Beyond Borders: AI Cough Analyses

This episode, we are featuring Dr. Alexandra Zimmer to discuss her riveting publication "Cough acoustic analysis using artificial intelligence for COVID-19 detection: A comparative study of patient cohorts from Lima, Peru and Montreal, Canada.” We dive into questions regarding COVID-19 detection and the usage of artificial intelligence on cough acoustics in both Lima, Peru and Montreal, Canada. Dr. Zimmer’s article is published in the June 2026 Issue (Vol. 118) of Annals of Epidemiology.

Read the full article here:
https://www.sciencedirect.com/science/article/pii/S1047279726000748

Episode Credits:

• Executive Producer: Sabrina Debas (Episodes 1-18) and Sofina Tran (19-)
• Technical Producer: Paula Burrows
• Annals of Epidemiology is published by Elsevier
  
  
  

Transcript

Patrick Sullivan 0:11

Hello, you're listening to EpiTalk Behind the Paper, a monthly podcast from the Annals of Epidemiology. I'm Patrick Sullivan, editor-in-chief of the journal, and in this series, we take you behind the scenes of some of the latest epidemiologic research featured in our journal. Today we're here with Dr. Alex Zimmer to discuss her article, Cough Acoustic Analysis Using Artificial Intelligence for COVID-19 detection, a comparative study of patient cohorts from Lima, Peru, and Montreal, Canada. You can read the full article online in the June 2026 issue of the journal at www.annals of epidemiology.org. Dr. Alex Zimmer is an epidemiologist at Heidelberg University with over eight years of experience in tuberculosis diagnostic, testing, and research. Her doctoral work investigated the potential of cough sound analysis for COVID-19 and TB screening. Currently, Dr. Zimmer's research focuses on evaluating novel tools and artificial intelligence-based approaches to optimize TB screening strategies in high burden settings. Dr. Zimmer, thank you so much for joining us today.

Dr. Alexandra Zimmer 1:24

Thank you for having me.

Patrick Sullivan 1:26

You know, I think just the title of your article catches attention because it has a lot of things that are interesting to public health people and epidemiologists. And I think many of us have been sort of struggling to understand what practical roles are for artificial intelligence in epidemiology and public health. And so your article really sort of converges on all those topics. And I'm excited to hear more about it and to talk with you about it today. So can you start maybe for folks who aren't familiar with this topic, just setting the stage a little by describing the problem that your research addresses and why it's important for public health?

Dr. Alexandra Zimmer 2:00

Yeah, so this topic, you have to go back to 2020, uh early days of the pandemic. And that's exactly when I started my A PhD. And so I was approached by my supervisor because I have an interest in infectious diseases and epidemiology with this specific topic, because there were a lot of articles coming out that centered on COVID, COVID detection, and then bringing AI into that picture. And so when you looked at this early literature on cough sound analysis and COVID detection, the papers were boasting really impressive performance. We were seeing 90% accuracy, sensitivity, specificity to detect COVID from a cough sound. The thing is, when you looked a little closer at those papers, you started to see that these were very much written from a machine learning angle. And so there was a lot of algorithm development work and you know using crowdsourced data sets, but there was no real investigation into some of the epi and biases that came about behind the scenes. And so, as someone who was doing an epidemiologist, I was interested in putting more of this epi lens on this topic and uncovering some of the challenges with this area of cost sound analysis.

Patrick Sullivan 3:20

So once you sort of realized that there was this lack of an epi approach, what are some of the main questions you had that you wanted to address with the analyses that you did?

