Exploring AI Matters
Our mission is to help the policy community understand the breadth and richness of AI and the potential for such technologies, wisely applied, to augment all sorts of human endeavors.
Some AI tools are able to assist humans in performing tasks faster, more accurately, or more efficiently. Some, however, are inaccurate and unreliable. Who or what we hold accountable for these flaws, and what incentives we do or do not create for their correction will influence AI’s hand in how we work.
In this series we will refine, sharpen, and clarify your understanding of AI.
Exploring AI Matters
Episode 8 - “STAT” AI Helps Stroke Specialists Speed Response
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Strokes are one of the most frightening of maladies, striking as they do at the brain, the center of personhood. In this episode of Exploring AI Matters we talk with the founder and leader of one of the leading stroke centers, Dr Stanley Tuhrim of the Mount Sinai Comprehensive Stroke Center in New York City. Mount Sinai’s Stroke Centers have been leading adopters of powerful AI techniques, the first ever approved by the FDA for clinical application. [2023-02-06]
Welcome to Exploring AI Matters. This podcast series, previously known as Mind the Gap Dialogues on Artificial Intelligence, will continue to appear in the ABA series to the extent that, in addition, all of the episodes, old and new, will now appear under our new podcast name, Exploring AI Matters. Thank you.
SPEAKER_04Stroke has historically been one of the most frightening of maladies, striking as it does at the brain, the center of personhood. A stroke may leave one aware but paralyzed or mute, failing to grant the mercy of mere death. Until comparatively recently, stroke was, like cancer, one of those things that medical science did not understand well and could do little to treat. In the last 40 years, however, the medical community has made tremendous strides in understanding and addressing stroke. Stroke centers have been established in great medical institutions around the world, and the doctors studying and working there have fought back the disease bit by bit. This fight has involved the adoption of modern technology, and the results have been tremendously encouraging. Hello, I am Mark Donner, a computer scientist.
SPEAKER_00And I'm Mama Adams, a national security lawyer. We are your host for this episode of Mind the Gap Dialogues on Artificial Intelligence. In addition, we have two more hosts.
SPEAKER_03Hello, I'm Roland Trump, a national security lawyer, and my colleague Charles Palmer is a computer scientist, but could not make it today for this recording.
SPEAKER_04Each episode will be led by two of us, with the other two adding impromptu questions and comments as the spirit moves them.
SPEAKER_00In this episode of Mind the Gap, we'll be talking with a leader in the study and treatment of stroke, Dr. Stanley Turn, who founded the Mount Sinai Comprehensive Stroke Center in New York City. Mount Sinai Stroke Centers have been leading adopters of powerful AI techniques, the first ever approved by the FDA for clinical application. So, Dr. Term, thank you so much for joining us today. I know Mark, Roland, and I are excited to dive into this informative and what I expect will be a fascinating discussion about strokes and AI. And to start off our discussion, I want to make sure that we and the audience are sort of all on the same page when it comes to having a baseline understanding of what a stroke is, because I think that's critical to understanding this conversation as it evolves. So could you tell us how you explain to your patients what's happening when they are having a stroke or have had a stroke?
SPEAKER_01Well, a stroke occurs when a portion of the brain does not receive sufficient blood flow to exist. And so that portion of the brain dies. It's really that simple. It's what we refer to as ischemia, not enough blood flow to a portion of the brain, typically because there's a blocked blood vessel. There are other types of stroke due to bleeding, but that's not really what we're going to focus on today.
SPEAKER_00Now, in terms of what we're going to focus on today, can you explain to us a little bit more what the impact on a patient is, or sort of an outcome for the patient if they have or experience that type of stroke?
SPEAKER_01Well, the outcomes from stroke are highly variable. People can have from as little in the way of permanent damage as one sees in something called a silent stroke. In other words, we pick it up only because we're imaging the brain and see that they've had a stroke in the past without there necessarily even being some historical uh carlet of the stroke that we see on the imaging. That's at one end of the spectrum. The other end of the spectrum, uh, ischemic strokes are fatal about 10% of the time. In between are really the vast majority of strokes in which people are left with some degree of neurologic impairment, that neurologic impairment being the result of that portion of the brain affected by the stroke no longer functioning. So if it's the part of the brain that is important in language function, then they may have difficulty speaking or comprehending what's being said. They may have trouble with writing, they may have trouble with reading, or all of the above. If the part of the brain that's affected is primarily concerned with motor function, then they may have weakness on the opposite side of the body, what we refer to as the contralateral side of the body. Same thing could apply to sensory deficits or to difficulty with vision, or it could be all of them: difficulty with language, difficulty with motor function, difficulty with sensory function, difficulty with vision. If the part of the brain affected is in the back, and what we often refer to as the posterior circulation, involving the cerebellum or parts of the cerebellar connections to the brain stem, then the problems may be with coordination or balance or the ability to walk. So the effects of a stroke are highly variable, both in terms of the particular dysfunctions that occur and the severity of those dysfunctions.
