Improving Hip-Worn Accelerometer Estimates of Sitting Using Machine Learning Methods | JoVE Visualize

Area of Science:

Biomedical Engineering
Physical Activity Epidemiology
Machine Learning in Health

Background:

Accurate measurement of sedentary behavior is crucial for understanding its health impacts.
Hip-worn accelerometers are commonly used in large cohort studies, but their accuracy in estimating sitting time is limited.
Existing methods may not adequately capture sitting and standing transitions.

Purpose of the Study:

To develop and validate a machine learning algorithm for improved estimation of sitting time using hip-worn accelerometers.
To compare the performance of the new algorithm against standard accelerometer cut-point methods.
To assess the algorithm's accuracy in detecting sitting, standing, and transitions in free-living conditions.

Main Methods:

Utilized data from 30 breast cancer survivors wearing both hip-worn accelerometers and thigh-worn activPAL devices for 7 days.
Developed a random forest classifier trained on activPAL data to identify postures and transitions in 5-second windows from hip accelerometer data.
Employed mixed-effect models to compare algorithm estimates with activPAL data and analyze differences across various bout lengths.

Main Results:

The machine learning algorithm demonstrated moderate accuracy in predicting postures (stepping 77%, standing 63%, sitting 67%) and transitions (sit-to-stand 52%, stand-to-sit 51%).
Errors in transition detection were primarily observed during short sitting bouts (≤ 2 minutes).
The standard accelerometer cut-point method significantly differed from the activPAL, overestimating short bouts and underestimating long bouts of sitting.

Conclusions:

This study presents one of the first algorithms for hip-worn accelerometers trained on free-living activPAL data for sitting and standing detection.
The developed algorithm shows promise in identifying prolonged sitting, a behavior linked to adverse health outcomes.
Further validation and refinement in larger, diverse cohorts are recommended to enhance the algorithm's generalizability and clinical utility.