|
Abstract:
|
In recent years, accelerometers have been widely deployed in large health studies to provide continuous and objective measurement of human physical activity. Due to limitation of hardware implementation, researchers used to be able to access only the reduced data (e.g. step counts, activity counts, etc.) that were processed by the on-device chips. While recent technology and data analysis advances have allowed the device to collect high frequency acceleration time series that provide rich information about device wearers’ detailed physical activity characteristics, many researchers were still interested in analyzing the count data because of the friendlier size, straightforward interpretability and off-the-shelf analytic tools. Therefore, multiple methods to convert the high frequency acceleration time series into count data were proposed. In this talk, I will discuss the implementation of a few methods for such conversion on a single accelerometry dataset. I will show how each method might affect the results of the data analysis through various statistical modeling. This study provides a general guideline of how raw accelerometry data are processed and reduced to count data.
|
