AAN 2026 | AI and sleep biomarkers for predicting cardiovascular and neurologic disease

So an area where I’ve been working quite a bit recently has been deep learning and using the sleep data to try to make predictions about health and also about sleep disorders. What’s really unique about sleep, if you think about it, is that you are captive. You know, you’re doing nothing. That’s a little time where people leave you alone. So basically, you’re sleeping and you go through an automatic program, you know, of your REM sleep, non-REM sleep. And that really changes all your physiology with these different sleep stages. And it’s really the best moment to really monitor your entire health if you think about it if we could do it in a non-invasive way. And of course we have as part of these sleep studies for sleep apnea for narcolepsy etc there are millions of people who have gone through this monitoring with polysomnography where we measure the brain waves, we measure for EEG, for sleep stages, but we also measure the EOG, the eye movements, the muscle tone, the heart rate, we have an EKG electrode, measure breathing for sleep apnea, measure all kinds of different physiology at the same time. And all those, of course, change depending on your circadian time and your sleep stage, everything is very very well synchronized. So we we started, you probably know that machine learning started with this idea of supervised machine learning so it’s basically trying to imitate humans you know and that we have a lot of because it’s technicians They look at these sleep studies and say, oh, people stop breathing. It’s a sleep apnea. And they write it down. Oh, people are going into REM sleep. And then they write it down. So it’s a very easy method to really kind of do what’s called a supervised machine learning to learn to teach a network, to learn how to score automatically and replace human. And pretty much, I don’t want to go in detail, but with that technique, it works very well. But then the next step, which you probably know, it’s moving so fast that you barely have time to breathe in this field. There is this foundation model, which they use, for example the large language model where instead of really training for a specific task like language for example you don’t train to to detect narcolepsy or to detect you know apnea you just teach it the language of sleep. And what you do is you feed the signal and you train a network to reconstitute the signal in some ways. So this way, it kind of concentrates the data in what we call auto-encoders in a relatively smaller amount of information. And by learning to redo the signal from the signal, I know it’s a bit absurd, it’s actually learning the content and simplifying it. And once it’s trained, you have this foundation model that you can use for all kinds of different applications. And that’s exactly what a large language model is for, you know, for speech. You know, it’s really, you have taught it the structure of the language, and then it can answer questions. It’s quite amazing. So here we have done the same thing for sleep. We have 30,000 sleep studies, and now we’re actually working with more, like 200,000 sleep studies. And we taught the system to really recognize and reconstitute the sleep signal from the sleep signal. We hide it and we try to teach the computer to reproduce it. And we also did something very cute, very interesting. We tried to train the network to, for example, recognize the brain waves from the heart and recognize how the heart should change as a function of brain waves. And by doing that, we actually discovered very interestingly that there are some cases where we could use that data to predict diseases, especially when we try to predict the heart from the brain or the brain from the heart or breathing from the heart. And this is because probably when you are sleeping, all these things have to be very synchronized. And when your heart is not doing well, your heart rate doesn’t follow what it should do across the different sleep stages. And it detects this kind of abnormalities using this form of learning that’s called contrastive learning. I know it’s very complicated. I’m sorry if I’m going in detail. But basically, with that kind of training, we were able to predict the occurrence of many, many diseases. I wouldn’t say that we can, you know, predict if you will die tomorrow, but we could kind of predict you have an increased chance of having a heart attack or having all kinds of different problems. We are very excited because we have now created this foundation model, we call that foundation model, that has learned the language of sleep in a way, and houses different channels from the brain activity to the heart activity to, you know breathing, you know work all together in different sleep stages how it can be abnormal how it can predict diseases and we hope to now add things like devices, genetics, proteomics to really create like a more a very general model that can predict very accurately all kinds of different diseases. I want to mention that it’s not like we can say to someone, you know, I had people call me, oh, yeah, please analyze my sleep and tell me if I will die tomorrow. Of course, it’s not like that. We can predict an increased risk of deaths or we can predict an increased risk of heart disease and myocardial infection or things of that nature. But it’s only an increased risk. But I think as the model becomes more and more sophisticated and we add more modalities, it could very well be a very nice way to even study yourself longitudinally. You could imagine a time where you will analyze your sleep every month and then if there’s something that seems to appear it’s maybe a trigger for something wrong with your heart or something that you are not yet detecting. So I think we’re very excited in pursuing this general area.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.