Children and adolescents spend, on average, 6–8 h of their time awake on sedentary behavior (SB) [1
], and engaging in excessive amounts of SB may be adversely associated with physical and mental health [2
]. Different determinants of SB have been investigated and active transportation seems to be one modifiable component that may reduce sedentary time in the transition from childhood into adolescence [3
]. The accurate and objective assessment of SB and biking activities with children and adolescents would be valuable to further investigate the determinants of SB in this population.
Assessing physical activity and SB in children and adolescents is challenging [4
], and objective assessment using activity monitors worn at the wrist or hip has been shown to be an attractive method when compared to both direct observation and self-report methods [5
]. The validity of using the hip [6
] or wrist [7
], respectively, for the assessment of SB has been investigated, and significant differences have been reported [9
]. Using the wrist has been shown to provide an excellent wear comfort, which increases subject adherence to the measurement protocol [10
], but thigh-worn accelerometers have also been shown to provide high compliance rates with tape-mounted devices [11
]. The ActivPAL (PAL Technologies Ltd., Glasgow, UK) activity monitors are taped to the thigh and have been shown to provide accurate estimates of SB based on the assessment of postural allocation with children and adolescents [13
]. The underlying algorithms for identifying SB with the ActivPAL device are not available to the researcher, and the lack of transparency makes it difficult for researchers to fully understand the strength and limitations of the selected methodology. Furthermore, the economic burden of using the ActivPAL device might make it more applicable with studies including a small number of subjects. These limitations with the ActivPAL device have previously been addressed by Skotte et al. [16
] proposing an open source method for the identification of the following body positions and activities: Sitting, standing, walking, running and biking with adults using a single ActiGraph GT3X (ActiGraph LLC, Pensacola, FL, USA) accelerometer worn on the thigh [16
]. However, the size and bulky design of the ActiGraph GT3X device does not make the proposed method by Skotte et al. [16
] applicable for use with smaller children. In the study by Stewart et al. [17
], a smaller open source Axivity AX3 (Axivity, Newcastle, UK) device was used for the identification of children’s activity types using accelerometers worn at the thigh and trunk in addition to the use of a more advanced machine learning method in the data processing [17
]. However, besides the use of an advanced method, the study only provided data for children 7–15 years of age and did not investigate the identification of biking. Moreover, the underlying implementation of this method, and also the method proposed by Skotte et al. [16
], have never been made public available, which is a major limitation. Currently, there are no simple and public methods available for the assessment of PA behavior—specifically SB and biking—with preschoolers, children and adolescents using a single thigh-worn activity monitor.
The aim of this study is to investigate the accuracy and validity of a simple method for the identification of six common activity types with preschoolers, children and adolescents using a single accelerometer worn on the thigh, and to make this method publicly available as open source.
In this study, we evaluated a method for identifying sitting, standing, moving, walking, biking and running using a single Axivity AX3 (Axivity, Newcastle UK) accelerometer worn on the thigh with preschoolers, children and adolescents. The results demonstrate that the proposed method provides an excellent sensitivity and specificity with all the proposed activities, with the only exception for biking on running bikes with pre-school children. The accuracy found in this study is comparable to the original study conducted on adults using the same method for the identification [16
Other studies have evaluated techniques to identify children’s time spent in different activity types from accelerometer data. Trost et al. [20
] used logistic regression to identify seven activity types (lying, sitting, standing, walking, running, basketball, and dancing) in 52 children and adolescents using accelerometers worn either on the hip or wrist. The hip- and wrist-based models achieved 91.0% and 88.4% accuracy, respectively. In a study by Stewart et al. [17
] a dual-accelerometry system was evaluated with both children and adults for the identification of six activity types. One device was worn at the thigh and another device worn on the lower back. A random forest algorithm and a total of 142 different signal features generated from the raw acceleration were used in the identification of activity type. The random forest algorithm is an ensemble learner and performs the identification of activity types from acceleration using multiple individual decision trees [21
]. The method used in the present study only uses a single decision tree and just six different signal features. Despite the large difference in algorithm complexity and number of features, the sensitivity and specificity are above 98% for most activities in both studies. Thus, using a more complex algorithm or additional features does not seem to improve the identification accuracy per se, suggesting that wear location and optimal selection of signal features might be more important than the algorithm and number of signal features.
