0285 Measuring Sleep with the Apple Watch: A Comparison of a Machine Learning versus Traditional Algorithms to Actiwatch

Abstract Introduction Consumer-based actigraphy has seen rapid adoption in the United States and presents an underutilized opportunity and resource for ecologically valid sleep-wake monitoring. Furthermore, with the discontinuation of Philips Actiwatch, a reliable and scalable alternative is required. Apple Watch may be a promising solution, with prior data indicating strong concordance for activity counts derived from the Apple Watch compared to the Actiwtach. The present study extends previous findings to sleep periods as an outcome of interest. Methods A community sample of 40 adults wore an Actiwatch and Apple Watch for 7 to 14 days with daily completion of the consensus sleep diary. Sleep periods from both wrist-worn devices were calculated and compared with sleep periods reported on the sleep diary. Sleep based on Apple Watch was derived using two approaches: one that mirrored traditional actigraphy (ie, sleep-wake classification using the Cole-Kripke model), and another than utilized machine learning with steps and heart rate as inputs. Agreement of sleep periods between the wrist-worn device and the sleep diary was operationalized as percent overlap. Performance of Apple Watch compared to the Actiwatch was then evaluated using the ratio of percentage overlap with sleep diary. A ratio of 1.00 represents perfect agreement. Results The agreement between Apple Watch derived sleep periods and sleep diary was comparable to that with Actiwatches. The ratio of percentage overlap between the two devices was 1.00 when using the machine learning algorithm, and 0.97 when using the traditional actigraphy approach. When averaged across the sample, sleep periods derived from Actiwatch overlapped by 80.7% (95% CI [84.1% - 77.3%]) with sleep diary periods, and sleep periods derived from Apple Watch overlapped by 82.0% (95% CI [84.7% - 79.4%]) and 81.2% (95% CI [84.9% - 77.5%]) for the machine learning and traditional actigraphy approaches, respectively. Total sleep period from the Apple Watch produced lower mean absolute errors compared to Actiwatch (machine learning: 24.8 minutes, traditional actigraphy approach: 19.1 minutes). Conclusion Sleep periods derived from Apple Watch data showed strong agreement with Actiwatch data, with the machine learning algorithm showing slightly stronger performance compared to the Cole-Kripke algorithm. Support (if any)