Comparison and Validation of Actigraphy Algorithms Using a Large Community Dataset: Algorithm Validation Study

Abstract Background For decades, the measurement of sleep and wake has relied upon watch-based actigraphy as an alternative to expensive, obtrusive clinical monitoring. At the time of this publication, we have relied upon a handful of algorithms to score actigraphy data as sleep or wake. However, these algorithms have largely been tested and validated with only small samples of young, healthy individuals. Objective This study aimed to establish the accuracy and agreement of conventional and traditional actigraphy algorithms against polysomnography, the clinical standard, using the diverse Multi-Ethnic Study of Atherosclerosis (MESA) sleep dataset. As a secondary objective, we examined algorithm and polysomnography agreement for key sleep metrics including total sleep time (TST), sleep efficiency (SE), and wake after sleep onset (WASO). Methods We assessed 5 well-established algorithms, including Cole-Kripke, University of California San Diego (UCSD) scoring, Kripke 2010, Philips-Respironics, and Sadeh, with and without rescoring across 1440 individuals (M age =mean 69.36, SD 8.97) from the MESA sleep dataset. We conducted epoch-by-epoch comparisons assessing accuracy, confusion matrix analyses, receiver operator characteristic curves (ROC), area under the curve (AUC), and Bland-Altman analyses for agreement. Results Primary results indicated all algorithms demonstrated accuracy between 78%‐80% with the highest accuracy by the Kripke 2010 (80%) algorithm followed closely by the Cole-Kripke (80%) and Philips-Respironics (80%‐79%) algorithms. In addition, moderate Cohen κ agreement and moderate positive Matthews correlations were demonstrated by all algorithms. Further, all algorithms demonstrated significant mean difference across sleep metrics. Conclusions The findings of this study establish that these traditional actigraphy algorithms can, with high accuracy, detect sleep and wake in large, diverse population samples, including older adults or populations at risk of health conditions. However, these algorithms may carry difficulty for precise assessment of sleep metrics, especially in cases of sleep disorders or irregular sleep.