The development of a wear-site detection algorithm of accelerometer data by applying machine learning techniques
Refereed conference paper presented and published in conference proceedings
CUHK Authors
Full Text
There are no full text file(s) associated with this record. |
Other information
AbstractPurpose:
Many research studies have adopted accelerometry as the measure for movement behaviors. Traditionally, the site in which the devices are attached to the body must be determined prior to deployment as it affects the choice of data extraction method and cut points, and it cannot be determined retrospectively. The ability to detect the device wear-site after collection could allow greater flexibility for deployment and less reliance of adhering instructed protocols. Thus, in this study, we developed and examined the efficiency of a machine learning-driven wear-site detection algorithm based on raw accelerometer data.
Methods:
60 participants (30 adults, 18-65 years; 30 youths, 9-17 years) wore tri-axial ActiGraph accelerometers at three sites (waist, wrist, and chest) and conducted a series of lab-based physical activities. Raw accelerations, at 100Hz, of activity segments were extracted and used for analyses. The Long Short-Term Memory approach was used for the development of the classification algorithm. 80% of the data was used for model training and the remaining 20% was used for validation.
Results/findings:
Models for classification between three (waist, wrist, and chest) and two (waist and wrist only) were developed and evaluated separately. Both models resulted in high accuracies in wear-site detection: 0.94 and 0.98 for three and two wear-sites, respectively. The performances of the algorithm was similar for adults and youths.
Conclusions:
The developed algorithm demonstrated promising accuracies in terms of detecting the wear-site of accelerometers. This implies that researchers can determine wear sites retrospectively and then apply appropriate cut points for activity classification. Furthermore, wear-site neutral algorithms can be developed, which would allow researchers to measure movement behaviors of participants more accurately.
Many research studies have adopted accelerometry as the measure for movement behaviors. Traditionally, the site in which the devices are attached to the body must be determined prior to deployment as it affects the choice of data extraction method and cut points, and it cannot be determined retrospectively. The ability to detect the device wear-site after collection could allow greater flexibility for deployment and less reliance of adhering instructed protocols. Thus, in this study, we developed and examined the efficiency of a machine learning-driven wear-site detection algorithm based on raw accelerometer data.
Methods:
60 participants (30 adults, 18-65 years; 30 youths, 9-17 years) wore tri-axial ActiGraph accelerometers at three sites (waist, wrist, and chest) and conducted a series of lab-based physical activities. Raw accelerations, at 100Hz, of activity segments were extracted and used for analyses. The Long Short-Term Memory approach was used for the development of the classification algorithm. 80% of the data was used for model training and the remaining 20% was used for validation.
Results/findings:
Models for classification between three (waist, wrist, and chest) and two (waist and wrist only) were developed and evaluated separately. Both models resulted in high accuracies in wear-site detection: 0.94 and 0.98 for three and two wear-sites, respectively. The performances of the algorithm was similar for adults and youths.
Conclusions:
The developed algorithm demonstrated promising accuracies in terms of detecting the wear-site of accelerometers. This implies that researchers can determine wear sites retrospectively and then apply appropriate cut points for activity classification. Furthermore, wear-site neutral algorithms can be developed, which would allow researchers to measure movement behaviors of participants more accurately.
All Author(s) ListNg J.Y.Y., Ha A.S., Zhang J.H., Cheng J., Jiang G., Hui S. S. C.
Name of ConferenceISBNPA XChange 2021
Start Date of Conference08/06/2021
End Date of Conference10/06/2021
Place of ConferenceNew Zealand
Country/Region of ConferenceNew Zealand
Year2021
LanguagesEnglish-United States
