This transcript has been edited for clarity.
Eric J. Topol, MD: Hello. This is Eric Topol, editor-in-chief of Medscape. I'm delighted to have Dr Aaron Neinstein from University of California, San Francisco (UCSF), on our Medicine and the Machine podcast today. Welcome, Aaron.
Aaron B. Neinstein, MD: Hi, Eric. It's great to be here.
Topol: I didn't realize all the great work you were doing as vice president of digital health at UCSF, a place I know well from when I did my residency training. But then I read your op-ed in Nature Medicine in October, "Our AI Future Is Better Than You Think," which said to me, whoa, he gets it. Let's start with that. This was before the craze over ChatGPT and large language models started in November and December. Could you summarize what you wrote about in that important Nature Medicine piece?
Neinstein: What I wrote about was my experience as a practicing endocrinologist. I spend a lot of my time working in digital health and realized just how wrong I was about the impact that artificial intelligence (AI) would have on medicine. We read all these stories about automation in other industries and how automation replaces people. Farm machinery has replaced farmers. Factory automation has replaced factory workers.
As an endocrinologist, I treat a lot of people who have type 1 diabetes, and about 5 years ago, I thought I was going to be next, that I would be replaced. Backing up for one second, for people who have type 1 diabetes, their care for many years involved close management by using of an insulin pump and a continuous glucose monitor. The insulin pump is delivering programmed amounts of insulin throughout the day, and the continuous glucose monitor is checking patients' blood sugars throughout the day.
During the beginning of my career, I would spend visits with patients downloading and printing out tens of pages of charts and graphs and data, and we'd pore over these, looking for patterns. My job was to look through these pages and pages of data to try to make recommendations for how patients might program their insulin pumps differently to improve their glucose management. That's what it meant to be an endocrinologist — sifting through data, looking for patterns.
Well, along comes 2017, when the first automated insulin delivery system came out, and that put an algorithm into the loop; this algorithm was between the continuous glucose monitor reading blood sugars and the insulin pump delivering insulin. We started talking in the hallways. Are patients with type 1 diabetes going to need an endocrinologist anymore? I would joke with patients, "Maybe you don't need me anymore." But I was so wrong. My practice of helping people with type 1 diabetes didn't disappear. It's been transformed for the better. We've realized that I was a poor substitute for a computer.
My patients now are having better outcomes than they were in the past with this algorithm in the loop. When people come to see me, rather than poring over pages and pages of data, we talk and engage with each other. I believe we've been able to develop deeper relationships because I talk with people about what life is like with type 1 diabetes — their challenges, their goals, their needs, their struggles — and I'm no longer trying to be a computer.
It has opened this world for me to peer into the future and see that AI and machines should augment the work we do as physicians and return us from staring at the computer or trying to be computers back to the bedside and back to spending time with patients. You wrote a whole book about this. I was inspired by your book and by my experiences in the clinic to write that piece that was in Nature Medicine.
Topol: The AI fields people jump to right away are radiology, dermatology, and even ophthalmology, but you brought out what AI can do for managing diabetes. It's a one-pager and I recommend that everyone who's listening should read it in Nature Medicine. The last sentence is, "What a joy it will be when this becomes a norm and when we can shed the role of being a poor substitute for a machine and reclaim the role of healer."
This is the essence of the ideal way we harness digital technology and AI analytics in the future. The depth you have in this area is extraordinary because it isn't just the use of AI that we're going to get into with large language models and foundation models, but antedating that, you were the co-founder of an organization called Tidepool that's involved in democratizing. A lot of people don't realize that when you digitize, you also democratize. Tidepool is exemplary for how to empower people with a given condition. Tell us about that.
Neinstein: Thanks, Eric. People can probably relate to the digital camera analogy. Years ago, in the early days of digital cameras, every camera company had its own proprietary file format. You would need a special chip or card from each digital camera. Sony had its own. Canon had its own. Nikon had its own. And the only way to download pictures was by using their software. Of course, now everyone uses one piece of software to capture information from every digital camera device.
The same thing has been true in the diabetes space. We have all these devices — continuous glucose monitors, insulin pumps, glucose meters — and for most of history, the data was and, in many cases, still is kept in proprietary silos. The manufacturer of each device creates the software. They force you to look at the data from the device within that software.
But the key thing you're touching on is that the software is built for doctors, not for patients. What we tried to do when we established Tidepool as a nonprofit a little over 10 years ago was to create one application where the person with diabetes would own their diabetes data. They could pull it in from any of the devices they were using, and then they are the locus for sharing that information with their medical team.
