So my portion of this talk is going to kind of set us up for the other great talks that we have coming up. So I will just kind of set the groundwork. So I'm a pediatric pulmonologist and sleep doctor at Stanford. So I have nothing to disclose. So my objectives are first to just identify the challenges in diagnosing sleep apnea, obstructive sleep apnea in adults and in pediatric populations and kind of define some of the next steps for us to consider as we, you know, think to address these diagnostic challenges. So we already know, I think all of us in this room are aware, sleep apnea is a problem and they think that the prevalence of this is only going to go up worldwide. It can have significant effects on our overall health, our quality of life, can have short-term, long-term effects, mental health, physical health effects, individual and community effects. So it's a big deal to the point where in the U.S. the Healthy People 2030 goals, which is kind of like the goals for the population health, is to actually increase the proportion of adults with symptoms of obstructive sleep apnea who actually seek medical evaluation. So I'm going to kind of set the stage of what are our current limitations of obstructive sleep apnea diagnosis within our current paradigm. So diagnosis is made by obviously first having some symptoms or daytime symptoms or some nighttime symptoms and then of course a sleep study either at home or in a lab. So this is just as a reminder we have the, you know, recently updated International Classification of Sleep Disorders. We look at AHI ranges. So this is just a reminder that adult and pediatric ranges are slightly different. So normal for adults 0 to 5, pediatric 0 to 1, some centers use 0 to 1.5. That AHI is a whole other talk in and of itself, but just wanted to remind you of that. So the limitations of screening for obstructive sleep apnea, right? So we have limited knowledge by a lot of in our community, primary doctors and community partners, so people who actually first interface with a lot of our patients. The patient's lack of awareness of need for screening or seriousness of the condition. I'm sure many of you have seen people as like, oh yeah, we all snore. I've snored for years and I, you know, I wake up and I gasp and then go back to sleep. No biggie, right? Or the patient's negative framework around even the diagnostic process or treatment. They're like, oh, I got to go get that sleep study and it's going to take months. Got to make an appointment. Got to drive the doctor. Just the thought of even screening for it is a big deal. And then that negative framework is actually justified by our poor coordination of our health system and screening for chronic issues. Many people will say, well, it took me two months to see the doctor and then another two months to get the study and then another two months to get the results and then another three months to get the machine, right? And then the pediatrics, I get the added bonus of, again, people having limited knowledge. A lot of people think it's very normal for kids to snore. There's also a limited number of pediatric sleep specialists. So, you know, we have just my, I'm already booking into March and April right now. Our pediatric patients also have different symptoms of OSA. So it's not that they're necessarily sleepy, but they're actually hyperactive. So it even looks differently and if you're not really aware, you may miss it. Now, our other limitation is access to testing. So in lab testing, as we know, can have long wait times. It's only one night. The environment can be challenging for those with special needs. So for adults or pediatric patients who are in a new place, you know, it can be really challenging to even sleep in such an environment. Home sleep testing often has no EEG or CO2 testing. And I know there's probably some great talks during this conference are looking at research on improving home sleep testing, but right now it's limited. And then it's often not recommended for our patients with comorbid conditions, those who have neuromuscular disease, lung disease, heart failure, et cetera. And then, again, the pediatrics part of it, a lot in lab is actually recommended for all pediatric patients. There are exceptions that apply. But generally that EEG and that CO2 data is really important to diagnose obstructive sleep apnea appropriately in this patient population. There's actually limited pediatric certified labs. A lot of labs and maybe some of that you work in actually will only see patients up to a certain minimum age. It's a burden for the family, right? Because it's not just one patient, it's two people, right? The parents have to take time off work or someone has to get a babysitter for the other kids and those kinds of things. And then even if they get there, will they tolerate the equipment? You have to put things up in their nose. Our techs work gallantly to try to get kids just to tolerate that nasal cannula and thermistor in their nose. And then there's also the concern of like, okay, well even if we diagnose obstructive sleep apnea, what are we going to do? So my kid has to have surgery? I'd rather not know. Or we have to do CPAP? Well, they're not going to tolerate it anyway, so what's the point? And then there's also the changing physiology and needs of kids, right? They're always growing, so maybe they're having sleep apnea symptoms at three, but will they have it at four? Maybe we should just watch and wait. What do we do? So there's added burdens of pediatrics. And now other limitation, we have utility and application of the diagnostics, right? So I think a lot of you in this room might be even thinking, well, does the AHI even paint a clear picture? What does the AHI really even tell us? Is one night enough, right? There's some research that's showing, well, if we do multi-nights, that's actually where we get a really good sense of what's going on with the sleep. And what about home studies? So I know insurance really likes home sleep studies, but is it sufficient to actually diagnose obstructive sleep apnea in most of our patients? And should we expand it to our pediatric patients? There's an argument for some that we should. And the thing I often am answering probably just about every week in my clinic is, well, what does mild, moderate, and severe sleep apnea in pediatrics even mean, right? So I think we have an idea of like the extremes, like the very mild case versus a very severe, but, you know, an AHI of 6.5, what do you have to do with that? People always wonder. And then treatment outcomes. So there's a known correlation between poor sleep and cardiovascular sequelae, metabolic dysfunction, neuropsychiatric disorders. However, if we look at a lot of the literature, we don't have a clear line back that says, well, if you treat it, all of those things get better, right? We're still struggling to actually show that. I think a lot of us believe that and know that and see it, but I think the data are really mixed on that. And again, there