Predicting Activity Type from Accelerometer Data in Older Adults

Introduction Assessing time spent in different activity types may be important for early detection of mobility limitations in older adults. To date, accelerometer-based activity type prediction using machine learning algorithms have not been validated for this segment of the population. Therefore, the aim of this study was to use Random Forest (RF) models to predict activity type from accelerometer data in older adults.

Methods Thirty-five healthy older adults (mean ± SD age = 70.8 ± 4.9 years) wore 3 ActiGraph GT3X + accelerometers. The monitors were initialized to collect data at 80hz and were positioned on the dominant wrist, hip and ankle. Participants performed one of two activity routines (7 activities each, 5 min/activity) including sedentary (SED), locomotion (LOC), household (HOU), and recreational (REC) activities. Accelerometer data were downloaded and transformed to 1-second epoch data using the Actilife 5 software. For each monitor, the 10th, 25th, 50th, 75th, and 90th percentiles of the vector magnitude counts corresponding to each minute of activity were calculated. These features along with the corresponding activity type label were used to train seven RF models (hip, wrist, ankle, hip + wrist, hip + ankle, wrist + ankle, and hip + wrist + ankle) for prediction of SED, LOC, HOU, and REC activity type. A leave-one-out method was used to test the accuracy of each model.

Results Overall accuracy of the RF models in detecting activity type ranged from 82% to 88% using single monitor data, and from 92% to 95% when combining data from two or three monitors. The RF model with the greatest accuracy (hip + wrist + ankle) correctly classified SED, LOC, HOU and REC activities 94%, 99%, 94%, and 91% of the time, respectively.

Conclusion The RF models in this study accurately predicted activity type from a single or multiple accelerometers. Using machine learning models such as the RF method to detect activity type in free-living older adults may be useful for identifying mobility limitations.

Funded in part by NIH R01 CA121005