Transition Detection for Automatic Segmentation of Wrist-Worn Acceleration Data: A Comparison of New and Existing Methods

Abstract

Introduction: Data from wrist-worn accelerometers often has an inherent natural segmentation that reflects transitioning from one activity to another. The aim of this study was to develop an activity transition detection method to realize this natural segmentation. Methods: Data was gathered from 16 participants who wore triaxial wrist accelerometers in a lab-based protocol and 47 participants in a free-living protocol. Change point detection was used to create a method for detecting activity transitions. The agreement between observed and predicted transitions was assessed by the Matthews Correlation Coefficient (MCC), Root Mean Squared Error (RMSE), and two additional metrics created for this task; the Ratio of Minimum Mean Distance (RMMD) and the Ratio of Sensitivity (RoS). The effects of varying combinations of acceleration axes were also investigated to determine the most effective set of axes. A novel post-processing technique was developed to mitigate a major limitation identified in current transition detection methods. Results: The developed transition detection method achieved a MCC of 0.763, a RMSE of 3.17, a RoS of 2.40, and a RMMD of 3.21, outperforming existing techniques. The post-processing technique developed improved the performance of all methods when identifying transitions. It was found that using solely the y-axis (vertical acceleration) allowed for optimal performance. Conclusion: Change point detection is a valid method for identifying transitions in activity using wrist-worn accelerometer data. The new post processing technique developed improves the performance of transition detection methods.