We tried to push forward several important concepts: that the person with the condition is the locus of control of their data, and also the importance of using device data as opposed to the traditional types of data we use in health systems, the data that are in electronic health records. So much of care today relies on data that live outside our electronic health records. I believe that was another important piece. Tidepool helped support telehealth because even before the pandemic, most of the visits I would do with patients with diabetes would be on Zoom, looking at their data on the web in Tidepool. It was an exciting platform that we built around those concepts.
Topol: We talk a lot about wearable sensors, but it really was the glucose sensors that ushered in the true digital medicine, not just steps and sleep metrics or heart rate. The glucose sensors had a transformative impact, and being able to have people own their data and have it interoperable between their sensor and their watch or their phone or whatever, this is important stuff.
On February 9, you were part of a terrific UCSF Grand Rounds with Bob Wachter and your colleagues Sara Murray, Atul Butte, and Dan Lowenstein. The four of you ganged up and took on the field of ChatGPT and large language models. I want to get into that with you because in so many ways, your Nature Medicine op-ed was prescient because that was even before these foundation models got the buzz. I want to get your views about how this could affect medicine even more than you've already written — the algorithms, the ability to have virtual scribes, the ability to deal with insurance providers, and all this kind of thing.
Neinstein: I am very excited about this. There are a few moments in history where the common person can pick up and see a new technology and realize that the world after may never be the same as the world before. Many people had a moment like that with the iPhone, and before that, the web browser was probably a moment like that. I believe we're in another one of those moments, and it is equal parts exciting, thrilling, and terrifying to think about the consequences and the risks and benefits that are going to come.
To your question about how this will change practice in medicine, for people who are practicing physicians, we all know the work we love doing the most is the administrative work, right? It's filling out prior authorization forms, filling out school forms, and durable medical equipment forms. This is why we went to medical school, right?
Topol: Oh yeah. Oh, sure.
Neinstein: To sit late into the night writing notes. So much of medical practice has become consumed by this administrative work and bureaucracy. And electronic health records, while they've done so much good, they've also become a vessel for putting more administrative overhead onto physicians', nurses', and care providers' plates.
The huge opportunity with AI is to take that work off our doctors' plates. I believe my experience in endocrinology is a microcosm for what we'll see, in that we need to return doctors from the computer to the bedside. We are facing an epidemic right now of burnout in our health systems. As you look across the country over the 3 years of the pandemic, doctors, nurses, frontline care teams, people are burnt out. They are overworked.
There have been a lot of articles recently about the moral distress people feel by working within a system where all the administrative work is not adding up to better care for our patients. There are enormous opportunities with AI. But I think GPT and large language models are a massive accelerant as we think about the tasks of hunting through the electronic health record and this massive amount of data to find those few key pieces of data that are important for decision making. Or filling out prior authorization or administrative forms, time that we spend at night documenting to put notes in the electronic health record. So, I see a huge opportunity.
There are many ways the large language models can accelerate things. But traditional AI requires a precise question being asked of it. It requires you to build a workflow or a solution that you put into care delivery, and then you have to train an algorithm to do a specific thing for a specific purpose.
This is democratizing AI. Of course, we're going to have to work through creating HIPAA-compliant versions for health systems, where medical information is safe to put into these models. That doesn't exist yet. But the average doctor can sit down, and they have this prompt — Microsoft is calling it a copilot — at their disposal to help them do their work.
We're already seeing TikTok videos of doctors saying, "Write the letter for me to United Healthcare." I have experimented with this myself, and I have to say it feels really good as a practicing physician to get that support and that help. Part of the burnout is that modern medicine too often ends up pitting the needs of the patient against the needs of doctors. So, a patient's need becomes more work for the doctor, and I think it's so unfortunate how this contributes to burnout and even resentment.
Patients ask their doctor a question via email that ends up as a patient portal message, and we hear about how doctors are staying up late at night, forgoing time with their families, answering these messages. We didn't get into medicine to be upset that our patients are sending us messages and trying to engage. We want more interaction like that. We want patients engaging in their care.
The problem today is that the technology is not supporting us in providing that level of care we got into medicine to provide. I see enormous potential in many of these use cases for large language models to accelerate, augmenting what we do as physicians and taking some of this administrative work off our plates.
Topol: I couldn't agree with you more, Aaron. I didn't know about the TikTok video and the preauthorization, but I have seen some others that show how well this can be done.
During your Grand Rounds, your colleague Sara Murray, presented a couple of actual interactions with ChatGPT. It was almost like it exhibited compassion; that is, conditioning the written text to reflect the style of the physician, of the person.
ChatGPT is not really understanding language, but it can try to simulate a real doctor's response. A recent New Yorker article had an excellent discussion about lossless vs lossy compression of language. Any comments about that?
Neinstein: That was a wonderful article, "ChatGPT Is a Blurry JPEG of the Web" — a great cautionary story about not getting over our skis with this new technology as we're caught up in our excitement.
Today's GPT is generalized based on the training data it has. We don't yet have versions of this that are trained on our health data within our HIPAA-protected environments. I imagine a world where I sit down, and I have my own large language model copilot that is trained on all of my charts from the past 10 years and how I care for people with diabetes, how I write, and how I think. So when I sit down in the afternoon to do my documentation, it's pulling in a version of me, essentially. Could it do that for each practicing physician and be trained on the way they practice, their style, as well as best practices?
I think you can envision more personalized copilots that are augmenting and drafting things for us. You see a lot of talk on the web and on social media about how there's going to be a premium placed on our ability to prompt these chatbots and AI and how we write the right questions and the right prompts to pull information forward.
It's analogous to using a calculator, right? A calculator can't do algebra for you. You have to know the concepts. You have to know what to input and how to place the outputs into context. I see this in much the same way in that, as you said, the accuracy may not be perfect with these large language models today, but if, as a trained physician, I'm able to look at the draft and edit it, I think it mitigates some of that inaccuracy.
Chatbots are not ready to do the work of medicine for us. At least today, I believe it will require the expert in the middle. But there's enormous promise in drafting language to support our work that is personalized around the way we might practice.
Topol: It's critical to point out that we're just seeing the beginning of large language models. GPT 4 is coming out soon, and Sparrow, and all these other new models that go to a much higher level of parameters, in the trillions. It's just going to get better. It's true that we have liabilities here, but we also have tremendous opportunities to harness in the future.
This is a prototypical Medicine and the Machine story. Medicine isn't ready for it in one respect. Have you put AI into the curriculum at the UCSF medical school? Has any medical school done this? Are we going to change the selection criteria for medical students in the future? What are the adjustments in education and training that may be well suited to this?
Neinstein: We in the healthcare system need to acknowledge that we need this. We're facing huge challenges as a healthcare system. Hospitals are overfull. People are having a hard time getting access to care. Primary care is becoming hard to find. Specialty care is hard to find.
As a healthcare system, we need to look in the mirror and acknowledge that we can't just work our way out of this, meaning that the one-to-one model of care — where a synchronous interaction between a doctor and a patient is the only way medical care happens — is unsustainable. Acknowledging that is necessary before we take the next steps forward. There was a lot of hope that telehealth would solve this. But, realistically, a video visit — yes, it eliminates real estate costs and maybe eliminates geography as a barrier, but at the end of the day, it's still a synchronous interaction.
It is critical that we look at the amount of care that's needed out there, the problems with disparities, the problems with costs, the problems with access, and say that we have to have new care models that rely on broader care teams, technology, and asynchronous engagement if we're going to deliver the quality and the value of care that's needed in the country.
We need to acknowledge that first because so many changes are going to be required in how we think about medical education, how we think about training, how we think about developing systems for care delivery, and how we think about what research needs to get done downstream of that acknowledgment. To me, that's the most important first step.
I still teach our first-year medical students. I still work with our residents and our fellows. The good news is, the next generation is ready for this.
In many ways, we need to get out of their way and let them lead on these new-use cases because they're fluent in what this technology means. They've been raised on these technologies. They understand the implications, and I believe they're going to have phenomenal ideas about what's possible. The more we can harness the ideas and enthusiasm of the new generation and include them in developing new systems of care, the better, because they're an important part of the path forward.
Topol: To your point about having to embrace this, step number one is essential because we haven't yet seen that across the medical community — in particular, reconfiguring education and training.
You're responsible for a lot of digital health at UCSF, which, as I understand it, has lots of digital groups that are trying to change the world in many respects. Are you gearing up for a "hospital at home?" Because, as we get all these sensors and multimodal AI and large language models that can cross between text and images and speech, we're going to have the power to keep people in their homes instead of in hospital rooms at the UCSF Medical Center. Are you gearing up for that?
