#50 - AI Caught The Heart Failure Nobody Saw

What if a five-minute EKG could reveal more than a rhythm problem or heart attack? EKGs are among the most common tests in medicine, but they’re rarely thought of as windows into the heart’s structure. That assumption changes with a remarkable case: a 45-year-old arrives in the ER with cough and trouble breathing, improves with treatment, and seems ready to go home. But an AI model reading the EKG detects something unusual—triggering a deeper workup that uncovers a dangerously weakened heart and ultimately leads to a heart transplant.

We break down the medicine in plain language, from what the spikes and waves on an EKG actually mean to what an echocardiogram can show that an EKG usually cannot. Along the way, we explore why structural heart disease can be so difficult to catch early, especially when symptoms don’t follow the classic heart failure script.

Then we turn to the technology behind the alert. EchoNext is trained on massive paired datasets of EKGs and echocardiograms, allowing convolutional neural networks to detect subtle patterns across multiple leads that human eyes might miss. But the promise of clinical AI comes with real-world challenges: how much interpretability clinicians need, what tools like saliency maps actually explain, and how false positives can strain healthcare systems through extra scans, staffing needs, and follow-up care.

For anyone interested in AI in healthcare, cardiology, patient safety, or what it really takes to deploy medical AI responsibly, this episode connects the math, the medicine, and the messy reality in between.

References:

A case of artificial intelligence-enhanced diagnostics leading to heart transplantation
Hartman et al.
Nature Medicine (2026)

Detecting structural heart disease from electrocardiograms using AI
Poterucha et al.
Nature (2026)

Deep Learning Electrocardiographic Analysis for Detection of Left-Sided Valvular Heart Disease
Poterucha et al.
JACC (2022)

rECHOmmend: An ECG-Based Machine Learning Approach for Identifying Patients at Increased Risk of Undiagnosed Structural Heart Disease Detectable by Echocardiography
Ulloa-Cerna et al.
Circulation (2022)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

Transcript

A Routine Test Saves A Life

SPEAKER_02 0:00

A routine test, an unexpected AI alert, and a clinical journey that ended in a heart transplant.

SPEAKER_00 0:17

Hello and welcome back to Code and Cure, where we discuss decoding health in the age of AI. Um, my name is Vasant Sarthi. I'm a cognitive scientist and AI researcher, and I'm with Laura Hagopian.

SPEAKER_02 0:28

I'm an emergency medicine physician.

SPEAKER_00 0:30

This is episode 50, Laura.

SPEAKER_02 0:32

I know, it's crazy. We're like almost at a year.

SPEAKER_00 0:35

Yeah, and I'm so thankful to all our supporters and all our fans and all the love you guys have shown. And I'm very excited for our next almost next year of stuff. So um today we're gonna be talking about matters of the heart.

SPEAKER_02 0:50

Oh, stop. Like that is just such a dad joke. I told you not to make it.

SPEAKER_00 0:53

I know, I couldn't help myself.

SPEAKER_02 0:56

Uh, but uh this is actually a really interesting case study of an application where AI literally saved someone's life. Literally, right?

SPEAKER_01 1:05

Yeah.

SPEAKER_02 1:06

Um, and so we'll walk through the case of the patient, but then we can talk a little bit about kind of the theory behind it and um how this is potentially being applied more broadly

The AI Alert And What Followed

SPEAKER_02 1:17

too.

SPEAKER_00 1:17

Sounds good. Yeah, let's do it.

SPEAKER_02 1:19

Okay, cool. So this has been all over the news. Um, it was in the New York Times. Uh it was there's articles published by American Medical Association. Uh, there's a correspondence article in Nature Medicine that details exactly what happened. But basically, um a 45-year-old man went into the emergency department because he had some trouble breathing and had a cough. And um he he basically was noted to have some tobacco use and um and maybe an exposure to some wildfire smoke. And they they did an EKG on him, which is a test for his heart. Um, and they did some other tests like a chest x-ray, and they gave him some medication for asthma, some breathing treatments, and he seemed to improve, and so they sent him home. Seems pretty routine, right? Like you have trouble breathing, you're exposed to some smoke, let's get you breathing better. Okay. Well, what happened was that EKG that was done on him went through this trial that was being done at this hospital where they were trying out an AI model to detect structural heart disease from EKGs alone. Normally, structural heart disease is detected on echocardiograms. And we'll get into we'll get into L'Aside. We'll get into that, like an L just out of the heart. But the EKG was what was done. It's very quick, it's very common for patients to get this. And so what happened was this AI model flagged his EKG and said, hey, something is off. We're worried about this EKG. Let's, and so they decided, let's get this patient in to see a better view of the heart with an echocardiogram with an ultrasound. And let's see what's going on. And so what happened with this patient was he was found to have a very weak heart. One of his valves wasn't working well, his heart wasn't pumping very well, and so his heart was basically in failure. Wow. Yeah.

