Lee Schwamm: Onboarding AI at the Hospital

Transcript

Harlan Krumholz: Welcome to Health & Veritas. I’m Harlan Krumholz.

Howard Forman: And I’m Howie Forman. We’re physicians and professors at Yale University, and we’re trying to get closer to the truth about health and healthcare. Our guest today is Dr. Lee Schwamm. But first, we always like to check in on current or hot topics in health and healthcare. What do you got today, Harlan?

Harlan Krumholz: Well, today I want to talk about a story that’s been emerging. About vaccines, scientific intransparency, and a paper that almost disappeared.

So the New York Times reported this week that the FDA held back publication of several vaccine safety studies, including two COVID-19 vaccine safety studies, that had been accepted by journals. And two abstracts on shingles vaccine that reportedly were not cleared in time for presentation at a scientific meeting. So according to this report, these were studies that were conducted by FDA scientists and contractors using large health datasets. And the findings were largely reassuring. Serious side effects were rare. So this had a directionality to it.

Now, the key thing is that this story, I think, should raise concerns. I will say it’s not the only administration that’s held back papers. For a long time, federal agencies have been in a position of approving studies that are done with federal dollars. So I don’t think people should think this is entirely new. I will say that at our place, at the Center for Outcomes Research and Evaluation, we have contracts with CMS. And we insist that, when we’re publishing papers, that they will never be held back for content. They can be screened for whether or not they have any confidential information, but that we can’t work with them if they’re going to be in a position where they would censor a paper because they didn’t like the result.

Now, that’s not exactly what’s being said here, because no one’s saying why these were held back, but it seems like the implication was that there was some issues about what they were finding. But one of the papers that they held back actually didn’t disappear. It had been posted on medRxiv. Now, medRxiv’s a preprint server. In disclosure, I, with Joe Ross, we’re one of the co-founders of medRxiv. It’s a place where scientists can put papers before they’ve been peer-reviewed so that people can discuss them, feedback can be given. They shouldn’t be necessarily acted on, in the sense that they haven’t been through scientific scrutiny yet, but they are an opportunity for people to put them out and make them freely available. And the fact that this preprint is out there actually creates a little bit of transparency and lets us look at one of these papers that was held back.

Now, they said that this paper was at the journal, so there likely was a draft or two beyond what was posted on the preprint server at the time that it was being considered for publication. But let me just tell you a little bit about this, because we can know about this paper because it was posted on medRxiv.

This study examined the safety from 2023 to 2024 of the COVID-19 vaccines among Medicare fee-for-service beneficiaries, people 65 and older. It included more than 7.5 million older adults who received the updated COVID-19 vaccine. They looked at 14 pre-specified serious outcomes, including heart attack, stroke, myocarditis, which is inflammation of the heart, or pericarditis, the lining around the heart. Guillain-Barre syndrome and neurologic syndrome, pulmonary embolism, thrombosis with thrombocytopenia, anaphylaxis.

The design was a self-controlled case series, which means that people essentially served as their own controls. The investigators compared the rate of events during a defined post-vaccination risk window, with rates among other periods, during other periods after vaccination. The approach sort of helps reduce confounding. It’s not perfect, but it allows us to make a comparison: same person, during a period where they might be vulnerable due to the vaccine, versus how did they stand before they got the vaccine.

And they reported no new safety signals among these older adults. There was a possible rare increase in anaphylaxis among people who got the Pfizer-BioNTech vaccine, but the estimated attributable risk was very small, meaning the risk was extraordinarily small. Maybe less than 0.1 cases per hundred thousand, like one case per million. And after accounting for some potential outcome misclassification, it was no longer statistically significant. So it wasn’t a robust finding.

Now, these studies aren’t perfect. Claims data are powerful but pretty crude. Diagnoses can be miscoded, some events are rare, which means they’re very difficult to estimate. The choice of the risk window matters, the way seasonality was handled matters. But with many outcomes and vaccine brands being tested, it gives you some information that may be useful.

Now, there’s lots of things that they could have improved, based on the thing that had been posted as a preprint. There are issues that could be addressed in peer review about this, points about what they were looking at, what kind of other analyses they might have done, but this is the whole point, is that science is progressing—“progressive and self-correcting,” as Francis Collins has said. We need a chance to be able to debate these studies. Any study that’s published should be debated, but to hold it back from public view, whether it’s this administration, another administration, isn’t in the best interest, I think, of the public.

Now, our friend, Akiko Iwasaki, who’s been on this program, a close colleague of mine, world-renowned immunologist, wrote a piece in Nature Reviews Immunology where she started arguing that science must protect space for rigorous inquiry into controversial questions, and that we shouldn’t be retreating from this. And in fact, she was focusing on vaccine safety. And her point was, on one hand, we can’t give oxygen to misinformation. But we should allow studies, serious science to be published, even though it may go against the grain, even though it may be in areas where people think they know the answers but maybe they don’t. And that public trust depends on our willingness to ask difficult questions openly and answer them with evidence and debate them. And debate them, really.

