Abstract

Background: Objective measures using accelerometer-based activity monitors have been extensively used in physical activity (PA) and sedentary behavior (SB) research. To measure PA and SB precisely, the field is shifting towards machine learning-based (ML) approaches for calibration and validation of accelerometer-based activity monitors. Nevertheless, various parameters regarding the use and development of ML-based models, including data type (raw acceleration data versus activity counts), sampling frequency, window size, input features, ML technique, accelerometer placement, and free-living settings, affect the predictive ability of ML-based models. The effects of these parameters on ML-based models have remained elusive, and will be systematically reviewed here. The open challenges were identified and recommendations are made for future studies and directions.

Method: We conducted a systematic search of PubMed and Scopus databases to identify studies published before July 2017 that used ML-based techniques for calibration and validation of accelerometer-based activity monitors. Additional articles were manually identified from references in the identified articles. Results: A total of 62 studies were eligible to be included in the review, comprising 48 studies that calibrated and validated ML-based models for predicting the type and intensity of activities, and 22 studies for predicting activity energy expenditure.

Conclusions: It appears that various ML-based techniques together with raw acceleration data sampled at 20–30 Hz provide the opportunity of predicting the type and intensity of activities, as well as activity energy expenditure with comparable overall predictive accuracies regardless of accelerometer placement. However, the high predictive accuracy of laboratory-calibrated models is not reproducible in free-living settings, due to transitive and unseen activities together with differences in acceleration signals.