are limited outcomes in pediatric data for sleep apnea, or for positive airway pressure specifically. And really just more studies are needed to understand the potential directional link between OSA and other medical conditions. So I just mentioned all these problems. So how do we address these limitations? So first, we need to really understand the limitations of our screening, our testing, our diagnostic interpretation and our clinical application. And then now we need to consider the next steps of how to address them. This is where my wonderful colleagues will actually do this for us. So first, we need to consider maybe new strategies to redefine obstructive sleep apnea in both populations. Maybe explore alternative and emerging methods for diagnosis and explore application of machine learning in the diagnosis of obstructive sleep apnea. We have all this technology. Why don't we see if it can work for us? So in conclusion, so again, my part is really to set the stage for what we're going to talk about. But it's important to acknowledge the limitations for diagnosis of obstructive sleep apnea in both children and adults and to understand the barriers specific to pediatric patients. I'm always, always trying to advocate for my pediatric patients. And really for us to consider how to modify our current paradigm in order to build on current and past science and advance our ability to diagnose patients in a timely manner. Thank you for your attention. Good morning. I'm Dave Balachandran. I'm from the MD Anderson Cancer Center in Houston, Texas. Great to be with you all this morning. Our learning objectives for this brief session will be to describe a little bit more the limitations of the current definition of OSA. We're going to spend a lot of time understanding the importance of OSA phenotypes in the management of OSA. We're going to examine how cluster analysis helps us define these phenotypes and how we can use them. And I've never used the word imagine in a learning objective before, but we're going to imagine the future, where we want to go, how do we provide value to our patients by doing the best for them. Well, to define OSA, we've used the apnea hypopnea index for a very long time. For those of you who remember the Rechoff and Kales Manual from 1968, there have been several updates on this from 1999 to 2007. And most recently, in 2012, the ASM redefined a hypopnea as a 3% desaturant with an arousal. So if you use these different definitions, you can come to very different conclusions about whether the patient has OSA and how severe it is. And you can see just right here, if you use the CMS definition of OSA at 4%, you're going to lose a lot of patients that might have previously been characterized with OSA or with a more severe finding of OSA. And you can see the numbers right here. Severe OSA went from 33 to 28%. And absolute numbers of not having OSA went up to about a third in this cohort. So it's really important that we define OSA properly so that we don't misdiagnose or don't misunderstand the severity of a person's disease. So there is a need for change. And here are some of the reasons why we need change. We cannot use a one-size-fits-all model anymore for defining OSA and trying to treat it. We all know that there's challenges of using just one particular therapy, CPAP. It's not a one-size-fits-all. And when we use that model, we end up with suboptimal adherence, suboptimal effectiveness, and we define our patients as failing CPAP. What does that even actually mean? Maybe it wasn't the right treatment for them in the first place, or maybe we didn't go about it the right way. So we need to think about other treatments, and how do we know which treatments will work? That's what we're going to get into by thinking about phenotypes. And this really gets to the heart of the problem that if we don't define OSA well, we're going to continue to have trial after trial that doesn't show good outcomes in terms of hypertension, cardiovascular disease, heart failure. And when it comes to showing value to third-party payers, Medicare, or private insurance, they're going to push back at us and say, well, why are we spending all these billions of dollars on CPAP and testing if you aren't showing us value to the patients? We could use our resources better elsewhere. So that's really the imperative for us to change how we do things. So how do we do this? How do we evaluate OSA differently? Well, we can start by using pathophysiological characteristics. We can start by using phenotypes. And how we get there is by using things like cluster analysis, which we'll talk about in a little bit, and you're going to hear much more about it later, using big data and AI to help us define these things. So let's start with some of the earlier studies that use pathophysiology to help us better define OSA and link it most importantly to treatment outcomes because that's really the crux of the matter. We wanted to find a strategy which, if we find this, it impacts treatment this way. One of the early studies was using something called the POM criteria, where they looked at PCRIT, arousal threshold, loop gain, and muscle responsiveness as a way to try to characterize obstructive sleep apnea in these patients. And what they found is by doing different variations of this, you could actually impact treatment outcomes. It could actually predict a particular outcome. And that was really the start of this. But it goes much more than just pathophysiology. We really need to take a more holistic view of how we view a disease, and that's where this definition of phenotypes come in. So just to define phenotypes, it's a subtype of disease defined by a unifying, consistent, natural history, clinical and physiological characteristics, underlying biology, using biomarkers and genetics to find a predictable response to affect relevant patient outcomes. And that's really important. I stress that in the next bullet point that we really need to anchor these phenotypes to relevant clinical outcomes for them to be really meaningful. And one of the ones that is best understood right now is if you find complete concentric collapse on endoscopy, that really does predict failure of hypoglossal stimulation. So there you can see, here's a particular characteristic. It defines an outcome, and that's what we're trying to achieve. So what are some of the characteristics of phenotype? These are just some of them there. You could have a much longer list, but I just couldn't fit all of them on a slide. There's demographics. You can look at age, BMI, race, ethnicity, sex, education, all of those factors, symptoms, excessive daytime sleepiness, aneurysis, napping, cognitive issues. We could also look carefully at the PSG, like the POM criteria. Look at AHI, look at O2 sat, arousal index, T sat 90, oxygen saturation index, and even other things related. I don't even have things about the EEG that might be predictive. Comorbidities, depression, anxiety, cardiovascular disease, neurologic conditions, all