SPEAKER_00 3:17

Wow.

SPEAKER_02 3:18

And it turned out that he had a genetic mutation, which was like basically causing his heart to dilate and not pump well. And the thing about this genetic mutation in particular is often it doesn't really even present with symptoms until you like die. Wow. Yeah. So it is it is a huge risk factor for sudden cardiac death. It's not something you're just like testing the whole population for. So it's like one of those needles in a haystack. Right. And the AI model found it. That's incredible. Yeah. And so what happened was that they tried to do like a regular regimen of medical therapy for him. It didn't work. He got worse. They had to put him on like a heart pump. And eventually he had a heart transplant and he's now home. Um, and this is all because the AI detected something on his EKG and flagged it as part of this trial that was being done.

SPEAKER_00 4:20

That's incredible. Yeah. No, this story is absolutely incredible. And, you know, we're gonna dive deep into all of these different pieces here, but you know, at the very high level, it's it's an example of a success story, right?

SPEAKER_02 4:33

Absolutely. It's like one of those like safety nets or needles in a haystack that the AI was able to find. It was trained on all this data. He happened to be at the right hospital where the trial was happening, and so his EKG happened to get flagged and found something that he could have died from.

SPEAKER_00 4:51

It's kind of the opposite of the human in the loop situation. We have an AI in the loop here that is fine, you know, looking for gaps and misses that we might we might otherwise miss. And so that's fantastic.

SPEAKER_02 5:04

Yeah, exactly. And so then like humans can review it and be like, who who should move on to to go ahead and get further testing so that we can do something about it if there's something truly wrong, like there was for this gentleman.

SPEAKER_00 5:17

Yeah, yeah, yeah, yeah. So so let's talk

What An EKG Actually Measures

SPEAKER_00 5:20

a little bit about a whole bunch of different concepts that we just threw out there, right?

SPEAKER_02 5:24

Like all the medical terms that I was like casually.

SPEAKER_00 5:26

Let's let's just lay them out, right? So for instance, we mentioned a person came in with something wrong with something with the heart and breathing, whatever, and then an ECG or an EKG was taken. So, first of all, you know, what is an EKG or ECG? And you know what? Let me before that. I think we all know what an EKG is in the sense that we've seen it in movies, we've seen it in maybe we've been to the hospital ourselves. We know that it's a little chart that has kind of like pointy, you know, has a has a line, has a waveform that is kind of pointy. And this is a great description. Keep going. No, no, no. I I mean, frankly, to me as a layman, it seems I I always am always amazed at doctors making sense of what seems like just spikes on a waveform. And and let's just so let's start there. Let's talk about EKGs and ECGs. Are they the same thing, first of all?

SPEAKER_02 6:12

Yeah, they're the same thing. It's an electrocardiogram. Why is it like an EKG then?

SPEAKER_00 6:16

Okay, where's the K from?