So anyway, I wanted to bring this up because, it’s being talked about. Like I said, I don’t think it’s specific only to this administration, although there’s a lot of instances that are being reported around the administration. And interestingly, many people in the administration, like Jay Bhattacharya, head of NIH, has actually been championing the idea that he thought he was canceled, when he was publishing studies that were out of favor at the time that he was a citizen. And so what we need is for people like Dr. Bhattacharya to stand up for publishing these things and letting them get out and be debated. Vaccine confidence is not strengthened by saying, “Trust us. We looked into it.” People need to see the work, they need to be able to evaluate it, we need to be able to debate it.

And so, I think I’m going to come out strong advocacy here, for saying that if the studies are done, they should be shared. They shouldn’t be pulled back from the journals because there are some issues. Let us debate them, especially if they’ve been through peer review. And preprint servers can serve a very valuable purpose, as they have done here, to give us a view into the work before it goes through peer review, but allows us to debate it, see it, be openly and freely available for people to debate, and contributes to our scientific dialogue.

Howard Forman: Isn’t part of the problem, though, that the journal editor community, of which you’re a member of now, but you obviously weren’t until recently, that they need to do a better job of being able to say, “This is controversial, and so, I may allow somebody to write a opinion piece or otherwise give feedback to it, but that to squash or quash a publication from appearing is just antithetical to what we’re trying to achieve here.” Because that seems to be the stumbling block. It’s not generally the FDA that is censoring; it’s generally the editors.

Harlan Krumholz: Yeah. Well, you’re exactly right. So this was a situation where people have been reporting that the agency pulled these back, but maybe a bigger problem is that our scientific community doesn’t really allow controversial views to surface, and that there’s censoring that’s occurring at the level of both review and editorial decision-making. You can see it and feel it. It even comes up in some debates that we have, which is, people say, like, “Well, people are going to misinterpret this” or “This is going to be bad for the public.”

And my stance is that, this is science. And if it goes for peer review, if people are clear about it, you can’t hold it back because you’re afraid that some bad actors are going to leverage it in ways that go beyond what the science says. What we need to be able to do is to put it out there and allow it to be debated, because where does that end? I mean, if you say, “I need to be paternalistic and protect—”

Howard Forman: Right.

Harlan Krumholz: “...the public from...” or “the science...” I mean, even worse, “As a journal, I need to protect other scientists from this information, because they may misinterpret it.”

I don’t know where that ends. I mean, that means “Only the things I’m comfortable with get published.”

Howard Forman: I’m just concerned that the peer reviewers.... Look, peer reviewers may give a low score to a paper because they feel like it’s somehow in opposition to their belief system, but their responsibility to the editor is to lay out reasons why they think this paper is wrong. And then it’s for the editor to say, “You know, these are actually addressable issues, and these are issues you can put into the limitations. But beyond that, you can write a letter to the editor responding to this if you want to, or write a different paper altogether, and I’ll consider it.” We need to do more of that.

Harlan Krumholz: There is some judgment though, Howie. I mean, if someone does a study that says injecting bleach is good for you, because the prior on that would be so low, your belief that that’s true is so low. The study has to be a little bit stronger before you would put something out that—

Howard Forman: That’s reasonable. That’s a reasonable point.

Harlan Krumholz: ... could be harmful.

Howard Forman: Yeah.

Harlan Krumholz: So there are judgments around saying, “It’s not that I’m against this, but this could be harmful if people do it, and it’s not right. And the scientific basis for it is very weak. So therefore, the study has to be that much stronger in order for us to feel like it’s moved us from where we started to where we would go out there.” So we can both champion being brave, and publishing things that we might not agree with, but there are some things we have to make judgments on about. Is it plausible? If it gets out there, could it be harmful? So then, how strong does the evidence need to be for us to push it forward? And those are all things that end up in our discussions.

Howard Forman: Yeah. But I think that’s a big part of the solution, is getting editors to be able to provide the type of leadership to reviewers to understand that there is a wide window of opportunity to present evidence, and that we shouldn’t be shutting down half the window.

Harlan Krumholz: Yeah. And you don’t have to yield to reviewers, either. I mean, if reviewers are basing their recommendations on invalid assumptions or arguments, then—

Howard Forman: Or biased.

Harlan Krumholz: ... it’s the editor’s job to make that judgment.

Howard Forman: Right.

Harlan Krumholz: All right. Hey, let’s get on to our interview. I know this is going to be great.

Howard Forman: Dr. Lee Schwamm is the senior vice president and chief digital health officer for the Yale New Haven Health System, and the associate dean for digital strategy and transformation for Yale School of Medicine. He’s a professor in the biomedical informatics and data science department, as well as the neurology department.

In his clinical practice, he is a stroke neurologist. Dr. Schwamm is an internationally recognized expert in the prevention, diagnosis, and treatment of stroke and transient ischemic attack, with his research interests focusing on stroke in the young, and cryptogenic strokes. Under his leadership, the American Heart Association’s “Get With the Guidelines” Stroke Registry has grown into the world’s largest stroke registry. He’s also, and this will be germane to today’s discussion, a leader in digital health and digital transformation, having helped shape the landscape of telestroke care in the United States.

He graduated from Harvard and got his medical degree from Harvard Medical School and completed his residency in neurology, including serving as chief resident at the Massachusetts General Hospital, plus a fellowship in stroke and neurocritical care. He spent three decades in senior leadership roles at the Mass General Brigham Health System before joining Yale just several years ago.

This is his third appearance on the podcast, and we are grateful to have you back. A lot is going on.

Harlan Krumholz: By popular demand, by popular demand.

Howard Forman: No, and really so relevant to everything we talk about on the podcast lately, because AI and digital health tech come up every day.

I just want to start off by mentioning that as I walk through the emergency room, and as I walk through my own office suite today, new software is being installed across the health system in an array of places. Some of them are large installations like Abridge, which is the ambient AI, some of them like in radiology, our new transcription service, which is huge for us, maybe smaller for other people. And then there are things that I come across where I meet clinicians and they’re like, “Let me show you something really cool about how UpToDate now integrates with ambient AI in the emergency room,” and so on.

Tell us a little about, from your perch, what you see as the transformation of the health system with regard to AI installations and tech in general.

Lee Schwamm: Well, first of all, thank you for that lovely introduction. My mother would have been delighted to listen to that introduction. I think each time I appear, you should give me a shorter and shorter intro.

I think we are all in the middle of a transformation. I suspect that the Industrial Revolution felt a lot the same way. You’re out there with your oxen plowing your field, and some guy has just got a tractor, and it looks dangerous. So we’re in the middle of a massive transformation.

There’s an important filter on the front end. Not anything that anyone wants to do should be able to happen in healthcare. We need guardrails that first and foremost prioritize the safety of our patients, but also prioritize, quite frankly, the burden on our staff. We have to figure out how to do more with less, and do more of what humans do best, and do less of clerical and administrative burdensome things that could be very effectively done, or even better done, by automation.

But I see three big buckets. The first big bucket is really the care team delivery model, and how can we just smooth around all the rough edges, and just make it easier and more delightful to do the thing you always wanted to do. You and Harlan and I, we are trained in an age before electronic health records. So you listen to the patient, you put notes on a piece of paper, you wrote a note that was usually only one or two pages long, not 15 pages long. And you spent most of your time at the patient’s bedside and not that much time documenting. That is not the case today, so a lot of the stuff that you see hitting the front line are things that alleviate administrative tasks or make certain things easier. Like you said, they’re not that exciting to me. They’re helpful, it makes people’s day better, but they don’t excite me.

What excites me are when we can completely turn the delivery model upside down, do something we couldn’t do before, because we have technology now in a way and in a place that creates a new opportunity that did not exist. I think that’s really exciting. Places where I see that happening: imaging sciences, and by that, I mean radiology, pathology, cardiology, where the AI can see something in the material, the scan, the slide, that you can’t see. Or that you might see once in 100 times, and it sees 99 out of 100. So that, I find those really interesting.

And then the last piece, that we spend very little time on, is the digital patient experience. And how do we make the patient’s experience better and better with the use of these tools, so they have the joy of making something possible rather than the frustrating experience of shouting “Representative!” into the telephone repeatedly, because the IVR response prompts just can’t help them answer their question.

Harlan Krumholz: One thing I wanted to ask you, just to get started here. There’s a lot of people who listen to us, actually have some entrepreneurial aspirations. And we are co-sponsored by the School of Management. There’s this idea about fostering change. What they’re seeing is, increasingly, that the big dogs are trying to elbow out smaller innovative groups that are coming in. And health systems are moving away from point solutions, “Here’s the solution to this pain point,” to saying “I want a platform. I want something that’s going to do something more than having to bring in a bunch of point solutions.”

Can we give them hope, that if they come up with good ideas and good work, that there’s still space for people who aren’t Google, who aren’t Abridge, who aren’t people who’ve become established. All those started small, as startups, at one time. But that they can compete with the big dogs and find a spot. What do you say to people who come to you and talk about that?

