Continuous Movelet Transformation in Application to Individual Walking Strides Segmentation in Accelerometry Data

• Activity Number: 571 - Statistical Signal Processing Applied to Physical Activity Research
• Type: Topic Contributed
• Date/Time: Wednesday, August 1, 2018 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Marta Karas* and Jaroslaw Harezlak and Marcin Straczkiewicz and William Fadel and Ciprian Crainiceanu and Jacek K Urbanek
• Companies: Johns Hopkins Bloomberg SPH and Indiana University Bloomington and School of Public Health-Bloomington, Indiana University and Indiana University and Johns Hopkins University and Johns Hopkins University
• Keywords: Wearable Computing; Accelerometry; Physical Activity; Walking Segmentation; Pattern Recognition

Abstract

Quantifying gait parameters and ambulatory monitoring of changes in these parameters has become increasingly important for epidemiological and clinical studies. Wearable accelerometers can provide objective high-density measurements of human gait dynamics through recording its acceleration. Due to complexity and volume of accelerometry data, automatic and unsupervised methods for precise walking segmentation are needed. We propose to employ the continuous dictionary learning framework to identify strides (two subsequent steps) from sub-second level accelerometry data of walking. We define data-derived baseline patterns, which we name as movelets, representing a population-specific stride. Next, we perform two-step strides segmentation by combining pattern-recognition with a maxima-detection approach to precisely identify beginnings and ends of individual's strides. We demonstrate the proposed method using accelerometry data collected during a 450-meter outdoor walk of 32 study participants wearing accelerometers on a wrist, hip and both ankles. We validate the performance of the method and discuss individual-specific gait characteristics.