of those could be factors in creating a phenotype. And this last category I have here is multi-domain. We'll get back to that in a little bit. But it's really important to tie all these phenotype characteristics back to treatment outcomes. Does it affect efficacy? Does it affect adherence? Does it affect which clinical outcomes? And does it affect survival? And that's really the crux of the matter. Let's go back to that green bar, because I think it's really instructive to show how we can combine some of these demographics, symptom, PSG, and comorbid factors, and really try to create phenotypes. Here are some of the ones that have been suggested recently. There's a path of phenotype of excessive daytime sleepiness in OSA. REM predominant OSA. Complete concentric palatal collapse in OSA. Excessive daytime sleepiness, insomnia in OSA. Younger females with insomnia, mild OSA. And you can see there's a variety of ways you can combine these different factors to try to create a more richer phenotype with multiple domains. And how do we know that we're doing this correctly? That's where understanding outcomes and cluster analysis comes in. Here are just some examples of where you take some of those multi-domain strategies and try to link them to outcomes. So we talked about excessive daytime sleepiness in OSA. Well, we know that patients treated with CPAP who have excessive daytime sleepiness and OSA have decreased blood pressure, better control their hypertension, decreased cardiovascular risk, and improved quality of life. Another example would be patients with supine OSA. About 50% of our patients who referred for OSA actually have supine OSA where the AHI is about twice what it was compared to the total. And there's some data to show that those patients may respond better to autopap strategies for their treatment. And that's been shown to improve their sleepiness, improve their wakefulness as well. This is a really interesting story where we have patients with REM predominant OSA. This tends to be, especially in the cohort of younger women who have a decreased sleep time and sleep efficiency and REM duration with desaturation. What they found is if they have a REM AHI greater than 15, that if you treated them, they had better hypertension control. And this was true even if their absolute AHI was less than five. So you would have called them no sleep apnea. But actually, if you treated their REM AHI, you may actually find that they benefit. And although it may be very hard to get the CPAP paid for in the circumstances, this is the kind of thing that we need to have data so that we can push for these kinds of strategies to take place. And it's really gonna come down to looking at these clusters to help us define these phenotypes. I know this is sort of a busy heat map slide, but let me just take you through a couple of the clusters so you can understand how this goes. So if you look at this cluster one right here, you have obese patients who are predominantly female. The red color means that they're about 100%. They have loud snoring. They feel unrest and awakening. You can see all these red bars kind of cluster in one sort of aspect. Here's another cluster, your general patient with obesity, elderly patient, loud snoring, stop breathing, unrested with hypertension and comorbidities of cardiovascular disease. That's another cluster. A third cluster could be obesity, mixed gender, could be male or female, loud snoring, stop breathing, excessive dietary sleepiness, but with a lot, not with a lot of cardiovascular disease, which may be related to their age or the predominance of female sex. So you can see by doing this kind of analysis, you can find these clusters that may have a meaningful clinical import. This is another way of kind of looking at this idea of cluster analysis. This table, you're actually gonna wanna read horizontally. And so what you're gonna try to find is like, for example, we already told you a little bit about REM-related OSA. And you can see that there's clinical features, there's a pathophysiology, and there's an outcome. And by mapping it this way, it actually tells us where the gaps are in our evidence, that where we need to do work to find the phenotypes that actually have clinical impact. You can see right here, outcomes is pretty weak right here, and genetics and genomics is really empty. And that's kind of the paths that we need to go forward to help really make robust phenotypes that have clinical meaning. And this really nice review paper, if you are interested in OSA phenotypes, I really asked you to look at this when it came out in CHEST in 2020. They were really able to define certain phenotypes that were meaningful. And they came up with these four based on these, the classic phenotype that we all see, patients who are predominantly male, younger than 65, who have significant sleepiness, they have high HI's, and some oxygen saturation. That correlates with their having a risk for drowsy driving, for having incident cardiovascular disease, and they really are the patients that we've shown to some extent have the best benefit for being treated with CPAP. There's also a second subtype, obese males with significant oxygen saturation, but because these are elderly patients, they already have a lot of coincident morbidity with hypertension and cardiovascular disease. And paradoxically, we haven't been able to show much benefit in this cohort, possibly because they already have comorbid disease. And so perhaps there needs to be more of a focus on treating the comorbid disease, rather than just the sleep apnea. Another subtype, women who have insomnia, who are overweight, we know that they have a lot of difficulty with CPAP adherence, and that they have lower incidences of cardiovascular disease and because they also have insomnia, maybe we need a better strategy. We can't just give them CPAP. Maybe we also need to think about using CBTI, cognitive and behavioral therapy for their insomnia, so we get the best impact for those patients. And then we can have another cohort, younger patients with a lot of upper airway disease. They can have significant hypoxemia. They don't have cardiovascular risk at this time because of their age perhaps, but we also know that they have low CPAP adherence and we need to come up with other strategies. Maybe we can move them to oral appliance therapy if they're not so hypoxic. Maybe we can think of other things. And this is the way I think we need to start to think in the future about defining these subsets of patients so that we can provide the best clinical outcomes for them. We can also use our PSG. These are two phenotypes that come out of the PSG, one being with severe hypoxemia and a lot of apneas. See this in our older patients. These probably will do really great with CPAP, but there's another group, lots of disturbance from their sleep apnea, but not significant hypoxemia. These patients are at risk for neurocognitive issues and maybe using CPAP, maybe using oral therapy or using other