Algorithms for using an activity-based accelerometer for identification of infant sleep-wake states during nap studies

Abstract Objective To determine the accuracy of using different algorithms on the output from an Actical accelerometer, a device normally used to measure physical activity, to distinguish sleep from wake states. Methods Thirty-one infants aged 10–22weeks wore the accelerometer on the shin for a daytime nap recording in tandem with polysomnography. Sleep–wake epochs were identified using four computations/algorithms: the zero-threshold computation, two common algorithms used for wrist-based devices (Sadeh and Cole), and a new algorithm developed for this study (count-scaled). Accuracy was examined in direct epoch comparison with polysomnography using 15-, 30- and 60-s sampling epochs. Results Overall agreements (accuracy) for sleep–wake states were >80% for all computations. The count-scaled algorithm sampling 15-s epochs gave the highest accuracy, with sensitivity (sleep agreement) at 86% and specificity (awake agreement) at 85%. Other computations yielded higher sensitivity at the expense of specificity. Another way to assess the accuracy of identification of sleep–wake states was to compare sleep parameter outputs. All computations and sampling epochs were significantly correlated with total sleep time ( r =0.76–0.88), sleep latency ( r =0.70–0.93), sleep efficiency ( r =0.76–0.87), and wake time after sleep onset ( r =0.41–0.53). The number of awakenings after sleep onset was overestimated by accelerometry. Conclusions The Actical accelerometer, designed to measure physical activity, can reliably identify sleep in infants during napping, with the count-scaled algorithm showing some advantages over other methods for accurate identification of sleep–wake epochs.