Including biking in the identification of activity type is primarily due to the known health benefits of this activity, but also due to the important assessment of transportation mode which is great importance in many countries. The lower sensitivity and specificity demonstrated for the identification of biking with preschool children is clearly caused by the use of running bikes rather than actual bikes with pedals. Running bikes do not have pedals as normal bikes do, and driving the running bike forward is carried out by either “running”, or double pushing both legs on the ground, or simply by having a break resting the legs on the footrests, although the children were encouraged not to do so. The accuracy for identifying the running bike as biking could be increased by decreasing the forward/backward angle threshold in the identification of biking from locomotion. Decreasing the angle threshold, however, would also potentially increase the misclassification of some running activities as biking, which will decrease the sensitivity of identifying locomotion in real data. The acceleration not identified as biking with the preschoolers using a running bike is primarily identified as running, but also as walking. Considering the actual movement with the running bike, it is more correct to the actual movement performed rather than as an indented biking activity.
The activities included in the present study and the study by Stewart et al. [17
] are performed in controlled environments, which simplifies the data processing and analysis. However, the frequent transitions between activities—which are an important element of children’s common movement behavior during free-living—are not included, suggesting that the sensitivity and specificity estimated by Stewart et al. [17
] and in the present study might not accurately reflect the performance of the algorithms in a free-living environment. In an attempt to address this, we included a basketball activity with the children and adolescents, and a playground activity with the children and preschoolers. The basketball and playground activities included movements such as standing still, moving, walking, running and jumping, varying in both duration and organization. We did not assess the amount of time spent with the different activity categories with the basketball and playground activities, but the estimated duration of the individual activity types seems to reflect the overall nature to be expected of these activities. The method proposed by Stewart et al. [17
]was evaluated in a free-living environment, and the results provide excellent sensitivity and specificity with most activities [22
]. However, transitions were excluded from the analysis and the results also further indicate that the identification of dynamic standing and movement is challenged with children. Increasing the number of available features and the complexity of the algorithm will increase the risk for overfitting and thus misclassification of some activity types in a free-living environment. The complex and sporadic movement behaviors, especially of younger children, seem to suggest that a robust identification (balance between sensitivity and specificity) of common activity types with children is most optimally approached using a limited number of signal features. The children enrolled in the free-living evaluation by Steward et al. [17
] were at the age of 10 years, and further studies seem to be required to evaluate the classification accuracy with even younger children in a free-living environment. Moreover, the implementation of the proposed method described by Stewart et al. [17
] is not publicly available, which makes it difficult for other researchers to use, replicate and modify.
The identification of stair walking was included in the original method proposed by Skotte et al. [16
]. The identification of stair walking was implemented using an individual defined threshold, determined using the median value for the forward/backward angle (Θ) below 5 degrees, and adding 4.5 degrees (Θd
+ 4.5). The threshold for the identification of stair walking is therefore <9.5 degrees, and thus sensitive to the misclassification of walking and running as stair walking. Running and walking movements with children are commonly performed in a complex environment and, combined with children’s short legs, it seems to increase the risk for the children to generate backward/forward angles that resemble stair walking. The question of whether to include or exclude stair walking is a balance between the misclassification of stair walking as normal walking or the misclassification of walking and running as stair walking. Most children do perform stair walking, but considering the nature of movement in children, we might introduce a systematic bias by the misclassification of walking and running as stair walking rather than obtaining an accurate estimate of children’s stair walking. In some environments, stair walking might be an important element of children’s movement behavior, which requires the accurate quantification of stair walking. However, only including the forward/backward angle in the identification of stair walking seems inadequate for obtaining a robust identification with younger children. This could argue for using additional features, as with the study by Stewart et al. [17
] for the identification of stair walking. However, as previously mentioned, increasing the number of signal features increases the risk of overfitting, and thus the misclassification of the actual activity performed. Currently, there is no single robust signal feature available for the accurate identification of stair walking in children, and further investigation of the biomechanical properties of children’s stair walking in relation to the acceleration measured at the thigh seems to be required.
