Research Study Abstract

Integrated Movelets Approaches For Predicting Human Movement Type Based On Multiple Accelerometers

  • Presented on June 17, 2013

Introduction Tri-axial accelerometers record the acceleration of people’s daily activity on three orthogonal directions. One fundamental question is how to decipher and interpret the acceleration signals into meaningful information such as types of human movement.

Purpose We provide statistical methods for predicting activity type and answer the following questions: 1) how well do accelerometry data reflect a given program of activities; and 2) given simultaneous accelerometry data from multiple sites of the human body, how could we effectively integrate the combined information?

Methods We propose the integrative movelet methods for predicting activity types. A movelet is a time series collected in a window of given length. The sets of movelets constructed from the accelerometry data with annotated labels are organized by activity types, i.e., “chapters”, which play the role of the accelerometry dictionaries for different activities. Predictions of accelerometry data without labels are provided through identifying the chapter that is most similar to the data in terms of mean squared error. Information from multiple accelerometers is integrated in three ways: 1) by building separate movelet dictionaries and combining predictions using voting (movelets voting); 2) by designing a joint dictionary (expanded movelets); and 3) by establishing a decision tree (movelets tree).

Results We tested our methods with the Aging Research Evaluating Accelerometry (AREA) study. 20 older subjects were instructed to perform 15 types of activities while wearing Actigraph GT3X at the hip, right and left wrist in lab sessions. The prediction precision of the proposed methods is shown in Figure 1.

Conclusions Based on our results, we propose that observational studies involving accelerometers could dramatically improve data quality by incorporation of a set of standardized “life” activities that could be performed at the time the devices were initially placed. While the home setting is preferable, if not feasible, then even a clinic setting would still result in enhanced data quality.

Acknowledgement This research was supported, in part, by the Intramural Research Program of the National Institute on Aging. He, Bai, and Crainiceanu were supported by Grant R01NS060910 from the National Institute of Neurological Disorders and Stroke. This work represents the opinions of the researchers and not necessarily that of the granting organizations.


  • Bing He 1
  • Jiawei Bai 1
  • Annemarie Koster 2
  • Paolo Caserotti 3
  • Nancy Glynn 4
  • Tamara B. Harris 5
  • Ciprian M. Crainiceanu 1


  • 1

    Johns Hopkins University, Baltimore, MD

  • 2

    University of Maastricht, Maastricht, Netherlands

  • 3

    University of Southern Denmark, Odense, Denmark

  • 4

    University of Pittsburgh, Pittsburgh, PA.

  • 5

    National Institute on Aging, Bethesda, MD

Presented at



, ,