S/o to my amazing co-author who is the actual sleep scientist for turning me into this and taking me on this ride! 💤🧠❤️🤖

We also interrogated the structure of sleep that the model learned and found that it learned more a complex structure of sleep, than standard approaches and could maybe be useful for answering how to optimally organize sleep stages overnight and over the lifespan.

It basically learned the gross physiology of sleep (as we currently understand it) on its own. But along with that, it learned lots of other features about sleep physiology that we could use to predict clinical and cognitive outcomes, even with limited data.

We trained a fancy neural network model to learn the multimodal structure of sleep from a minimal set of standard sleep study sensors (but EEG is probably doing a lot of the work for now) *without any labels* from > ten of thousand of nights of sleep.

When people were making these rules decades ago, they probably couldn’t have imagined the huge quantities of sleep that are publicly available to analyze today as well as the tools we now have to interpret them.

The stages that we use to label sleep are central, but we’ve known for a long time there are problems with them. There’s also lots of nuanced health information revealed by sleep that probably goes under used because of this.

Our maybe-dweeby special sauce is really strong baselines - the same powerful model trained on the same data to predict standard sleep stages. Turns out you can fine tune that to do lots of cool stuff too! Just not as well as our model we trained to learn sleep structure on its own ;)

TLDR we thought we could help start to address some core challenges in sleep science using deep learning (and self-supervision!) and along the way found that we can better predict standard and potentially novel clinical outcomes using this model.

Such a complex pattern