SPEAKER_02 6:18

Cardio with a K? I don't I actually don't know the answer to that. Uh, I guess ECG makes more sense, but EKG just rolls off my tongue better. So that's what I say. That's fair. Um but EKGs are really common. It's a quick test. It's done in like five minutes or less. It's simple, it's painless. Um, and what it does is it can check a lot of things about the electrical signal in your heart, right? Which is what's squeezing and pumping your blood. And so it you can check, you know, how fast someone's heart is beating, whether their heart rhythm is normal or irregular. Um, it can look at the exact electrical signals, which is what you were talking about, all the little waves and spikes and all of that. And so anytime people come in with like chest pain or trouble breathing or palpitations or dizziness or something, we'll look at these things to be like, hey, is there an irregular heartbeat? Is there a heart attack happening? Is there something else going on? Um, and and because it's quick and simple and easy, and and you can do it almost anywhere. Like you mentioned doing it in the hospital, but like you can do it in a primary care clinic. These machines are are relatively inexpensive. It's an easy test to get. You you basically like lie down somewhere and they pay place electrodes on you, on your chest, on your arms, on your legs. They connect it to the machine, you lie still, and it takes uh a reading over the course of a few seconds. And the whole thing takes like five minutes to do or less. Right. So it's easy. There's no preparation, really. You just you just get it done. And what it shows, uh, and so what it ends up showing, you were talking about the sort of waveform of it, is it shows this printout, these waves, these lines that show the electric electrical activity going on in your heart for each heartbeat. And so the first, I mean, I can go, do you want detail? Do you want detail?

SPEAKER_00 8:08

I want a little bit more detail, I think. Yeah.

SPEAKER_02 8:10

Yeah. So the waves that it shows correspond to like what your heart is doing at that moment. So we we call it a PQRST. So that I don't know why we start with that and not like ABCDE. There's probably a reason, but the P wave is the first like little bump, and that shows the electrical signal going through the atria, which are the upper chambers of the heart. Okay. And so when they squeeze, that's when you would see that P wave. Okay. Then the QRS complex, so that big spike that you were talking about, is that like tall? Like it's the really noticeable part. When you're looking at the EKG, that's that's the part that you're like, what's that spike? That's the spike.

SPEAKER_00 8:52

Is this are these two things the the two things that sort of make the noise? Like they, you know, you know what I'm talking about when you talk when you have a heartbeat. I wonder if that's that's what they correlate to. But um, but you know, regardless, I think you're right. I think you have the small sort of small bump followed by uh the more familiar big spike, and and that's a it's called the QRS complex because it has QRS, all three pieces to it, like three separate pieces to it.

SPEAKER_02 9:20

Yeah, but they're all sort of together. And the the heartbeat is actually like the valves of the heart snapping shut. It's not exactly the same thing, but the but the the blood is moving through those. So the QRS complex is more the electricity that's happening in the heart. Got it. Whereas the sounds that you hear in the stethoscope are the valves, um, which connect the atria and the ventricles, so the upper and the lower chambers. So the QRS complex um is the electrical signal moving through the lower chambers of the heart, the ventricles. Understood. And that's when they pump blood out to your lungs and the rest of your body. Okay. And it's usually sharp and narrow. It's usually that nice spike that you see. Sometimes it's not, and there's an abnormality with it.

SPEAKER_01 10:05

Okay.

SPEAKER_02 10:05

Right? And then the final part is the T wave, and that comes when everything sort of resets and recharges in the electrical system. Got it.

SPEAKER_00 10:14

And then repeats that cycle.

SPEAKER_02 10:16

Over and over and again. And then there's intervals. So like you see a wave, and then in between, like for example, in between the P wave when the atria contract and that QRS complex when the ventricles contract, there's a little interval. There's usually a flat, like a line there. That's in that example I just gave, that's the PR interval. How long does it take for the electrical signal to travel from the atria, the upper chambers, to the ventricle, the lower chambers? Because if the time is too long, that means something. Okay. And so there are segments throughout. There's like a segment between the S and the T. Um, that's often where we might see like a heart attack if it's higher than it should be, if it's raised up. And then there's an interval between the Q and the T, which is, you know, how long it takes for the ventricle to squeeze and then recharge. And if it's taking too long to do that whole thing, then you're prone to arrhythmias. Got it. Okay. So there's a lot we look at. I mean, there's a lot we look at. I'm going into a lot of detail here, but like it's it's looking at what's the electrical system of the heart doing. And there are certain flags that we look for with disease processes, like, is it beating too fast or too slow? Um, are do the are the P waves even there? Sometimes they're missing, right? Is it is it irregular? Uh, you know, are there problems with some of the segments where we think someone might be prone to an arrhythmia? Or are there problems with some of the segments where we think someone's having a heart attack?

SPEAKER_01 11:44

Right.

