0302 Sleep Staging Classification from Wearable Signals Using Deep Learning

Abstract Introduction Typical data derived from a wrist worn device include accelerometer and photoplethysmogram (PPG) sensor signals . These reflect underlying movement, heart rate, and vascular dynamics that contain sleep stage information. We investigated the ability of a deep learning network to map raw data from such sensors to estimated sleep stages defined by full polysomnography scoring. Methods A convolutional neural network (CNN) was proposed for application to raw PPG (green light at 25 Hz) and 3D accelerometer data (also sampled at 25 Hz). The CNN had 70 hidden layers and output labels were mapped to four classes (wake, light sleep, deep sleep, and REM sleep) where light sleep is defined as Stages N1 and N2. The CNN was pretrained using 1654 records of finger PPG data from the Multi-Ethnic Study of Atherosclerosis (MESA) sleep records. The system was then further trained and evaluated on an internal set of 184 records obtained from adults (mean age = 68) with corresponding scored PSG sleep stage labels. Data augmentation techniques were used to create additional training data. The system was then tested using a withheld data set of 16 records. The overall performance of the system was evaluated by calculating two stage (wake versus sleep) and four stage accuracy and Cohen’s kappa values (𝜅). Results The overall performance for two-stage wake/sleep classification was an accuracy of 0.94 and 𝜅=0.79. For four stage classification, the accuracy was 0.79 and 𝜅=0.66. A comparable figure for expert human scoring four-stage class is accuracy of 0.8-0.85 and 𝜅=0.7-0.75. Conclusion Raw accelerometer and PPG signals contain a significant amount of information related to underlying sleep stages, and can be trained to produce hypnograms which approach the accuracy of human scorers. This may provide utility for both multi-night clinical use and underlying research in sleep science. Support (if any) This research was funded by Google Inc.