Lee Schwamm: Yeah. Oh no, those are great. Those are all great comments. But I want to get back to what you said about, are we at risk of being elbowed out of the way? And I think we face that risk on all sides. There’s a paper that was just published last week by Google, their DeepMind team called “Exploring the feasibility of conversational diagnostic AI in a real-world clinical study.” They put Google’s DeepMind algorithms up against primary care docs in a simulated setting with scripted, simulated patients. Basically showing, and it shouldn’t surprise any of us, that for common problems, at least... many of these are protocol-driven, and that a conversational AI can ask the same questions that a doctor can ask. It doesn’t mean that they can perceive the emotional context and all the other nonverbal cues that are happening in an office or a bedside setting, but we will face competition on that side of things.

But at the health system here, what we have done is, we do believe that platform approaches are important and that modularized approaches that allow us to continue to innovate and bring new services on board at low cost is key. The way we’ve found I think the perfect compromise is, we have something called the computational health platform, which is a secure compute environment, on premises, that sits in the same server space as all of our Epic data. So our live Epic environment and all the databases that back Epic are supported with a compute and storage layer on top. Why is that important?

What that means is that we can take an algorithm developed by one of our colleagues, put it in a container, drop that container inside the computational health platform, and then build the pipes to bring data from the organization or Epic into the algorithm, send the results back out of the algorithm, and then deposit them wherever it’s appropriate. Is that an email? Is that an in-basket message? Is that an action, or an order, or some other activity in Epic?

I’m really excited to say we are about to start silent validation on a project of one of your colleagues, Harlan, Dr. Rohan Khera, who has developed an automated AI ECG algorithm for detecting structural heart disease and basically recommending who might need an echo, who doesn’t need an echo. And that silent validation goes live later this month. And if that is successful, we’ll then proceed to true clinical integration. Why is that important?

Well, we have just created a platform delivery model inside the health system that creates a box, a slot for you to drop in a Yale-invented algorithm without having to go through all of the costly expense of getting all the security certifications, raising millions of dollars, going out and getting security certifications, and building out all the apparatus that most of our inventors don’t want to build. They want to still be in the health system. They want to be physician-investigators, but they want to bring these opportunities live in the health system. We have this with pathology; we’re going to have more of it with radiology.

So our vision is, create a fast lane for Yale inventors, remembering that it’s a huge step from having a great idea and even building a retrospective algorithm to building an algorithm that works in real time, where the data that the algorithm needs is actually available at the time the algorithm runs. So many people don’t think about it a lot. They use the discharge summary to power the algorithm, but they want the decision made during the admission. Well, you got to have the data at your fingertips when you need to make the decisions.

The other thing to remember is, we have the Center for Healthcare Innovation Innovation Awards. So we do have a pathway for inventors to get some seed funding and some support, and increasingly what we’ve been doing is trying to raise external funds so that the winners of those competitions have a glide path and some funds to support silent validation, prospective silent validation, in the health system data.

So I think there is a risk. Epic is releasing AI tools every hour, they’re knocking vendors off left and right. The other big players are all trying to move into healthcare. We have to be the stewards. We have to be the ones to ask the question, “Is good enough good enough in this scenario, or is excellence what is needed in this scenario?” And I like to think that we’ll use algorithms that save us money, contracts, efficiencies, getting rid of fax machines, automatically converting documents into electronic tasks. Use those savings to sidecar the innovations that make life better for physicians, make life better for patients but don’t actually generate revenue. So, use that two for one to advance the goals that we all trained for and the mission that we all adopted, which was to improve the care of our patients. And run it on the backbone of cost saving.

Howard Forman: Going off of that, as you know, there’s a backlash about AI society-wide. It’s mostly about losing jobs, I think, but there’s also a legitimate fear of anything new. You mentioned that at the top of the hour, when we were talking about tractors and farms. What is your role and what does the institution do, both in terms of governance as well as developing guidelines, to reassure the public about the data integrity, the data safety, and the general fears that they might otherwise have about somebody intruding on their office visit?

Lee Schwamm: Yeah. I think we’ve gone beyond the place where, if you’re a patient and you came to Yale New Haven Hospital to be admitted, you could say, “I don’t want any AI in my care.” But that moment is gone. It’s baked into our care in a way that we can’t extract. You won’t be able to create a radiology report without using AI. Many of our scanners now use AI to actually acquire the images, at a faster speed, in a way that’s better for patients. So we’re beyond that.

I think what we have to do is to make sure that patients understand, when they come into a clinical encounter, the clinician is still in charge, the clinician makes the decisions. The clinician is the interface between technology and the patient. They are the integrator. It’s like the attending when you’re a patient in the hospital. There may be medical students, there may be residents, but the final voice is the voice of the attending, and they’re the ones held accountable for the decisions.