All signal features generated in the method proposed by Stewart et al. [17
] are determined using a 5-s non-overlapping window, whereas only a 2-s 50% over-lapping window (providing second by second resolution) is used in the method proposed by Skotte et al. [16
] and the present study. The need to use a longer time window with the method proposed by Stewart et al. [17
] is most likely to provide sufficient resolution with the frequency-related signal features, in order to accurately detect the cyclic or non-cyclic nature of some activities. The dominant frequency is commonly included in many machine learning algorithms [17
], and estimated using the Fast Fourier transformation (FFT). The resolution and minimal detectable frequency is coupled with the total number samples used to estimate the feature. The sporadic nature of children’s movement behavior in a free-living environment seems to suggest a general increased misclassification, with longer time windows as compared to shorter time windows. Window size or epoch length has previously been investigated extensively with intensity-derived measures from acceleration, and it is clear that the intermittent nature of children’s activity has to be analyzed with short-duration epochs [24
]. Another interesting difference between the present study and the study conducted by Steward et al. [17
] is the use of non-overlapping and overlapping windows. Non-overlapping windows determine the activity type for each window independently, whereas over-lapping windows consider a smoother transition between windows. Preschoolers and children often perform a transition between different activity types lasting less than 5 s, and most likely not in synchronization with the window by window classification [26
]. This will cause the rate of misclassification to follow the number of activity transitions performed by the subject. However, using over-lapping windows might also be challenged with identifying the actual onset and offset between activities. The optimal selection of window size and overlapping windows seems to be an important aspect of the accurate classification of activity types in children, which seems to require further investigation in the future.
The lying posture is commonly interpreted as an indirect measure of sleep and the identification of lying using accelerometry has been approached using both single and multiple devices worn on both wrist, thigh, and hip [17
]. For many children and adolescents, it is not uncommon in the late afternoon or evening, during weekdays or generally during weekend days, to lie on the couch or in bed watching TV, using a tablet or their cell phone. This seems to suggest that with the indirect measure of sleep from the identification of lying, as validated with laboratory conducted experiments, there is a substantial risk of misclassifying sedentary behavior as sleep. The lying posture allocation associated with evening sedentary behavior is clearly not sleep and should not be included as such. The accurate identification of lying and sleep with 24-h free-living recordings is challenging and only including a laboratory lying activity, even though it is in different positions, does not provide sufficient information for distinguishing lying as sedentary behavior from time in bed and thus sleep. If accelerometry is to be used to provide an indirect assessment of time in bed and potentially sleep, it is of utmost importance to distinguish between these behaviors. The accurate identification of lying and time in bed/asleep has to be performed using measurements conducted with subjects during their free-living behavior similar to in previous studies [29
], and thus including the important temporal information regarding circadian rhythm and essentially sleep behaviors. Currently, there is only one study investigating the assessment of sleep using free-living recordings and a thigh-worn device. However, this study relies on the proprietary ActivPAL device and the validity was assessed with adults [28
]. Future studies should investigate the accurate identification of children’s lying from time in bed with thigh-worn accelerometers using free-living recordings, rather than only relying on laboratory data or standardized protocols.
Strength and Limitations
A major strength of this study was that all activities were performed in the subject’s natural environment, as well as the inclusion of subjects across multiple age groups. Although the subjects followed a strict protocol during the field validation, we allowed some natural movement adjustment during most activities—for example, adjusting sitting position when sitting. The use of short and overlapping windows in the generation of the signal features is a strength of the method, and provides a resolution which is required with the sporadic nature of this population. The method implementation is made publicly available and open source.
A major limitation with the present study that a free-living validation is not included. Conducting a true free-living validation is challenging, and the sensitivity and specificity estimated in this study do not reflect the true sensitivity and specificity with the method using real recordings. In the discussion, we addressed the possible limitations of the current method in comparison to more advanced methods and, considering the complexity of human behavior and movement, it seems to suggest that selecting a simpler and more robust method might perform better with real recordings. It is a minor limitation to use running bikes to perform biking for preschool children. However, the number of preschool children capable of using a real bike with pedals is likely to be very small, suggesting that this activity is not common with real recordings. It is a limitation that the method is implemented in the commercial and costly software Matlab®
. However, GNU Octave (https://www.gnu.org/software/octave/
) is a free and open-source alternative to Matlab. We implemented the described method using standard functions which are also available with Octave. Moreover, using a standard function also provides an easy replication of the method with common statistical software such as R or Python, which are freely available.