Dr. Alexandra Zimmer 3:30

Yeah, so I think looking at the data and looking at these publications, a lot of the times the data sets were relying on crowdsourced data sets. So, you know, people were confined at the time at home and asked to turn on their computer, record their cough sounds, and self-report if they had COVID. So there's a lot of problems there already. Who's able to record their cough sounds? Is there a selection bias and who's recording the cough sounds? And then the self-reported COVID, obviously, it's not a gold standard. What is this misclassification that's coming about that is then trickling down into these AI models? Another thing that was more interesting and maybe not as intuitive for the machine learning side is that at the time that the main circulating pathogen was COVID-19. So if someone's coughing and they have an abnormal cough sound, it's most likely COVID. So the years after that, there's other pathogens that have come back: flu, RSV, tuberculosis, had a resurgence. And so when you kind of mix the epidemiology, how well do these models perform in that scenario? So that's another angle. And finally, in machine learning, it's quite standard to do an internal validation where you have, you know, your whole collected data set of cough sounds, you take 80% of those to train the model, the remaining 20% are validation. And so they come from the same source, but there's no real external validity or transportability of these models in other settings. So, how does a model that is trained in cough sounds in one part of the world do in another part of the world? So these were some of the questions that I was asking myself looking at these early papers and seeing how I could address it in my research.

Patrick Sullivan 5:15

Yeah, there's a lot of great epi questions in your answer there, which we might have time to come back to. But let's first just talk about what the major findings were from your paper. What did you conclude after going through that process?

Dr. Alexandra Zimmer 5:29

So, in my study, I chose, or I did my research in two studies. So I was at McGill at the time in Montreal. So I was doing a cough sound collection in Montreal. And then I was also working with collaborators in Lima, Peru, obviously, very different epidemiologies, different physiological differences and how people cough between the settings. Also, different smartphone microphones are, you know, what microphone is available in one part of the world versus another affects the quality of the cough sound. And so by having all this at this point in time, it feels kind of obvious that it's not going to be compatible, that an AI model developed using cough sounds from people who are copying Montreal will not perform well in Peru and vice versa. It felt obvious at this point in time, but it was still not quite explored, I'd say, in the literature. So I think the main finding overall was that the external validation of cough sounds from one setting and another performs really poorly, and AI is not really able to get over that hurdle.

Patrick Sullivan 6:33

One of the ways that you report this is an area under the curve or AUC for those models. So, and that is obviously like a good sort of description of the data and can be used for comparisons, but is there sort of an intuitive interpretation of what the area under the curve means in this setting?

Dr. Alexandra Zimmer 6:53

Yeah, I'd say the the AUC here is trying to evaluate. So the AUC is for the machine learning model and trying to determine if it can discriminate between people who are truly COVID positive versus not COVID positive. And so I think if I look back at my results, the AUC for the external validation, so taking the model from Lima and testing in Montreal, or taking the model from Montreal and testing Lima, we're all hovering around 0.5, which is equivalent to a coin toss. The model doesn't really know or can't infer if the person has COVID or not in these other settings. So it just reinforces that there's very, very poor external validity of these models between each sites.

Patrick Sullivan 7:39

And practically, then it means that a model would need to be developed for each site in which you wanted to use it. Is that a fair interpretation?

Dr. Alexandra Zimmer 7:47

I think that's more in theory. I think in practice, the idea of going to every site and collecting because AI needs a lot, a lot of cough sounds. And so that's the appeal of these more global approaches is okay, let's collect cough sounds from anyone who can cough. We don't really care about the distribution that much. But then you're missing some people in other areas that it'll perform poorly, or you under-sample from other areas and that doesn't perform as well in those populations. So ideally in an ideal world, yeah, we could get cough sounds from everyone in every part of the world and have hyper-local algorithms that are tailored to specific populations. But you know, the resource and reality of doing something like that is quite limited, I'd say. I don't think it's a it's a reality that we could do at this point, I'd say.

Patrick Sullivan 8:33

Yeah. And I think it it's it is you alluded to this, but it's as we think about this as epidemiologists and about data sources. You sort of alluded to the fact that AI is information hungry. And it seems like that it generalizes across other settings where we may be familiar with AI and when you're doing a research question like this, and you're responsible for generating all that data that has to feed in, it's not a trivial thing to say, like we're going to generate data and then analyze it with AI. And I think that's probably an underappreciated, you know, aspect of using AI and research. So then just at a high level, do you have a sense of why the performance might might differ between two cities? Just uh, and it could be hypotheses or it could be things that you inferred from your data, but you know, would you expect it to differ in different places? Are coughs different? You know, why might the technology perform differently in different parts of the world?