SPEAKER_00Now, in terms of this variability, you know, I think one thing that we as lay persons always hear is that, you know, when it comes to diagnosing a stroke and getting treatment for a stroke, that time is critically important and of the essence. So understanding as you kind of just walk through the different disruptions and the variability of those disruptions, can you give us some insight into, you know, sort of why that timing factor is so critical to the outcome of some of these disruptions and their long-term impact?
SPEAKER_01Right. So it turns out that the stroke doesn't occur in an instant, but rather over a period of time. And that period of time may be as little as five minutes or as long as a day, and perhaps even longer than that, depending upon how much blood flow remains after whatever vessel it is that's affected, is occluded, blocked. This, by the way, is only true in the past 25 years or so. Prior to that, there really wasn't much that could be done in terms of restoring blood flow. It either happened naturally or it didn't happen. And so the notion that time is of the essence, or as we in the profession say time is brain, is really only about a quarter of a century old and really has come about because of the advent of uh what we call thrombolysis, being able to break up the blood clot that is locking the offending vessel. Since the amount of time that it takes for the tissue that is not receiving enough blood flow to actually die, there is the opportunity for intervention. And the sooner you can intervene, the more likely it is to affect the outcome. There are various ways, and I think we'll talk about them a little bit later on, to determine just how much of the brain is permanently damaged and how much of the brain is potentially salvageable, and how much of the brain really isn't at risk at all in the course of an acute stroke. And that's where we get the notion that time is brain, that uh time is of the essence.
SPEAKER_00And in your work and in your studies around stroke, can you give us a little bit of background in how you came to be interested in AI and its technology and kind of in your field of medicine?
SPEAKER_01Well, this is a long and and winding tale. So I did my stroke fellowship at the University of Maryland in the 1980s. And the University of Maryland 1980s was actually one of the centers for the development of things that were uh referred to as AI. And it so happened that one of the neurologists with whom I trained clinically was actually a computer scientist. And to give you an idea how times have changed, um, he was basically doing neurology at that point, not so much because he was really in love with neurology, he really had um developed a passion for computer science, um, but because that's how he made a living was being a doctor. You couldn't really make any money as a computer scientist. I think that's changed a little bit. Just a little, I think. In the 40 years or so since we got started. But what his interest was was really in developing computerized medical decision-making systems and sort of more broadly developing systems to, if not mimic uh the way in which uh physicians use diagnostic reasoning, at least mimic the results. Um, and this I found fascinating. And so uh, even though the fellowship was really clinical, and my involvement with AI was literally happenstance, uh, because that happened to be what he was interested in. And he wasn't the the primary person with whom I was training. He was a relatively junior person, and as I say, really only peripherally interested in stroke. Nevertheless, that became a great interest of mine.
SPEAKER_00And in terms of this interest that you developed in AI, and you you mentioned, you know, computerized medical decision making and sort of platforms, were there particular elements or facets of AI that caught your attention as being potentially helpful to diagnosis and sort of future outcomes?
SPEAKER_01Uh, you know, I I think Mark, for one, could speak more to the different facets of AI or what came to be known as AI than I could. I really uh was primarily a clinician who had some background in using computers, you know, perhaps more so than than most physicians at the time. But I wasn't an academic uh computer scientist by any means. But really, it was the application of computer programming to these particular kinds of problems that interested me.
SPEAKER_03Can I can I just uh ask, because I think uh some of us in the audience anyway would would wonder, was there some point where you became aware of AI as opposed to other areas of computer science? And in becoming aware of it, you thought this may or may not be useful in the work that you did in stroke therapies?
SPEAKER_01Well, what I'm trying to convey is the notion that I wasn't really so interested in in computers in general. So it wasn't really a question of AI versus something else. Um, it was gee, here's an interesting problem. How can we go about solving it? Um and oh, isn't that interesting? They call it AI.
