Application of Raw Accelerometer Data and Machine-Learning Techniques to Characterize Human Movement Behavior: A Systematic Scoping Review

Abstract

Background: Application of machine learning for classifying human behavior is increasingly common as access to raw accelerometer data improves. The aims of this scoping review are (1) to examine if machine-learning techniques can accurately identify human activity behaviors from raw accelerometer data and (2) to summarize the practical implications of these machine-learning techniques for future work. Methods: Keyword searches were performed in Scopus, Web of Science, and EBSCO databases in 2018. Studies that applied supervised machine-learning techniques to raw accelerometer data and estimated components of physical activity were included. Information on study characteristics, machine-learning techniques, and key study findings were extracted from included studies. Results: Of the 53 studies included in the review, 75% were published in the last 5 years. Most studies predicted postures and activity type, rather than intensity, and were conducted in controlled environments using 1 or 2 devices. The most common models were support vector machine, random forest, and artificial neural network. Overall, classification accuracy ranged from 62% to 99.8%, although nearly 80% of studies achieved an overall accuracy above 85%. Conclusions: Machine-learning algorithms demonstrate good accuracy when predicting physical activity components; however, their application to free-living settings is currently uncertain.

The Use of Inertial Measurement Units for the Study of Free Living Environment Activity Assessment: A Literature Review

“…Of all studies using more than one sensor, 67% use four or less sensors. [ 27 , 28 , 54 , 61 , 65 , 74 , 86 ]). Moreover, only three studies use more than ten IMUs—these studies use the IGS-180 suit consisting of seventeen IMUs [ 18 , 63 , 70 ].…”

“…The participants of these studies had to perform all of these three behaviors. A second slice of activities also seems to emerge and is composed of the following movements; running [ 17 , 78 , 93 ], lying down [ 54 , 62 , 67 , 94 ], going up and down a staircase [ 65 , 90 ]. It is notable that several studies are interested in the detection of postural transitions (sit-to-stand, stand-to-lie, etc.)…”

“…Others focused on younger cohorts, with participants between adolescence and under 30 years of age [ 38 , 43 , 73 , 74 , 78 , 80 , 93 , 100 ]. The remaining studies used the age, gender, and BMI of the participants as a criteria [ 62 , 65 , 76 , 88 , 99 ], or only the average age of the included participants [ 16 , 54 , 60 , 108 ]. One study in particular took only the best survivors [ 85 ].…”

“…Some work can be put in place in order to define which features are either redundant (two features are redundant when their values are correlated) or irrelevant and which are thus to be kept to perform the classification: this is feature selection. Almost half of the studies classifying activities (18 of them) used feature selection processes such as Principal Component Analysis (PCA) [ 41 , 99 ] or others [ 54 , 83 , 93 ]) before setting up their classification system. The vast majority of these studies obtained classification accuracies (accuracy or sensitivity) beyond 85 percent.…”

“…Vienne et al [ 19 ], Yang and Hsu [ 52 ], and Tedesco et al [ 48 ] also focused on the use of wearable sensors in clinical settings, but considered any kind of cohort. Attal et al [ 53 ] and Narayanan et al [ 54 ] analyzed articles related to activity classification algorithms and classifiers. Schwickert et al [ 55 ] and Henriksen et al [ 56 ] focused on other kinds of studies (studies based on fall detection and studies using one specific brand of sensors, respectively).…”