Power-efficient real-time approach to non-wear time detection for smartwatches

There is a growing interest in the use of smartwatches to collect real-time physical activity and health care data. Much of this information is useful only when the person is wearing the smartwatch. Collecting data during other times consumes precious battery life and potentially communication bandwidth to relay the information. An approach that can correctly classify temporal regions during which the watch is not worn has the obvious advantages of both improving battery life and minimize communication bandwidth. There have been several methods proposed to determine wear time using passive devices such as accelerometers. However, they are offline procedures where data are collected continuously during data collection phase and non-wear times are excluded later and before analysis. Smartwatches' computational power allow us to perform several preprocessing steps instantly and concurrently with the data collection. We propose a power-efficient and real-time approach to non-wear time detection method which is capable of providing labels for 15-sec epochs accurately (accuracies for wear time and non-wear time are 97.68% and 96.78%, respectively). We also show that detecting non-wear periods in real-time ameliorates the battery requirements by as much as 50%.