SPEAKER_04Right. I mean, uh the the this this harkens back to uh Stu Feldman's comment, which is that you know, AI is whatever we don't really know how to do. Um and I I think uh and and and Stan is being a little modest. He's got a number of, he showed me some uh a uh publication with a number of papers that he uh was a co-author of that uh involved applying computers in uh in in actual practice or in uh in developing practice. So uh like I said, uh Stan's pretty pretty modest about this. So so Stan, your center uses an AI system from a company called Viz.ai. Um, could you describe a little bit about how it how it is uh used in practice and and what benefits it brings?
SPEAKER_01Certainly so this is really uh a far cry from what we were doing 40 years ago and really involves image uh analysis. I now have a an in I think sort of an interesting perspective on how it's used because I practice at two different types of medical centers. For many years I was at Mount Sinai uh Medical Center in New York, and that's a very sophisticated place with lots of stroke neurologists and trainees and uh neuroradiologists who are uh expert in analyzing brain imaging and brain vessel imaging. And we should probably take a step back and talk about some fairly recent developments, meaning in the last decade, that are particularly important with regard to the use of VizAI. So, around about 10 years ago, people started being able to put a catheter into the larger blood vessels in the brain or on the surface of the brain, and actually remove the blood clot that was blocking this major artery and causing the stroke. And we refer to that as thrombectum, uh, take literally taking out the thrombus of the clot. In 2015, there were several trials published that demonstrated that this was a very effective way of treating a large stroke with uh potentially devastating consequences, um, and if people in many instances literally walked out of the hospital. What this required, however, was recognizing, first of all, that a stroke was occurring. Secondly, that the stroke was due to a blockage of one of these large arteries, and then thirdly, being able to intervene in a time frame which would allow the salvage of the part of the brain that was at risk. So the technology evolved and proved to be highly effective. The ability to recognize that that was what was going on was really limited to a relatively small group of people: stroke neurologists, neurointerventionalists, neuroradiologists, and not necessarily the frontline people who would first encounter the patient. So these would be patients who would come into an emergency room and typically be seen first of all by an emergency room physician, who would probably recognize that it was a stroke, and then the they would call a consult, typically from a neurologist who might be uh a general neurologist or might be a resident, depending on the locale. Um, and eventually somebody would be involved who was a you know a stroke expert. And then the appropriate studies would get done, and and somebody would look at those studies and say, aha, there's a large vessel occlusion that might be uh remedied by one of these clot and mool devices. Time. And in a place like Mount Sinai, there is the expertise that exists, so that even if it's not handled in the most efficient way imaginable, the expertise eventually is brought to bear on the problem. In 98% of the country, however, not to mention the rest of the world, that expertise really isn't available, and certainly not in the time frame that would be necessary. So an automated way of recognizing that that's what the problem is would obviously be highly advantageous. Now, as I mentioned, I practice really in two locations, Mount Sinai being one of them, and that's where this system was first put into use. But I also practice at Berkshire Medical Center, which is a very high-quality community hospital that essentially services all of Berkshire County in western Massachusetts and some of the surrounding area as well. Um, nevertheless, it's a small hospital, and there isn't necessarily that level of expertise because it's a highly verified kind of problem. Um, and there isn't the demand just in terms of the sheer volume of patients to support that level of expertise being uh present. So if you could accomplish pretty much the same thing with a technological solution, you'd be way ahead of the game. And so the the way in which the Viz AI software was uh first employed at Mount Sinai is one thing. The way in which a similar system is employed at Berkshire Medical Center and really throughout the country is something a little bit different. So to talk first about uh Viz AI and what it does, and then how it was employed at Mount Sinai, and then we can talk a little bit about how uh more general use might be very advantageous. So a patient, and this was a this is really a um a protocol that was developed over a period of years as we rolled out both clinical processes to efficiently manage an acute stroke, and in terms of the software that was developed to aid in that in that process. So patient comes into the hospital, it's thought to be a stroke, and that can is not necessarily a difficult thing to recognize. It is something that is often recognized by the triage nurse who first encounters the patient. And if not the nurse, then the ER doc. And either way, a stroke code is called, much akin to any other kind of emergency uh process in a hospital, say cardiac arrest, and the patient is immediately sent for imaging, literally before there's even a detailed neurologic exam. At the same time, the stroke team is alerted. The first step is a CT scan, the the with which I think most people are familiar, and that gives us imaging of the brain. That CT scan in most acute strokes is normal. You don't see a stroke, you don't see anything that would tell you that there's an abnormality. And