Search

Few methods for classifying physical activity from accelerometer data have been tested using an independent dataset for cross-validation, and even fewer using multiple independent datasets. The aim of this study was to evaluate whether unsupervised machine learning was a viable approach for the development of a reusable clustering model that was generalisable to independent datasets. We used two labelled adult laboratory datasets to generate a k-means clustering model. To assess its generalised application, we applied the stored clustering model to three independent labelled datasets: two laboratory and one free-living. Based on the development labelled data, the ten clusters were collapsed into four activity categories: sedentary, standing/mixed/slow ambulatory, brisk ambulatory, and running. The percentages of each activity type contained in these categories were 89%, 83%, 78%, and 96%, respectively. In the laboratory independent datasets, the consistency of activity types within the clusters dropped, but remained above 70% for the sedentary clusters, and 85% for the running and ambulatory clusters. Acceleration features were similar within each cluster across samples. The clusters created reflected activity types known to be associated with health and were reasonably robust when applied to diverse independent datasets. This suggests that an unsupervised approach is potentially useful for analysing free-living accelerometer data.

Background: Technological advances have seen a burgeoning industry for accelerometer-based wearable activity monitors targeted at the consumer market. The purpose of this study was to determine the convergent validity of a selection of consumer-level accelerometer-based activity monitors., Methods: 21 healthy adults wore seven consumer-level activity monitors (Fitbit One, Fitbit Zip, Jawbone UP, Misfit Shine, Nike Fuelband, Striiv Smart Pedometer and Withings Pulse) and two research-grade accelerometers/multi-sensor devices (BodyMedia SenseWear, and ActiGraph GT3X+) for 48-hours. Participants went about their daily life in free-living conditions during data collection. The validity of the consumer-level activity monitors relative to the research devices for step count, moderate to vigorous physical activity (MVPA), sleep and total daily energy expenditure (TDEE) was quantified using Bland-Altman analysis, median absolute difference and Pearson's correlation., Results: All consumer-level activity monitors correlated strongly (r > 0.8) with research-grade devices for step count and sleep time, but only moderately-to-strongly for TDEE (r = 0.74-0.81) and MVPA (r = 0.52-0.91). Median absolute differences were generally modest for sleep and steps (<10% of research device mean values for the majority of devices) moderate for TDEE (<30% of research device mean values), and large for MVPA (26-298%). Across the constructs examined, the Fitbit One, Fitbit Zip and Withings Pulse performed most strongly., Conclusions: In free-living conditions, the consumer-level activity monitors showed strong validity for the measurement of steps and sleep duration, and moderate valid for measurement of TDEE and MVPA. Validity for each construct ranged widely between devices, with the Fitbit One, Fitbit Zip and Withings Pulse being the strongest performers.

Objectives: The GT1M ActiGraph has been replaced by the triaxial GT3X which includes an inclinometer to detect postures. The purpose of this study was to investigate comparability of the GT3X to the GT1M and to develop activity intensity cut-points for the vector magnitude of the GT3X (VM(GT3X)) in children. Additionally, the study aimed to assess the validity of the GT3X inclinometer for detecting postures., Design: Forty-nine children aged 10-15 were tested during semi-structured activities in a laboratory setting (lying, sitting, standing, Nintendo Wii boxing, walking and running)., Methods: Expired respiratory gases were measured continuously using the Cosmed K4b(2) portable metabolic system. Simultaneously, ActiGraph counts by a GT3X and a GT1M were recorded., Results: Significantly higher counts were found for GT3X vertical mean counts per second (vcps) and antero-posterior counts per second (apcps) during running, relative to the GT1M. Cut-points for the VM(GT3X), developed using Receiver Operator Characteristics (ROC) curves (development group N=32, validation group N=17), were <3 counts s⁻¹ for sedentary (cross-validation: 87% correctly classified), 3-56 counts s⁻¹ for light (cross-validation: 70% correctly classified) and >56 counts s⁻¹ for moderate to vigorous (mvpa) (cross-validation: 86% correctly classified). The inclinometer correctly classified standing 20%, lying 15%, sitting 94% and the off position 45% of the time., Conclusions: The vcps from the two monitors differ for certain activities. Cross-validated cut-points for the classification of VM(GT3X) counts into sedentary, light and mvpa are presented. Posture classification by the GT3X should be interpreted with care, as misclassifications are common., (Copyright © 2012 Sports Medicine Australia. Published by Elsevier Ltd. All rights reserved.)

