Gaurav Choudhary, MD - phaware® interview 557

Jan 28, 2026

How an AI Stethoscope Could Transform Global Healthcare

What if diagnosing PH didn’t require an echo or heart catheterization—but just a AI powered stethoscope? Dr. Gaurav Choudhary talks real-world use cases, validation studies, and the global potential of portable, AI-powered diagnostics in under-resourced settings.

My name is Gaurav Choudhary. I'm a cardiologist. I'm in Providence, Rhode Island. I'm a Professor of Medicine at Brown University, Alpert Medical School of Brown University. As well as I'm the Director of Cardiovascular Research at Brown University Health. I got interested in pulmonary vascular disease because of some incredible mentors that I found here who, were working on pulmonary vascular endothelial cells. With my background in cardiology , I got interested in pulmonary vascular remodeling, pulmonary vascular endothelial cells, and pulmonary hypertension. As a practicing cardiologist, I used to read echocardiograms all the time. We'd see people come in with elevated tricuspid regurgitation jets or elevated pulmonary systolic pressures on echocardiogram. We would do this often and then put it in our report and hope something happens with the results. More often than not, those observations, this is 15, 20 years ago, were there and reported, but not acted upon, in most of the cases. It was one of the things that was part of the echocardiography report, but nobody really paid attention. 

That got me thinking that, okay, we are doing this. We know at least from some of the literature that these PASP is bad, elevated pulmonary systolic pressure is bad, but there's nothing happening. I dabble both in epidemiological studies and clinical studies, as well as basic science studies. My day job is actually running a lab, looking at right ventricular dysfunction, mechanisms to improve right ventricular function using a lot of preclinical models and biological specimens. But in the epidemiological side, that's where we started off, because first we wanted to define the problem in some ways. We looked at databases with collaborators from here and Boston and then saw, okay, how common is elevated PA systolic pressure? And what is the impact of having high elevated PA systolic pressure?

We found in the VA population, first, that there was a large number of people who were undergoing echocardiograms, had elevated PA systolic pressures. Now, we cannot diagnose pulmonary hypertension by echocardiogram, as all of us know, we need a right heart catheterization for that. But it's a pretty good suspicion that if you have high PA systolic pressure, your probability of having elevated mean PA pressure is high. Thresholds kept getting debated on and we have contributed our own to the literature about the threshold that may be meaningful to diagnose pulmonary hypertension. So, it's very under recognized. That was some of our initial work. Then, we used national databases to look at, well, what's the impact of elevated mean PA pressures on outcomes?

We found that as early as a mean PA pressure of 19, we start seeing poor outcomes. Then, we expanded some of that work with PVR and wedge pressure. So, that was kind of the epidemiological area. But the other problem of under recognition when we talk about is once you have an echo, you see an elevated PA pressure and nobody acts on it. But not everybody gets an echo. We have all these people who have dyspnea. People in the PH community are well aware of the delay in diagnosis in pulmonary hypertension, because of the nonspecific symptoms that come before. If you go to the primary care doctors, you go to specialists and there is a well-known documented delay. So, how can we see if we can act earlier? One of the things that I was always curious about is using tools that we use every day, but using it better for that particular purpose.

The stethoscope has been around for 100 years. Unfortunately, what goes between the stethoscope is getting worse and worse as far as the skills concerned from our trainee perspective, and that's also well documented. There has been a real decline in understanding what you're hearing through a stethoscope. All the major societies have realized that and actually published data on it. Unfortunately, and again, this is not a generalizable statement, so not all people are that, but in primary care it's a little bit worse than in specialty care, which is understandable. Especially, when it comes to cardiac auscultation. Cardiac auscultation is one of those things that has been used clinically to suspect diseases like pulmonary hypertension. The idea was, okay, can we now use cardiac auscultation and see if the new technology of deep learning and convolutional neural networks can develop an algorithm which can suspect elevated PA pressures?

I was interested in it and I was trying to say, "Okay, how can we do that?" I looked up online to see if there were any things already there in the published literature. I came across a company which was working on digital stethoscopes. They had some early algorithms out on murmur detection already. I just reached out to them. I said, okay, this is something that we want to do. This is important. They realized the importance of pulmonary hypertension in general. Now, I'm not talking PAH, I'm talking about pulmonary hypertension in general, and got interested in it. We did a pilot study first, which we just published I think last year in Journal of American Heart Association. We first developed an algorithm. We used about 6,000 patient data points where they had an echocardiogram and phonocardiograms.