that's the good thing because that means that there's no bleeding, there's nothing that's mimicking a stroke, it's probably a stroke. At the same time, contrast is then administered. And here protocols vary. It may be that if the patient is clinically a candidate for intravenous thrombolysis, the intravenous thrombolytic is administered even before the additional imaging is obtained, or in some instances, since the additional imaging takes a very short period of time. That is also administered before the medication is administered. In any event, the next imaging step is uh what we call a CT angiograph. So there's intravenous contrast material injected, and the blood vessels leading up to the brain and in the brain itself are imaged with the same computer technology, the same uh CT scanner. And now the sort of more recent uh development is an additional uh imaging technique called CT perfusion, in which the amount of blood and the rapidity with which the blood gets to the various parts of the brain can be measured. What Viz AI does is analyze those images uh in an automated fashion. So this will say, aha, there is not enough blood getting to a certain part of the brain. That's what you see in a stroke. And furthermore, that there is an occlusion blockage of one of the vessels which are amenable to intervention. And this can is done in a completely automated fashion. The the technician simply presses literally presses a button and all of this rolls out. Furthermore, the software then says, aha, I've identified a large vessel occlusion. That's what these things are referred to as, and sounds the alarm to whoever is holding the appropriate software, and uh that literally is on a phone who works, you'll be happy to know, on both Apple and non Apple products. Um and it really is a very loud alarm that. Will waken you from sleep for sure. And the advantage of that is that it happens almost instantaneously. So it whereas before you would wait for the images that are on the scanner console to be sent over to the uh where the radiologist can uh read the images, and the that radiologist may or may not be a particularly expert one, depending on the time of day or night. And then that radiologist recognizes the large vessel occlusion and the fact that there is a stroke occurring, and then contacts the person in the ER that ordered the study, who then contacts the stroke neurologist, who then contacts the interventionalist, who then starts to get dressed. Needless to say, this takes a certain amount of time. Instead, what happens with Viz AI is the interventionalist has his or her phone by the bedside or in the OR or wherever they happen to be, and this alarm goes off. Now, is is Viz right 100% of the time? No. But you can set the sensitivity and specificity for detecting these large vessel occlusions such that you can miss virtually none of them and get alerted to a certain number of false alarms, or miss more of them and not have any false alarms. And not surprisingly, we opt for a high degree of sensitivity. And if it so happens that somebody is interrupted from whatever it is that they're doing the false alarm, that's a relatively small price to pay. So what we're really talking about is tremendous efficiency, even though all of the expertise is readily available in a place like Mount Sinon. In a place like Berkshire Medical Center, where that level of expertise just doesn't exist, this kind of software performs an even more important function, which is literally to alert the people that are available to take care of the patient that this may be the problem. And then what happens is something that happens in a large part of the country now, which is that some central facility, in the case of Berkshire Medical Center, it's Mass General, Massachusetts General Hospital, but there are companies that do this literally nationwide. There are academic centers that do it. Mount Sinai has a system for doing it within its own health system, which now involves eight hospitals. Then there are commercial companies that may do it for 40 hospitals. So this, the expertise then to look at these scans and decide what should be done can now reach out to small community hospitals that may be geographically distant, you know, quite remote from where these so-called experts are located. And then there's a process for either administering thrombolytic treatment uh intravenously locally, or getting the patient to a place where the kind of intervention can be implemented, perhaps by helicopter, which is what we do here in Virtue Medical Center. But it all starts with the recognition of what's going on with the imaging.
SPEAKER_03So how much of how much time have have you been able to compress or save by this procedure? Because you said time is of the essence. The way you described it initially, it sounded like it could take hours before somebody could make a decision to intervene. It sounds much shorter in what you're describing, but I'm not sure by how much and what how much of a difference that makes to the success of the treatment.
SPEAKER_01All of that is highly variable. I suspect there may be statistics about that in a particular, you know, from a particular location. I can tell you at a place like Mount Sinai, it makes a difference of, you know, perhaps an hour in terms of when that patient actually gets onto the angiogram table and gets uh that intervention. In a place that's more remote, it's not even a question of saving time. It's simply making possible something that was previously impossible. Absent this sort of technology and a telemedicine arrangement, remote places, you know, whether it's in the middle of uh Montana or on the outskirts of Cleveland or in western Massachusetts, simply couldn't get a patient to an appropriate location where they could have this kind of intervention in the time frame that would allow it to be effective.