Introduction: Most accelerometer-based activity monitors are worn on the waist or lower back for assessment of habitual physical activity. Output is in arbitrary counts that can be classified by activity intensity according to published thresholds. The purpose of this study was to develop methods to classify physical activities into walking, running, household, or sedentary activities based on raw acceleration data from the GENEA (Gravity Estimator of Normal Everyday Activity) and compare classification accuracy from a wrist-worn GENEA with a waist-worn GENEA., Methods: Sixty participants (age = 49.4 ± 6.5 yr, body mass index = 24.6 ± 3.4 kg·m⁻²) completed an ordered series of 10-12 semistructured activities in the laboratory and outdoor environment. Throughout, three GENEA accelerometers were worn: one at the waist, one on the left wrist, and one on the right wrist. Acceleration data were collected at 80 Hz. Features obtained from both fast Fourier transform and wavelet decomposition were extracted, and machine learning algorithms were used to classify four types of daily activities including sedentary, household, walking, and running activities., Results: The computational results demonstrated that the algorithm we developed can accurately classify certain types of daily activities, with high overall classification accuracy for both waist-worn GENEA (0.99) and wrist-worn GENEA (right wrist = 0.97, left wrist = 0.96)., Conclusions: We have successfully developed algorithms suitable for use with wrist-worn accelerometers for detecting certain types of physical activities; the performance is comparable to waist-worn accelerometers for assessment of physical activity.

The purpose of this study was to assess the relationship of accelerometer output, in counts (ActiGraph GT1M) and as raw accelerations (ActiGraph GT3X+ and GENEA), with ground reaction force (GRF) in adults. Ten participants (age: 29.4 ± 8.2 yr, mass: 74.3 ± 9.8 kg, height: 1.76 ± 0.09 m) performed eight trials each of: slow walking, brisk walking, slow running, faster running and box drops. GRF data were collected for one step per trial (walking and running) using a force plate. Low jumps and higher jumps (one per second) were performed for 20 s each on the force plate. For box drops, participants dropped from a 35 cm box onto the force plate. Throughout, three accelerometers were worn at the hip: GT1M, GT3X+ and GENEA. A further GT3X+ and GENEA were worn on the left and right wrist, respectively. GT1M counts correlated with peak impact force (r = 0.85, p < 0.05), average resultant force (r = 0.73, p < 0.05) and peak loading rate (r = 0.76, p < 0.05). Accelerations from the GT3X+ and GENEA correlated with average resultant force and peak loading rate irrespective of whether monitors were worn at the hip or wrist (r > 0.82, p < 0.05, r > 0.63 p < 0.05, respectively). In conclusion, accelerometer count and raw acceleration output correlate positively with GRF and thus may be appropriate for the quantification of activity beneficial to bone. Wrist-worn monitors show a similar relationship with GRF as hip-worn monitors, suggesting that wrist-worn monitors may be a viable option for future studies looking at bone health., (Copyright © 2011 Elsevier Ltd. All rights reserved.)