We knew the PA systolic pressure in these patients from all over the world. We had around, I would say 170,000, 169,000 just phonocardiograms without known PASP. These inputs were used to kind of train an algorithm to detect PH. We took a threshold of PASP of 40. So again, we took the phonocardiograms. Basically, these are little recordings, audio recordings. They're converted into what's called a Mel Spectrogram, which is an easier way for the computer to read it. They're small stretches of it that are cropped, and after they're cropped they're put into an algorithm to train. We basically take 80% of the data to train it and then 20% to validate it. This happens in five-fold cross-validation. Then, the output comes in whether they say, "Okay, did they meet the threshold of PASP or not?" Once we trained the algorithm, we got about an average ROC curve of about 0.8, so about 80% area under the curve.

Then, we wanted to validate it, because it can function really well on the data that you give it, because it diagnoses perfectly because of that. We had about 200 patients who came in for regularly scheduled echocardiogram. We did the recordings on them and we put it through this model. Again, we got about a sensitivity of about 78% and the specificity of about 70%. So, pretty reasonable. Now, these were pretty small numbers, 196, 200 patients. So, it is feasible. It's doable. The output right now would come as either normal or suspect PH. There were several limitations in our initial study. Obviously, it was at a single center. Most of our patients because of the nature of the population that we, had were men, so very few women. They were older. We did it a number of ways. We did the analysis. We did a per patient analysis where you take five recordings on a patient and then average it out and see PH or no PH. Or you can do it per recording. In each of the four or five areas, whether it's PH or no PH.

There was still some work to be done, because at the end of the day, we don't care which area it is coming from. We care whether the patient has PH or not, or suspect PH in the patient. The other issue is our gold standard in this case was an echocardiography. We all know the gold standard for diagnosis of pulmonary hypertension is the right heart cath. Based on this preliminary data and these new questions, we actually submitted a SBIR grant to the NIH to get funding for a larger study, which is currently ongoing. I'm an MPI on that because it's an SBIR grant. Eko, which is the stethoscope company, they have a PI on that. So, it's a collaborative project. We are enrolling about 2000 patients who are coming in for echocardiogram, plus we are enrolling patients who are just coming in for right heart cath. So, we have both cohorts.

And then, further refining these algorithms, because one of my concern always has been that if you take a threshold, whatever threshold you pick for PASP to train the algorithm, you don't want the algorithm to over-diagnose, because what happens is it's like the false alarm. If I'm a busy clinician, primary care practice, and I'm seeing a patient and they keep saying every third patient I that suspect PH, and I send a few of them for the echo and nothing comes out of them, I'll stop using it. What we also want to know is not only the accuracy, but also maybe the severity. We want to know is okay, maybe mild, moderate or severe, because that kind triggers different things. Maybe in mild you do certain other things, as well if it's severe they may need urgent evaluation.

One of the things we are going to try to do is refine the algorithm so it can give us some categorization of the severity of the disease. We will validate it with the gold standard right heart catheterization, and then also we will have a larger number of patients to further refine and improve on what we are using. That's the goal. The study's ongoing. It's in the second year. We probably have another year, year and a half to do. Obviously, one of the great things about partnering with a company that has a product like this is it can come to fruition really, really fast, because they can go to the FDA, they can get it incorporated into the current suite of algorithms and we can see the light of the day.

This was really exciting to me because not only we can get something to the point of care in the hands of primary care patients, this is also direct to consumer too. You can buy a stethoscope yourself, though I wouldn't recommend interpreting what you hear from it. But you can look at it in remote care settings. You can look at places which are under resourced. There's a lot of PH in sub-Saharan Africa, South America. The echocardiograms are not available at all places. So, how can we further ensure that people who really need echo are identified in the field or in practices out there, and then they can go to specialty centers. I think there could be a lot of potential for a technology like this to be implemented to improve the diagnosis and detection of pulmonary hypertension.

But this is not the only technology in AI. People are using AI and pulmonary hypertension for detection purposes like we are. They have used ECG. There are very nice algorithms coming out of Mayo and other places where they have looked at diagnostic accuracy or detection for PH.