SPEAKER_02 11:44

Um, is your EKG different from your last one? So maybe your T wave in the current EKG is upside down and it used to be right side up and you're having chest pain, and now I'm getting nervous that something is going on. Right, right. So there's a lot that we look for. We don't tend to look for necessarily

Why Echo Beats EKG For Structure

SPEAKER_02 12:03

structural heart disease on an EKG. If we think someone has something going on with the structure of the heart, we might see it on the EKG, but we'll often say, hey, we want to get an echocardiogram because that lets us actually visualize the structure of the heart. And when I talk about structural heart disease and what they did in this um deep learning model that we'll get into called Echo Next, is they were like, hey, we want to understand how we can flag who is having problems with the structure of their heart. Not the necessarily the electrical system, but the structure, like the walls or the chambers. Is their heart failing? Is it not pumping well enough? Is it um is the wall thicker than it should be? All of those things.

SPEAKER_00 12:48

I got it. And then is that the kind of thing that the echocardiogram is meant to capture?

SPEAKER_02 12:52

Exactly. Because you're getting like a direct so think about it as like, okay, an ultrasound. When I say ultrasound, you probably think like babies. Yes. Babies, right? So you can like tell the gender of a baby. Oh, right? Because you can see the baby. You can see if there's a penis, right? You can see the head. You can see how long the arms and legs are, like you can see the heartbeat of the baby on an ultrasound. Well, an echocardiogram is similar in nature to that, except it's done over your heart. And so you're able to look at all these things. You're able to look at the valves, how thick the walls are. You can look at um, you know, how well the heart is pumping. All of those things are available to see on an echo. But an echo is like a much bigger test than an EKG.

SPEAKER_01 13:34

Yeah.

SPEAKER_02 13:35

It's not like you can just like walk into your doctor's office and get an echo. It's done at specialized places. It might be done at the cardiologist's office. It's not done in the ER, for example. I'd have to order one and then they'd have to go separately to the echo lab. It takes uh a longer amount of time to do an echo. It's not done in in five minutes or 10 minutes. It can take an hour to do.

SPEAKER_00 13:54

Yeah, and it uses a lot of resources, a lot of people, a lot of systems, a lot of map machines, everything, right? So yeah.

SPEAKER_02 14:00

Exactly. But what it does do is it gets you a real picture of what your heart is doing, what the size is, what the shape is, how well it's pumping, how well the heart valves are moving, what the blood flow looks like, what the lining around the heart looks like, and the blood vessels that are connected to it. And so you get a lot of information from it, and it's different information than you get from the EKG.

SPEAKER_00 14:25

Right, right, right, right. Okay, so we have um the benefit of a quick read with the ECG, but it it it only tells you so much about the electrical system, which is but it doesn't tell you enough about the structure directly. It may tell you enough about the structure, but it's a secondary signal because looking at the structure, you want to look at the structure directly, and that's where the echo is. But of course, doing an echo cardiogram requires a lot more resources and is not something that you can do quickly, and uh it's not something that you do on every single patient that comes in, anyways. So that's so that's the tension right there because structural heart disease is the type of thing where something is wrong with the heart structure, and that's something that is caught with an echo, but is hard, but but but but some of these, and and it seems like with this patient especially, by the time it manifests itself, the patient's dead.

SPEAKER_02 15:16

So that it's it's like an extreme case, right? But yeah, the thing is, like, you know, echoes are much more expensive. You need expertise to do them, you need expertise to read them, and um, and so the total number of patients who get echoes is much smaller than the number of patients who get EKGs. But the question they were raising in this trial is okay, we normally look at EKGs to see what the electrical system of the heart looks like. And there's other things we look at too, but that's like, but could we use them? Could we train our AI systems on the combination of someone's EKG and echo to figure out, oh, hey, does the EKG show us when there's a structural heart disease?

SPEAKER_00 16:01

Or some abnormality that requires an echo, right?

SPEAKER_02 16:04

Exactly. Like, can it flag the ones that are need to be echoed? That need to be echoed. Right. Um, as clinicians, we rely on other information to do that. The EKG is part of it, right? Yeah. But we might rely on a lab test or the patient's medical history where they're coming in like this patient w was short of breath after exposure to wildfire smoke and improved after a breathing treatment. That doesn't scream heart disease, right? Right, right. It would be very different if we said, hey, this patient is short of breath. They've got uh swelling in their legs, which may be backed up from their heart, and their BNP lab test was elevated, and they said, you know, that it's been progressive over the last couple weeks, and now they're having uh more shortness of breath with exertion climbing stairs. That's like a very different history that might flag a clinician to say, hey, let's figure out what's going on here.

