Discrimination of sleep and wake periods from a hip-worn raw acceleration sensor using recurrent neural networks

Abstract

ABSTRACTThe use of accelerometers has become an established method in population research. Accelerometers are small body-worn sensors that can monitor movement patterns and cycles over several days under free-living conditions. A key requirement for any accelerometer-based analysis is the reliable discrimination of sleep and wake episodes. However, many studies using hip-worn sensors either instruct the participants to remove the sensor over night or use algorithms developed for wrist placement which are known to perform poorly on hip data. Here we present a new algorithm to differentiate sleep from awake time in raw hip acceleration data using a machine learning approach. Validated on sleep estimates from a subset of the Tromsø Study, the proposed algorithm outperformed the standard algorithms by discriminating sleep and wake correctly in 93.81% (95% CI: 0.72) of the time (F1 score = 0.93, 95% CI: 0.01). The value of our algorithm lies in the improved recognition of sleep and movement in large studies where manual scoring is unfeasible due to the amount of data. Even though the algorithm was developed to learn characteristic hip movements, it can easily be adapted for other sensor placements or study populations.