0068 ESTIMATION OF SLEEP STAGES USING CARDIAC AND ACCELEROMETER DATA FROM A WRIST-WORN DEVICE

Abstract

We investigated the ability of a wrist-worn tracker to estimate sleep stages in normal adult sleepers. Such a device could be useful in simplifying sleep research and in increasing public knowledge of sleep. Movement and cardiac data was collected from 60 adult subjects (36 M: 24 F, ages 34 ± 10 yrs) wearing two wrist worn devices (left and right hand) containing a 3D-accelerometer and an optical photoplethysmogram (PPG), while undergoing a sleep stage study using a Type III home sleep testing device. The accelerometer was used to generate various features of movement; the PPG records cardiac peaks generated by each heartbeat, and can be used to determine heart rate and heart rate variability metrics. The sleep study was scored independently by two registered PSG technicians, using consensus AASM scoring rules. Using these labels, an automated classifier and post-processing rule was developed to label 30-second epochs as one of Wake/Light/Deep/REM (note that Stages N1 and N2 were combined into a single “Light” classification). The estimated performance of this automated classifier system was calculated using a leave-one out validation method. The performance metrics were Cohen’s kappa (measures the level of agreement greater than chance) and per-epoch accuracy (percent of epochs correctly labeled). The estimated Cohen’s kappa was 0.52 ± 0.14 for left hand wear, and 0.53 ± 0.14 (right hand). The per-epoch accuracy was 69%. Across the population, there was no statistically significant bias in the estimated durations of the wake, light, deep and REM stages versus the gold standard measurements. These results suggest that a wrist worn device with movement and cardiac sensors can be used to determine sleep stages with a reasonable degree of accuracy in normal adult sleepers, but without the cost and artificial sleep environment of a sleep laboratory. The reported performance figures are similar or better than previously reported results from non-EEG based sleep staging using combinations of cardiac, respiratory and movement information. This work was supported by Fitbit.