Connor Landgraf, Co-founder and CEO, Eko 
Moderator: Michal Lev-Ram, FORTUNE 


Brainstorm Health 2023 - An A.I. Enabled Stethoscope 

Conferences

Brainstorm Health 2023 - An A.I.-Enabled Stethoscope 

The National Institutes of Health say that valvular heart disease VD is a leading cause of cardiovascular morbidity and mortality worldwide. And the stats are all just going in the wrong direction. Negative outcomes from V H D are only expected to increase in the decades ahead. Millions of Americans have this disease. Millions of Americans don't know they have it. So detection is key. And so now what we're going to do is show you a really innovative platform that's been described as Shazam for heartbeats. Shazam is that app that you can get on your phone and play any music you want. It will hear it and tell you what the song is. Well, that may get your heart pumping with music. But the cardiovascular disease detection platform echo E K O Echo is tapping into a different set of beats. Uh The system utilizes machine learning A I powered stethoscopes and a library of nearly 80,000 heart sounds to identify murmurs which are a key sign of V H C. So let's see it at work. Please welcome back. Fortunes Mikal Lev, along with Connor Landgraf co-founder and CEO of Echo, they're gonna give us a little show and tell here they are. Ok. So we're going to play a heartbeat for you guys and you have to guess who it is. Just kidding. Tell us about echo. Yeah, absolutely. So the stethoscope is the world's most ubiquitous medical device. Probably the world's most iconic as well used hundreds of millions of times each year by clinicians to do critical decision making around patients, risk of cardiovascular pulmonary disease. And yet it's, you know, incredibly subjective, highly open to physician interpretation. They have to hear those differences between health and disease and that's, that's really challenging for them to do. And I think if you ask most physicians, they'd say they're not super comfortable using the stethoscope as a diagnostic tool even though it is so iconic and ubiquitous. And so what we'll demonstrate here is we'll give an example of the way that A I is being used to inform this critical clinical decision making on the front lines of care. We'll take this stethoscope, place it on the simulator here. This is my friend, Manny Manny is a mannequin. He is missing, missing some appendages today, but he does simulate a heart sound of a patient who has valvular heart disease. And we'll demonstrate how that A I can inform that clinical decision making so quickly. OK. And let's take a closer look on the screens at what we're looking at here. Then go hold still, Manny. All right. So what's happening there? We did a quick recording of Manny's heart sounds here. We sent it up to the A I algorithm in the cloud. Did the analysis sent the results back down and they're displayed here. We see a structural systolic murmur likely highly associated with a vaguer heart disease called aortic stenosis. Uh This is definitely a situation where Manny should be referred to cardiology probably urgently. Um But that there is highly likely uh pathology here uh that we could quickly screen in just 25 seconds in a patient exam room, right at the front lines of care. Um So amazing. OK, I wanna ask you a few more questions. Let's make our way over there and leave Manny to his space over there. By the way, is your heartbeat in the database, Connor? It is uh I've done a lot of the recordings early on. So my heart is probably over represented in the database, but it is one of several million heart sound recordings that does encompass this machine learning algorithm. What is pretty exciting to us is that, you know, it's it surpasses almost kind of any clinician experience that they would be able to be able to see. I mean, the clinician will see maybe 30,000 patients in their career and this data set has been informed off of millions now. Um And you know, it's not that dissimilar to Siri or Alexa, these kind of voice assistant technologies just applied to a new clinical application. How how does one actually go about generating this data set though of millions of heartbeats, we've done a lot of work over the last several years doing prospective clinical studies, working with academic institutions to sit in their cardiology labs when patients come in and get an echocardiogram or a CT scan or you know, an EKG will capture their heart sounds alongside that. So we have a really strong gold standard reference point and everybody has a unique heartbeat. They do and and unique ECG as well. We do find that there are really signature fingerprints within these signals that help us, you know, hear those subtleties, we're hearing the flow of blood inside the heart itself and when that flow becomes turbulent, when there are kind of abnormalities, they have very kind of consistent signatures to them. And so is that process of just kind of learning what the signatures are so that the algorithm can provide that accurate result? Have you learned anything specific to your own heartbeat? I it's funny you ask, I've definitely struggled with palpitations in the past. And it's been pretty neat to use the technology to understand my own condition, understand that, you know, it's not necessarily something that I need to be worried about, but having that knowledge is pretty, pretty comforting, I would