Challenge of Detecting Personal Deviations and Trends in Sensor Based Activity Data.

INTRODUCTION: Physical activity and health are closely linked. Therefore, monitoring movement behavior is of great interest e.g., to monitor a patient's physical state. Nowadays it is easy to record movement with a smartphone. The aim of this work was to develop a concept to detect trends based on personalized movement behavior recorded with a smartphone.

METHODS: A first prototype with a control chart was designed. Since this approach did not prove suitable for analyzing activity data for trends in practice, a second prototype was subsequently developed with a statistical trend test (Mann-Kendall test (MK test)). It was extended by the Yue-Wang correction approach to be able to deal effectively with serial correlation. Furthermore, the traditional trend modeling using Theil-Sen slope was extended by three additional models to be able to represent non-linear trend shapes.

RESULTS: Movement behavior can be highly variable, which leads to wide control limits when using control charts. As the lower control limit was always in the negative range the use of a control chart was impossible for this use case. The evaluation results of the second prototype confirm the choice of a non-parametric test, as well as the decision for the Yue-Wang correction factor. Furthermore, it could be determined that the MK test is robust against outliers. The number of detected trends increases with increasing significance level. The MK test is also suitable for detecting step-like trends.

CONCLUSION: Live trend detection is not straightforward with the MK test but can be simulated by overlapping time periods. In the future, trend modeling should be extended even further, as it plays a major role in the concept. The sensitivity of the test can be increased by means of various parameters.