SPEAKER_04So, what I think I heard you say is that however important this AI technology was and has been to Mount Sinai, it's even more important to the community medical centers that represent the vast bulk of the country because the recognition and the you know the strategy of getting to an interventionalist was not even feasible before that without this kind of recognition process.
SPEAKER_01That's exactly correct. Now we've skipped over a couple of processes in the whole developmental sequence. With those initial studies that were published in 2015, basically the structure of the clinical trial was anybody that was thought clinically to have a large vessel occlusion was potentially eligible. And they often received intravenous thrombolysis, but not necessarily because intravenous thrombolysis is approved for administration only within three hours of the onset of stroke. It's actually commonly used up to four and a half hours from the onset of the symptoms of stroke, but a relatively narrow uh time period. These trials, the early trials that were successful in thrombectomy, keeping in mind that that's only a subset of the patients that present with acute stroke. The initial trials were patients were eligible mostly within six hours of the onset of symptoms, um, in one case up to eight hours from the onset of symptoms, but still a relatively short period of time. More recently, as a result of the use of similar kinds of imaging processing and the use of CT perfusion, which I alluded to earlier, that time window has been expanded literally up to 24 hours. Now, a lot of that is not time from after the patient presents to the hospital. It may be the time from when the symptoms began when the patient gets to the hospital, because everybody recognizes a stroke as occurring instantaneously, and certainly not everybody acts upon it immediately. Right. And so often we don't see patients within the first few hours even of the onset of their symptoms. We may see them the next morning. Right. And indeed, if the stroke occurs while they're asleep, they don't recognize that there's anything wrong until they awaken. So those more recent studies that allow for at least some percentage of patients with the appropriate kind of picture on the CT perfusion to be treated really expanded the window dramatically for at least those that subset of patients.
SPEAKER_04So so this uh these techniques are now approved by the FDA for integration into clinical processes. How long did it take from FDA approval to actual integration in terms of real day-to-day operations?
SPEAKER_01The short answer is decades, but let me take a step back and say that actually the FDA doesn't approve the procedure. The FDA approves the device. Um, and that's a whole nother set of issues. So the FDA approves medications and the FDA approves devices. In general, the FDA requires that a medication or a device is effective for some treatment. And then it's up to the clinician to employ that approved device or that approved medication as he or she sees fit. In terms of intravenous thrombolysis, that was around for a very long time, I think dating back to the 1950s, and was in fact not proven to be effective until the pivotal trial in the mid-1990s. For the devices, the early devices just weren't that good. They maybe restored blood flow in half of the arteries that were blocked. These are the thrombectomy. Thrombectomy devices, yes. The initial devices were literally a corkscrew that you put into the clot and tried to pull the the cork out. And some, you know, it's enough to go with a wine bottle. Um, it turns out that it just wasn't the best device possible. Uh it was at the time. And indeed, and other intra arterial approaches, including putting that same sort of clot-busting medicine directly into the clot, were tried, and they also didn't didn't work, or at least weren't proven to work until, as I mentioned, 2015, they were around probably for 20 years or so. So part of that is the the process of getting approval, and part of it is that the earlier devices simply weren't good enough. Uh, they weren't effective enough to actually produce the reduct desired results in a sufficient number of patients.
SPEAKER_03Some of us tend to draw analogies to things that we have a little more familiarity with, but not much knowledge actually. And as you were describing that, I was wondering how is that procedure, uh, the profusion procedure, different from like an angioplasty, which is trying to clear a blockage in in the heart? And it sounds like they're quite different. Um, and I'm probably wrong to draw the analogy, but having thought of it, I just thought I'd ask you.