Purpose: The study aims were: 1) to assess the technical reliability and validity of the GENEA using a mechanical shaker; 2) to perform a GENEA value calibration to develop thresholds for sedentary and light-, moderate-, and vigorous-intensity physical activity; and 3) to compare the intensity classification of the GENEA with two widely used accelerometers., Methods: A total of 47 GENEA accelerometers were attached to a shaker and vertically accelerated, generating 15 conditions of varying acceleration and/or frequency. Reliability was calculated using SD and intrainstrument and interinstrument coefficients of variation, whereas validity was assessed using Pearson correlation with the shaker acceleration as the criterion. Next, 60 adults wore a GENEA on each wrist and on the waist (alongside an ActiGraph and RT3 accelerometer) while completing 10-12 activity tasks. A portable metabolic gas analyzer provided the criterion measure of physical activity. Analyses involved the use of Pearson correlations to establish criterion and concurrent validity and receiver operating characteristic curves to establish intensity cut points., Results: The GENEA demonstrated excellent technical reliability (CVintra = 1.4%, CVinter = 2.1%) and validity (r = 0.98, P < 0.001) using the mechanical shaker. The GENEA demonstrated excellent criterion validity using VO2 as the criterion (left wrist, r = 0.86; right wrist, r = 0.83; waist, r = 0.87), on par with the waist-worn ActiGraph and RT3. The GENEA demonstrated excellent concurrent validity compared with the ActiGraph (r = 0.92) and the RT3 (r = 0.97). The waist-worn GENEA had the greatest classification accuracy (area under the receiver operating characteristic curve (AUC) = 0.95), followed by the left (AUC = 0.93) and then the right wrist (AUC = 0.90). The accuracy of the waist-worn GENEA was virtually identical with that of the ActiGraph (AUC = 0.94) and RT3 (AUC = 0.95)., Conclusion: The GENEA is a reliable and valid measurement tool capable of classifying the intensity of physical activity in adults.

Accelerometry is frequently used to assess physical activity in children. The number of articles in this area has increased dramatically in the last 10 years. The aim of this article is to provide a contemporary overview of accelerometry research in children. Specifically, the review addresses the choice of monitor, choice of epoch, when and for how long activity should be measured, the lack of consensus over how to define a complete day of activity data, issues of compliance, methods for the calibration of activity output, and the richness of data available from accelerometry.

Purpose: Studies have reported strong linear relationships between accelerometer output and walking/running speeds up to 10 km x h(-1). However, ActiGraph uniaxial accelerometer counts plateau at higher speeds. The aim of this study was to determine the relationships of triaxial accelerometry, uniaxial accelerometry, and pedometry with speed and step frequency (SF) across a range of walking and running speeds., Methods: Nine male runners wore two ActiGraph uniaxial accelerometers, two RT3 triaxial accelerometers (all set at a 1-s epoch), and two Yamax pedometers. Each participant walked for 60 s at 4 and 6 km x h(-1), ran for 60 s at 10, 12, 14, 16, and 18 km x h(-1), and ran for 30 s at 20, 22, 24, and 26 km x h(-1). Step frequency was recorded by a visual count., Results: ActiGraph counts peaked at 10 km x h(-10 (2.5-3.0 Hz SF) and declined thereafter (r=0.02, P>0.05). After correction for frequency-dependent filtering, output plateaued at 10 km x h(-1) but did not decline (r=0.77, P<0.05). Similarly, RT3 vertical counts plateaued at speeds > 10 km x h(-1) (r=0.86, P<0.01). RT3 vector magnitude and anteroposterior and mediolateral counts maintained a linear relationship with speed (r>0.96, P<0.001). Step frequency assessed by pedometry compared well with actual step frequency up to 20 km x h(-1) (approximately 3.5 Hz) but then underestimated actual steps (Yamax r=0.97; ActiGraph pedometer r=0.88, both P<0.001)., Conclusion: Increasing underestimation of activity by the ActiGraph as speed increases is related to frequency-dependent filtering and assessment of acceleration in the vertical plane only. RT3 vector magnitude was strongly related to speed, reflecting the predominance of horizontal acceleration at higher speeds. These results indicate that high-intensity activity is underestimated by the ActiGraph, even after correction for frequency-dependent filtering, but not by the RT3. Pedometer output is highly correlated with step frequency.