But I think what I want all of our patients to know, and our staff, is that we have built an extraordinarily effective and comprehensive evaluation framework for any new AI that comes into the system. That starts with an intake process, which is seamlessly integrated with our digital intake process. Any new software that comes in gets evaluated; if it’s got AI, it goes through a secondary layer of review. We call it a multidimensional assessment that has domains of compliance and contracting, impact and value, and a deep technical assessment. So privacy and security, compliance and billing compliance, conflict of interest, legal and regulatory supply chain. Those are happening on every contract.

But we also look at ethics and equity with a dedicated equity evaluation. We look at workforce implications, and that factors into our decision about whether or not a more in-depth review needs to be done. Is this going to displace workforce? Is this going to reassign work within that workforce? And also, we look at the impact on the revenue cycle. And then we look very diligently about, how is the model built? How has it performed in its retrospective data? Is there any data sharing with an outside entity? If so, what is it doing with our data? And we have very strict guardrails around what can be done with our patient’s data without our permission.

And so, all of that gets wrapped into a multidimensional assessment that then goes to something called the Digital AI Implementation Advisory Committee, or DAIAD. And in that, we have representation across both the health system and the School of Medicine to ensure that what we’re doing is responsible, fair, trustworthy, and effective. And then on top of that, we have a healthcare AI Governance Committee, a small committee of senior leaders, which reports up to the Audit and Compliance Committee of our board, as well as the executive leadership council, which is Dean Brown, Chris O’Connor, the CEO of the health system, and their most senior direct reports.

So it took a lot of—I’ve spent six months building this framework, and it took a lot of time and thought, but I’m very proud of what we have. And it makes me feel that we could defend our selection process to any critic who wanted to understand what we reviewed, what decisions were made, and why they were made.

Harlan Krumholz: Of course, the balance there is not over-engineering it so it slows everything down. What I’ve been impressed by what you’ve done, because I’ve seen you at work, is to be able to create this level of scrutiny but not to slow down innovation so that it’s going to take years to get through the process. So kudos to you for that, because it really is about that balance of being able to stand behind the layers of assessment but not create it so that it takes five years to complete the process. So a lot of places are having trouble figuring that out.

Lee Schwamm: Yeah. I mean, a lot of our peers have a process, but they do eight applications a year. They do eight reviews, and each one is like a dissertation, and that’s not viable.

The other thing is that, I think it’s... I’m very proud. It’s months, not years. Things go through our process in months. And the reason is that we have put a lot of effort into, what does it mean to submit a proposal? Submitting a proposal is not submitting an idea. It’s actually submitting a piece of software, or a third-party vendor, or an algorithm that’s already built and is ready to go. In this venue, we don’t entertain, “I have a great idea for how to automate something, and I want you to do it.” We’re not in that business. We don’t offer that. We partner with inventors and investigators in other ways, but not in this process. This is a mature product process.

But the other thing that I would say is that we also have a fast track in this process. So if you already have an approved vendor who’s already providing services, and now there’s an AI enhancement, well, we’ve already done most of the diligence around that to begin with—that can move through incredibly quickly. So we’re not putting everything through this fine-toothed comb. It’s risk-proportionate. Where is the tool being used? How many are exposed to it? Does it have any patient-facing attributes? Is there reputational risk? Is there malpractice risk?

Those data feed into an overall assessment of risk, and for almost purely automation, with low-risk, non-patient facing, we can move a lot of those through very quickly in a package and therefore increase the efficiency of the process.

Harlan Krumholz: I just want to get one other point in, that just strikes me as you’re talking. One of the problems on the university side is that if you have a product that you want to study, and it’s based on an LLM, it becomes enormously difficult. Because you can vet if you want, but you can’t tell them exactly what the thing is going to say, for example, if there’s an interaction with a patient. And that puts the IRB [institutional review board] in a difficult position, because historically they’re saying, “Tell me exactly what you’re going to... show me scripts of what you’re going to say to the patient in a particular instance.” And you go to them and you say, “But this intervention is not script-based. It’s based on an interaction reasoning model. It’s a new thing.” And they’re, I believe, struggling a bit, with how do we give IRB approval to something that we can’t know exactly what the interaction is going to be? It’s going to be adaptive and leverage off of the response of the patient.

And this is going to also slow innovation if we can’t do rapid-cycle, randomized evaluations of these products, but these products are not going to be like any other product they’ve ever seen before, because they’re not going to be able to know exactly what’s going to happen when the interaction occurs. Now, we can say we have a safety layer. That study you talked about, that occurred at Beth Israel and Boston, that Google DeepMind did. They did have doctors sort of in the background, but they progressed to not doing it in real time, to do it retrospectively. But we’re going to need guidance. And I think you’re got to be one of the best people to figure out, how do we do this so we can be safe but also encourage innovation?

Lee Schwamm: I think you’re alluding to, you didn’t say it directly, but you’re alluding to what I consider the problem of hyperscale. We can do these things safely with small cohorts of individually consented patients, and you have a doctor kind of listening in every time the patient has an interaction with the LLM. It’s recorded, there are real-time surveillance and triggers. If it says anything that sounds suspicious, the doctor gets notified right away. We can build those, but we can’t test those at scale, those kinds of models, at scale. I do think that we have to understand what does safe and responsible research look like in this domain? I will tell you also, I don’t think the IRB, as it is currently staffed, has the expertise to understand some of the risks with these models. These models are so sophisticated, they change so frequently. I just think we’re in a weird place.

And let me just make a comment that’s sort of analogous to this, that I’ve been saying a lot, as I talk about this.

We think about regulation of medical products in two big categories: devices and drugs. And drugs, we regulate in very heavily. You’ve got to prove you’re effective, not just that you’re safe—very high standards. And I think part of the reason we do that is we recognize, once a drug is approved, it disseminates very rapidly. Once it’s out there in a pharmacy, you can just write a piece of paper or write a little digital thing and boom, lots of patients have it.

Devices, on the other hand, have a much lower regulatory bar. It’s much more about, does it do the thing it says it does? Very few devices get regulated and approved as actually having clinical outcomes associated with them. And we do that in part because we know that devices disseminate very slowly. An approved device, if it’s a catheter, I’ve got to train Howie on how to use it, he can only see a certain number of patients a day. There’s much more opportunity to step in if there’s problems, and there’ll be another device six months later, much more turnover.

Well, we regulate AI like a device, low bar, and we have this sort of low expectations around performance, but it disseminates faster than any drug on the planet. So this is the problem. You build that model, and then you turn it on, and it could affect a hundred thousand patients tomorrow. You couldn’t do that with a device. I was saying, you can’t opt out of AI. We have a program here called eCART, which is built by a company called AgileMD, that is constantly reviewing 75 to a hundred data elements that are found in the electronic health record, over time, to identify patients who are at increased risk of a major clinical deterioration, cardiac arrest, or transfer to the ICU. That is going on continuously. All day, all night, that thing is running, and then it posts a level of elevated risk with a color coded yellow or red.

So the beauty of that is, it’s instantaneously disseminated. The problem with that is, it’s instantaneously disseminated. Now, an elevated risk indicator that’s not actually changing directly the care of the patients is probably a very safe intervention. There are the risk that you might be paying attention to the wrong patients, and the other patients who have other kinds of deterioration are getting less attention. We haven’t seen that signal here, and what we have seen is remarkable reductions in mortality and in serious clinical deterioration.

But that to me is the problem, Harlan. How do we create environments where we can test at micro scale and believe that the results will be valid when they hyperscale and the edge cases emerge, in very large numbers, very quickly.

Howard Forman: As we get to the end, I just want to ask you one quick question, really. Hopefully quick. Ten years ago, or a little more than 10 years ago, Geoffrey Hinton prophesied the disappearance of radiologists. It hasn’t been true at all, and in fact, we probably have a greater shortage now than we ever did. And you mentioned that early on in the podcast, the fact that opportunistic screening, as well as other added-value components of AI in imaging analysis, are absolutely popping up in big ways and have a lot of promise.

Is there an opportunity for our health system or others to be able to more rapidly adopt those possibilities to the benefit of patients, without necessarily having a commercialized product be sold to us? What are your thoughts about that?

Lee Schwamm: I think, what you need for that perfect stew that you want to make for dinner is, you need well-curated data; you need large, secure storage; and you need access to high-performance computing. We built that stack in that computational health platform on top of our EHR, and it is now surrounded by a private Azure cloud that allows us to flexibly expand into additional GPUs, or additional storage as needed, and attach that to the CHP environment. So we have all the ingredients, and I think what we need to do is recruit and train our faculty in how to start to build those models, using frontier models that exist today and fine-tuning them in our data, or actually doing additional kinds of experiments. And I think there’s a lot of opportunity there. It will be the new generation that are natively digital, who have computer backgrounds, engineering backgrounds I think, who go into medicine, who are going to help drive that revolution.

Before we end, Howie, I want to read you... I was recently asked to just comment on what makes an excellent digital patient experience, and I have four sentences. I just want to read them out loud because I really think that it would be nice for us to remember this.

So I’m just going to read them.

So, “an excellent digital patient experience is one that moves complexity away from the patient and makes it easy for patients to access, navigate, and coordinate care across the continuum, without needing to fragment their care across multiple healthcare delivery systems.”

That would be like urgent care, or online this and that.

“It provides seamless longitudinal continuity of identity, so patients are never asked twice for the same information, and their preferences endure between sessions and across experiences. This intelligent hyperpersonalization enables care journeys that are meaningful, effective, and seamlessly intermingle traditional and digital care.”

So it’s not a digital experience. It’s a care experience.

