Our office will remain closed through Friday, September 18th as we continue to assess the damages caused by Hurricane Sally. ActiGraph team members are working remotely, however shipping delays should be expected at this time. We expect to resume regular business hours on Monday, September 21st. If you need immediate assistance, please contact us by email at email@example.com and we will respond as quickly as possible. Thank you for your continued support.
An Algorithm to Distinguish Mail Time From Accelerometer Wear Time in the Women’s Health Study
- Presented on May 29, 2013
Assessing physical activity (PA) with wearable sensors in epidemiologic studies often requires initializing activity monitors at the study site and mailing them to participants. In these studies, it is important to discern true PA data from mail transit data. Typically, participants are asked to log the date and time they start and stop wearing the monitor. Using an algorithm to identify “wear days” will reduce researcher burden when participants forget to log their information and provide a standardized method for eliminating mail transit data.
Purpose To develop and validate an algorithm to distinguish “mail days” from “wear days” for the ActiGraph GT3X+ accelerometer.
Methods GT3X+ monitors were initialized and mailed to 11 female volunteers (age = 62.8 ± 10.9 y) living in different locations in the US. After receiving the monitors, participants wore the GT3X+ for 1-5 days and recorded: a) date and time monitor was received in mail, b) times monitor was worn and removed during data collection, and c) date and time monitor was mailed back. Accelerometer data were plotted to identify features that discriminated mail days from wear days. These features were used to develop and train a linear discriminant algorithm and derive a linear classifier score. The algorithm and score were validated on accelerometer data in an independent sample of 10 participants in the Women’s Health Study (WHS) by calculating the classification accuracy of mail and wear days.
Results The mail time algorithm was: (7.908204e-06)(total daily counts) + (1.015293e-02)(total number of minutes with non-zero counts) + (-5.960652e-04)*(total counts between 3AM and 5AM). A score < 4.014542 predicted a mail day, while a greater score predicted a wear day. Using this algorithm, the percent correct classification of mail and wear days in the independent sample was 100%.
Conclusions The algorithm and linear classifier score were valid in discriminating mail days from wear days for the GT3X+ in a small sample. We aim to further validate the algorithm using a larger sample from the WHS. This will establish the usefulness of the algorithm in automatically removing GT3X+ mail transit data from analyses in large epidemiological studies. Funded by NIH R01 CA154647