve care – both at the individual and population level. Also, the excitement, enthusiasm and the amount of talent that’s going into the approach. All this is great. But I think what we need to be careful about is that, just because things are promising, it doesn’t mean that we get to skip the work that needs to happen in the translational and implementation space.
We need to make sure there are no unintended negative consequences of the technology that we are introducing. We need to be aware that, when a model performs well in an experimental setting, that doesn’t mean it will perform in the real-world setting and – even if it does perform well – we need to know how it affects downstream physician and patient behaviors and how it affects the larger health system in general. We don’t know any of that yet. This is a very early stage in the evolution of healthcare AI.
I think we have reason to be excited and optimistic about what the future holds. At the same time, we also need to be cautious that the very technologies we are producing to impact patient care may not help patients and may make health inequities worse. There are already so many inequities in our healthcare system, we need to make sure we are not introducing additional barriers that will further disadvantage individuals – especially those in marginalized racial and ethnic minority groups that will not have access to this technology, or even worse, be harmed by the technology.
If we could predict a patient’s chance of going to the emergency ward or the hospital prospectively using electronic health record (EHR) data without relying on claims, then we could potentially do a lot of good and make a dent in one of the biggest problems in population health, which is the vast amount of human suffering and economic pain that’s being generated by preventable hospitalizations and emergency ward visits. We have lots of amazing work going on now but there’s also evidence to show that, if you just focus on EHR data alone, which represents only people with access to care, you are leaving out everybody who doesn’t have access to care.
So, predictive models that don’t include things like SDoH data will often underpredict the risk of racial and ethnic minorities who are poorly insured or aren’t insured and don’t end up in the emergency ward until they have a catastrophic illness. We need to make sure we are incorporating things like SDoH and community-level data, and not throwing technology into the wild without meaningfully thinking about the health equity implications to avoid harming the people we are trying to help.
Is HEA3RT doing any work to overcome some of the mentioned challenges?
The whole reason I created HEA3RT was to fill that translation gap between basic science and real-world implementation. We have developed a tripartite strategy. The first is to make sure that use cases are always clinically relevant and human-centered. Secondly, we focus on translational research techniques that tend to be better for AI implementations than traditional research techniques. Finally, we advocate for accountable and equitable AI.
In terms of making sure that use cases are relevant, there are a lot of cool AI applications that are possible that don’t necessarily fill a genuine need. We partner hand in hand with technology companies, nonprofit organizations, academic researchers, and startups to identify the right problems that fill a genuine need in the care of patients, and that is meaningful to health providers, health systems and society.
AI technologies are not the same as pharmaceutical drugs or traditional medical devices. Every single model needs to be validated at the local level to make sure that they work. So, AI implementations are better served with rapid iterative cycles of improvement, rather than multi-year long randomized controlled clinical trials. That doesn’t quite work for this type of technology. At HEA3RT, we focus on strategies and techniques such as quality improvement, design thinking, human factors engineering, and others to fix current healthcare processes. I believe this approach will ultimately serve AI better.
How do you think AI will progress in the next decade?
I think there are a couple of different spaces that people are working on and it’s quite exciting. For example, the work looking at risk modeling – thinking about not just how you benefit patients at the individual level but at the population level – is especially important as healthcare moves away from volume-based care and fee-for-service care to value-based care and accountable care. We need to keep people healthy at home and prevent them from getting sick and having to go to the emergency ward and hospitals. We need better ways to triage with limited resources. How do you increase access to traditional healthcare services in clinics for those who need it, at the same time deliver care in new, innovative, and virtual ways to those who don’t necessarily need to come into the clinics? These AI augmentations of digital health delivery are exciting.
I think there are bright spots where AI can augment healthcare providers to deliver better care and patient experiences in every single domain. If AI can help us to increase access to care for people who currently don’t have it, that’s a tremendous benefit, going against the traditional thinking about technology being something only for the rich and privileged.
You briefly mentioned the value of AI in primary care and population health. Could you please elaborate?
I’d like to build on what I have talked about regarding the relationship between patients and healthcare providers. AI offloads cognitive, administrative and clerical burdens so that healthcare providers can pay more attention, and deliver more emotional energy and cognitive energy, to their patients in a way that strengthens relationships that lie at the heart of healing.
Today, primary care physicians are overwhelmed with data and the burdens of navigating through electronic health records, finding the right information to communicate via patient portals, and handling the incredible amount of documentation – so much so that they barely spend time with patients. That’s a terrible thing for medicine because I truly believe that the human-to-human connection is what drives healing. If we can use AI to offload that burden from physicians, they will have more emotional and cognitive bandwidth to deliver better patient-centered care. Diagnostic ability also goes up, with fewer errors, and long-term, therapeutic relationships can also be forged. I hope to see more AI applications in this space in the next five to ten years.