treatments for them may be the right strategy. So we really need to think about how we define it, how we do this. So where do we go from here? And this is what my colleague is gonna help us understand a lot better than this, but we need to start structuring getting all this data. And if you have noticed, the ASM and their latest quality imports is asking practitioners to list some of these factors as part of their QI metrics. And I think creating databases like that is gonna be really important for us as going forward, knowing risk factors, putting in clinical details, putting in pathophysiology. When we get to this point, biomarkers for obstructive sleep apnea and outcomes and understanding the genetics behind it, if we combine all that information in a proper way, we will start to find other phenotypes, I'm sure, that may be instructive and help us. And so it's by doing this iterative process, defining the phenotypes, testing it, looking at the symptoms, finding biomarkers, looking at treatment outcomes, looking at the genetics behind it and continually improving that process that we're going to make progress in really defining OSA and in really getting the best outcomes for our patients. And so I'll just conclude with this. Right now, what we're trying to do is create these OSA phenotypes so that we can tailor our treatment to the patient to get them the best outcome and the best value. And with these strategies that we're headed towards, with cluster analysis, big data, neural networks and artificial intelligence, what I hope one day is that we'll actually be able to find genotypes which will help us direct a therapy and that's really our goal in the end. So thank you for your attention today. I'm gonna make one plug if you don't mind. We have a session on sleep wellness and yoga and traditional Hawaiian medicine at 7.15 on the terrace here. Tomorrow at 7.15, I hope you can all attend. Thank you. Good morning, everyone. Is coffee kicking in? It's funny about the last time when CHEST happened in Hawaii, I had to give a talk about jet lag. And I was teaching people how to fight it, how to adjust. And I think now we probably, I was completely wrong. We shouldn't be fighting it. We should take advantage of it and be in early morning sessions here. So thank you for your attention. I'm gonna make one plug. Tomorrow at 7.15, I hope you can all attend. Thank you. Early morning sessions here. So I'm Tavio Arkemascu. Thanks for inviting me to give a talk about the current and emerging methods for diagnosing obstructive sleep apnea. I don't have any significant disclosures. I will only discuss FDA-cleared devices. And the real disclosure is that this is really hard to cover within 12 to 15 minutes, but I will try. So my objectives today are to review the current methods and standards for diagnosing obstructive sleep apnea. And you already heard a little bit from my colleagues about the current standards. So I'm not gonna insist too much about it. We will look through a few like in a whirlwind about the new and emerging technologies and devices coming in our field. And I would like to add another objective here, which is imagine. So we all know the gold standard for diagnosing obstructive sleep apnea is still polysomnography, PSG. And you've heard about certain limitations and advantages of the PSG from the previous talks. It does include a lot of signals. I listed there. And this slide has a lot of words, I understand. And I'm just gonna go through a couple of major limitations. One, you see how many signals we're collecting, how many wires are out there. And the information we're collecting, sometimes we don't use it all. So there is some redundancy. But there is also a lot of fatigue for visual inspection by technologists and sleep practitioners. And there is some fatigue and some inter-rater reliability problems. Now we do have issues also with night-to-night variability and first night or lab night variability or effect. And that's kind of the major component of the intrinsic variance of what we measure from polysomnography. And of course, because of all these limitations you've heard also from Caroline, we see polysomnography less and less. So it's become less and less available. It's costly. It's laborious. Now you've also heard that the best we came up with is AHI, apnea hypopnea index. And that's kind of in a way structuring part of the limitations we see with our technology. So if we spoke about polysomnography as the gold standard for diagnosing sleep apnea, high-call oligosomnography or home sleep apnea testing or out-of-center testing or ambulatory testing, the silver standard because it's becoming more and more accepted that this is one alternative way of diagnosing obstructive sleep apnea. It records fewer channels. Hence the made-up name oligosomnography. And rarely we record EEG, but there are some devices that do. Some use nasal pressure transducers or thermistors. Some just use other types of signals. And you're probably all familiar with peripheral arterial tonometry-based devices, photoplatysmography, basic technology there. And now we see more emerging devices that use repetitive movements of the mandible, for example. Patches with all types of sensors in the pillows, in the night sheets and so forth. There is less night-to-night variability if we collect data over multiple nights so we can overcome a little bit of that source of intrinsic variance. And we don't see as much as the first night effect. Maybe the earlier devices that were bulkier and harder to attach to the body, maybe we would see more of that initially. But now we see less and less of that. And the best we came up for HSAT is the respiratory event index, REI. So we've been trying over many decades to diagnose obstructive sleep apnea using different tools in the clinic. And we started out with can we predict presence and severity of obstructive sleep apnea by using simply questionnaires. And we went from Berlin, stop, stop, bang. And I wanted to list there one of the more recent ones which is called BASH-GN. It's based on BMI, age, snoring, hypertension, gender, and neck circumference which seems to have a nice performance. If you look at the area under the receiver operating characteristic curve there, it seems to be superior to the other ones. Of course, it has to be validated. Why I point out about this questionnaire is it has been developed using artificial intelligence, unsupervised, completely agnostic methodology. So we may be able to come up with better tools in terms of questionnaires. We've always seen certain trends and certain phenotypes that include demographic data, gender, age, smoking status. I keep reminding our fellows we should not forget about smoking status. Even ex-smoking, it's actually still a significant factor and we should collect that as part of our clinical impression overall. Of course, anthropometric measurements, if one looks at the literature over time, neck circumference and BMI have been kind of the basis of many models. And of course, the