Can we use the Apple Watch to measure sleep reliably

Introduction: Individuals and researchers are increasingly interested in using consumer‐grade wearable devices to gain insight into health and wellbeing, including sleep. Sleep researchers are interested because of the potential to monitor much larger numbers people than polysomnography or clinical‐grade systems devices would permit. This study aims to investigate the feasibility of using an Apple Watch for sleep/wake monitoring and compare its performance to the clinically validated Philips Actiwatch Spectrum Pro, under free‐living conditions. Methods: We recorded 26 nights of sleep was from 13 healthy adults (9 male, 4 female). Each participant was asked to wear both devices on their non‐dominant wrist for two consecutive days. We extracted activity counts from the Actiwatch, and classified 15‐second epochs into sleep or wake using the Actiware Software (version 6.0.9). With the Apple Watch, we extracted raw acceleration data (at 50Hz), calculated Euclidean Norm Minus One (ENMO) for the same epochs, and classified them using a similar algorithm (with a sleep/wake threshold based on data from another participant). We used a range of analyses, including Bland‐Altman plots and linear correlation, to visualize and assess the agreement between Actiwatch and Apple Watch. Results: The Apple Watch had high overall accuracy (97.3%), high sensitivity (99.1%) in detecting sleep, and adequate specificity (75.8%) in detecting wakefulness. The total sleeping period is highly correlated with a Pearson Correlation coefficient of 0.99 between the two devices. On average, the Apple Watch over‐estimated the total sleep time by 4.65 min. Discussion: Popular consumer‐grade devices like the Apple Watch (of which over 30 million have been sold) provide an opportunity to conduct longitudinal studies at very large scales. Our initial results indicate promising performance in comparison to a clinically‐validated device. Optimization of the classification algorithm for high‐resolution acceleration data is likely to further improve the results. Further study is also needed to assess performance over a long period of time in ecological settings, and against polysomnography in a clinical environment.