Automatic Segmentation of Walking Strides in Wearable Accelerometry Data with Adaptive Empirical Pattern Transformation

Abstract

Quantifying gait parameters and ambulatory monitoring of changes in these parameters has become increasingly important for epidemiological and clinical studies. Wearable accelerometers provide objective high-density measurements of human gait dynamics through recording of acceleration. Many studies use accelerometry to objectively measure physical activity using the activity counts, vector magnitude, or number of steps. These measures use just a fraction of the information in the raw accelerometry data as they are typically summarized at the minute level. To address this problem, we focus on raw, sub-second level accelerometry data and define a set of gait characteristics based on these data. Additionally, to overcome the analytical challenges of these complex and voluminous data we develop automatic and unsupervised methodology for precise segmentation of stride patterns. PURPOSE: We propose Adaptive Empirical Pattern Transformation (ADEPT) and maximization-tuning procedure for automatic identification of individual walking strides from raw accelerometry data that uses data-derived baseline patterns, representing a population-specific strides. METHODS: Data were collected as a part of the study on Identification of Walking, Stair Climbing, and Driving Using Wearable Accelerometers, funded by the Indiana University CTSI grant and conducted at the Department of Biostatistics, RM Fairbanks School of Public Health at Indiana University. The study enrolled 32 healthy participants between 23 and 52 years of age. Participants wore accelerometers on a wrist, hip and both ankles during a 450-meter outdoor walk. RESULTS: ADEPT yields results that are in most cases visually indistinguishable from manual segmentation and reduces strides segmentation time radically. The average absolute deviation of estimated stride duration across study participants was 4.74, 1.42, 1.28 and 1.31 percent, for wrist, hip and both ankles respectively. CONCLUSIONS: : Our results indicate that the errors are small relative to the signal for all body locations suggesting that ADEPT is a robust and universal tool for segmentation of strides in accelerometry data.