SPEAKER_00 16:54

Got it, got it,

Training EchoNext On Paired Data

SPEAKER_00 16:55

got it, got it. Um, all right. So what these guys did then is also did a very comprehensive study.

SPEAKER_02 17:03

It was a huge study, right?

SPEAKER_00 17:04

Yeah, it was a huge study. Um they actually so they they they got so one of the challenges they identified with building a system for looking at ECGs for the purposes of extracting out potential um structural issues with the heart. One of the challenges they said was um that there's there's a fundamental data challenge, which is people patients come in with different clinical contexts, there's different patient demographics, there is a range of disease states. It's not just one thing, it could be a range of overlapping disease states. And a lot of the data is proprietary. So they went ahead and they built the data set of um if you want to think about it this way, it's sort of ECG and echo pairs of data. So they uh and this was they had 1.2 million such pairs for over 230,000 patients um in a lot of them in some of the New York hospitals. Uh-huh. And it was over the period of time from 2008 to 2022, and it was mostly it was all adults. Uh, and they had a range of demographics and all of that stuff. Um but but the idea is that you get this pairing, which is you get the ECGs and you get the corresponding echocardiogram, and you get labels over them. So uh, and this is something you could correct me here, but it seems like the structural heart disease, um structural heart disease changes things like the heart muscle, the chamber size, uh the pressure, timing, all these types of things which appear as potential faint signals in the ECG. And that's the whole theory behind it, right? And so um but you have these um so but then there's like clinical ways, uh like a there's a there's a checklist, I don't Aaron Fist checklist, but there's a formula for looking at an echo and looking at various thresholds and deciding if this patient has structural heart disease or not, right?

SPEAKER_02 18:48

Yeah, yeah.

SPEAKER_00 18:48

And so so they they had people uh code that so they had people look at the echoes and actually identify which of these patients had structural heart disease and which of them did not.

SPEAKER_02 18:58

Yeah, and they had to make some binary decisions. No, those are never perfect, right? They had to be like, oh, here's the threshold. But they followed like established guidelines. They did follow established guidelines for doing that to kind of flag. This is a yes or a no, right?

SPEAKER_00 19:11

So essentially what they had was a data set that had um ECGs and a zero or one, yes or no, true or false, for structural heart disease. Now, there are other more details in this paper which people can read that had they also did out some other stuff and and made the model more rich. But the core that's the core idea. So now the question is what machine learning model

Why Convolutional Nets Fit Waveforms

SPEAKER_00 19:32

did they use? And they used what's called a convolutional neural net. Now, convolutional neural nets are what people use for image analysis. It was a big breakthrough in the early 2010s, and it basically was what led to a big revolution in computer vision. So, like identifying images, identifying cats and images or whatever, right? And all of your major, uh a lot of the major um Facebook and all of these different websites that actually, you know, put little boxes around your um pictures and identify people and things like that are often based off of convolutional neural nets. The idea behind convolutional neural nets is very similar to how the human vision system works, which is the idea that you start off by identifying um abstract features that you then compose. So maybe you see edges and shapes first, and then you see something more complex, and then you see ears, and then you see um the whole animal, the whole cat or whatever, right? And so it it builds up this sort of um system where it filters through these different types of things. Think of them as stencils, right? Looking at are there any edges here? Oh, these edges are here. Okay, now are there any shapes here and so on. They apply that same idea. Now, instead of images, it's uh you know a waveform, right, in an ECG, it's just a line. Yeah. But if you think about it, um, a line, uh if you think about your number line in a similar way, you can sort of see that there are these windows. And if you move that window across, the convolutional uh network allows you to filter those areas, those regions. And by doing so, what it's able to do, and it does this in in sort of in a cascading manner, what it's able to do is it's able to look for those peaks and those, you know, those little bumps, but then also look at a higher level of what two peaks do, what three peaks do, what you know, what the distance of together. So they can do all of that, right? So it captures the lower level stuff and higher level stuff. So it actually kind of is a good model for how one would want to read um an ECG for those kinds of higher level issues. Because in some sense, the structural heart disease is a higher level issue. It's not like one peak, right? It could appear across the 12 different channels of the ECG in different ways that is much more complex than just a single bump.