“At a macro level, these experiences are delivered equitably and adjust to the patient’s level of digital engagement, engendering trust and conveying empathy. An equitable digital front door is one that opens for everyone easily, allows access to needed services, and improves clinical care and operational efficiency, rather than simply digitizing existing ineffective or byzantine processes. And lastly, excellent digital patient experiences are not just built but are maintained and curated through continuous measurement, iteration, and alignment to the needs of patients rather than the organizational structure of the health system.”

Harlan Krumholz: And where’d that come from?

Lee Schwamm: I just wrote that.

Howard Forman: Oh, those are good guideposts. Yeah.

Harlan Krumholz: Why don’t you send that to us, and we’ll post it on the notes for the podcast, and we’ll attribute it to you. I think it would be great. I have a final question. By the way, when you meant “next generation,” you meant us, right? The next generation that’s going to come. I just want to be clear.

Lee Schwamm: Howie, sorry. Harlan, sorry.

Harlan Krumholz: Okay. I just—

Lee Schwamm: We’re the old generation.

Harlan Krumholz: ... I just wanted to be clear.

So one of the things that’s most ... Oh, there’s so many things, but one of the things that’s most remarkable about you is, there are many people in jobs and positions like yours who are not walking into work every day with joy. Not sort of feeling the possibilities. How, amidst... you know, there’s a lot of headwinds right now in healthcare, in hospitals, and there’s a lot of Eeyore-ish-ness, but you’re able to maintain a Tigger approach where, positive energy, looking forward, trying to get stuff done. How do you maintain it when people are all talking about retrenchment? Lots of people talking about, are fearful of new policies, what’s coming down. You seem to have a positive outlook on the future. How do you maintain that, in this position within health systems? What lessons are there for the rest of us?

Lee Schwamm: I will confess, one of our sons, when he was a teenager, accused me and my wife, Lisa, of being “relentlessly optimistic.” Which was meant, that was an insult.

I think it’s a worldview, honestly. I mean, I think the world is changing around us incredibly fast. And what I say to all of my team members, and what we talk about at our retreats is, you need to adjust yourself so that change is the expected occurrence, and resiliency and adaptability are, like, that’s your superpower. Because the world is not going to stop changing just because you don’t want change. So I think that’s number one, that’s really important. And that draws on all sorts of personal characteristics. You can learn, you can train yourself to be more adaptable, but some people are just built more adaptable than others.

I just sort of feel like, you know, Descartes this great response to when he was asked, “Do you believe in God?” And so he made a two by two, “Believe in God, Don’t believe in God, God exists, God doesn’t exist.” So if you believe in God and God exists, you go to heaven. If you believe in God and he doesn’t exist, you’re disappointed. If you don’t believe in God, and he doesn’t exist, you get to be smug if you still exist, if there’s any afterlife. But if you don’t believe in God, and he exists, you go to Hell. Based on that calculation, you might as well believe in God, because that’s how the equation plays out. I think you can look at the world and be disappointed, or you can look at the world and think about all the ways in which you can make it better today than you found it yesterday, and that’s what motivates me. And it’s not out of altruism. It just is the kind of world I want to live in.

Howard Forman: That’s a great way to end. And I share your enthusiasm for the future. I do not think it’s a bad future, but I do think we all have to be a part of it, so thank you for doing your part.

Harlan Krumholz: Thank you, Lee. Thanks. Really great to have you.

Lee Schwamm: Pleasure. I’m waiting for my jacket that says “3” on it, or my bathrobe.

Howard Forman: We’re getting you... your fifth time on the show, we get you the fifth time.

Harlan Krumholz: Like Saturday Night Live. We’re going to see whether you can be the host with the most.

Howard Forman: Right. He’s always great. We’re so lucky to have him.

Harlan Krumholz: Hey, that was terrific. Lovely. That was just amazing.

But okay, let’s get to this next part, one of my favorite parts of the podcast. What’s on your mind this week, Howie?

Howard Forman: Yeah. Well, obviously, this is big news. This is changing.

There’s an active and deeply concerning situation unfolding right now on a ship called the MV Hondius, which departed Argentina, and really the southernmost type of Argentina, in early April on a polar voyage through Antarctica and the South Atlantic and has been crossing toward Africa. As of today, as of a few minutes ago, at least eight people, passengers and crew, have confirmed or suspected Andes hantavirus. At least three have died. The situation remains fluid. And we now have the key piece of information that changes this story significantly. Health authorities have confirmed the Andes strain of hantavirus, the one strain known to transmit between people.

So what is hantavirus? It’s a family of RNA viruses spread primarily by rodents, mice, and rats, through their urine, droppings, and saliva. And what makes this so dangerous is that those particles, you don’t even have to touch the rat or the mice, it’s their particles that can become aerosolized. You don’t need to touch the rodent; just disturbing the contaminated dust in an enclosed space can be enough.