What advice would you give to someone starting their medical AI career?
I think one of the things I often hear from students and residents who are interested in the space is “Oh, do you mean I have to learn how to code? Do I have to get a degree in computer science or informatics?” The short answer is, no, you don’t have to do that. I don’t have any computer science background and I don’t know how to code. What I tell my students and residents is that technology partners are not looking for a physician who knows computer science. They are looking for a physician who understands medicine: how the healthcare system works, clinical workflows, the genuine problems that need to be addressed and how AI can play a role in solving those problems. In other words, I think the way we can best further AI is to break down the silos of our disciplines and allow healthcare professionals to work alongside developers in a synergetic way.
What’s been the best piece of advice you’ve ever received?
That’s a hard question. I didn’t have a mentor to guide me in the AI work that I do, as I sort of stumbled into the space. However, I’ve had many mentors who have encouraged me to pursue my interests broadly. They gave me the space to explore my interests and I am deeply grateful for that. I think the best piece of advice I’ve ever had is not to be afraid of not knowing everything about the technical side of AI because that’s not what makes me valuable to the work. It’s the medical side of things that makes me valuable to my work.
Yes, there may be an entirely new lexicon that I must learn, but I don’t have to be scared that I don’t know all that. I also don’t pretend that I know all that. I don’t need to feel inferior when I have nothing to add technically, I should feel like I have a lot to add clinically in terms of my experiences as a physician and healthcare system leader. I feel empowered and confident about what I have to offer while being humble about all the things that I don’t know about in AI. If one can strike that balance, I think it allows you to succeed in the field.
If we insist that coding ability and AI knowledge are prerequisites for clinicians doing medical AI, then we are decreasing all the potential talent, knowledge, and great perspectives in the world to only a small fraction of individuals that can participate and engage in this important work. I think that’s detrimental to the whole field. So, we need a diverse talent pool, we need to have different sets of experiences in the conversation. That’s why I encourage others not to be scared about not knowing.
What’s the best way for clinicians to work with developers and ultimately, work with AI-driven tools?
I think the best practices are still being discovered because we haven’t been doing this long enough to know what they all are. What I see in the field is people operating in silos, because that’s what we are used to and it’s convenient. Thus, I feel we need to incorporate partnerships and conversations between clinicians, developers, and other stakeholders at the very early stages of ideation. Then everyone knows all the relevant clinical workflows, potential pitfalls, and identifies the right problems, even before the model-building part starts. I believe this is a better way to create technologies that meaningfully serve patients, providers, and health systems.
What’s the greatest challenge you’ve ever overcome?
Understanding how to engage with industry has been a big challenge, and educational. It’s not anything that I was ever trained for or learned during medical school and residency. Appropriately engaging with startups and big technology companies, and speaking with people who speak a different language when it comes to business and technology is a huge challenge. I have made some missteps along the way, as I didn’t know how to effectively partner with industry collaborators. Sometimes I wonder if it would have been helpful to attend business school. I feel that this has been the greatest challenge I have overcome.
What’s left to conquer?
I think there’s so much work to be done in this space. I created HEA3RT because of all the barriers I see related to the translation of AI technologies from code to bedside. One thing which worries me are some of the high-profile setbacks. For example, the Amazon-Berkshire-JPMorgan venture to disrupt healthcare is disbanding and IBM is giving up Watson Health as unprofitable. I am afraid if we don’t have tangible, scalable, real-world success in implementing AI on the frontline, we may see enthusiasm and investment fade. This will slow the momentum required to keep the space from advancing. That’s why I think translation and implementation is so crucial.
We need to fix this translational problem in the same way that the pharmaceutical companies had to do several decades ago when there was a big gap between scientists in the lab making breakthrough discoveries and physicians on the frontline treating patients. It wasn’t until they invested in translational science that those breakthrough discoveries became lifesaving drugs that are used on the frontline. I think the same problem is happening in healthcare AI right now. We need to start moving those innovations to the frontline, meaningfully evaluate their implications and effectiveness in real-world settings, and then scale and disseminate that technology.
If you could go back to the past, what would be that one thing you wish to change or do differently?
I wish I’d had a bit more business training and I think that’s super helpful coming from the perspective of an academic and engaging with industry. I think that real progress in healthcare AI can only occur when there are meaningful academic and industry collaborations, and so that is what I really hope for. I don’t regret not knowing computer science, but I do feel often that I wish I had more business expertise and understanding of the industry in general. But, as I said, I stumbled on this path accidentally and unintentionally, so it’s difficult to look back and say what I would do differently. I am just grateful that I get to work in a space that I think is one of the most exciting things happening in all of healthcare today.
Dr Lin is a speaker for our Clinician Series on primary care and population health. Click here to find out more.