SPEAKER_02 21:41

And structural heart disease isn't just like one thing either. There are lots of things that could be wrong with the structure of the heart. Like one person might have a mitral valve stenosis, whereas another person might have a left ventricular hypertrophy causing outflow tract obstruction. Those are different things.

SPEAKER_00 21:57

Yes. Um, very complex words. Also, but yes, there were other things. But um, but yeah, so that's what the convolutional neural net is good for, and they trained that on all of this data and they tested it across different um, you know, pretty diverse data sets. Um, and that's kind of what was the basis for the model that actually worked in practice, where they put it sort of behind the scenes and was able to detect the person's issues.

Explainability And What Doctors Can Learn

SPEAKER_02 22:22

So, question though, like when we talk about uh neural nets or what we have in the past and LLMs, it's like one of the questions that I always have is like, oh, can we work to understand why it flagged this EKG? Or like how explainable is it? Like, can we learn something from it? Or why was in particular, like why was this patient's I can see this patient's EKG, I can see some abnormalities in it myself. But like, why was this patient's EKG flagged? Can we understand why the AI system, Echo Next, flagged this one or that one?

SPEAKER_00 22:56

Yeah, so that's an interesting question because it's sort of a it depends question answer, which is that it might at some level you might say it um is providing an explanation. So what you can do with these systems is say, okay, which um, and especially in a convolutional neural net, there's been some you know developments where you can do what's called saliency maps and chap values and things where you can look at the part of the image that was most influential in making this particular prediction. So in an EKG case, it might be able to identify the window in the or or highlight the part of the AKG that was most influential in its deciding whether or not this person had um the structural heart disease. So that's something, right? Because that tells you some it might be. It was the QRS complex or whatever. Right. It might it might it might not be quite something that is meaningful like that. It might just be like a time 0.1 to time 0.5, you know, it might just be something. And the question then remains is okay, if if a if a doctor looked at that explanation or that that region, will they make any sense of it? It might be the case that it just identifies a whole bunch of different things and you're just like, okay, I don't why is this, you know. So you can't ask any more questions. You can't. So that's the challenge, right? With these convolutional neural nets and in general with these deep learning deep learning systems, you don't get the sort of decision tree type thinking because that's not how the thinking worked, right? It's pattern matching, it's fantastic at finding patterns at different levels of abstraction.

SPEAKER_02 24:24

Patterns that like as physicians, we might not even know exist or might not even see. That's what's crazy to me. Is like it's flagging staff that like I don't like maybe You might not even see the pattern. I might not even know that that pattern means something. Yes, because that's not part of my like the training that we receive as physicians on EKGs. Like, no, maybe nobody knows, right?

SPEAKER_00 24:43

Right, nobody knows. And you've gone through and it's gone through 1.2 million of these pairs. You haven't, right? And so it's maybe identified a set of patterns. Now, there is a question whether it's identified the right patterns, and that relates to understanding structural heart disease, understanding how structure in the heart manifests in electrical signals. That's the explanatory framework that one would need. It doesn't provide that.

SPEAKER_02 25:07

But but what they did do in like the testing phase of this is they they went through and they found people who had EKGs that were flagged and said, Oh, let's send them on for echoes. And like, I forget the exact number, but I feel like it was like at least half of them who actually did have abnormalities. Yeah. So it's not like it's a perfect system, but you're flagging people who would otherwise be missed.

SPEAKER_00 25:28

Yes. Yes. I mean, that's from the performance standpoint, but from an explanation standpoint, I don't think it's going to be satisfactory for people because, you know, even if you wanted to use it as a training tool to improve the physician's ability to detect these things. Yeah, that's what it's hard to do that because you don't beyond, I mean, if the examples are uh narrowing down regions that are of meaningful value, then you can say, oh yeah, you know what? This certain type of QRS complex um is morphology, yeah, is meaningful in some fashion. But it's not doing that. It's just identifying a region. And it can't do any more introspection beyond that, right? Yeah. It's like asking you why you look at a you look at you know a blue mug and I'm asking you, why is this blue? There's no uh you can't provide me an answer for why this is blue to you.

