Accelerometer-Based Machine Learning Categorization of Body Position in Adult Populations.

Published

Conference Paper

This manuscript describes tests and results of a study to evaluate classification algorithms derived from accelerometer data collected on healthy adults and older adults to better classify posture movements. Specifically, tests were conducted to 1) compare performance of 1 sensor vs. 2 sensors; 2) examine custom trained algorithms to classify for a given task 3) determine overall classifier accuracy for healthy adults under 55 and older adults (55 or older). Despite the current variety of commercially available platforms, sensors, and analysis software, many do not provide the data granularity needed to characterize all stages of movement. Additionally, some clinicians have expressed concerns regarding validity of analysis on specialized populations, such as hospitalized older adults. Accurate classification of movement data is important in a clinical setting as more hospital systems are using sensors to help with clinical decision making. We developed custom software and classification algorithms to identify laying, reclining, sitting, standing, and walking. Our algorithm accuracy is 93.2% for healthy adults under 55 and 95% for healthy older adults over 55 for the tasks in our setting. The high accuracy of this approach will aid future investigation into classifying movement in hospitalized older adults. Results from these tests also indicate that researchers and clinicians need to be aware of sensor body position in relation to where the algorithm used was trained. Additionally, results suggest more research is needed to determine if algorithms trained on one population can accurately be used to classify data from another population.

Full Text

Duke Authors

Cited Authors

  • Jarvis, L; Moninger, S; Pavon, J; Throckmorton, C; Caves, K

Published Date

  • September 2020

Published In

  • Comput Help People Spec Needs

Volume / Issue

  • 12377 /

Start / End Page

  • 242 - 249

PubMed ID

  • 33047112

Pubmed Central ID

  • 33047112

Digital Object Identifier (DOI)

  • 10.1007/978-3-030-58805-2_29

Conference Location

  • Germany