The different gold standards echos or right heart cath; CT, Echo and MRI, not only for detection prognosis purposes, but also further subdividing which group of PH it may be. Or, in some cases, automatically coming up with an EF or right ventricular function or segmentation in some of these things, looking at lung vasculature using AI and CT and segmenting it. So, there's a lot of work that's being done in other technologies, as well. The reason I like the stethoscope is because you don't need any big equipment. It's something that you carry with yourself and most physicians have it or clinicians have it. You don't need any setup for that. You do need internet connection for it to go to the cloud and tell you what the results are. That's why I kind of got interested in this AI project, and that's where we are.

I have always believed that AI is going to be here to stay. We have to embrace AI. We not only have to embrace AI, we have to work with the technology companies to make tools that are meaningful for our patients. Because we are in the best position and the patients are in the best positions to help develop the tools that are useful for them. We have to define the questions and we have to understand the limitations. It's really, really important. That was the example I gave you earlier is, if every second patient in primary care has suspected PH or suspects murmur or flags it, we have to be really making sure we are asking the real clinical question. Half of the echocardiograms have pulmonary hypertension or elevated PASP. The large databases have shown that 80% of heart caths have pulmonary hypertension. Now, they don't have pulmonary arterial hypertension, but they have pulmonary hypertension. So, how do we make sure that we are getting the right kind of people for the right kind of therapy?

I think this is the first step. I think we need to be very careful about moving too fast in it. This is the first step, but this is exciting step, because we are able to do what we weren't able to do before. I'm hoping the next steps are going to be faster. It will need a larger population because won't it be great, and this is speculation, if we put this stethoscope or we use an AI tool on somebody and it'll tell, well, this looks like pulmonary arterial hypertension? High mean PA pressure is suspected. We are not there yet, but right now we are using only one tool. But just imagine how if we combine what we're seeing on one AI tool and combine it with a wearable tool, combined with the data from your EMR which a large language model has already processed, and you put it all together. That would increase your confidence.

Things that we cannot see in a 15-minute visit that may be available to a primary care doctor and is flagging it from these multiple things. Or you have a chest x-ray which had an enlarged PA that nobody paid attention to, and then, suddenly, that also becomes pieces of it. The promise of this technology is that it'll bring together data, synthesized data from multiple data streams for you to look at. Then, the final decision I think has to be with the clinician and the patient, because at the end of the day, AI is not going to replace the clinician. I think it'll augment the clinician and I think there will be clinicians who will be using AI to help provide better care. That's the way I think about it.

Now, a patient-provider relationship is something that cannot be replaced by a machine. It's so nuanced. It doesn't matter what an algorithm says. It doesn't matter what the next titration dose needs to be. It matters on what's going on that you need to take care of. So, taking care of patients is more than just following a protocol. I think that will not get replaced. I think these will be tools that will assist us. Yes, that is a promising future. I think we will get there, I have no doubt. But I think we have to really be careful and engage with the companies and the people who are making these tools, incorporate feedback not only from clinicians but even from patients, especially from patients.

One of the studies we proposed, it didn't get funded, but basically looking at implementation; because you right now will have so many tools in AI that everybody's coming up with their diagnostic tool for disease A, B and C. I gave you an example of what we are working on, but there are other examples even in the PH space that they’ve come up with. They have excellent diagnostic accuracy in the studies, but then how do we implement it in a patient care setting? What is the acceptability from the patient? What is the acceptability from the providers? What is the acceptability from the nursing staff that's going to be using it? When you implement it, do you really change the trajectory of the course of that patient's disease? Do you detect early? If you do, what are the extra tests that happen? How important it is, and do you start treatment early? Because at the end of the day, that's what you want to see. It's not that you have a tool that gives you 99% diagnostic accuracy. You want a tool that gets the patient treatment faster.

The second phase of studies have to be moving beyond just testing area under the curves, but testing if I implement this tool in a practice setting, am I making a difference to the patient? Those are difficult studies to do. Where we are at the first step of just defining these tools. But I think we have an exciting time ahead as we implement them in the future.

Thanks for listening. My name is Dr. Gaurav Choudhary, and I'm aware that my patients are rare.

Learn more about pulmonary hypertension trials at www.phaware.global/clinicaltrials. Follow us on social @phaware Engage for a cure: www.phaware.global/donate #phaware
Share your story: info@phaware.global Like, Subscribe and Follow us: www.phawarepodcast.com. #phawareMD @BrownMedicine