clinical characteristics such as retrognathia, macroglossia, and so forth. Now we've used imaging studies also to assess upper airway statically or dynamically. We started to use metabolic inputs to define these phenotypes and maybe endotypes too. And of course, we're going to use different physiologic systems and signals to one, detect sleep apnea. Is it present or not? To stratify it, how severe is it? What's the burden? What's the association? And I think that's the most important part, looking at what's the comorbidity that is significant and is associated with obstructive sleep apnea such as coronary artery disease, stroke, type 2 diabetes. And of course, you've heard a little bit about also there are new methods, new models developed to assess how effective one therapeutic modality may be in a particular patient. So when I started to prepare this talk, I had about 60 slides and I had a completely different plan on how to approach this. And this was kind of the backbone of how I would go about it. But then I thought I would start putting some photo icons there in case you want to take pictures and see all the information from some slides because I'm not going to be able to go that systematic. But the backbone is we're using physiologic inputs and these go as polysomnography signals or different types of signals like repetitive mandibular movements or PPG-based signals. We have different sources and these are important for what models we develop. Where do you get your data? Is there any particular bias? And those are the studies that we use data. And then there are different signal processing methods that include preprocessing, decomposition of the signal, feature extraction, feature selection, and processing. And I'm not gonna go through details about these. These are more of a kind of cookbook for how do you assess this for your mathematical model that you're trying to approach it in the model. I touched a little bit earlier and we'll talk more about the different hardware, the different devices that we're using for picking up some of those signals. And of course, what classifier in the end do we use? Do we use a specific mathematical model? Do we use artificial neural network? Do we use the latest and the newest advances in computational science so that we can actually pick up signals that we don't necessarily see with our naked eye and you hear more about AI. I like this recent articles that went back and looked at how do we assess sleep stages and looked at all the studies that have been published only in the past five years and found out that sometimes we use models where we combine different signals or sometimes we rely exclusively on one signal. I see EEG is still going strong and it's probably not surprising because we have now the computational methods. We have the big data capabilities where we can actually use the big data from the EEG. As far as sources, the same paper reviewed and find out that a lot of times, it's this mixed bag of one individual study but then there are available databases and studies that are out there that provide the data we can actually train and test different models. And physioapnea is one of them, for example, sleep heart health study, MESA and so forth. I mentioned about classifiers, what methodology, mathematical methods you use for dividing that entity, creating those phenotypes and this is a full alphabet soup. I'm not gonna go into details but at least you have there the legends for some of those major methods that are used nowadays in artificial intelligence. So let's go into some examples. This is one interesting paper recently that looked at can we use that oligosomnography concept, few signals to analyze but to pick up certain entities, certain phenotypes that we are not necessarily able to refine further with our previous methodologies. And here it's looking at the entropy of the snoring. So the microphone is collecting a signal and just looking at that pattern of signal using artificial intelligence method. In this case, they use the support vector machine. What it is, it's really creating the best hyperplane between the different data points and this is not only 2D, there are more than two variables so it becomes almost unimaginable if you want in terms of how many dimensions you're doing this hyperplane but it's you're dividing up the different points and then you find out what's the performance of this model. So in this article, in this author's hands, they found about area under the receiver operating characteristic curve of about 6.68. So not bad using snoring entropy and then they used EKG and then the respiratory effort. As it was mentioned earlier, we probably should focus less about is the AHI less than one or less than five or more? Is it five to 15 or we should actually look at the comorbid burden and the next two papers I'm gonna show you as examples, they looked at that particular thing which is what's the impact of obstructive sleep apnea? Can we find out who is gonna have cardiovascular risk? And this is one paper as I mentioned, they looked at different variables, fasting glucose, probably cannot see very well there but I think the important segment here is that we can use different methods, you see on the right hand side logistic regression all the way to support vector machine and assess the performance of the model by just doing the kitchen sink approach. You're throwing all the variables you have there and maybe you'll be able to find a phenotype or be able to classify the condition. And this is a second paper also looking at cardiovascular risk and mortality prediction, again, important heart outcomes where they developed three models. One looked at all the sleep parameters and came up with a sleep model that only included four features in the end, this ODI, the mean heart rate, heart rate area, area under the heart rate curve and the low frequency to high frequency ratio. So perhaps not surprising since these came up in multiple models in the past is significant. They looked at clinical parameters and developed a clinical model that included gender, age, hypertension, type two diabetes and systolic blood pressure. So it's interesting, right? Without artificial intelligence, you would say hypertension and systolic blood pressure are highly collinear, you shouldn't be using them if they're not independent of the model but with AI methods you can. And then the mixed model that includes all of those and they found area under the receiver operating characteristic curve for the mixed model of about 0.78. So very encouraging. Again, we're looking at different methodologies and different parameters. Of course, the authors were smart enough to look at are these variables as predictive in all ages or all genders and they found out that not. Their contribution was different as you would probably expect it as a clinician. I don't know why they did a cutoff of 60 for both men and women. I would probably have gone a little bit differently about it but this is what they found and they