How lethal is it? For the Andes strain and related hantaviruses causing pulmonary syndrome, somewhere in the range of 35% to 50% of people who develop respiratory symptoms will die. There is no specific treatment and no vaccine. Many Americans heard about hantavirus just a year ago, when Betsy Arakawa, that’s Gene Hackman’s wife, the famous actor, died of it at their home in Santa Fe, New Mexico. That was a different strain, Sin Nombre virus, the dominant North American type. And crucially, it does not spread person to person. This is where the Andes virus is different, and why this ship outbreak has grabbed the attention of global health authorities.

Andes virus is found primarily in Argentina and Chile, exactly where this voyage originated. It is the only hantavirus known to transmit between people, though even then, only among very close contacts. People sharing a cabin, a household, a bed. World Health Organization officials have said that they believe some human-to-human transmission may be occurring among the closest contacts on the ship. The pattern of illness, starting with a Dutch man who died on April 11th, and his wife fell ill and died in South Africa, and cases then appeared in a British passenger and a Swiss national. These are consistent with a chain of close contact transmission. The rodent exposure during island excursions along the route remain another possibility.

The bottom line, local and global health authorities are doing the epidemiology in real time. They are contact tracing, genotyping, and monitoring everybody on board. WHO has said the risk to the general public remains low, and this is emphatically not a pandemic-level threat, but this is exactly the kind of cluster that warrants the close attention that it is currently getting.

Harlan Krumholz: Howie, this is a hot story, lots going on. I was just trying to piece this together. What’s the usual incubation period for this, once you get the exposure?

Howard Forman: Within rounding error, nine to 40 days.

Harlan Krumholz: Forty? Up to 40, up to 40.

Howard Forman: Forty. And there are some that have gone even more than 40 days, but I’m using the range in one publication I saw. So this is also why we’re concerned right now, because there were 88 passengers, and I think 59 crew on this boat, and most of them are not beyond the incubation period by any stretch. None of them are, in effect. So we’re waiting to see if more people get infected over the ensuing weeks.

Harlan Krumholz: And the treatment for this?

Howard Forman: It’s treating the symptoms. There’s no treatment to the virus, there’s no vaccine, there’s no specific antiviral. And look, the fatality rate, the number that I’m going to give you is 38%. You could be correct if you said 25% or 50%. Thirty-eight percent is the number that’s being thrown around.

Harlan Krumholz: Yeah. The only other thing I was wondering was, is it that maybe a lot of other people have had this and they’ve just attributed it to flu, or something like that. They never show up to medical attention. And so, this thing is not as serious as we think—

Howard Forman: That’s a great question.

Harlan Krumholz: ... because we only see it when it flares into something that kills someone.

Howard Forman: It’s a great question. There is a paper in The New England Journal of Medicine that came out several years ago, in the middle of COVID in fact, that talked about a super-spreader event. And it does seem that when it infects people, you get real symptoms. Some go really bad, like respiratory distress into death very quickly, some resolve on their own, but most people seem to get symptoms. But nobody’s done any of these large-scale screening studies, where you do seroprevalence the way we did during COVID, to try to assess who might have had an asymptomatic infection.

I think this boat, because it’s so self-contained, is going to be a seroprevalence study. I think we’ll get more information on it.

Harlan Krumholz: We’ll learn a lot, unfortunately—

Howard Forman: Yes.

Harlan Krumholz: ... from this.

Howard Forman: Yep. That’s exactly right.

Harlan Krumholz: Well, thanks Howie, thanks for bringing this forward.

You’ve been listening to Health & Veritas with Harlan Krumholz and Howie Forman.

Howard Forman: So how did we do? To give us your feedback or to keep the conversation going, email us at health.veritas@yale.edu, or follow us on any of social media, and in particular, our Instagram account.

Harlan Krumholz: Yeah, and give us some feedback. Let us know how we’re doing. If you put it on any of these platforms, it helps people find us. We appreciate it.

Howard Forman: Health & Veritas is produced with the Yale School of Management and the Yale School of Public Health. To learn about Yale SOM’s MBA for Executives program, visit som.yale.edu/emba. And to learn about the Yale School of Public Health’s Executive Master of Public Health program, visit sph.yale.edu/emph.

Harlan Krumholz: And hat tip to our superstar students, Gloria Beck, who prepared us wonderfully for this interview today, and to Donovan Brown, Tobias Liu, Tobias about to graduate. To our great producer, Miranda Shafer. And Howie, I get to work with the best in the business every week, thank you to you. It’s been great working with you today.

Howard Forman: Yeah, love working with all of you. It’s such a pleasure.

Harlan Krumholz: Yeah. Talk to you soon, Howie.

Howard Forman: Thanks, Harlan. Talk to you soon.