Dr. Alexandra Zimmer 9:30

Yeah, I think having worked on this for so long, I feel so obvious now, but it's true at the beginning of my PhD, it's like, well, a cough is a cough, right? So, you know, if we're collecting costs from different people, we should be able to do some cool things with this data point. But in reality, I mentioned some of this previously, you know, microphones and the availability of smartphones. We were recording on smartphones. So in Montreal, we were using, I think it was Android or Samsung, I can't quite remember. And then in Peru, the most widely affordable tool was uh the Xiaomi uh smartphone, which different smartphones may introduce different artifacts into the cough sound that then directly influenced the model's performance. Another thing I observed when I was at each field site is the environment where cough sounds are recorded. In Montreal, we were able to have private rooms where people were coughing in isolation from background noises, whereas in Peru, in Lima, it was outside intense, and there was some sound in the background that may have interfered. That's another aspect. And then finally, I do think you know, epidemiology, there's different circulating pathogens and different trends in the cough sounds that are not the same. I think that's one of the findings of the paper is that in Montreal there's a lot of viral pathogens. And so differentiating COVID from flu from RSV, these very viral pathogens that may have a more similar way of infecting the lungs is a little more challenging, potentially, and this is still a hypothesis compared to in Lima, where I collected cough sounds from a population with a very high burden of TB, which is bacterial, and infects the lungs in a different way and may also influence the cough sound.

Patrick Sullivan 11:21

Right. Or maybe even history of TB infection or chronic TB infection might alter things.

Dr. Alexandra Zimmer 11:26

Yeah.

Patrick Sullivan 11:26

Yeah. So it does seem like a testable hypothesis, though, if you could get the same devices in both places. You could rule out one, you have a really good hypothesis there, which is that maybe the technology, you know, alters what's actually captured, or certain technologies lose some ends of the sound spectrum or you know, have greater data density. I don't know what it might be, but it that seems like a testable hypothesis if you synced up the devices. We hope we don't have like a COVID pandemic, but uh, but just as the technology. Very cool. So I think there's just so much interest in AI. And I think when I read this, and I think when when people read this, they're gonna just think about the opportunity. Like there's a signal here. There's a bunch of questions. And um, as a great epidemiologist, you've put out all the caveats and put them into buckets of problems. And but if they're, you know, if there's a signal here, AI-powered diagnostic tools really might be deployed at a fraction of the cost of traditional testing and think about in a future pandemic. Think the shortages of diagnostic tests that we experience in many countries. So I'll just ask, you know, either specific to access to COVID-19 detection or more broadly, two questions. Like one, what are next steps that would need to be done to make it useful and proven clinically? But you know, how might this expand access to COVID-19 detection? And depending on what you've learned, maybe also to TB detection, particularly in low resourced or underserved communities.

Dr. Alexandra Zimmer 12:56

Yeah, my focus has been mostly tuberculosis. And so I my mind goes straight to tuberculosis because that's you know, currently the leading infectious disease killer in the world. And it got passed by COVID, and then now that COVID kind of went down, TB is back to number one. There's over 2 million people every year who have active TB disease who don't get diagnosed, and there's a huge challenge in the TV screening and diagnostic community of reaching these populations that are more remote, and for them, accessing healthcare and getting to the health facility is can be challenging in certain settings. And so the idea, and I think this is also why when I was starting my PhD, I thought this was so cool the idea of taking out a smartphone, which is becoming more widely widely available and coughing into it and having it give you a sense of your risk of certain pathogens, would be amazing to have. You know, it's a screening tool in your back pocket. And so I think in terms of access, in terms of making people in underserved communities have more resources to have better control or understanding of their symptoms and their disease, I think this would be something that'd be great to do. And then terms of the path to getting towards that, I think in the cough community, we've talked a lot about this now. And I think what happens, we kind of jumped ahead of a couple of steps. We went straight from can an AI detect a very specific disease in the cough? And I think this came about because we were in a pandemic and everyone was focused on COVID, and so there was a need for it. And I still think this would be the ultimate, ultimate goal. But realistically, and this is what you know the field feels about these days, is to take it more gradually, just looking at is there any pathology? Is it infectious versus chronic? If it's infectious, is it bacteriological or is it viral? And then kind of working, you know, as we get more data, as we get more cough sounds to try and infer specific diseases.