show you the contribution of each one of these parameters in the different categories. So shifting gears a little bit about new technologies. So speaking about new technologies, in the figure here on the left in A, it's the depiction of the first photoplethysmograph, PPG technology in an article that was published in 1937. And then you see in the middle, the PPG of the watch path device that you're probably all familiar with that came up with the idea of inflating and deflating the cuff there. And the latest we'll talk a little bit about is also the night owl device, which has been miniaturized and it's just a button on the finger there. But the bottom line of the photoplethysmography signals is that it picks up the pulsatile nature of the arteries in the periphery. And these are the four devices or technologies, I should say, cleared by FDA at this time using PPG-based signals. One of them, the sleep image ring, is really a software as a medical device approved, not the device per se. You're probably all familiar with a watch path device based on PPG, watch path one, the disposable one. Ectosense, this company from Belgium, which was acquired recently by ResMed, came up with a night owl. I showed you an earlier image about it. And they miniaturized it because they were able to use the signal to go directly into the cloud. So you only need the signal detection methodology. Everything else is miniaturized because of that. And then the newest, the kid on the block, the Belong, the ring, which is using two different AI algorithms. One is to pick up sleep and the other one is to define respiratory events. Shifting a little bit gears, although it's about the sleep image basis, it's cardiopulmonary coupling. We've looked at heart rate variability for a few decades now and found out that if you just look at RR variability and you decompose that as a frequency analysis and then you look at the R amplitude or QRS amplitude changes because of the impedance of the chest changing with respiration and multiply those two signals by different frequency bins, you come up with a cross-spectral power. You multiply it by the coherence, which is really looking at the time series, looking at the fact that one signal that you're collecting now does have actually an association with the previous signal. You come up with these ways of categorizing the heart rate variability and the EKG-derived respiration. This is what EDR stands for. And I'm showing you here two examples from two different papers from the same group from Robert Thomas's initial work, looking at somebody with high-frequency coupling periods. This is stable sleep, if you want, and you see some desaturations here. And this is kind of low-frequency coupling so that they picked up these patterns that you can define unstable sleep, associate with sleep apnea with the desaturations and call low frequency. And then very low is during wakefulness. And a more kind of dramatic example on the right-hand side where I look at multiple desaturations and somebody with very, very little stable sleep. The majority of it, this is low-frequency coupling, unstable sleep, and some wakefulness with very low frequency. So shifting gears a little bit about, I showed you the night owl picture earlier. So they pick up the PPG signals from the periphery, from the finger, miniaturized device. But I think this is a major game changer for our field because they came up with a very interesting idea. They said, if you have an REI or an HI of 14.9, that may not be extremely different from a 15.1. And it may be actually within the intrinsic variability or inter-score reliability. So what they did when they validated their technology, they looked at, against polysomnography, but they had two, actually three score sets for the polysomnography and looked, what's the intrinsic variability? And they said, if you were just to flip the coin, it would be 50% variability. But if you look at these areas around cutoffs of five, 15, and 30, we would accept only lower than 33%. And that's an arbitrary cutoff. And they called these near-border zones. So they did near-border labeling. In other words, you have more than three or four categories, not only normal, mild, moderate, and severe, but you have normal or mild in between. And you'll see the HI band for those, mild or moderate, or moderate and severe. And when they added the near-bording labeling or near-bording zone, either for 3% or 4%, their accuracy of categorizing the disease improved significantly. So in other words, you're gonna put normal or mild in both normal and mild and assess this way. And the performance was pretty outstanding when you compare it to previous devices and previous classification systems. Another technology is using the repetitive mandibular movements as correlates of esophageal pressure changes, which would correlate with sleep apnea events. And they use also machine learning, pretty miniaturized chin-based sensor. And it goes into the cloud and then get assessed. And then they found actually great performance. And on the next slide here, I'm showing the performance on the left-hand side. This is the confusion matrix. So if you look on the diagonal, these are the performances, accuracies for those four categories. No OSA, mild, moderate, and severe. And they used the same near-border labeling that was published a few months before by the NiteOwl folks. And they found that the accuracy improved even more. Interestingly, the cutoffs are different because the variability for their data set was different. So this is more work for us to see what's gonna be the cutoff. But you've probably done this in your clinical practice many times before. When somebody has an HI of 14.9, is that really different than a 15.1? And don't you look at symptoms and don't you look at comorbidities and the entire clinical picture? So I think we're moving into that direction using also the data science capabilities. And I couldn't leave you without picking up some extra technologies coming into the field. This is one study that looked at radar and the 60 gigahertz frequency continuous wave radar that is placed in the ceiling of the sleep lab room and with a detector on the chest. And it can be used for both PSG and HSAT. It turns out the accuracy and the kappa statistics are very good, so very promising. Could this be as a sole signal used in the past? Possibly. Or it could be in combination with other things. And I think, imagine the sky is the limit. Thank you so much. All right, aloha everyone. This is the final presentation of our session, Applications and Limitations of Machine Learning in Diagnosis of OSA. I'm Miranda Tan from the Stanford University School of Medicine. The objective of today's talk will be to learn all about machine learning in 15 minutes. Just kidding. We'll explore ways machine learning can potentially overcome diagnostic challenges in OSA by reviewing a few of its applications in OSA, understand how ML can be leveraged to better define it and highlight its current