SPEAKER_01Well, you you're not so wrong, actually. Well, look, but let me clarify: the CT perfusion is the is the diagnostic imaging test to determine whether there's a part of the brain that is under-perfused but not so underperfused as to be irreparably damaged. And we often refer to that as the ischemic penumbra. That is the tissue at risk and potentially salvageable, and that is the target for the intervention. In terms of the procedure, angioplasty is sometimes actually part of the reopening of the blood vessel. Angioplasty just refers to the process whereby the narrowed artery is made wider. Um, and that can be done in a number of different ways. It's often done in the heart with the deployment of a stent, and it's often done in blood vessels bringing blood to the brain, but only in conjunction with the actual removal of the thrombus, which is what really we've been talking about now. So sometimes that thrombus is the result of a blood clot that formed elsewhere, typically in the heart, and then embolized, that is, broke off from that location where it was formed, and travel downstream and block an artery, a smaller artery in the brain. And then the removal of that clot is what we refer to as the thrombectomy. And so the vessel that it's blocking may actually be relatively normal. It's just too small to let the clot pass through it. Sometimes the clot forms in situ. The problem is atherosclerosis, which we all have to some degree, and the clot may form the novo right in uh that location. And when that happens, the parent vessel may be problematic. And so once the clot is removed, there may be a high degree of narrowing still present, and that's when an angioplasty may be performed. In other instances, there may be what we call tandem lesions. So a large blood vessel, maybe in the neck, even not yet even in the brain, may be blocked, and then a blood clot forms there, and then a piece of that blood clot actually goes further into the into the brain and blocks a smaller blood vessel. And so the first step is to open up the larger blood vessels so that the catheter can then be passed to the smaller blood vessel and a thrombectomy performed there. And in the process, that larger blood vessel may have to be stented to keep it open. So there are all variations on a theme, but it's not so different from what goes on in the heart in certain circumstances. And in fact, as I was mentioned earlier, the expertise isn't always available the way it is at Mount Sonar. So if you go to places that are much less densely populated, there may not be the volume to support a neurointerventionalist, per se. And in fact, it's the cardiac uh interventionalists, the interventional cardiologist that actually performs the thrombectomy in the brain as well. Those of us in the Ivory Tower say, well, you know, that's that's clearly not what we would like to see. But in fact, um it's far better than not doing it. And uh those there are those that would argue it's just as good. Um, and that's sort of one whole area of uh research that's uh that's being conducted now to see uh what sort of training and and who is best able to do this, and and what is the trade-off between having a somewhat less expert person doing it versus the additional time that it would take to get the patient to the sort of ideal person to do it.
SPEAKER_03Well, can I take that just one step further? Is this leading to some change in what medical schools are doing or what residencies are doing to train cardiologists to be capable of doing that kind of intervention?
SPEAKER_01Well, certainly not medical schools. This is not something that it goes on at the level of a medical student. This is something that goes on after medical school residency fellowship. Okay. And I don't know that interventional cardiology fellowships are training people in doing cerebral thrombectomies. I don't think that that has happened as yet. But people can do some additional training, um, and it's not necessarily formal training, it may be hanging out in the uh cat lab with somebody that's doing it for some period of time. So there's sort of many fellowships, and that sort of thing is going on. I I don't know how frequently that occurs, but clearly there are a lot more arteries that need to be cleared than there are people capable of clearing them.
SPEAKER_04So, now how widespread is the adoption of the AI technology we've been talking about in the overall stroke treatment community?
SPEAKER_01Yeah, I I don't know the specific statistics. I can tell you that it's a lot more common than it used to be. And I think Berkshire Medical Center is a good example of that because all of the academic medical centers have adopted, whether it's Viz AI or its main competitor. I I would guess that virtually all of the uh sort of major academic centers have this at this point. It's the smaller places that are now using it the way that I described, and the way that we use it here in the Berkshires, as an adjunct to what they're currently doing, with the idea that then they can uh partner with a place that has the thrombectomy technology available if they can arrange to get the patient to that location quickly enough. And since now, with the advent of uh CT perfusion-based criteria for the extended window, that becomes possible, at least for that small fraction of uh stroke patients who have the appropriate perfusion mismatch, that it's it's feasible literally anywhere in the country to get somebody there quickly enough.
SPEAKER_04Okay, that's fascinating. Before we go on, is this is this sort of stuff involved in teaching stroke docs at the primary levels, or is it uh sort of in practice that one uh learns the latest stuff here?
SPEAKER_01Well, it's it's part of a stroke fellowship for sure. And in fact, it's part of a neurology residency, and I would presume part of a neurosurgery residency, or you know, for people that are interested in going on to interventional, uh to learning interventional tech techniques. Literally every patient with a presumed stroke or a possible stroke at either Mount Sinai or Berkshire. So, in other words, either in the academic ivory tower type place or in the community hospital will get uh these kinds of images that will be obtained in those settings, and it's becoming increasingly widespread. So that I mean, I don't mean to say that every community hospital uses uh CT perfusion, but it's becoming increasingly frequent, and I would guess within the next few years it will be a very high degree of penetrance. You can use something very similar with an MR machine, it's just that there are fewer MR machines, and so that the technology that's used primarily is CT. And so basically, every radiology residency will train uh radiology residents in how to interpret these images. And neurologists, there are of course fewer neurology residencies, but they will also be trained in how to interpret these images. It's actually not that difficult to do, especially since it's automated. And the automation is sort of the gold standard, if you will, because the studies used this software in an automated fashion. And so if I disagree with the mismatch between the territory that's embarked in the territory that's at risk, if I disagree with the machine, I'm not really sure who you should believe, since the the studies were done with the machine's interpretation, not mine.