Patrick Sullivan 15:01

It's so interesting. I also think like you raised this interesting issue about the technology and the quality. And there might also be a place for, you know, with a different kind of investment and not in the middle of a pandemic, um, you know, trying to get maybe even recordings with technology that is, you know, very sophisticated so that you can figure out which nuances are important. And then you could back step to say, okay, what kind, what's the sort of lowest cost device that would pick up what we need to pick up to make these important distinctions. And I think your point about diseases like TB that really are like so prevalent in areas of the world that are so resource deprived, is a very exciting prospect in terms of thinking about the potential good of this and the population impact. I can just sort of see writing a federally funded grant research proposal that talks about the public health benefit or the leverage of this just seems like it's tremendous. So let

Behind the Paper

Patrick Sullivan 15:58

me just ask, you sort of talked a little bit about, like at the time, whatever we were doing in our research careers, we worked on COVID. I think that's just the experience of everybody who's there. But in terms of this particular topic, what drew you to this aspect of COVID research? And you've alluded to this a little bit, but I always think it's interesting how people end up, you know, doing a study in Lima and Montreal. Like sometimes it's because of opportunity or collegial connections or you know, personal connections, or like where we've lived. So just how did this come together for you in this way in the context of like now we're in COVID? What are we going to do? And how did you get to this topic?

Dr. Alexandra Zimmer 16:38

Yeah, so I think like a lot of PhD students who started their PhD during COVID, we kind of had to adjust to the reality that COVID was everywhere. So I, you know, went into this PhD hoping to do something TB in diagnostics, and then COVID happens, and this opportunity came about, and I'm someone who had an interest in diagnostics, and this seemed like an interesting opportunity to get my foot in the door for AI and machine learning, which was new and still kind of feels new in the area of Epi and clinical health and well healthcare in general. So it's kind of an I saw it as an opportunity for me to bring this epi lens to machine learning and AI that I felt was lacking in the field at the time. And I still kind of take with me to this day my own research. And then the opportunity to collaborate with Lima was, I'd say, more of an opportunity. So my supervisor, Dr. Madukar Pai at McGill University, he was well connected with you know researchers in other parts of the world, but COVID kind of shut down a lot of the field sites, or it was very hard to get things up and running. So there was an opportunity to align ourselves with Peru. And I'm very grateful for all the support I got from the Peruvian team. Dr. Cesar Ugarte -Gil was a great co-supervisor and mentor in this process.

Patrick Sullivan 17:59

So you sort of talk about how whatever we were doing, we pivoted to COVID. And I wonder, having done this and having moved through your program, what are you excited in public health about right now? Either your own work or you know what you're looking forward to or other things that are going on. What are you excited about right now in our field of public health?

Dr. Alexandra Zimmer 18:20

I I'm excited for many things. I think my time in AI has made me excited and also cautious. I think I saw what happens when people are overly optimistic with these early COVID cough detection models that didn't really pan out. So I in on the side of AI, I'm kind of navigating this mixture of optimism and also caution. And working with difficult data sets in my research and applying it to AI has been an interesting challenge. If I take a step back and talk more broadly about tuberculosis and the research fields at large, there's a lot of new innovations in diagnostic tools for tuberculosis that just got released and WHO made a press release or a uh a policy statement about this literally a week ago. So, right now my whole mind is centered on the idea that we can now sample tuberculosis with alternative samples in sputum, because sputum is a very difficult sample to produce for some people. And so there's innovations in the tongue swab area that have been really exciting to see happen in the field of tuberculosis diagnostics.