limitations. So first, let's start with why we should even reconsider inventing this, reinventing this proverbial OSA wheel, right? So as Caroline articulated very beautifully this morning, the entire continuum of OSA diagnosis has its limitations. There are barriers to screening, access to testing and issues with what truly means to have OSA, particularly in the absence of strong, consistent evidence for associations and comorbid conditions as all of these panelists have discussed. So basically the square wheels on our OSA bike will not move forward unless we reconsider our approach. So ML may help us improve upon what we already know about obstructive sleep apnea. Sleep medicine is in fact primed for machine learning with its large labeled data sets acquired through PSGs, ongoing nightly monitoring information with our PAP data and EHR information. But the data are complex. There's heterogeneity in people interpretation, spatial temporal relationships. The EHR data is not always integrated with sleep data as you may know. And analysis often requires the expertise of an expensive data scientist. But we shouldn't give up. We should embrace the complex layers of data available. It's increasingly recognized that sleep apnea is caused by multiple pathophysiologic mechanisms. And AHI fails to capture this breadth of physiologic heterogeneity. So traits that contribute to OSA include neuromuscular collapsibility, arousal threshold and loop gain to just name a few. So the combinations of these factors define a unique group of patients with OSA. And importantly, they may associate with comorbidity much more closely. ML thus may be useful technique to further identify these unique subtypes to target patient outcomes. But at the very minimum, automation with machine learning will improve the sleep study process since tech scoring is expensive, labor intensive and imperfect with the inter-score reliability. So let's see how machine learning can be used to facilitate OSA diagnosis. Before we diagnose OSA, we should start to try and screen for it. Cheng's group aimed to screen for obstructive sleep apnea using an ML approach using anthropometric features only. They were able to develop an accurate model for moderate to severe sleep apnea and severe sleep apnea. So the anthropometric features studied are all listed here, as you can see, they have many. Of these features, visceral fat was found to be the most heavily weighted feature in the risk screening models in this Asian population. So this study suggests that machine learning can improve screening for obstructive sleep apnea using anthropometric features only and can this potentially improve access to care. In continuing this theme of using objective characteristics to determine the presence of obstructive sleep apnea, the Sagitt group studied the ability of age, sex, BMI and race to predict obstructive sleep apnea in a global cohort of about 17,000 patients. Following the development of their ML algorithm using the Sagitt cohort, the training model was subsequently used prospectively in comparison with SOP-BANG in both the Sagitt and the sleep heart health patients. What they found in their study was that the AUC of the models were higher than the standard logistic regression in both training and validation sets and that the ML could predict OSA without patient-reported symptoms better than prior screening questionnaires with questions. So am I sensing a trend here? If machine learning can better predict sleep apnea with only objective measures as opposed to traditional symptom questionnaires, then can machine learning improve screening and thus access to care, TBD? So machine learning is not restricted solely to adult applications. It has also been shown to be reliable in diagnosing pediatric OSA in this recent systematic review and meta-analysis. So 19 machine learning studies met the inclusion criteria and a total of 4,700 unique pediatric patients were included. So this study found that decreasing pool sensitivities and increasing pool specificities as OSA severity worsened, thus reflecting the well-known threshold effect of diagnostic test accuracy meta-analysis. But as you can see here, very high specificity was reached for the severe OSA group, which was accompanied by moderate sensitivity. The result was also accompanied by an AUC of .94 in severe OSA. So although the ML studies included in this meta-analysis mostly used SpO2 as part of their training algorithm, if we only look at ML studies that use the SpO2 as part of their training algorithm, then the sensitivity and specificity of the severe OSA group notably increased to 75% and 96% respectively. The moderate sleep apnea also improved to a sensitivity of 75% and 90% respectively. So overall, this has been a really good start to using machine learning to diagnose OSA for pediatric patients. This study highlights how automating machine learning to diagnose pediatric OSA may improve time to diagnosis and thereby access to care. This is highly encouraging findings since children with moderate to severe OSA are at higher risk of cardiovascular and neurocognitive morbidity. So they would really benefit from this early diagnosis if possible. So now let's shift gears. As we've heard from our excellent speakers this morning, the AHI is imperfect. It's unidimensional, it's imprecise, and definitely imperfect from an outcomes perspective. We need better ways to understand obstructive sleep apnea rather than these arbitrary cutoffs that we're currently using. So the next several slides will focus on ways to redefine obstructive sleep apnea using machine learning. First, we'll start with the hypoxic burden. Ali Azerbaijan's group sought to develop a severity measure for obstructive sleep apnea using OSA-related hypoxemia to understand if this can significantly reproduce CBD-related mortality. So a crude explanation of hypoxic burden is that it's just basically the area under this curve. As we can see in the diagram, the decrease in SpO2 related to the sleep event, respiratory event here, is used to calculate the area under the curve. So the total hypoxic burden was defined as the sum of all the individual burdens over total sleep time. So their study analyzed two cohorts, the Mr. Oz cohort and the Sleep Heart Health cohort. And what did they find? That the hypoxic burden strongly predicted CBD mortality and all-cause mortality in the Mr. Oz cohort. They also found that the hypoxic burden in the highest two quintiles had a hazard ratio of 1.8 and about 2.7. Similarly, the hypoxic burden in the highest quintile had a hazard ratio of 1.96 in the Sleep Heart Health study. Interestingly, hypoxic burden was variably associated with the other PSG parameters that we're very familiar with. So AHI, ODI, time spent under 90%. So separately, Zipzer's multicenter clinical-based cohort in 2021 also found that hypoxic burden predicts incident cardiovascular events and all-cause mortality, which is in line with the previous population studies with hypoxic burden. So if hypoxic