SPEAKER_04Fascinating.
SPEAKER_01Wow. I'm saying that a little bit tongue-in-cheek, because I'd I'd like to believe that I'm more accurate than the machine when we disagree, but your comment proven as yet.
SPEAKER_03Your your comments about the adoption of the technology, was that speaking of the practice of medicine in the United States or more globally?
SPEAKER_01I I'm really speaking now about the United States. I can tell you that uh Western Europe is very similar and perhaps not as enamored of technology in in certain parts as others, but the the Viz AI company is actually an Israeli company. So this technology is is uh you know readily available there. And uh Western Europe is is using that as well. I I don't know that it's been quite as fully adopted yet there. In the rest of the world, not so much. You know, these are very sophisticated kind of temperamental machines that need a fair amount of upkeep. And also the ability to do thrombectomy is not as rampant, certainly in less uh westernized countries. And there's no point in doing the imaging if you can't do the intervention. Right.
SPEAKER_00Dr. Term, one thing that we've learned as we've gone through this podcast series and had an opportunity to talk to different specialists, um, computer scientists and the like is that AI is continuing to evolve and change, right? And there's no sort of one definition of what it is or what it could be. And I think it would be really interesting to know how doctors are keeping on top of AI innovations as that continues to expand.
SPEAKER_01Yeah, you know, I don't know that doctors are looking specifically at AI innovations as a category. I think that in general, the notion of uh lifelong learning is something that's been getting more and more attention in uh medical fields and in uh the way in which you maintain certification. And I think that certainly the practical applications are something that are part of the continuing education process. I don't think doctors are particularly interested in AI per se. I can tell you that way back when I was at the University of Maryland and and we were Wanting very much to uh test what we had developed in real life situations, we tried to bribe the faculty uh to participate in a study. So we literally gave them all desktop machines. We said, here, this is yours, you can use it. And that was at a time when not everybody had a machine, you know, a computer on their desktop. We said, here, this is yours. You you can do it, do it with it what you will. All you have to do is um use it for our, you know, it came with the software that we had developed, use it on a few cases. That's that's all we want you to do. And we want you, we actually had had built software to help them develop their own rudimentary uh decision-making systems. We we just want you to to to try to do this, use the software to develop a decision-making system. You don't have to learn how to program or anything, you know, quite so complicated as all of that. This is English language, it's virtually natural language. You just have to put in a comma here and you know, maybe a semicolon, but nothing more complicated than that. And this is what you get. This is a $5,000 machine. That's what they cost in those days. Yours for the taking, just spend a few hours. They wouldn't do it. They said yes, but then they didn't do it. And the whole we eventually, this was am I allowed to say it was an IBM project? Charles is not here, sorry. You you can edit it now, it doesn't mean and they they gave us these uh these AI workstations, which also could be used just as ordinary uh personal computers, and that it was too high a price to pay. We took the machines back from them. So, you know, and I don't know that that much has changed. The doctors are not particularly enthusiastic about using computers in general. They may, you know, use Game Boys, but I'm not sure that they can. But if an application happens to be something that's labeled AI, that's fine, you know. But in general, I don't think there's an approach to we need to keep up with AI, the development of AI technology.
SPEAKER_00Yeah, it may be it may be more situation whereas AI continues to evolve and become of our everyday lives and application, right? There might be more of an ease of the intersection between what doctors are doing in AI as opposed to doctors proactively going out and seeking it.
SPEAKER_01One area in which I think uh there has been a fair amount of adoption of sophisticated computer usage is actually doing database analyses. So it's really statistical analysis, machine learning, which I I guess Mark, maybe you can it would do we consider machine learning to be a branch of AI? Absolutely. All right. So if if we want to include that, then um, and its application to developing systems based on a machine analyzing a database and learning from it on its own without a human doing a whole heck of a lot. Certainly, researchers have have bombed onto that in a big way.