Patrick Sullivan 19:32

Yeah, and I think as we are in a time of uncertainty about global funding and commitment to a lot of diseases where we have made significant progress, it makes it all the more relevant to think about new tools that may be one time effect.

Dr. Alexandra Zimmer 19:47

Yeah.

Patrick Sullivan 19:48

You know, the time, the lag time between sort of testing someone. And if TB is like many other infectious diseases, when you get a result back, it may be hard to find that person. It may be hard to link them back into care. So there are all kinds of other advantages. And I think this is a time when we have to be thinking about the reality that we may have to do public health with more constrained resources for some period of time than we've had in the past. So these kinds of innovations are more important than ever, I think. And the time is right. I'm so grateful to get to talk to you about it as I'd say an earlier career colleague who's coming into this with real facility and talking about what these tools are. I appreciate that you've really articulate what their limitations are because I think often we dream about, um, we soar about the possibilities of AI, but the realities of what it takes to do that are left behind. And you've really given us a great insight into all that. So thank you for the work that you've done. Thank you for sharing with Annals. And we look forward to what you do next in this work and in your career. Thank you so much. That's very kind. Before we wrap up, is there anything else you want to add or anyone you want to acknowledge in terms of this work and who helped you along this journey?

Dr. Alexandra Zimmer 21:02

Oh, I mean, it takes my supervisor always says it takes a village for a PhD student. So I have a whole list of people in the back of my mind, but I think I'll just specifically thank my supervisor, Dr. Madukar Pai at McGill, my co-supervisor, Dr. Simon Grandjean Lapierre at Université de Montréal, of course, my team, my collaborators in Peru, Cesar Ugarte-Gil. It would not be possible without all of them. And I think one thing I really appreciate about my research on cough was when I look at my author line, I just see so much diversity in terms of the people who are working on it, both in terms of where they come from, but also, you know, their fields. I have machine learning experts, I have people who worked on the digital tool side and helped develop the app, clinicians, other epidemiologists, statisticians. And I think that really helped build a really nice narrative and picture of the research that may have been lacking before.

Patrick Sullivan 22:02

I mean, I just think it's a truth about discovery that the interesting things happen at the margins of different disciplines. And so, on the one hand, I appreciate that as an articulation of gratitude from you because we all get where we are because of folks who support us, but also as a sort of aspiration in our careers to combine, as you did. Yeah, like it's a real skill to be able to bring together these people who talk different languages and might say, like, I don't know anything about what she does, I do this, and see how those things come together. Like that, I think is the magic of doing research that's innovative and really moves things along because you're you're figuring out a way to get parts to work together that when it isn't necessarily obvious. And then I'll just like close from my perspective with that sense of gratitude and humility about what we don't know, what we do know and what we don't know. And the coolest stuff gets done when people can say, like, I know this, but I really like need some help about that. And this seems like bringing together Epi, but also some very cool technology and infectious disease expertise to move things along. So that brings us to the end of the episode. This is a great conversation. I appreciate so much your making kind to join us today and to share this these pieces about the science, but also about the process. I think many of our colleagues who are in doctoral training or in postdoctoral training really benefit from just hearing how people in this phase of their career put things together and do work that's really impactful. So it was a pleasure to have you and thanks for sharing this research and the story around it.

Dr. Alexandra Zimmer 23:38

Thank you so much for having me. I had- this was a great conversation.

Patrick Sullivan 23:42

I'm your host, Patrick Sullivan. Thanks for tuning in to this episode and see you next time on Epitalk. Brought to you by Annals of Epidemiology, the official journal of the American College of Epidemiology. For a transcript of this podcast or to read the article featured on this episode and more from the journal, you can visit us online at www.annals of epidemiology.org.