burden is a stronger predictor of incident CV and mortality, can we improve patient outcomes by targeting hypoxic burden instead of the AHI? So many alternative metrics have been studied to either redefine or define obstructive sleep apnea in addition to the hypoxic burden. In the interest of time, I'll just list the ones that are currently being studied here for future reference, and I'll briefly discuss a couple. So we talked about the hypoxic burden. Also what's being investigated is the odds ratio product. So what this does is it quantifies sleep death. It's derived from EEG, and it's used to predict the differences between wakefulness and sleep. And the idea is that the odds ratio product can measure susceptibility to adverse neurocognitive outcomes for obstructive sleep apnea. Arousal intensity is also very hot. Arousals from sleep can vary in intensity, and that the idea is that these arousals may be associated with different neurocognitive outcomes as well as cardiovascular events. And then finally, the respiratory event duration. Have you ever seen some respiratory events on the PSG last longer than others? Well, the idea is that could this be related to a lower arousal threshold, or could this be related to maybe a high loop gain because of ventilatory control? Food for thought. And then as Tavi mentioned earlier, cardiopulmonary coupling, mandibular movement, snore signals, these are all currently being investigated, and likely, it's 2023, many more will come. So utilizing machine learning comes with its share of logistical, security, ethical, and legal obstacles. So machine learning needs large volumes of quality data. So not garbage data, quality data. We can't achieve these data sets until there's standardization of the data that we're trying to understand, and interoperability between the healthcare systems are required to collect all that data together. Data should also be representative of the population. Because ML programs learn from the data provided, healthcare inequities that already exist may be further amplified. So it's called like the gold standard paradox, where we think we know something, and we use that as a standard. But in reality, what we thought was the gold standard was completely false. So for machine learning, we try to step away from that. And then finally, there are some machine learning models that may not be explainable or transparent. This is called the black box methodology of deep learning or deep neural networks. I'm sure some of you in the room may have heard of deep learning. So it takes even scores of data and says, well, this is your answer, this is sleep apnea. But then you, as a doctor, are like, well, why, how does this make sense? So that is the black box. This is in contrast to explainable AI. What we really want is to understand how did the AI algorithm derive at this answer? What is my probability that this is correct? So we kind of need to further unravel this black box of deep learning, and hopefully we'll get there at some point. Okay, so where do we lie for the immediate future for machine learning and diagnosis of obstructive sleep apnea and overcoming all these challenges? So first of all, there's adoption of data sharing standards. So now the EHR is recognizing that, hey, we have to help these doctors and health systems. Let's make more interoperability between all the healthcare systems so that this data can be properly aggregated and studied. And then recently in 2021, the FDA realized that machine learning is here to stay, like it or not, and they released this good machine learning practice. So what this is is guiding principles of how machine learning algorithms should be evaluated. So I'll go over briefly these 10. The first one is multidisciplinary expertise is leveraged throughout the total product lifecycle. What this means is we shouldn't be working in isolation. We should be thinking about the problem from start to finish. So from the patient to the doctor to the lab tech to everyone involved, the social worker, et cetera. So anyone and everyone in the process should be involved and at the stakeholder table. Then the next principle is that clinical study participants and datasets are representative of the intended patient population. This is what we touched on earlier where historic studies were previously poorly sampled for like minorities and women in general. So now we have to make a conscious effort to make sure that datasets that we're using to interpret are very like unbiased. Next we have selected reference datasets are based upon best available methods. So what this basically means is make sure the data that you're choosing to study is actually clinical relevant, right? So does the color of the hair is that clinically relevant or do you really think it could matter for obstructive sleep apnea? So these are the things to think about. Next focus is placed on the performance of the human AI team. So what this means is don't just say, AI algorithm, go figure it out. It's kind of like the patient going to chat GPT and saying, do I have obstructive sleep apnea? So what this really begs is that there should always be a physician there and it's not to replace a physician but to work in conjunction with them. And I'll just go over the last one, which is users are provided clear essential information. So meaning that the data shouldn't be, I mean, the answer shouldn't be muddy. It should be very clear in line with what we're trying to intend to do and in which cases this would be sleep apnea. Okay, so I know many of you must be pumped to learn more about AI since you're here since seven o'clock in the morning. So here are some articles that I recommend for you to start your AI journey and in sleep. I also highly recommend the Coursera course by the Stanford professors. So we have in the upper left-hand corner by JAMA, how to read ML articles and then bottom right, yeah, right. It's an application of AI in the diagnosis of sleep apnea. And then in summary, you know, understanding the multi-level heterogeneity of the data we receive is of high clinical and translational importance in OSA. And, you know, mindfulness of current limitations in ML will help us develop equitable algorithms in the future. So, and this concludes our discussion portion of the panel. I'd like to thank my awesome co-panelists here, Caroline, Dave, and Tavi. It was extremely smooth to work with all of you. You're great and you truly are the dream team. And I'll also be remiss not to acknowledge Saadia Faiz, who is our shadow graphics editor and Keita Paul, who's a data scientist who worked on our last recent, our most recent ML project together. And she ingrained in me garbage in, garbage out. So you always want to make sure the data that you have is really good quality. And thank you all for coming. Thank you.

Machine learning

obstructive sleep apnea

diagnosis improvement

pediatric populations

alternative metrics

patient outcomes

large volumes of quality data

data standardization

transparent ML models