SPEAKER_02 26:13

I mean, I painted it blue.

SPEAKER_00 26:15

Well, that's an explanation that's that sort of um historically defines why something was the way it is, but it doesn't define for you why this thing, why you decided this was blue, right? You decided this was blue because you that is blue to you, right?

SPEAKER_02 26:30

My vision told me so.

SPEAKER_00 26:32

Yeah, so there are some um types of um cognition that are kind of unconscious in a way for humans. And the idea is that those kinds of cognition is what uh these neural nets kind of capture. It's not like you're not reasoning about blueness of something, you're just saying that's blue or not, right? And so the idea is that maybe there are patterns in here that are so deeply um that exist that maybe some physicians do actually see it. Maybe they don't even know why, but they just see they're like something is off here, right? And that's enough for they don't have to explain it, they just know that something's off here.

SPEAKER_02 27:04

Well, a lot of EKG stuff is pattern recognition, right? It's like, oh, if the ST segment is elevated and this morphology, then we're worried about a heart attack, right? And so it would be interesting, and it sounds like we can't do this, at least not yet, to be like, oh, well, we should be looking for X, Y, or Z also. Yes, because that could flag structural heart disease in general.

SPEAKER_00 27:24

Yes, what would be super cool is if we're able to take this device, this tool, AI tool, and figure out rules, right? Right. That it it has identified, it's identified all these patterns. But then once we have those patterns, are there rules that that we can then apply so that humans can be

Workflow Burden False Positives And Scale

SPEAKER_00 27:41

trained to also look at it uh quickly?

SPEAKER_02 27:43

Yeah, I mean, in a way, it's like I I kind of want to this is is this weird to say I kind of want to like run all my EKGs through this system. But then of course you have to figure out okay, what do we do if it flags that EKG? If I'm working in an ER or if I'm working in a primary care clinic and I run all my EKGs through Echo Next system, what is the next step after that? Right? Yeah, yeah.

SPEAKER_00 28:05

Yeah. I mean, the next step is identifying the abnormalities and then taking echoes for those patients who you've identified because they're all possibilities, right?

SPEAKER_02 28:13

Yeah, within what time course and like, you know, say someone is coming in with shortness of breath, do I need to and it flags that person? Do I need to then send them to the ER to get, you know, uh stat lab testing done or whatever, yeah, in addition to the echocardium.

SPEAKER_00 28:27

Yeah, and I don't remember what the false positive rates are because there is a an issue about just having too many false positives uh resulting in an increased burden on the in the healthcare system because now more potentially more people are getting echoes now. I don't know if that's actually true, but more people might be getting echoes if more things are being flagged.

SPEAKER_02 28:44

Well, if you think about this at the population level, that would be a concern, right? It's like you have this new thing you're deploying. It's um it's it's pretty good, right? But it's not perfect. And but what it's gonna do potentially is flag patients that may not have otherwise been flagged, and now you're going to have to arrange for echoes on all of them. And does your system allow for it, or how can we get the system to allow for that if we think it's an appropriate thing to do? Right. You need, you know, specialized um uh echocardiogram technicians to do the testing, you need um physicians to read the echocardiograms, you need uh care coordination in order to, you know, bring those patients back in, tell them their results, and decide what to do next. Right. So there's like multiple layers of systematizing something like this, um which may be worthwhile, right? Given the case that we talked about right at the beginning, but would be an additional layer to think about. It's not something that that you can like operationalize with the snap of your fingers when there are so many EKGs out there, you could be flagging a lot more people than we currently are.

SPEAKER_00 29:54

Right. Exactly. Exactly.

Final Takeaways And Closing

SPEAKER_02 29:58

All right, well, um, I hope you enjoyed your tutorial on EKGs and echocardiograms.

SPEAKER_00 30:03

I did, I really did. I feel like I learned a lot. I always had uh, you know, a vague understanding of it, but I definitely did learn a lot from that, from this. And um, I think it's super exciting work. I I think there's so much so much more to be done.

SPEAKER_02 30:17

I absolutely do too, and I think there are a lot of potential applications for this moving forward, and I'm excited to see where this work goes next. And I'm so excited to see it save someone's life. Yes. We'll see you next time on Code and Cure. Bye bye.