Neinstein: We are. Like many health systems, we think hospital at home is going to be an important part of the future. Early data coming out of pilot projects from other systems around the country have shown benefits with respect to outcomes, cost, and patient experience, so we are also excited about the hospital at home potential.
I believe that as you look at the overall care ecosystem, it's the nonhospital care at home, the rest of the care at home, that is bigger. Chronic illness is a huge burden of disease in this country that, unfortunately, is only going to get worse as long COVID numbers rise. We continue to see metabolic disease and obesity on the rise. Hospital at home is a wonderful tool.
The number of people it may touch is small. It should have a larger, outsized impact on reducing costs, given that hospital costs are such an important driver. But as I think about the care that touches most people day-to-day, I think it's this ability to do more continuous care, like we used Tidepool years ago to interact with people regarding their chronic conditions on a more continuous basis, that's going to have a lot of meaning and impact, moving away from the synchronous doctor-patient visit as the only way we do care. I'm looking forward to transformations in that part of care.
Topol: The last thing I want to zoom in on is empowerment of the patient. I know you've already mentioned that with the work that you did in Tidepool and how it's essential that people have their data and work with them. But increasingly, we're going to see a lot of diagnoses at least screened for through patients capturing their own data, whether it's heart rhythm, urinary tract infections, skin cancers and lesions, ear infections in children, and a long, long list of common conditions that otherwise would now require a doctor or clinician diagnosis. Then we also see all these coaching entities using the person's data for diabetes, hypertension, depression, obesity, and every chronic condition, and then chatbots with humans in the loop as backup.
What about this? We tend to only think about the clinician's side of things. You have had a much broader engagement, knowing the importance of the patient side of AI data analytics and data capture. Obviously, not every patient wants to be empowered or wants all their data. But what should be our best position going forward to help promote that?
Neinstein: It's a great question, Eric. I think several things need to happen, and a lot of them are in the policy realm. One important thing that has been made even more important by the Supreme Court's unfortunate Dobbs decision are protections around privacy. People are quite familiar with HIPAA, but I believe there's a false assumption that all health data are protected by HIPAA, and that, unfortunately, is untrue. A lot of data out there in the consumer world are not protected by HIPAA. So I believe comprehensive privacy reform is an important part of this story.
I'm encouraged to see that it tends to be a bipartisan issue. There are some proposed policies coming through various congresspeople, so I think in the next couple of years we will get to privacy reform around consumer data, and that will be important for health data.
Another important part of this story is interoperability policy through Health and Human Services and the Office of the National Coordinator and other agencies. We passed the 21st Century Cures Act in 2016. The core message of 21st Century Cures was that all patients, all people in the United States, should be allowed to access their health information without any special effort. That is making huge differences in people being able to access information from their electronic health records.
But as you were implying, what we think of as health data — coming out of the electronic health record — are the labs you have done: the radiology and imaging studies, x-rays, and MRIs. But health data is much larger than that.
It's the data that are coming from your Apple watch or your Fitbit. It's the genomics data you obtain. It's your continuous glucose monitoring data. Those other data sources are growing in size and importance. And just as the privacy policies don't really cover those data, neither do our interoperability policies in a very robust way.
I believe we need a more systemic re-envisioning of what the scope of healthcare means, because we need to draw much broader borders and boundaries around what healthcare is and that it is occurring in the home and on a continuous basis for people who are using devices. All these different policies and technologies we're building need to apply not just in the hospital setting and not just in the acute care setting, but also in the home and outside our walls if we're really going to be successful.
Topol: I couldn't agree with you more. Just to wrap up, I know that you graduated medical school 28 years after me at University of Southern California, and so you are a next-generation leader of digital health and what will be just called "medicine and health" and forget the "digital" eventually. We're counting on you, Aaron, to keep up this great work.
I have such a tremendous regard for you. Getting your thoughts today has been so helpful to provide some anchoring on the concerns as well as the excitement we can look forward to.
Thanks so much for being with us. Let's get on to healing, and let's get away from all this burden of keyboards, and papers, and authorizations, and all that hullabaloo.
Neinstein: Thank you, Eric. Great to be with you.
Follow Medscape on Facebook, Twitter, Instagram, and YouTube
Medscape © 2023 WebMD, LLC
Any views expressed above are the author's own and do not necessarily reflect the views of WebMD or Medscape.
Cite this: How AI and Chatbots Can Make Us Healers Again - Medscape - Feb 21, 2023.
Comments