Development of Activity Type Classification Algorithms in Older Adults from Laboratory and Free-living Data

Purpose: To compare activity type recognition rates of machine learning algorithms trained on laboratory versus free-living accelerometer data in free-living older adults.

Methods: Thirty-seven older adults (21F and 14M ; 70.8 ± 4.9 y) performed selected activities (total of 35 min) in the lab while wearing three ActiGraph GT3X+ activity monitors (dominant hip, wrist, and ankle). Monitors were initialized to collect raw acceleration data at a sampling rate of 80 Hz. Fifteen of the participants (9F and 6M; 70.0 ± 4.3 y) also wore the GT3X+ in free-living conditions and were directly observed for 2-3 hours. Time- and frequency- domain features from the accelerometer signals of each monitor were used to train random forest (RF) and support vector machine (SVM) models to classify five activity type categories: sedentary, standing, household, locomotion, and recreational activities. All algorithms were trained on lab data (RFLab and SVMLab) and free-living data (RFFL and SVMFL) using 20 s and 30 s time intervals. Prediction accuracy of both types of algorithms was tested on free-living data using a leave-one-out technique.

Results: Overall classification accuracy using 20 s intervals for the lab-based algorithms was between 49% (wrist) to 55% (ankle) for the SVMLab algorithms, and 49% (wrist) to 54% (ankle) for RFLab algorithms. Overall classification accuracy of SVMFL and RFFL algorithms ranged from 58% (wrist) to 69% (ankle) and from 61% (wrist) to 66% (hip and ankle), respectively. Using 30 s intervals improved classification accuracy up to 71% (SVMFL ankle). Significant improvements in classification accuracy were observed for RFFL hip, RFFL wrist, and RFFL ankle algorithms (76%, 70%, and 76%) when three activity type categories were used: sedentary behavior, moving intermittently and locomotion.

Conclusion: Algorithms developed in free-living settings are more accurate in classifying activity type in free-living older adults than lab-developed algorithms. Our results suggest that future studies should consider using free-living accelerometer data to train machine-learning algorithms in older adults