SPEAKER_00One thing that strikes me is we spent a lot of time sort of talking about the benefits of AI, some of the impacts that it's had that that have been positive, um, particularly in sort of the efficiencies of diagnoses. I'm wondering if there are any limitations that you've seen or or that you foresee with respect to how AI can be used either in the context of stroke diagnoses or sort of the medical context more broadly.
SPEAKER_01There are two issues that I've thought a bit about. One of these is an old issue and was readily apparent even when we were first developing these systems. And that's that there seems to be a human tendency to believe what a computer says, to the point of if it's if the computer computers are infallible, therefore, if the computer says it's true, it must be true. If the computer tells me to do this, even though it doesn't make any sense to me, I'm gonna do it. And we try to address that in in some of our earlier work. And I don't know to the extent to which that that that's a a problem now, but I suspect that it will continue always to be a problem where we have this, to some extent, undeserved reliance on the accuracy of the computer, right? I mean, who actually totals up their bill in the restaurant, right? We used to do that all the time. We did they do the arithmetic right? Well, now that it's now that it just comes to you printed out and it was on the computer, it must be right. Who checks? Yeah. I mean, in fact, you're considered to be kind of, you know, a country pumpkin if you do, uh, or worse. Anyway, the other area that I'm a little bit concerned about is the inability of the computer systems to tell you how it is that they got to the conclusion that they got to. So there's really no way to quote, check your work, right? You they you can't they can't tell you how it is that they said that this is what the diagnosis is. So you it it's it's it's kind of a double-edged sword, right? You you trust them because they're a computer. On the other hand, if you don't trust them, there's no way to figure out if what they're thinking makes sense. And what one other interest of mine over the years has been chess. And of course, at this point, the computer is a better chess player than any human, but they don't go about it in the same way. So are they really playing chess? Uh, you know, in some sense, it doesn't matter. If if the whole point is to win the game, uh make the correct diagnosis, treat the patient as effectively as possible. Maybe it doesn't really matter if you don't exactly understand what's going on. But that's a little scary to me if you if you don't understand. And of course, they make science fiction movies about this sort of thing, computers gone wild. But it it is a little bit worrisome if you if you can't understand why the computer is telling you to do so.
SPEAKER_03It almost sounds like there's a fourth area that you're implying that we ought to be concerned about, and that is humans when they're working on something, even when they're fairly far along in solving a problem, they have a tendency to ask themselves from time to time, did I make a mistake? Or they may suddenly realize I didn't make a connection or I left something out. I don't know that AI is recursive in that way. Does that worry you?
SPEAKER_01Sure. Anytime you're reliant upon someone or something that you don't really understand, it becomes a leap of faith rather than of comprehension. And that's worrisome.
SPEAKER_00So the the last question we like to ask all of our guests on the podcast is in an ideal world, what would you like to see AI do in the future? Right? In sort of the work that you do or the medical field more broadly, you know, what what would you like to see it be used to do?
SPEAKER_01Well, I I think the sort of the the the broadest answer is to accomplish the sort of analysis or furtherance of knowledge that can't be done by humans alone. And I think um machine learning is an example of of areas in which uh technology which had has been applied to areas in medicine that um have resulted in insights that could not have been obtained by human thought alone, in part because of the sheer volume of analysis that's required, and in part because, and this is the part that's both intriguing and at least in my very limited understanding of such things, a little scary in that um the machine is seeing connections that humans don't. You know, where does that all lead? Who knows? But uh it's potentially very exciting.
SPEAKER_00Agree, potentially very exciting, and I think that's an excellent note to wrap this up. So I want to thank you, Dr. Term, very much for the generosity of your time and sharing kind of the illuminating thoughts that you have about AI and how it's being used in your field. So thank you very much for joining us.
SPEAKER_01My pleasure.
SPEAKER_04We thank the business law section of the American Bar Association for their generous sponsorship of the production of this podcast. We welcome questions and comments from listeners. Send email to comments at mindthegapdialogues.com. We read all comments and questions and will try to respond in the letters section of a future episode. If you are writing about a particular episode, please do mention the specific episode number. Please also do include pronunciation tips to help us properly say your name when we reply in a subsequent episode. See you next time on Mind the Gap Dialogues on AI.
SPEAKER_05Thank you for listening to the AVA Business Law Sections Podcast Series to the extent that the section offers a robust collection of content. To explore more about this topic, or to learn about joining the section, visit ambar.org slash bizlaw. That's B-I-Z L A W.