say and talk a little bit more about the stethoscope. Why is it, does it need to be a stethoscope? It doesn't have to necessarily be that clinician form factor, you know, that that familiar form factor. But one of the things we recognize based upon kind of our learnings with clinicians was that to have the impact we wanted to on primary care and to get the utilization, it needed to be comfortable, it needed to be familiar to the clinician. It needed to look like the tool that they expected it to, it needed to fit within their existing workflows. It couldn't be something that would take inordinate amounts of time in the patient exam room, it wouldn't be utilized otherwise. Um and it needed to be fast and non invasive. You know, I think one of the things that we've recognized is that primary care is going to take so much more responsibility in the coming years to support a holistic integrated care. Primary care will be more and more of a gatekeeper and having technologies that can fit seamlessly into primary care workflows is critical for their success in adoption. So it's a little bit like how the camera on the iphone makes that clicking sound, which is completely irrelevant, doesn't need to be there, but we're just familiar with that. Yeah, that, yeah, that comfort. OK. Um And talk a little bit about, you know, what else can you do with this? What else can you really decipher glean from heartbeats? Where, where do you hope to take it next? Yeah, I mean, we're, we're building what we think of as a digital biomarker, an acoustic digital biomarker for disease. And you know, the stethoscope has survived for 200 years of medicine simply because of the fact that there is a tremendous amount of data in sound in our breath, sounds in our lung sound, the crackles, the wheezes associated with asthma or COPD or pneumonia. These are all signature clues to helping us understand decipher the physiologic status of these patients and we can continue to evolve the A I to to work well, even in these applications outside cardiology, our recent clinical data is showing that we can double the sensitivity of primary care physicians using the A I that, you know, there is so much opportunity to improve the timeliness and accuracy of cardiovascular disease detection at the, you know, at the front lines of care at that first opportunity with these patients. So they don't have to wait to get a cardiology referral. They don't have to, you know, go down a long chain that's likely to be broken before they get, before they get confidence and clarity. Um Can you give a sense? I mean, you, you shed some light on this but, but give us a better sense of just how much more effective the A I is and at what scale than human beings. Yeah. So our devices, we now have more than 300,000 providers using our devices to serve in their physical exam tools in that primary care setting. And we're seeing tremendous results. As I said, you know, we are more than two xing the sensitivity for identification of substantial murmurs associated with Faul or heart disease and other common conditions. And what, what's exciting to us and what motivates us is thinking about the patients who might not likely be, be, you know, be risk factors for cardiovascular disease or have suspicion, think of, you know, pregnant mothers, for example, or pregnant women who have substantial risk of cardiovascular disease. But not every physician is going to be thinking about that. And so having these non-invasive tools that can help us uh quickly ascertain whether there are signs of disease and then get those patients to the appropriate testing. Um The clearance pro FDA clearance process for you, I hear was pretty quick. It is, you know, the actual submission process itself is, is relatively quick. But the work that goes into those applications is strenuous and it took us, you know, more than three or four years to gather enough clinical data to demonstrate safety and accuracy validation before we submitted to the FDA. So it's a lot of lead up to a lot of materials. Think our submission packet was close to 6, 700 pages of material to really show that, that, that the algorithm could perform the way that we, the way that we wanted it to. And last question for you, obviously, a lot of we'll be talking about this later on as well. But a lot of interest in generative AI open A I all the, you know, competing products does, does, does that excite you? What else could you do with that involved? We do think. Yeah, there's a lot of interesting corollary with large language models and chat GPT. I think what we, we have stayed really focused on is just being that, you know, the acoustic analysis tool. The thing that machine learning is particularly effective at when there's a large data set powering it and where we can kind of map, you know, a single input to a single output for disease detection. And we're excited about the impact we can have and even global health too. OK. I'm gonna ask you actually one other thing. Could you recognize your heartbeat? That's a good question. I don't know if I can actually recognize my own heartbeat. Uh I, I don't think my ears are that so. No, no. You know, I mean, that's why we have a, I, I guess so. Thank you so much Connor. This is um exciting. We'll be excited to, to watch what you do in the, in the coming years. And um hopefully Manny is doing ok. I hope so too. Thank you so much.