Some Reflections on Better Sleep

Using RF reflection imaging, BodyCompass can track sleep position for better health while preserving privacy and comfort.

Nick Bild
4 years agoMachine Learning & AI

Sleep position is more than just a matter of comfort for many. Being in particular positions can prevent post-surgical bedsores, reduce sleep apnea events, or even be fatal in the case of epilepsy or sudden infant death syndrome. Moreover, tracking sleep positions over time can provide pertinent information about one’s health, such as in the case of tracking Parkinson’s disease progress.

Current solutions for tracking sleep positions typically make use of cameras or on-body or mattress sensors. These methods have drawbacks, however — cameras are considered an unacceptable invasion of privacy for most, and sensors tend to be uncomfortable and hinder good sleep.

Getting lots of mileage out of the radio frequency (RF) imaging techniques from MIT CSAIL professor Dina Katabi’s lab (see here and here), the same lab applies the technique to the problem of tracking sleep position with BodyCompass. By using RF reflection imaging, issues with privacy and discomfort are resolved.

BodyCompass captures time-sequenced data from an FMCW radio equipped with an antenna array over the course of the night. This data is fed into a fully-connected neural network. A convolutional neural network, very commonly used for image analysis, was not chosen because BodyCompass requires global pixel comparisons to account for how RF signals bounce around in space before reaching the radio. Training data was collected by having volunteers sleep wearing accelerometers. This ground-truth data provided the model with the information it needed to associate RF imagery with sleep positions.

For a new user, the BodyCompass model requires some additional training to achieve the best accuracy in sleep position prediction. The model was found to be 94% accurate when trained with a full week of data, however, even with a single night of data, 87% accuracy was achieved. The data collection burden need not be high — 16 minutes of data collection was sufficient for a very respectable 84% accuracy.

The primary reason that the model needs additional training for each user is due to room layout. RF waves reflect all about the room before returning to the radio, so the environment will significantly change the results. So, if any user decides to reorganize their bedroom, BodyCompass would need a tune-up.

Nick Bild
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles