Hip and Wrist Accelerometer Algorithms for Free-Living Behavior Classification

Abstract

Purpose Accelerometers are a valuable tool for objective measurement of physical activity (PA). Wrist-worn devices may improve compliance over standard hip placement, but more research is needed to evaluate their validity for measuring PA in free-living settings. Traditional cut-point methods for accelerometers can be inaccurate, and need testing in free-living with wrist-worn devices. In this study we developed and tested the performance of machine learned (ML) algorithms for classifying PA types from both hip and wrist accelerometer data. Methods Forty overweight or obese women (mean age = 55.2 ±15.3 yrs; BMI = 32.0 ± 3.7) wore two ActiGraph GT3X+ accelerometers (right hip, non-dominant wrist) for seven free-living days. Wearable cameras captured ground truth activity labels. A classifier consisting of a random forest and hidden Markov model classified the accelerometer data into four activities (sitting, standing, walking/running, riding in a vehicle). Free-living wrist and hip ML classifiers were compared to each other, to traditional accelerometer cut points, and to an algorithm developed in a laboratory setting. Results The ML classifier obtained an average of 89.4% and 84.6% balanced accuracy over the four activities using the hip and wrist accelerometer, respectively. In our dataset with an average of 28.4 minutes of walking or running per day, the ML classifier predicted an average of 28.5 minutes and 24.5 minutes of walking or running using the hip and wrist accelerometer, respectively. Intensity-based cutpoints and the laboratory algorithm significantly underestimated walking minutes. Conclusions Our results demonstrate the superior performance of our PA type classification algorithm, particularly in comparison to traditional cut-points. While the hip algorithm performed better, additional compliance achieved with wrist devices might justify using a slightly lower performing algorithm.

Measuring Children's Physical Activity

“…We made this decision based on the lower validity of transforming data from a wrist-worn accelerometer obtained in conditions resembling free-living conditions into PA intensities, energy expenditure, and activity type classification compared with a hip-worn one in current data transformation methods (10,17,23,30). However, the literature is not conclusive in this area, and new data processing techniques currently emerge at a rapid pace, which does not seem to slow down any time soon given the recent large surveillance studies like UK Biobank and NHANES using wrist accelerometers.…”

“…In the current literature, wrist placement has generally been found to be less accurate than waist placement in PA behavior classification (9,10,17,23,24,30,33,35), although certain activities, such as basketball and dancing, are more accurately recorded by wrist-worn accelerometers than hipworn ones (9,30). Thigh-placed accelerometers accurately classify time spent lying down, sitting, standing, and moving for use in sedentary behavior analyses (18,22) and PA types, including cycling and walking up and down stairs (26).…”

An Activity Recognition Framework Deploying the Random Forest Classifier and A Single Optical Heart Rate Monitoring and Triaxial Accelerometer Wrist-Band

“…For instance, the survey of National Health and Nutrition Examination (NHANES) changed the location of accelerometer measurements from the hip to the wrist in 2011 [10]. One of the most prominent advantages of using wrist-worn sensors for activity recognition is better compliance compared to other measurement locations [9]. …”

“…The same author presented another wrist-based activity recognition framework in which they used the same device, but a multilevel classification approach using RF and HMM. They reached an 85% accuracy in classifying sitting, standing, walking/running and riding in a vehicle [9]. ActiGraph GT3X+ was also exploited in the study done by Trost et al in which an average classification accuracy of 88% was achieved for activities including sitting, standing, walking, playing basketball, lying down, running and dancing.…”

“…The framework covers some of the principal building blocks required for an activity monitoring system that works only based on a single PPG-ACC wristband. The system operates based on a well-known supervised classification algorithm, random forest (RF) [7], which has been widely tested in the literature [2,8,9]. The data processing comprised of label correction, signal segmentation, feature extraction, parameter examinations and feature reduction is comprehensively explained, and the corresponding strengths and weaknesses are discussed.…”